On Mon, May 23, 2022 at 07:43:08PM +0800, Julien Rouhaud wrote:\\n> > That being said, I'd gladly drop that enum and only handle a single error\\n> > message, as the rest of the error context (including the owning file name and\\n> > line) should provide enough information to users.\\n> > \\n> > If so, should I use \\\"included authentication file\\\" everywhere, or use something\\n> > else?\\n> \\n> With the file name provided the context, \\\"authentication file\\\"?\\n\\nSo you mean having an error message like that (having an \\\"include myconf\\\"\\nin the HBA file):\\n\\nLOG: could not open authentication file \\\"myconf\\\" as \\\"/path/to/myconf\\\": No such file or directory\\nLOG: pg_hba.conf was not reloaded\\n\\nThis would be somewhat consistent with how it's done for the postgresql.conf,\\nassuming that we do fix that hardcoded \\\"pg_hba.conf\\\":\\n\\nLOG: could not open configuration file \\\"/path/to/toto\\\": No such file or directory\\nLOG: configuration file \\\"/path/to/postgresql.conf\\\" contains errors; no changes were applied\\n\\nWith this message it's clear that the file that couldn't be opened is an\\nincluded file.\\n\\nHowever those two cases are slightly different, and technically the \\\"secondary\\nfile\\\" is not an authentication file, just a list of tokens, so I'm still a bit\\nworried about ambiguity here.\\n\\n> > The detail should be provided in the logs so it should disambiguate the\\n> > situation. While you're mentioning this message, AFAICS it can already be\\n> > entirely wrong as-is, as it doesn't take into account the hba_file\\n> > configuration:\\n> > \\n> > ALTER SYSTEM SET hba_file = '/tmp/myfile';\\n> \\n> Hmm, indeed. And we load the GUCs before the HBA rules, obviously.\\n> This could be made less confusing. Do you think that this can be\\n> improved independently of the main patch, building the include\\n> structure on top of it?\\n\\nSure, I can split that in another commit.\\n\\nAfter a bit more digging, I think that this comes from the fact that there's no\\n\\\"official\\\" name for this file. Even the documentation just says \\\"the\\npg_hba.conf file\\\" [1]. So using pg_hba.conf can either means explicitly\\n$PGDATA/pg_hba.conf or the instance's HBA file in general, whatever its\\nlocation.\\n\\nI think it would be good to improve this, including in the doc, but I'm\\nassuming it's entirely for HEAD only, including the error messages?\\n\\nIf so, should I also change the doc to replace \\\"pg_hba.conf\\\" with something\\nelse when it's not referring to the file default name?\\n\\nI'm thinking of using \\\"HBA file\\\" to replace pg_hba.conf, and using\\n\\\"authentication file\\\" when it can be either the \\\"HBA file\\\" and the \\\"User Name\\nMaps file\\\", would that be ok?\\n\\n> > But the same problem already exists for the postgresql.conf's include_dir.\\n> > \\n> > Having an hba rule masking another isn't really better than having a GUC\\n> > value overloaded by another one. Usually people rely on a sane naming (like\\n> > 01_file.conf and so on) to get a predictable behavior, and we even document\\n> > that very thoroughly for the postgresql.conf part. Improving the\\n> > documentation, as you already pointed out a bit above, should be enough?\\n> \\n> Ah. ParseConfigDirectory() does a qsort() before processing each\\n> file included in a folder. Unless I am missing something, your patch\\n> reads the entries in the directory, but does not sort the files by\\n> name to ensure a strict ordering of them.\\n> \\n> While on it, it looks like you should have sanity checks similar to\\n> ParseConfigDirectory() for CRLFs & friends, as of:\\n> if (strspn(includedir, \\\" \\\\t\\\\r\\\\n\\\") == strlen(includedir))\\n> \\n> All this logic is very similar, so this could be perhaps refactored.\\n\\nIndeed, I will move that in a common function to be consistent.\\n\\n> >> + <para>\\n> >> + Rule number, in priority order, of this rule if the rule is valid,\\n> >> + otherwise null\\n> >> + </para></entry>\\n> >> This is a very important field. I think that this explanation should\\n> >> be extended and explain why relying on this number counts (aka this is\\n> >> the order of the rules loaded across files). Btw, this could be added\\n> >> as a separate patch, even if this maps to the line number when it\\n> >> comes to the ordering.\\n> > \\n> > Agreed, I will improve the documentation to outline the importance of that\\n> > information.\\n> > \\n> > Do you mean a separate patch to ease review or to eventually commit both\\n> > separately? FWIW I don't think that adding any form of inclusion in the hba\\n> > files should be done without a way for users to check the results. Any test\\n> > facility would also probably rely on this field.\\n> \\n> Yeah, agreed that rule_number is important for the sake of the\\n> inclusions. Still, I was wondering if rule_number makes sense on its\\n> own, meaning that we could add it before working on the inclusion\\n> logic. Having it without the inclusion is less interesting if you\\n> have the line numbers to order the HBA rules, of course, as this is\\n> enough to guess the priority of the HBA entries.\\n\\nAh I see. Sure I could split it in another commit too, but this isn't\\nbackpatchable (and pg15 isn't branched yet anyway) so I'm not sure how useful\\nthat is. I will go head and do that anyway, it will be easy to merge the\\ncommits if needed.\\n\\n\\n[1] https://www.postgresql.org/docs/current/auth-pg-hba-conf.html\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 May 2022 15:26:57 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Thu, May 26, 2022 at 03:26:57PM +0800, Julien Rouhaud wrote:\\n> So you mean having an error message like that (having an \\\"include myconf\\\"\\n> in the HBA file):\\n> \\n> LOG: could not open authentication file \\\"myconf\\\" as \\\"/path/to/myconf\\\": No such file or directory\\n> LOG: pg_hba.conf was not reloaded\\n> \\n> This would be somewhat consistent with how it's done for the postgresql.conf,\\n> assuming that we do fix that hardcoded \\\"pg_hba.conf\\\":\\n> \\n> LOG: could not open configuration file \\\"/path/to/toto\\\": No such file or directory\\n> LOG: configuration file \\\"/path/to/postgresql.conf\\\" contains errors; no changes were applied\\n> \\n> With this message it's clear that the file that couldn't be opened is an\\n> included file.\\n\\nYes, consistency would be good here.\\n\\n> However those two cases are slightly different, and technically the \\\"secondary\\n> file\\\" is not an authentication file, just a list of tokens, so I'm still a bit\\n> worried about ambiguity here.\\n\\nHmm. Yes, using only this term could lead to some confusion.\\nThinking more about that, perhaps you are right to use fully separated\\nterms for the ident and HBA parts, particularly once the inclusion is\\nin place as the file names could be anything, so we would lose context\\nabout where a file for used for what.\\n\\n> After a bit more digging, I think that this comes from the fact that there's no\\n> \\\"official\\\" name for this file. Even the documentation just says \\\"the\\n> pg_hba.conf file\\\" [1]. So using pg_hba.conf can either means explicitly\\n> $PGDATA/pg_hba.conf or the instance's HBA file in general, whatever its\\n> location.\\n> \\n> I think it would be good to improve this, including in the doc, but I'm\\n> assuming it's entirely for HEAD only, including the error messages?\\n\\nYes, that would be a set of changes only for HEAD, once 16~ opens for\\nbusiness. FWIW, the acronym \\\"HBA\\\" is defined as \\\"Host-Based\\nAuthentication\\\", so we could use that as a base for the description of\\nthe file, using simply HBA in the follow-up paragraphs for simplicity,\\ntelling that pg_hba.conf is the default.\\n\\n> If so, should I also change the doc to replace \\\"pg_hba.conf\\\" with something\\n> else when it's not referring to the file default name?\\n>\\n> I'm thinking of using \\\"HBA file\\\" to replace pg_hba.conf, and using\\n> \\\"authentication file\\\" when it can be either the \\\"HBA file\\\" and the \\\"User Name\\n> Maps file\\\", would that be ok?\\n\\nUsing \\\"HBA file\\\" in the docs is fine by me, knowing that the acronym\\nis already defined. The modified parts of the docs should perhaps\\nmention once something like \\\"Host-Based Authentication file (or HBA\\nfile)\\\" for clarity. For the error message, I think that we tend to\\navoid those acronyms, don't we? \\n\\n> Ah I see. Sure I could split it in another commit too, but this isn't\\n> backpatchable (and pg15 isn't branched yet anyway) so I'm not sure how useful\\n> that is. I will go head and do that anyway, it will be easy to merge the\\n> commits if needed.\\n\\nThe rule_number shines with the inclusion logic in place, still\\nsplitting that into a separate commit makes the review and the commit\\nhistory cleaner IMO. Perhaps that's just me being overly-pedantic in\\nkeeping a clean commit history :)\\n--\\nMichael\",\n \"msg_date\": \"Thu, 2 Jun 2022 10:08:15 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Thu, Jun 02, 2022 at 10:08:15AM +0900, Michael Paquier wrote:\\n> On Thu, May 26, 2022 at 03:26:57PM +0800, Julien Rouhaud wrote:\\n>\\n> > After a bit more digging, I think that this comes from the fact that there's no\\n> > \\\"official\\\" name for this file. Even the documentation just says \\\"the\\n> > pg_hba.conf file\\\" [1]. So using pg_hba.conf can either means explicitly\\n> > $PGDATA/pg_hba.conf or the instance's HBA file in general, whatever its\\n> > location.\\n> >\\n> > I think it would be good to improve this, including in the doc, but I'm\\n> > assuming it's entirely for HEAD only, including the error messages?\\n>\\n> Yes, that would be a set of changes only for HEAD, once 16~ opens for\\n> business. FWIW, the acronym \\\"HBA\\\" is defined as \\\"Host-Based\\n> Authentication\\\", so we could use that as a base for the description of\\n> the file, using simply HBA in the follow-up paragraphs for simplicity,\\n> telling that pg_hba.conf is the default.\\n\\nOk.\\n\\n> > If so, should I also change the doc to replace \\\"pg_hba.conf\\\" with something\\n> > else when it's not referring to the file default name?\\n> >\\n> > I'm thinking of using \\\"HBA file\\\" to replace pg_hba.conf, and using\\n> > \\\"authentication file\\\" when it can be either the \\\"HBA file\\\" and the \\\"User Name\\n> > Maps file\\\", would that be ok?\\n>\\n> Using \\\"HBA file\\\" in the docs is fine by me, knowing that the acronym\\n> is already defined. The modified parts of the docs should perhaps\\n> mention once something like \\\"Host-Based Authentication file (or HBA\\n> file)\\\" for clarity. For the error message, I think that we tend to\\n> avoid those acronyms, don't we?\\n\\nI don't have an extensive knowledge of all the user facing error messages, but\\nafter a quick grep I see multiple usage of OID, PID, GIN and other defined\\nacronyms. I also see multiple occurrences of \\\"only heap AM is supported\\\",\\nwhile AM isn't even a defined acronym.\\n\\nIt doesn't seem that my proposal would be inconsistent with existing messages\\nand will help to reduce the message length, but if you prefer to keep the full\\nname I'm fine with it. Those should be very rare and specialized errors\\nanyway.\\n\\nWhile on the bikeshedding part, are you ok with the proposed keywords (include\\nand include_dir), behaving exactly like for postgresql.conf, and to also add\\ninclude_if_exists, so that we have the exact same possibilities with\\npostgresql.conf, pg_hba.conf and pg_ident.conf?\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Jul 2022 14:57:21 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Fri, Jul 08, 2022 at 02:57:21PM +0800, Julien Rouhaud wrote:\\n\\nMy apologies for the late reply.\\n\\n> I don't have an extensive knowledge of all the user facing error messages, but\\n> after a quick grep I see multiple usage of OID, PID, GIN and other defined\\n> acronyms. I also see multiple occurrences of \\\"only heap AM is supported\\\",\\n> while AM isn't even a defined acronym.\\n\\nA lot depends on the context of the code, it seems.\\n\\n> It doesn't seem that my proposal would be inconsistent with existing messages\\n> and will help to reduce the message length, but if you prefer to keep the full\\n> name I'm fine with it. Those should be very rare and specialized errors\\n> anyway.\\n\\nSo you mean to use \\\"HBA file\\\" instead of pg_hba.conf and\\n\\\"authentication file\\\" when it can be either one of an HBA file or a\\nmapping file? That would be okay by me. We would have a full cycle\\nto tune them depending on the feedback we'd get afterwards.\\n\\n> While on the bikeshedding part, are you ok with the proposed keywords (include\\n> and include_dir), behaving exactly like for postgresql.conf, and to also add\\n> include_if_exists, so that we have the exact same possibilities with\\n> postgresql.conf, pg_hba.conf and pg_ident.conf?\\n\\nOkay, agreed for consistency. With include_dir being careful about\\nthe ordering of the entries and ignoring anything else than a .conf\\nfile (that's something you mentioned already upthread).\\n--\\nMichael\",\n \"msg_date\": \"Mon, 11 Jul 2022 10:16:44 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Mon, Jul 11, 2022 at 10:16:44AM +0900, Michael Paquier wrote:\\n> On Fri, Jul 08, 2022 at 02:57:21PM +0800, Julien Rouhaud wrote:\\n> \\n> My apologies for the late reply.\\n> \\n> > I don't have an extensive knowledge of all the user facing error messages, but\\n> > after a quick grep I see multiple usage of OID, PID, GIN and other defined\\n> > acronyms. I also see multiple occurrences of \\\"only heap AM is supported\\\",\\n> > while AM isn't even a defined acronym.\\n> \\n> A lot depends on the context of the code, it seems.\\n> \\n> > It doesn't seem that my proposal would be inconsistent with existing messages\\n> > and will help to reduce the message length, but if you prefer to keep the full\\n> > name I'm fine with it. Those should be very rare and specialized errors\\n> > anyway.\\n> \\n> So you mean to use \\\"HBA file\\\" instead of pg_hba.conf and\\n> \\\"authentication file\\\" when it can be either one of an HBA file or a\\n> mapping file? That would be okay by me.\\n\\nYes, it seems to me like a good compromise for not being overly verbose and\\nstill being understandable.\\n\\n> We would have a full cycle\\n> to tune them depending on the feedback we'd get afterwards.\\n\\nAgreed.\\n\\n> > While on the bikeshedding part, are you ok with the proposed keywords (include\\n> > and include_dir), behaving exactly like for postgresql.conf, and to also add\\n> > include_if_exists, so that we have the exact same possibilities with\\n> > postgresql.conf, pg_hba.conf and pg_ident.conf?\\n> \\n> Okay, agreed for consistency. With include_dir being careful about\\n> the ordering of the entries and ignoring anything else than a .conf\\n> file (that's something you mentioned already upthread).\\n\\nOk! All those that should be covered by new regression test so it should be\\nclear what is being implemented.\\n\\nWhile on the regression tests topic, I started to implement those and faced\\nsome problems quite fast when trying to workaround the problem we previously\\ndiscussed (1).\\n\\nSo first, even if we can test 99% of the features with just testing the views\\noutput, I think it's should use the TAP framework since the tests will have to\\nmess with the pg_ident/pg_hba files. It's way easier to modify the auth files,\\nand it uses a dedicated instance so we don't have to worry about breaking other\\ntest that would run concurrently.\\n\\nAlso, if we want to test the views error reporting we have to use a persistent\\nconnection (as in interactive_psql()), otherwise tests will immediately fail on\\nWindows / EXEC_BACKEND builds. Adding the ability to run queries and wait for\\ncompletion on top of interactive_psql() doesn't seem to cause any problem, but\\ninterpreting the results does.\\n\\nSince it's just manipulating the psql's stdin/stdout, we retrieve the prompt\\nand executed query too. So for instance, if you just want to test \\\"SELECT 1\\\",\\nyou will have to cleanup something like\\n\\ndbname=# SELECT 1;\\n1\\ndbname#\\n\\nThat may still be workable by splitting the output per newline (and possibly\\nremoving the first prompt before sending the query text), and remove the first\\nand last entry (assuming you want to test somewhat sane data, and not e.g. run\\nthe regression test on a database containing a newline), but then you have to\\nalso account for possible escape sequences, for instance if you use\\nenable-bracketed-paste. In that case, the output becomes something like\\n\\ndbname=# SELECT 1;\\n[?2004l\\n1\\n[?2004hpostgres=#\\n\\nIt could probably be handled with some regexp to remove escape sequences and\\nremove empty lines, but it seems like really fragile, and thus a very bad idea.\\n\\nI'm not really sure what should be done here. The best compromise I can think\\nof is to split the tests in 3 parts:\\n\\n1) view reporting with various inclusions using safe_psql()\\n2) log error reporting\\n3) view reporting with various inclusions errors, using safe_psql()\\n\\nAnd when testing 3), detect first if we can still connect after introducing\\nerrors. If not, assume this is Windows / EXEC_BACKEND and give up here without\\nreporting an error. Otherwise continue, and fail the test if we later can't\\nconnect anymore. As discussed previously, detecting that the build is using\\nthe fork emulation code path doesn't seem worthwhile so guessing it from the\\npsql error may be a better approach.\\n\\nDo you have any better idea, or do you have comments on this approach?\\n\\n[1]: https://www.postgresql.org/message-id/YkFhpydhyeNNo3Xl@paquier.xyz\\n\\n\\n\",\n \"msg_date\": \"Mon, 18 Jul 2022 15:11:51 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 18, 2022 at 03:11:51PM +0800, Julien Rouhaud wrote:\\n> So first, even if we can test 99% of the features with just testing the views\\n> output, I think it's should use the TAP framework since the tests will have to\\n> mess with the pg_ident/pg_hba files. It's way easier to modify the auth files,\\n> and it uses a dedicated instance so we don't have to worry about breaking other\\n> test that would run concurrently.\\n\\nAgreed.\\n\\n> That may still be workable by splitting the output per newline (and possibly\\n> removing the first prompt before sending the query text), and remove the first\\n> and last entry (assuming you want to test somewhat sane data, and not e.g. run\\n> the regression test on a database containing a newline), but then you have to\\n> also account for possible escape sequences, for instance if you use\\n> enable-bracketed-paste. In that case, the output becomes something like\\n> \\n> dbname=# SELECT 1;\\n> [?2004l\\n> 1\\n> [?2004hpostgres=#\\n> \\n> It could probably be handled with some regexp to remove escape sequences and\\n> remove empty lines, but it seems like really fragile, and thus a very bad idea.\\n\\nHmm. Indeed, that sounds fragile. And I don't really think that we\\nshould make more testing infrastructure a requirement for this patch\\nas being able to maintain a connection while running the tests is\\nsomething we've bumped on for a bit of time.\\n\\n> I'm not really sure what should be done here. The best compromise I can think\\n> of is to split the tests in 3 parts:\\n> \\n> 1) view reporting with various inclusions using safe_psql()\\n\\nYou mean in the case where the HBA and indent files can be loaded,\\nso as it is possible to peek at the system views without the\\nEXEC_BACKEND problem, right?\\n\\n> 2) log error reporting\\n\\nThis one should be reliable and stable enough by parsing the logs of\\nthe backend, thanks to connect_ok() and connect_fails().\\n\\n> 3) view reporting with various inclusions errors, using safe_psql()\\n> \\n> And when testing 3), detect first if we can still connect after introducing\\n> errors. If not, assume this is Windows / EXEC_BACKEND and give up here without\\n> reporting an error. Otherwise continue, and fail the test if we later can't\\n> connect anymore. As discussed previously, detecting that the build is using\\n> the fork emulation code path doesn't seem worthwhile so guessing it from the\\n> psql error may be a better approach.\\n\\nYeah, we could do that. Now we may also fail on other patterns, so we\\nwould need to make sure that a set of expected error outputs are the\\nones generated? I'd be fine to give up testing the error output\\ngenerated in the system views at the end. Without a persistent\\nconnection state with the same kind of APIs as any of the drivers able\\nto do so, that's going to be a PITA to maintain.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 19 Jul 2022 15:13:12 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Mon, Mar 28, 2022 at 04:22:32PM +0900, Michael Paquier wrote:\\n> On Mon, Mar 28, 2022 at 04:20:07PM +0900, Michael Paquier wrote:\\n> > See the attached, for reference, but it would fail with EXEC_BACKEND\\n> > on WIN32.\\n> \\n> Ditto.\\n\\nWhile working on the full regression test coverage for the file inclusion\\nthing, I discovered an embarrassing typo in the pg_ident_file_mapping\\ninfrastructure, which was using the hba file name rather than the ident file\\nname in one of the calls.\\n\\nIt doesn't have much impact most of the time. The filename is reported if\\nthere's an IO error while reading the already opened correct file. The real\\nproblem is if the hba_file and ident_file are stored in different directory,\\nany secondary file (@filename) in the pg_ident.conf would be searched in the\\nwrong directory. With the pending file inclusion patchset, the problem is\\nimmediately visible as the view is reporting the wrong file name.\\n\\nSimple fix attached. I'll add a v15 open item shortly.\",\n \"msg_date\": \"Tue, 26 Jul 2022 13:04:02 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 01:04:02PM +0800, Julien Rouhaud wrote:\\n> It doesn't have much impact most of the time. The filename is reported if\\n> there's an IO error while reading the already opened correct file. The real\\n> problem is if the hba_file and ident_file are stored in different directory,\\n> any secondary file (@filename) in the pg_ident.conf would be searched in the\\n> wrong directory. With the pending file inclusion patchset, the problem is\\n> immediately visible as the view is reporting the wrong file name.\\n\\nOops, obviously. I'll go and fix that as that's on me.\\n\\n> Simple fix attached. I'll add a v15 open item shortly.\\n\\nThanks. Better not to lose track of it.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 26 Jul 2022 14:14:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Tue, Jul 19, 2022 at 03:13:12PM +0900, Michael Paquier wrote:\\n> On Mon, Jul 18, 2022 at 03:11:51PM +0800, Julien Rouhaud wrote:\\n>\\n> > I'm not really sure what should be done here. The best compromise I can think\\n> > of is to split the tests in 3 parts:\\n> >\\n> > 1) view reporting with various inclusions using safe_psql()\\n>\\n> You mean in the case where the HBA and indent files can be loaded,\\n> so as it is possible to peek at the system views without the\\n> EXEC_BACKEND problem, right?\\n\\nYes, testing the general behavior when there are no errors in the auth files.\\n\\n> > 2) log error reporting\\n>\\n> This one should be reliable and stable enough by parsing the logs of\\n> the backend, thanks to connect_ok() and connect_fails().\\n\\nI meant testing the postgres logs reporting, something like generating entirely\\nbogus files, restarting, checking that the start failed and verify that the\\nlogs are expected. With a variation that if only the pg_ident.conf is bogus,\\nthe server will (and should) still start, so both needs to be tested\\nseparately.\\n\\n> > 3) view reporting with various inclusions errors, using safe_psql()\\n> >\\n> > And when testing 3), detect first if we can still connect after introducing\\n> > errors. If not, assume this is Windows / EXEC_BACKEND and give up here without\\n> > reporting an error. Otherwise continue, and fail the test if we later can't\\n> > connect anymore. As discussed previously, detecting that the build is using\\n> > the fork emulation code path doesn't seem worthwhile so guessing it from the\\n> > psql error may be a better approach.\\n>\\n> Yeah, we could do that. Now we may also fail on other patterns, so we\\n> would need to make sure that a set of expected error outputs are the\\n> ones generated? I'd be fine to give up testing the error output\\n> generated in the system views at the end. Without a persistent\\n> connection state with the same kind of APIs as any of the drivers able\\n> to do so, that's going to be a PITA to maintain.\\n\\nYes, or just checking that the error is expected on the psql side as the server\\nwill report it.\\n\\nI've been working on all of that and came up with the attached v8.\\n\\n- 0001: I modified things as discussed previously to report the real auth file\\n names rather than the hardcoded \\\"pg_ident.conf\\\" and \\\"pg_hba.conf\\\" in the\\n various log messages\\n- 0002 and 0003 are minor fixes to error logs more consistent\\n- 0004: the rule_number / mapping_number addition in the views in a separate\\n commit\\n- 0005: the main file inclusion patch. Only a few minor bugfix since\\n previous version discovered thanks to the tests (a bit more about it after),\\n and documentation tweaks based on previous discussions\\n- 0006: the pg_hba_matches() POC, no changes\\n\\nAbout the regression tests:\\nI added a new 003_file_inclusion.pl in src/test/authentication TAP tests to do\\nthings as previously described, ie test both extisting behavior and new\\nfeatures added, for working and non working scenarios. The view error\\nreporting are bypassed for Windows / EXEC_BACKEND build by detecting a\\nconnection failure with the expected error reported by the server. If the\\nerror doesn't match, the rest of the tests are still skipped but overall test\\nwill fail due to the mismatch in the reported error.\\n\\nIt's a bit troublesome to write such tests, as you need to keep in sync the\\nvarious rule and line number, format things differently depending on whether\\nyou want to check the logs or the view with each new tests added and so on.\\n\\nTo make the tests easier to write (and to maintain) I added some wrapper\\nfunctions to add lines in the wanted files that allow to automatically\\ngenerates the wanted regex (for the logs) or output (for the view). There are\\nstill things to be careful about and a bit of duplication, but this way I can\\neasily modify any test, or add new ones, without the need to modify everything\\naround. As written, it also uses the same included files for the log error\\nreporting and the view error reporting, so there's no need to define everything\\ntwice.\\n\\nThe tests work as expected locally on a normal build and an EXEC_BACKEND build,\\nand also on the CI.\",\n \"msg_date\": \"Sat, 30 Jul 2022 16:09:36 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 1:14 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Tue, Jul 26, 2022 at 01:04:02PM +0800, Julien Rouhaud wrote:\\n> > It doesn't have much impact most of the time. The filename is reported if\\n> > there's an IO error while reading the already opened correct file. The real\\n> > problem is if the hba_file and ident_file are stored in different directory,\\n> > any secondary file (@filename) in the pg_ident.conf would be searched in the\\n> > wrong directory. With the pending file inclusion patchset, the problem is\\n> > immediately visible as the view is reporting the wrong file name.\\n>\\n> Oops, obviously. I'll go and fix that as that's on me.\\n>\\n> > Simple fix attached. I'll add a v15 open item shortly.\\n>\\n> Thanks. Better not to lose track of it.\\n\\nFor the archives's sake, this was fixed shortly after as of\\n27e0ee57f68d27af68967759a2ff61a581f501dc on master and the open item\\nis closed.\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Aug 2022 09:12:52 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Sat, Jul 30, 2022 at 04:09:36PM +0800, Julien Rouhaud wrote:\\n> I've been working on all of that and came up with the attached v8.\\n> \\n> - 0001: I modified things as discussed previously to report the real auth file\\n> names rather than the hardcoded \\\"pg_ident.conf\\\" and \\\"pg_hba.conf\\\" in the\\n> various log messages\\n> - 0002 and 0003 are minor fixes to error logs more consistent\\n\\nAs a quick update from my side, I intend to look and apply 0001~0003\\n(not double-checked yet) shortly.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 2 Aug 2022 19:32:54 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 02, 2022 at 07:32:54PM +0900, Michael Paquier wrote:\\n> As a quick update from my side, I intend to look and apply 0001~0003\\n> (not double-checked yet) shortly.\\n\\nAnd a couple of days later, these look fine so done as of 47ab1ac and\\n718fe0a. 0002 and 0003 have been merged together.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 5 Aug 2022 09:56:29 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Fri, Aug 05, 2022 at 09:56:29AM +0900, Michael Paquier wrote:\\n> On Tue, Aug 02, 2022 at 07:32:54PM +0900, Michael Paquier wrote:\\n> > As a quick update from my side, I intend to look and apply 0001~0003\\n> > (not double-checked yet) shortly.\\n>\\n> And a couple of days later, these look fine so done as of 47ab1ac and\\n> 718fe0a. 0002 and 0003 have been merged together.\\n\\nThanks!\\n\\n\\n\",\n \"msg_date\": \"Sat, 6 Aug 2022 17:19:35 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Sat, Jul 30, 2022 at 04:09:36PM +0800, Julien Rouhaud wrote:\\n\\n>\\n> - 0001: the rule_number / mapping_number addition in the views in a separate\\n> commit\\n> - 0002: the main file inclusion patch. Only a few minor bugfix since\\n> previous version discovered thanks to the tests (a bit more about it after),\\n> and documentation tweaks based on previous discussions\\n> - 0003: the pg_hba_matches() POC, no changes\\n\\nAttached v9, simple rebase after multiple conflicts and part of the patchset\\napplied.\",\n \"msg_date\": \"Tue, 16 Aug 2022 14:10:30 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 16, 2022 at 02:10:30PM +0800, Julien Rouhaud wrote:\\n> Hi,\\n> \\n> On Sat, Jul 30, 2022 at 04:09:36PM +0800, Julien Rouhaud wrote:\\n> \\n> >\\n> > - 0001: the rule_number / mapping_number addition in the views in a separate\\n> > commit\\n> > - 0002: the main file inclusion patch. Only a few minor bugfix since\\n> > previous version discovered thanks to the tests (a bit more about it after),\\n> > and documentation tweaks based on previous discussions\\n> > - 0003: the pg_hba_matches() POC, no changes\\n\\nv10 attached to fix a few conflict with recent refactoring commits, no other\\nchange.\",\n \"msg_date\": \"Sun, 18 Sep 2022 01:06:12 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Sun, Sep 18, 2022 at 01:06:12AM +0800, Julien Rouhaud wrote:\\n> On Tue, Aug 16, 2022 at 02:10:30PM +0800, Julien Rouhaud wrote:\\n> >\\n> > On Sat, Jul 30, 2022 at 04:09:36PM +0800, Julien Rouhaud wrote:\\n> >\\n> > >\\n> > > - 0001: the rule_number / mapping_number addition in the views in a separate\\n> > > commit\\n> > > - 0002: the main file inclusion patch. Only a few minor bugfix since\\n> > > previous version discovered thanks to the tests (a bit more about it after),\\n> > > and documentation tweaks based on previous discussions\\n> > > - 0003: the pg_hba_matches() POC, no changes\\n\\nv11 attached, fixing a shadowed variable warning.\",\n \"msg_date\": \"Mon, 10 Oct 2022 10:51:32 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 10, 2022 at 10:51:32AM +0800, Julien Rouhaud wrote:\\n> On Sun, Sep 18, 2022 at 01:06:12AM +0800, Julien Rouhaud wrote:\\n> > On Tue, Aug 16, 2022 at 02:10:30PM +0800, Julien Rouhaud wrote:\\n> > >\\n> > > On Sat, Jul 30, 2022 at 04:09:36PM +0800, Julien Rouhaud wrote:\\n> > >\\n> > > >\\n> > > > - 0001: the rule_number / mapping_number addition in the views in a separate\\n> > > > commit\\n> > > > - 0002: the main file inclusion patch. Only a few minor bugfix since\\n> > > > previous version discovered thanks to the tests (a bit more about it after),\\n> > > > and documentation tweaks based on previous discussions\\n> > > > - 0003: the pg_hba_matches() POC, no changes\\n\\nv12 attached, fixing multiple conflicts with recent activity.\",\n \"msg_date\": \"Mon, 24 Oct 2022 13:33:12 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 24, 2022 at 01:33:12PM +0800, Julien Rouhaud wrote:\\n> v12 attached, fixing multiple conflicts with recent activity.\\n\\n typedef struct TokenizedAuthLine\\n {\\n List *fields; /* List of lists of AuthTokens */\\n+ char *file_name; /* File name *\\n\\nHmm. While putting my eyes on this patch set for the first time in a\\nfew months (sorry!), the addition of file_name to TokenizedAuthLine in\\n0002 stands out. This impacts all the macros used for the validation\\nof the ident and HBA lines, leading as well to a lot of bloat in the\\npatch with patterns like that in 20~25 places:\\n errcontext(\\\"line %d of configuration file \\\\\\\"%s\\\\\\\"\\\", \\\\\\n- line_num, IdentFileName))); \\\\\\n+ line_num, file_name))); \\\\\\n[...]\\n errcontext(\\\"line %d of configuration file \\\\\\\"%s\\\\\\\"\\\",\\n- line_num, HbaFileName)));\\n+ line_num, file_name)));\\n\\nDo you think that it would make sense to split that into its own\\npatch? That would move the code toward less references to HbaFileName\\nand IdentFileName, which is one step toward what we want for this\\nthread. This tokenization of HBA and ident files is already shared,\\nso moving this knowledge into TokenizedAuthLine looks helpful\\nindependently of the rest.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 24 Oct 2022 16:13:51 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Mon, Oct 24, 2022 at 04:13:51PM +0900, Michael Paquier wrote:\\n> On Mon, Oct 24, 2022 at 01:33:12PM +0800, Julien Rouhaud wrote:\\n> > v12 attached, fixing multiple conflicts with recent activity.\\n>\\n> typedef struct TokenizedAuthLine\\n> {\\n> List *fields; /* List of lists of AuthTokens */\\n> + char *file_name; /* File name *\\n>\\n> Hmm. While putting my eyes on this patch set for the first time in a\\n> few months (sorry!), the addition of file_name to TokenizedAuthLine in\\n> 0002 stands out. This impacts all the macros used for the validation\\n> of the ident and HBA lines, leading as well to a lot of bloat in the\\n> patch with patterns like that in 20~25 places:\\n> errcontext(\\\"line %d of configuration file \\\\\\\"%s\\\\\\\"\\\", \\\\\\n> - line_num, IdentFileName))); \\\\\\n> + line_num, file_name))); \\\\\\n> [...]\\n> errcontext(\\\"line %d of configuration file \\\\\\\"%s\\\\\\\"\\\",\\n> - line_num, HbaFileName)));\\n> + line_num, file_name)));\\n>\\n> Do you think that it would make sense to split that into its own\\n> patch? That would move the code toward less references to HbaFileName\\n> and IdentFileName, which is one step toward what we want for this\\n> thread. This tokenization of HBA and ident files is already shared,\\n> so moving this knowledge into TokenizedAuthLine looks helpful\\n> independently of the rest.\\n\\nIt would also require to bring HbaLine->sourcefile. I'm afraid it would be\\nweird to introduce such a refactoring in a separate commit just to pass a\\nconstant down multiple level of indirection, as all the macro will remain\\nspecific to either hba or ident anyway.\\n\\nI agree that there are quite a lot of s/XXXFileName/file_name/, but those\\naren't complicated, and keeping them in the same commit makes it easy to\\nvalidate that none has been forgotten since the regression tests covering those\\nmessages are in that commit too.\\n\\nAnd of course while double checking that none was forgotten I realize that I\\nmissed the new regcomp_auth_token() which introduced a couple new usage of\\nHbaFileName.\\n\\n\\n\",\n \"msg_date\": \"Mon, 24 Oct 2022 16:03:03 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Mon, Oct 24, 2022 at 04:03:03PM +0800, Julien Rouhaud wrote:\\n> It would also require to bring HbaLine->sourcefile. I'm afraid it would be\\n> weird to introduce such a refactoring in a separate commit just to pass a\\n> constant down multiple level of indirection, as all the macro will remain\\n> specific to either hba or ident anyway.\\n\\nPutting my hands on it, I am not really afraid of doing that\\nindependently. From what I can see, this is cutting 23kB worth of\\ndiffs from 0002, reducing it from 94K to 71kB.\\n\\n> I agree that there are quite a lot of s/XXXFileName/file_name/, but those\\n> aren't complicated, and keeping them in the same commit makes it easy to\\n> validate that none has been forgotten since the regression tests covering those\\n> messages are in that commit too.\\n\\nAnother advantage is that it minimizes the presence of the hardcoded\\nHbaFileName and IdentFileName in hba.c, which is one thing we are\\ntrying to achieve here for the inclusion of more files. I found a bit\\nstrange that IdentLine had no sourcefile, actually. We track the file\\nnumber but use it nowhere, and it seems to me that having more\\nsymmetry between both would be a good thing.\\n\\nSo, the key of the logic is how we are going to organize the\\ntokenization of the HBA and ident lines through all the inclusions..\\nAs far as I get it, tokenize_auth_file() is the root call and\\ntokenize_file_with_context() with its depth is able to work on each\\nindividual file, and it can optionally recurse depending on what's\\nincluded. Why do you need to switch to the old context in\\ntokenize_file_with_context()? Could it be simpler to switch once to\\nlinecxt outside of the internal routine?\\n\\nIt looks like GetDirConfFiles() is another piece that can be\\nrefactored and reviewed on its own, as we use it in\\nParseConfigDirectory()@guc.c.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 25 Oct 2022 15:43:21 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 25, 2022 at 03:43:21PM +0900, Michael Paquier wrote:\\n>\\n> Another advantage is that it minimizes the presence of the hardcoded\\n> HbaFileName and IdentFileName in hba.c, which is one thing we are\\n> trying to achieve here for the inclusion of more files. I found a bit\\n> strange that IdentLine had no sourcefile, actually. We track the file\\n> number but use it nowhere, and it seems to me that having more\\n> symmetry between both would be a good thing.\\n\\nIf IdentLine->linenumber is useless, why not get rid of it rather than tracking\\nanother useless info? Hba is much more structured so we need a more\\nspecialized struct, but I don't think ident will ever go that way.\\n\\n> So, the key of the logic is how we are going to organize the\\n> tokenization of the HBA and ident lines through all the inclusions..\\n> As far as I get it, tokenize_auth_file() is the root call and\\n> tokenize_file_with_context() with its depth is able to work on each\\n> individual file, and it can optionally recurse depending on what's\\n> included. Why do you need to switch to the old context in\\n> tokenize_file_with_context()? Could it be simpler to switch once to\\n> linecxt outside of the internal routine?\\n\\nIt just seemed the cleanest way to go. We could do without but then we would\\nhave to duplicate MemoryContextSwitchTo calls all over the place, and probably\\nhandling an optional memory context creation in the function.\\n\\n> It looks like GetDirConfFiles() is another piece that can be\\n> refactored and reviewed on its own, as we use it in\\n> ParseConfigDirectory()@guc.c.\\n\\nI'm fine with it.\\n\\n\\n\",\n \"msg_date\": \"Tue, 25 Oct 2022 15:08:59 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 25, 2022 at 03:08:59PM +0800, Julien Rouhaud wrote:\\n> On Tue, Oct 25, 2022 at 03:43:21PM +0900, Michael Paquier wrote:\\n>> Another advantage is that it minimizes the presence of the hardcoded\\n>> HbaFileName and IdentFileName in hba.c, which is one thing we are\\n>> trying to achieve here for the inclusion of more files. I found a bit\\n>> strange that IdentLine had no sourcefile, actually. We track the file\\n>> number but use it nowhere, and it seems to me that having more\\n>> symmetry between both would be a good thing.\\n> \\n> If IdentLine->linenumber is useless, why not get rid of it rather than tracking\\n> another useless info? Hba is much more structured so we need a more\\n> specialized struct, but I don't think ident will ever go that way.\\n\\nHmm. I would be tempted to keep track of the file name and the line\\nnumber as well in IdentLine. One reason is that this can become\\nuseful for debugging. A second is that this can reduce a bit the\\narguments of fill_ident_line() and fill_hba_line() in hbafuncs.c once\\nwe track these in HbaLine and IdentLine. And HEAD is slightly\\noverdoing it in its interface for the line number, actually, as we\\npass the line number twice: from {Ident,Hba}Line and the respective\\nfield from TokenizedAuthLine.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 25 Oct 2022 20:59:57 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Oct 25, 2022 at 08:59:57PM +0900, Michael Paquier wrote:\\n>\\n> Hmm. I would be tempted to keep track of the file name and the line\\n> number as well in IdentLine. One reason is that this can become\\n> useful for debugging. A second is that this can reduce a bit the\\n> arguments of fill_ident_line() and fill_hba_line() in hbafuncs.c once\\n> we track these in HbaLine and IdentLine.\\n\\nOk, I guess something like the attached v14 is what you want.\\n\\n> And HEAD is slightly\\n> overdoing it in its interface for the line number, actually, as we\\n> pass the line number twice: from {Ident,Hba}Line and the respective\\n> field from TokenizedAuthLine.\\n\\nThat wouldn't be overdoing anymore if we remove the line number / filename from\\nthe fill_*_line prototypes right?\",\n \"msg_date\": \"Wed, 26 Oct 2022 11:19:48 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 26, 2022 at 11:19:48AM +0800, Julien Rouhaud wrote:\\n> That wouldn't be overdoing anymore if we remove the line number / filename from\\n> the fill_*_line prototypes right?\\n\\nYeah, but there is a twist: HbaLine or IdentLine can be passed as\\nNULL when entering in fill_hba_line() or fill_ident_line() in the\\nevent of an error when tokenizing the line or when failing the\\ncreation of their Line, so attempting to read the line number from\\neither of them when filling in a tuple for their respective view would\\ncause a crash.\\n\\nSo, for now, I have taken the minimalistic approach with the addition\\nof the source file name in HbaFile and in TokenizedAuthLine. The\\nformer is actually necessary anyway as auth.c wants it in two\\nlocations (improved auth_failed() and set_authn_id()).\\n\\nIt is still true that the line number in IdentLine remains unused.\\nHence, do you think that the addition of the source file name and its\\nline number could be useful for the error reporting done in\\ncheck_ident_usermap()? The contents of the HbaLines are used in\\nauth.c for its own reporting magic, and it looks like this would be\\nthe only location where what's in the IdentLines is useful, aka\\nprovide more details about what is happening. Once users are able to\\ninclude ident files, that would likely help in debugging issues.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 26 Oct 2022 13:15:05 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 26, 2022 at 11:19:48AM +0800, Julien Rouhaud wrote:\\n> That wouldn't be overdoing anymore if we remove the line number / filename from\\n> the fill_*_line prototypes right?\\n\\nSo, I have spent a good portion of today looking at what you have\\nhere, applying 0001 and 0003 while fixing, rebasing and testing the\\nwhole, discarding 0002 (we could do more for the line number and\\nsource file in terms of the LOGs reported for a regexec failure). \\n\\nNow remains 0004, which is the core of the proposal, and while it\\nneeds a rebase, I have not been able to spend much time looking at its\\ninternals. In order to help with the follow-up steps, I have spotted\\na few areas that could be done first:\\n- The diffs in guc.h/c for the introduction of GetDirConfFiles() to\\nget a set of files in a directory (got to think a bit more about\\nParseConfigFp() when it comes to the keywords, but that comes to play\\nwith the parsing logic which is different).\\n- The TAP test, which is half the size of the patch in line number.\\nCould it be possible to make it more edible, introducing a basic\\ninfrastructure to check a set of rules in pg_hba.conf and\\npg_ident.conf without the inclusion logic? Checks for error patterns\\n(that I agree we strongly lack tests for) look like something we'd\\nwant to tackle independently of the inclusion logic, and it should be\\nbuilt on top of a basic test infra, at least.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 26 Oct 2022 15:56:07 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 26, 2022 at 03:56:07PM +0900, Michael Paquier wrote:\\n>\\n> So, I have spent a good portion of today looking at what you have\\n> here, applying 0001 and 0003 while fixing, rebasing and testing the\\n> whole, discarding 0002 (we could do more for the line number and\\n> source file in terms of the LOGs reported for a regexec failure).\\n\\nThanks!\\n\\n> Now remains 0004, which is the core of the proposal, and while it\\n> needs a rebase,\\n\\nHave you already done a rebase while working on the patch or are you intending\\nto take care of it, or should I? Let's no both do the work :)\\n\\n> - The TAP test, which is half the size of the patch in line number.\\n> Could it be possible to make it more edible, introducing a basic\\n> infrastructure to check a set of rules in pg_hba.conf and\\n> pg_ident.conf without the inclusion logic? Checks for error patterns\\n> (that I agree we strongly lack tests for) look like something we'd\\n> want to tackle independently of the inclusion logic, and it should be\\n> built on top of a basic test infra, at least.\\n\\nI don't mind taking care of that, but before doing so I'd like to have some\\nfeedback on whether you're ok with my approach (per my initial email about it\\nat [1]) or if you had some different\\nideas on how to do it.\\n\\n[1] https://www.postgresql.org/message-id/20220730080936.atyxodvwlmf2wnoc@jrouhaud\\n\\n\\n\",\n \"msg_date\": \"Wed, 26 Oct 2022 23:32:14 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 26, 2022 at 11:32:14PM +0800, Julien Rouhaud wrote:\\n> I don't mind taking care of that, but before doing so I'd like to have some\\n> feedback on whether you're ok with my approach (per my initial email about it\\n> at [1]) or if you had some different\\n> ideas on how to do it.\\n\\nPutting things afresh, there are two different things here (sorry I\\nneed to see that typed ;p):\\n1) How do we want to check reliably the loading of the HBA and ident\\nfiles on errors? EXEC_BACKEND would reload an entire new thing for\\neach connection, hence we need some loops to go through that.\\n2) How to check the contents of pg_hba_file_rules and\\npg_ident_file_mappings?\\n\\nThere is a dependency between 1) and 2) once we try to check for error\\npatterns in pg_hba_file_rules, because connections would just not\\nhappen. This is not the case for pg_ident_file_mappings though, so we\\ncould still test for buggy patterns in pg_ident.conf (or any of its \\nincluded parts) with some expected content of\\npg_ident_file_mappings.error after a successful connection.\\n\\nHmm. And what if we just gave up on the checks for error patterns in\\npg_hba_file_rules? One thing that we could do for this part is to\\ninclude all the buggy patterns we want to check at once in pg_hba.conf\\nin its included portions, then scan for all the logs produced after\\nattempting to start a server as the loading of pg_hba.conf would\\nproduce one LOG line with its CONTEXT for each buggy entry. The patch\\nchecks for error patterns with generate_log_err_rows(), but it looks\\nlike it would make the part 3 of the new test cleaner and easier to\\nmaintain in the long-term.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 27 Oct 2022 12:08:31 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Oct 26, 2022 at 11:32:14PM +0800, Julien Rouhaud wrote:\\n> Have you already done a rebase while working on the patch or are you intending\\n> to take care of it, or should I? Let's no both do the work :)\\n\\nSpoiler alert: I have not done a rebase yet ;)\\n--\\nMichael\",\n \"msg_date\": \"Thu, 27 Oct 2022 12:10:09 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Thu, Oct 27, 2022 at 12:08:31PM +0900, Michael Paquier wrote:\\n>\\n> Putting things afresh, there are two different things here (sorry I\\n> need to see that typed ;p):\\n> 1) How do we want to check reliably the loading of the HBA and ident\\n> files on errors?\\n\\nI guess you meant the failure to load HBA / ident files containing invalid\\ndata?\\n\\n> EXEC_BACKEND would reload an entire new thing for\\n> each connection, hence we need some loops to go through that.\\n> 2) How to check the contents of pg_hba_file_rules and\\n> pg_ident_file_mappings?\\n>\\n> There is a dependency between 1) and 2) once we try to check for error\\n> patterns in pg_hba_file_rules, because connections would just not\\n> happen. This is not the case for pg_ident_file_mappings though, so we\\n> could still test for buggy patterns in pg_ident.conf (or any of its\\n> included parts) with some expected content of\\n> pg_ident_file_mappings.error after a successful connection.\\n>\\n> Hmm. And what if we just gave up on the checks for error patterns in\\n> pg_hba_file_rules?\\n\\nWe discussed this problem in the past (1), and my understanding was that\\ndetecting a -DEXEC_BACKEND/Win32 build and skipping those tests in that case\\nwould be an acceptable solution to make sure there's at least some coverage.\\nThe proposed patch adds such an approach, making sure that the failure is due\\nto an invalid HBA file. If you changed you mind I can remove that part, but\\nagain I'd like to be sure of what you exactly want before starting to rewrite\\nstuff.\\n\\n> One thing that we could do for this part is to\\n> include all the buggy patterns we want to check at once in pg_hba.conf\\n> in its included portions, then scan for all the logs produced after\\n> attempting to start a server as the loading of pg_hba.conf would\\n> produce one LOG line with its CONTEXT for each buggy entry. The patch\\n> checks for error patterns with generate_log_err_rows(), but it looks\\n> like it would make the part 3 of the new test cleaner and easier to\\n> maintain in the long-term.\\n\\nI'm not sure what you mean here. The patch does check for all the errors\\nlooking at LOG lines and CONTEXT lines, but to make the regexp easier it\\ndoesn't try to make sure that each CONTEXT line is immediately following the\\nexpected LOG line.\\n\\nThat's why the errors are divided in 2 steps: a first step with a single error\\nusing some inclusion, so we can validate that the CONTEXT line is entirely\\ncorrect (wrt. line number and such), and then every possible error pattern\\nwhere we assume that the CONTEXT line are still following their LOG entry if\\nthey're found. It also has the knowledge of which errors adds a CONTEXT line\\nand which don't. And that's done twice, for HBA and ident.\\n\\nThe part 3 is just concatenating everything back, for HBA and ident. So\\nlong-term maintenance shouldn't get any harder as there won't be any need for\\nmore steps. We can just keep appending stuff in the 2nd step and all the tests\\nshould run as expected.\\n\\n[1] https://www.postgresql.org/message-id/YtZLeJPlMutL9heh%40paquier.xyz\\n\\n\\n\",\n \"msg_date\": \"Thu, 27 Oct 2022 12:26:25 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Thu, Oct 27, 2022 at 12:26:25PM +0800, Julien Rouhaud wrote:\\n> On Thu, Oct 27, 2022 at 12:08:31PM +0900, Michael Paquier wrote:\\n>>\\n>> Putting things afresh, there are two different things here (sorry I\\n>> need to see that typed ;p):\\n>> 1) How do we want to check reliably the loading of the HBA and ident\\n>> files on errors?\\n> \\n> I guess you meant the failure to load HBA / ident files containing invalid\\n> data?\\n\\nYeah.\\n\\n>> Hmm. And what if we just gave up on the checks for error patterns in\\n>> pg_hba_file_rules?\\n\\nOne part that I was thinking about when typing this part yesterday is\\nthat an EXEC_BACKEND build should work in non-WIN32 in TAP even if\\npg_ident.conf cannot be loaded, but I forgot entirely about the part\\nwhere we need a user mapping for the SSPI authentication on WIN32, as\\nset by pg_regress.\\n\\n> We discussed this problem in the past (1), and my understanding was that\\n> detecting a -DEXEC_BACKEND/Win32 build and skipping those tests in that case\\n> would be an acceptable solution to make sure there's at least some coverage.\\n> The proposed patch adds such an approach, making sure that the failure is due\\n> to an invalid HBA file. If you changed you mind I can remove that part, but\\n> again I'd like to be sure of what you exactly want before starting to rewrite\\n> stuff.\\n\\nI am still not completely sure what's the best way to do things here,\\nso let's do the following: let's keep the patch the way you think is\\nbetter for now (I may change my opinion on that but I'll hack that by\\nmyself anyway). Using what you have as a base, could you split the\\ntest and have it in its simplest to ease irs review? It would be able\\nto stress the buildfarm with a first version of the test and see how\\nit goes from there, and it is useful by itself IMO as HEAD has zero\\ncoverage for this area.\\n\\n> I'm not sure what you mean here. The patch does check for all the errors\\n> looking at LOG lines and CONTEXT lines, but to make the regexp easier it\\n> doesn't try to make sure that each CONTEXT line is immediately following the\\n> expected LOG line.\\n\\nHmm. Perhaps we'd better make sure that the LOG/CONTEXT link is\\nchecked? The context includes the line number while a generic\\nsentence, and the LOG provides all the details of the error\\nhappening.\\n\\n> That's why the errors are divided in 2 steps: a first step with a single error\\n> using some inclusion, so we can validate that the CONTEXT line is entirely\\n> correct (wrt. line number and such), and then every possible error pattern\\n> where we assume that the CONTEXT line are still following their LOG entry if\\n> they're found. It also has the knowledge of which errors adds a CONTEXT line\\n> and which don't. And that's done twice, for HBA and ident.\\n\\nOkay, so you do check the relationship between both, after all.\\n\\n> The part 3 is just concatenating everything back, for HBA and ident. So\\n> long-term maintenance shouldn't get any harder as there won't be any need for\\n> more steps. We can just keep appending stuff in the 2nd step and all the tests\\n> should run as expected.\\n\\nHmm. Okay.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 28 Oct 2022 10:24:23 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 28, 2022 at 10:24:23AM +0900, Michael Paquier wrote:\\n> On Thu, Oct 27, 2022 at 12:26:25PM +0800, Julien Rouhaud wrote:\\n>\\n> I am still not completely sure what's the best way to do things here,\\n> so let's do the following: let's keep the patch the way you think is\\n> better for now (I may change my opinion on that but I'll hack that by\\n> myself anyway). Using what you have as a base, could you split the\\n> test and have it in its simplest to ease irs review? It would be able\\n> to stress the buildfarm with a first version of the test and see how\\n> it goes from there, and it is useful by itself IMO as HEAD has zero\\n> coverage for this area.\\n\\nTo be honest I'd rather not to. It's excessively annoying to work on those\\ntests (I spent multiple days trying to make it as clean and readable as\\npossible), and splitting it to only test the current infrastructure will need\\nsome substantial efforts.\\n\\nBut more importantly, the next commit that will add tests for file inclusion\\nwill then be totally unmaintainable and unreadable, so that's IMO even worse.\\nI think it will probably either be the current file overwritten or a new one\\nwritten from scratch if some changes are done in the simplified test, and I'm\\nnot volunteering to do that.\\n\\n\\n\",\n \"msg_date\": \"Fri, 28 Oct 2022 11:49:54 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Fri, Oct 28, 2022 at 11:49:54AM +0800, Julien Rouhaud wrote:\\n> To be honest I'd rather not to. It's excessively annoying to work on those\\n> tests (I spent multiple days trying to make it as clean and readable as\\n> possible), and splitting it to only test the current infrastructure will need\\n> some substantial efforts.\\n\\nWell, I'd really like to split things as much as possible, beginning\\nwith some solid basics before extending its capabilities. This\\nreduces the odds of introducing issues in-between, particularly in\\nareas as sensible as authentication that involves not-yet-logged-in\\nusers.\\n\\n> But more importantly, the next commit that will add tests for file inclusion\\n> will then be totally unmaintainable and unreadable, so that's IMO even worse.\\n> I think it will probably either be the current file overwritten or a new one\\n> written from scratch if some changes are done in the simplified test, and I'm\\n> not volunteering to do that.\\n\\nNot sure about that either.\\n\\nAnyway, the last patch posted on this thread does not apply and the CF\\nbot fails, so it needed a rebase. I have first noticed that your\\npatch included some doc fixes independent of this thread, so I have\\napplied that as e765028 and backpatched it down to 15. The TAP test\\nneeds to be renamed to 0004, and it was missing from meson.build,\\nhence the CI was not testing it.\\n\\nI have spent some time starting at the logic today of the whole, and\\nGetConfFilesInDir() is really the first thing that stood out. I am\\nnot sure that it makes much sense to keep that in guc-file.c, TBH,\\nonce we feed it into hba.c. Perhaps we could put the refactored\\nroutine (and AbsoluteConfigLocation as a side effect) into a new file\\nin misc/?\\n\\nAs of HEAD, tokenize_inc_file() is the place where we handle a list of\\ntokens included with an '@' file, appending the existing set of\\nAuthTokens into the list we are building by grabbing a copy of these\\nbefore deleting the line memory context.\\n\\nYour patch proposes a different alternative, which is to pass down the\\nmemory context created in tokenize_auth_file() down to the callers\\nwith tokenize_file_with_context() dealing with all the internals.\\nprocess_included_auth_file() is different extension of that.\\nThis may come down to a matter of taste, but could be be cleaner to\\ntake an approach similar to tokenize_inc_file() and just create a copy\\nof the AuthToken list coming from a full file and append it to the\\nresult rather than passing around the memory context for the lines?\\nThis would lead to some simplifications, it seems, at least with the\\nnumber of arguments passed down across the layers.\\n\\nThe addition of a check for the depth in two places seems unnecessary,\\nand it looks like we should do this kind of check in only one place.\\nI have not tried yet, but if we actually move the AllocateFile() and\\nFreeFile() calls within tokenize_auth_file(), it looks like we may be\\nable to get to a simpler logic without the need of the with_context()\\nflavor (even no process_included_auth_file required)? That could make\\nthe interface easier to follow as a whole, while limiting the presence\\nof AllocateFile() and FreeFile() to a single code path, impacting\\nopen_inc_file() that relies on what the patch uses for\\nSecondaryAuthFile and IncludedAuthFile (we could attempt to use the\\nsame error message everywhere as well, as one could expect that\\nexpanded and included files have different names which is enough to\\nguess which one is an inclusion and which one is a secondary). \\n\\nAttached are two patches: the first one is a rebase of what you have\\nposted, and the second one are some changes I did while playing with\\nthe logic. In the second one, except for conffiles.{h.c}, the changes\\nare just a POC but that's to show the areas that I am planning to\\nrework, and your tests pass with it. I still need to think about all\\nthat and reconsider the design of the interface that would fit with\\nthe tokenization of the inclusions, without that many subroutines to\\ndo the work as it makes the code harder to follow. Well, that's to\\nsay that I am not staying idle :)\\n--\\nMichael\",\n \"msg_date\": \"Wed, 2 Nov 2022 16:46:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 02, 2022 at 04:46:48PM +0900, Michael Paquier wrote:\\n> On Fri, Oct 28, 2022 at 11:49:54AM +0800, Julien Rouhaud wrote:\\n> > To be honest I'd rather not to. It's excessively annoying to work on those\\n> > tests (I spent multiple days trying to make it as clean and readable as\\n> > possible), and splitting it to only test the current infrastructure will need\\n> > some substantial efforts.\\n>\\n> Well, I'd really like to split things as much as possible, beginning\\n> with some solid basics before extending its capabilities. This\\n> reduces the odds of introducing issues in-between, particularly in\\n> areas as sensible as authentication that involves not-yet-logged-in\\n> users.\\n>\\n> > But more importantly, the next commit that will add tests for file inclusion\\n> > will then be totally unmaintainable and unreadable, so that's IMO even worse.\\n> > I think it will probably either be the current file overwritten or a new one\\n> > written from scratch if some changes are done in the simplified test, and I'm\\n> > not volunteering to do that.\\n>\\n> Not sure about that either.\\n\\nMaybe one alternative approach would be to keep modifying the current test, and\\nonly do the split when committing the first part? The split itself should be\\neasy and mechanical: just remove everything unneeded (different file names\\nincluding the file name argument from the add_(hba|ident)_line, inclusion\\ndirective and so on). Even if the diff is then difficult to read it's not\\nreally a problem as it should have already been reviewed.\\n\\n> Anyway, the last patch posted on this thread does not apply and the CF\\n> bot fails, so it needed a rebase. I have first noticed that your\\n> patch included some doc fixes independent of this thread, so I have\\n> applied that as e765028 and backpatched it down to 15.\\n\\nYes I saw that this morning, thanks!\\n\\n> The TAP test\\n> needs to be renamed to 0004, and it was missing from meson.build,\\n> hence the CI was not testing it.\\n\\nAh right. This is quite annoying that we have to explicitly name every single\\ntest file there. For the source code it's hard to not notice a missed file and\\nyou will get an error in the CI, but a missed test is just entirely transparent\\n:(\\n\\n> I have spent some time starting at the logic today of the whole, and\\n> GetConfFilesInDir() is really the first thing that stood out. I am\\n> not sure that it makes much sense to keep that in guc-file.c, TBH,\\n> once we feed it into hba.c. Perhaps we could put the refactored\\n> routine (and AbsoluteConfigLocation as a side effect) into a new file\\n> in misc/?\\n\\nYes I was wondering about it too. A new fine in misc/ looks sensible.\\n\\n> As of HEAD, tokenize_inc_file() is the place where we handle a list of\\n> tokens included with an '@' file, appending the existing set of\\n> AuthTokens into the list we are building by grabbing a copy of these\\n> before deleting the line memory context.\\n>\\n> Your patch proposes a different alternative, which is to pass down the\\n> memory context created in tokenize_auth_file() down to the callers\\n> with tokenize_file_with_context() dealing with all the internals.\\n> process_included_auth_file() is different extension of that.\\n> This may come down to a matter of taste, but could be be cleaner to\\n> take an approach similar to tokenize_inc_file() and just create a copy\\n> of the AuthToken list coming from a full file and append it to the\\n> result rather than passing around the memory context for the lines?\\n> This would lead to some simplifications, it seems, at least with the\\n> number of arguments passed down across the layers.\\n\\nI guess it goes in the same direction as 8fea86830e1 where infrastructure to\\ncopy AuthTokens was added, I'm fine either way.\\n\\n> The addition of a check for the depth in two places seems unnecessary,\\n> and it looks like we should do this kind of check in only one place.\\n\\nI usually prefer to add a maybe unnecessary depth check since it's basically\\nfree rather than risking an unfriendly stack size error (including in possible\\nlater refactoring), but no objection to get rid of it.\\n\\n> I have not tried yet, but if we actually move the AllocateFile() and\\n> FreeFile() calls within tokenize_auth_file(), it looks like we may be\\n> able to get to a simpler logic without the need of the with_context()\\n> flavor (even no process_included_auth_file required)? That could make\\n> the interface easier to follow as a whole, while limiting the presence\\n> of AllocateFile() and FreeFile() to a single code path, impacting\\n> open_inc_file() that relies on what the patch uses for\\n> SecondaryAuthFile and IncludedAuthFile (we could attempt to use the\\n> same error message everywhere as well, as one could expect that\\n> expanded and included files have different names which is enough to\\n> guess which one is an inclusion and which one is a secondary).\\n\\nYou meant tokenize_auth_file_internal? Yes possibly, in general I tried\\nto avoid changing too much the existing code to ease the patch acceptance, but\\nI agree it would make things simpler.\\n\\n> Attached are two patches: the first one is a rebase of what you have\\n> posted, and the second one are some changes I did while playing with\\n> the logic. In the second one, except for conffiles.{h.c}, the changes\\n> are just a POC but that's to show the areas that I am planning to\\n> rework, and your tests pass with it. I still need to think about all\\n> that and reconsider the design of the interface that would fit with\\n> the tokenization of the inclusions, without that many subroutines to\\n> do the work as it makes the code harder to follow. Well, that's to\\n> say that I am not staying idle :)\\n\\nThanks!\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Nov 2022 21:06:02 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 02, 2022 at 09:06:02PM +0800, Julien Rouhaud wrote:\\n> Maybe one alternative approach would be to keep modifying the current test, and\\n> only do the split when committing the first part? The split itself should be\\n> easy and mechanical: just remove everything unneeded (different file names\\n> including the file name argument from the add_(hba|ident)_line, inclusion\\n> directive and so on). Even if the diff is then difficult to read it's not\\n> really a problem as it should have already been reviewed.\\n\\nI have not reviewed the test part much yet, TBH. The path\\nmanipulations because of WIN32 are a bit annoying. I was wondering if\\nwe could avoid all that by using the basenames, as one option.\\n\\n>> I have spent some time starting at the logic today of the whole, and\\n>> GetConfFilesInDir() is really the first thing that stood out. I am\\n>> not sure that it makes much sense to keep that in guc-file.c, TBH,\\n>> once we feed it into hba.c. Perhaps we could put the refactored\\n>> routine (and AbsoluteConfigLocation as a side effect) into a new file\\n>> in misc/?\\n> \\n> Yes I was wondering about it too. A new fine in misc/ looks sensible.\\n\\nconffiles.c is the best thing I could come up with, conf.c and\\nconfig.c being too generic and these are routines that work on\\nconfiguration files, so..\\n\\n>> Your patch proposes a different alternative, which is to pass down the\\n>> memory context created in tokenize_auth_file() down to the callers\\n>> with tokenize_file_with_context() dealing with all the internals.\\n>> process_included_auth_file() is different extension of that.\\n>> This may come down to a matter of taste, but could be be cleaner to\\n>> take an approach similar to tokenize_inc_file() and just create a copy\\n>> of the AuthToken list coming from a full file and append it to the\\n>> result rather than passing around the memory context for the lines?\\n>> This would lead to some simplifications, it seems, at least with the\\n>> number of arguments passed down across the layers.\\n> \\n> I guess it goes in the same direction as 8fea86830e1 where infrastructure to\\n> copy AuthTokens was added, I'm fine either way.\\n\\nI won't hide that I am trying to make the changes a maximum\\nincremental for this thread so as the final part is easy-ish.\\n\\n>> The addition of a check for the depth in two places seems unnecessary,\\n>> and it looks like we should do this kind of check in only one place.\\n> \\n> I usually prefer to add a maybe unnecessary depth check since it's basically\\n> free rather than risking an unfriendly stack size error (including in possible\\n> later refactoring), but no objection to get rid of it.\\n\\nThe depth check is a good idea, though I guess that there is an\\nargument for having it in only one code path, not two.\\n\\n>> I have not tried yet, but if we actually move the AllocateFile() and\\n>> FreeFile() calls within tokenize_auth_file(), it looks like we may be\\n>> able to get to a simpler logic without the need of the with_context()\\n>> flavor (even no process_included_auth_file required)? That could make\\n>> the interface easier to follow as a whole, while limiting the presence\\n>> of AllocateFile() and FreeFile() to a single code path, impacting\\n>> open_inc_file() that relies on what the patch uses for\\n>> SecondaryAuthFile and IncludedAuthFile (we could attempt to use the\\n>> same error message everywhere as well, as one could expect that\\n>> expanded and included files have different names which is enough to\\n>> guess which one is an inclusion and which one is a secondary).\\n> \\n> You meant tokenize_auth_file_internal? Yes possibly, in general I tried\\n> to avoid changing too much the existing code to ease the patch acceptance, but\\n> I agree it would make things simpler.\\n\\nI *guess* that my mind implied tokenize_auth_file() as of\\nyesterday's.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 3 Nov 2022 20:55:05 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 03, 2022 at 08:55:05PM +0900, Michael Paquier wrote:\\n> On Wed, Nov 02, 2022 at 09:06:02PM +0800, Julien Rouhaud wrote:\\n>>> The addition of a check for the depth in two places seems unnecessary,\\n>>> and it looks like we should do this kind of check in only one place.\\n>> \\n>> I usually prefer to add a maybe unnecessary depth check since it's basically\\n>> free rather than risking an unfriendly stack size error (including in possible\\n>> later refactoring), but no objection to get rid of it.\\n> \\n> The depth check is a good idea, though I guess that there is an\\n> argument for having it in only one code path, not two.\\n\\nSo, I have looked at this specific part, and I think that we should\\ntry to solve this issue on its own as the error message leading to a\\nfailure at startup is confusing now, giving:\\nFATAL: exceeded maxAllocatedDescs (166) while trying to open file \\\\\\\"blah\\\\\\\".\\n\\nIt took me some time to ponder about what kind of interface we should\\nuse for this job, but I think that it comes down to three things:\\n- The depth ought to be checked just before doing AllocateFile() for a\\nfile.\\n- We should have one code path calling AllocateFile().\\n- The same interface should be used for hba.c and hbafuncs.c.\\n\\nI have settled down to the following in the attached patch 0002:\\nopen_auth_file(const char *filename, int elevel, int depth, char\\n**err_msg)\\n\\nMore checks could be added in this code path (guc-file.l does more\\nwith immediate recursions, for example), but the depth level should be\\nenough anyway. I am aware of the fact that this reduces a bit the\\ninformation provided for what's called a \\\"secondary\\\" file in your\\npatch set, but I tend toward moving into something that's\\nminimalistic. Currently, we provide the file and its '@' shortcut,\\nand we don't mention at all the file where it comes from. This is a\\nminimal problem now because we only have pg_hba.conf to worry about\\nfor most users, still I am wondering whether it would be better in the\\nlong run to complete that with an error context that mentions the\\nfile(s) where this comes from and report a full chain of the events\\nthat led to this opening failure. At the end, I think that 0002 moves\\nus closer toward the goal of this thread. Not been a fan of\\nHbaIncludeKind since the patch was proposed, TBH, as well.\\n\\nAnother thing is that this removes the need for open_inc_file(), my\\nopinion is that it is overcomplicated, and actually it seems like\\nunder the non-strict mode (aka include_if_exists), we would generate a\\nLOG to mention that a configuration file is skipped, but there would\\nbe nothing in err_msg for the TokenizedAuthLine, which is a mistake I\\nguess?\\n\\n>>> I have not tried yet, but if we actually move the AllocateFile() and\\n>>> FreeFile() calls within tokenize_auth_file(), it looks like we may be\\n>>> able to get to a simpler logic without the need of the with_context()\\n>>> flavor (even no process_included_auth_file required)? That could make\\n>>> the interface easier to follow as a whole, while limiting the presence\\n>>> of AllocateFile() and FreeFile() to a single code path, impacting\\n>>> open_inc_file() that relies on what the patch uses for\\n>>> SecondaryAuthFile and IncludedAuthFile (we could attempt to use the\\n>>> same error message everywhere as well, as one could expect that\\n>>> expanded and included files have different names which is enough to\\n>>> guess which one is an inclusion and which one is a secondary).\\n>> \\n>> You meant tokenize_auth_file_internal? Yes possibly, in general I tried\\n>> to avoid changing too much the existing code to ease the patch acceptance, but\\n>> I agree it would make things simpler.\\n> \\n> I *guess* that my mind implied tokenize_auth_file() as of\\n> yesterday's.\\n\\nWhile thinking more about the logic, I saw that your approach to use\\nthe same MemoryContext \\\"linecxt\\\" and pass it down to the various\\nlayers with tokenize_file_with_context() has the advantage to remove\\nthe need to copy a list of TokenizedAuthLine coming from other\\nincluded files and/or dirs, which could easily lead to bugs especially\\nfor the sourcefile if someone is not careful enough. There is no need\\nto touch the existing tokenize_inc_file() that copies a set of\\nAuthTokens to an existing list in TokenizedAuthLine, of course.\\n\\nShould we do without tokenize_file_with_context() though? There is an\\nargument for making everything go through tokenize_auth_file(), having\\na MemoryContext as argument (aka if NULL create it, if not use it),\\nwhile still returning the memory context used? Similarly, it looks\\nlike we should have no need for process_included_authfile() at the\\nend.\\n\\nI have applied as a1a7bb8 the refactoring of the directory logic for\\nconfiguration files, and noticed that this could also be used for\\ntokenize_inc_file(). This reduces the whole patch by ~15%.\\n\\nAttached is a set of three patches:\\n- 0001 changes tokenize_inc_file() to use AbsoluteConfigLocation().\\nAbsoluteConfigLocation() uses a static buffer and a MAXPGPATH, but\\nwe'd rather change it to use a palloc()+strcpy() instead and remove\\nthe static restriction? What do you think? The same applies for the\\ncase where we use DataDir, actually, and it seems like there is no\\npoint in this path-length restriction in this code path.\\n- 0002 invents the interface to open auth files and check for their\\ndepths, simplifying the main patch a bit as there is no need to track\\nthe depth level here and there anymore.\\n- 0003 is the rebased patch, simplified after the other changes. The\\nbulk of the patch is in its TAP test.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 7 Nov 2022 15:07:15 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 07, 2022 at 03:07:15PM +0900, Michael Paquier wrote:\\n> Attached is a set of three patches:\\n> - 0001 changes tokenize_inc_file() to use AbsoluteConfigLocation().\\n> AbsoluteConfigLocation() uses a static buffer and a MAXPGPATH, but\\n> we'd rather change it to use a palloc()+strcpy() instead and remove\\n> the static restriction? What do you think? The same applies for the\\n> case where we use DataDir, actually, and it seems like there is no\\n> point in this path-length restriction in this code path.\\n> - 0002 invents the interface to open auth files and check for their\\n> depths, simplifying the main patch a bit as there is no need to track\\n> the depth level here and there anymore.\\n> - 0003 is the rebased patch, simplified after the other changes. The\\n> bulk of the patch is in its TAP test.\\n\\nCF bot unhappy as I have messed up with rules.out. Rebased. I have\\nremoved the restriction on MAXPGPATH in AbsoluteConfigLocation() in\\n0001, while on it. The absolute paths built on GUC or ident\\ninclusions are the same.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 8 Nov 2022 10:04:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 08, 2022 at 10:04:16AM +0900, Michael Paquier wrote:\\n> CF bot unhappy as I have messed up with rules.out. Rebased. I have\\n> removed the restriction on MAXPGPATH in AbsoluteConfigLocation() in\\n> 0001, while on it. The absolute paths built on GUC or ident\\n> inclusions are the same.\\n\\nRebased after 6bbd8b7g that is an equivalent of the previous 0001.\\nJulien, please note that this is waiting on author for now. What do\\nyou think about the now-named v18-0001 and the addition of an\\nErrorContextCallback to provide more information about the list of\\nincluded files on error?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 9 Nov 2022 09:51:17 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Wed, Nov 09, 2022 at 09:51:17AM +0900, Michael Paquier wrote:\\n> On Tue, Nov 08, 2022 at 10:04:16AM +0900, Michael Paquier wrote:\\n> > CF bot unhappy as I have messed up with rules.out. Rebased. I have\\n> > removed the restriction on MAXPGPATH in AbsoluteConfigLocation() in\\n> > 0001, while on it. The absolute paths built on GUC or ident\\n> > inclusions are the same.\\n>\\n> Rebased after 6bbd8b7g that is an equivalent of the previous 0001.\\n\\nThanks a lot!\\n\\n> Julien, please note that this is waiting on author for now. What do\\n> you think about the now-named v18-0001 and the addition of an\\n> ErrorContextCallback to provide more information about the list of\\n> included files on error?\\n\\nYes, I'm unfortunately fully aware that it's waiting on me. I've been a bit\\nbusy this week with $work but I will try to go back to it as soon as I can,\\nhopefully this week!\\n\\n\\n\",\n \"msg_date\": \"Wed, 9 Nov 2022 12:09:01 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 09, 2022 at 12:09:01PM +0800, Julien Rouhaud wrote:\\n> On Wed, Nov 09, 2022 at 09:51:17AM +0900, Michael Paquier wrote:\\n>> Julien, please note that this is waiting on author for now. What do\\n>> you think about the now-named v18-0001 and the addition of an\\n>> ErrorContextCallback to provide more information about the list of\\n>> included files on error?\\n> \\n> Yes, I'm unfortunately fully aware that it's waiting on me. I've been a bit\\n> busy this week with $work but I will try to go back to it as soon as I can,\\n> hopefully this week!\\n\\nFWIW, I have been playing with the addition of a ErrorContextCallback\\nin tokenize_auth_file(), and this addition leads to a really nice\\nresult. With this method, it is possible to know the full chain of\\nevents leading to a failure when tokenizing included files, which is\\nnot available now in the logs when reloading the server.\\n\\nWe could extend it to have more verbose information by passing more\\narguments to tokenize_auth_file(), still I'd like to think that just\\nknowing the line number and the full path to the file is more than\\nenough once you know the full chain of events. 0001 and 0002 ought to\\nbe merged together, but I am keeping these separate to show how simple\\nthe addition of the ErrorContextCallback is.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 10 Nov 2022 10:29:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 10, 2022 at 10:29:40AM +0900, Michael Paquier wrote:\\n>\\n> FWIW, I have been playing with the addition of a ErrorContextCallback\\n> in tokenize_auth_file(), and this addition leads to a really nice\\n> result. With this method, it is possible to know the full chain of\\n> events leading to a failure when tokenizing included files, which is\\n> not available now in the logs when reloading the server.\\n>\\n> We could extend it to have more verbose information by passing more\\n> arguments to tokenize_auth_file(), still I'd like to think that just\\n> knowing the line number and the full path to the file is more than\\n> enough once you know the full chain of events. 0001 and 0002 ought to\\n> be merged together, but I am keeping these separate to show how simple\\n> the addition of the ErrorContextCallback is.\\n\\nIt's looks good to me. I agree that file name and line number should be enough\\nto diagnose any unexpected error.\\n\\n\\n\",\n \"msg_date\": \"Sat, 12 Nov 2022 16:13:53 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 12, 2022 at 04:13:53PM +0800, Julien Rouhaud wrote:\\n> It's looks good to me. I agree that file name and line number should be enough\\n> to diagnose any unexpected error.\\n\\nThanks for checking. I have looked at 0001 and 0002 again with a\\nfresh mind, and applied both of them this morning.\\n\\nThis makes the remaining bits of the patch much easier to follow in\\nhba.c. Here are more comments after a closer review of the whole for\\nthe C logic.\\n\\n-MemoryContext\\n-tokenize_auth_file(const char *filename, FILE *file, List **tok_lines,\\n- int elevel, int depth)\\n+static void\\n+tokenize_file_with_context(MemoryContext linecxt, const char *filename,\\n\\nI really tend to prefer having one routine rather than two for the\\ntokenization entry point. Switching to the line context after setting\\nup the callback is better, and tokenize_file_with_context() does so.\\nAnyway, what do you think about having one API that gains a\\n\\\"MemoryContext *\\\" argument, as of the following:\\nvoid tokenize_auth_file(const char *filename, FILE *file,\\n List **tok_lines,\\n int depth, int elevel, MemoryContext *linectx)\\n\\nIf the caller passes NULL for *linectx as the initial line context,\\njust create it as we do now. If *linectx is not NULL, just reuse it.\\nThat may be cleaner than returning the created MemoryContext as\\nreturned result from tokenize_auth_file().\\n\\n+ /* Cumulate errors if any. */\\n+ if (err_msg)\\n+ {\\n+ if (err_buf.len > 0)\\n+ appendStringInfoChar(&err_buf, '\\\\n');\\n+ appendStringInfoString(&err_buf, err_msg);\\n+ }\\n\\nThis aggregates all the error messages for all the files included in a\\ngiven repository. As the patch stands, it seems to me that we would\\nget errors related to an include_dir clause for two cases:\\n- The specified path does not exist, in which case we have only one\\nerr_msg to consume and report back.\\n- Multiple failures in opening and/or reading included files.\\nIn the second case, aggregating the reports would provide a full set\\nof information, but that's not something a user would be able to act\\non directly as this is system-related. Or there is a case to know a\\nfull list of files in the case of multiple files that cannot be read\\nbecause of permission issues? We may be fine with just the first\\nsystem error here. Note that in the case where files can be read and\\nopened, these would have their own TokenizedAuthLines for each line\\nparsed, meaning one line in the SQL views once translated to an\\nHbaLine or an IdentLine.\\n\\nThis line of thoughts brings an interesting point, actually: there is\\nan inconsistency between \\\"include_if_exists\\\" and \\\"include\\\" compared to\\nthe GUC processing. As of the patch, if we use \\\"include\\\" on a file\\nthat does not exist, the tokenization logic would jump over it and\\ncontinue processing the follow-up entries anyway. This is a different\\npolicy than the GUCs, where we would immediately stop looking at\\nparameters after an \\\"include\\\" if it fails because its file does not\\nexist, working as a immediate stop in the processing. The difference\\nthat the patch brings between \\\"include_if_exists\\\" and \\\"include\\\" is\\nthat we report an error in one case but not the other, still skip the\\nfiles in both cases and move on with the rest. Hence my question,\\nshouldn't we do like the GUC processing for the hba and ident files,\\naka stop immediately when we fail to find a file on an \\\"include\\\"\\nclause? This would be equivalent to doing a \\\"break\\\" in\\ntokenize_file_with_context() after failing an include file.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 14 Nov 2022 14:40:37 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Mon, Nov 14, 2022 at 02:40:37PM +0900, Michael Paquier wrote:\\n> On Sat, Nov 12, 2022 at 04:13:53PM +0800, Julien Rouhaud wrote:\\n> > It's looks good to me. I agree that file name and line number should be enough\\n> > to diagnose any unexpected error.\\n>\\n> Thanks for checking. I have looked at 0001 and 0002 again with a\\n> fresh mind, and applied both of them this morning.\\n\\nThanks a lot!\\n\\n> This makes the remaining bits of the patch much easier to follow in\\n> hba.c. Here are more comments after a closer review of the whole for\\n> the C logic.\\n\\nAgreed.\\n\\n> -MemoryContext\\n> -tokenize_auth_file(const char *filename, FILE *file, List **tok_lines,\\n> - int elevel, int depth)\\n> +static void\\n> +tokenize_file_with_context(MemoryContext linecxt, const char *filename,\\n>\\n> I really tend to prefer having one routine rather than two for the\\n> tokenization entry point. Switching to the line context after setting\\n> up the callback is better, and tokenize_file_with_context() does so.\\n> Anyway, what do you think about having one API that gains a\\n> \\\"MemoryContext *\\\" argument, as of the following:\\n> void tokenize_auth_file(const char *filename, FILE *file,\\n> List **tok_lines,\\n> int depth, int elevel, MemoryContext *linectx)\\n>\\n> If the caller passes NULL for *linectx as the initial line context,\\n> just create it as we do now. If *linectx is not NULL, just reuse it.\\n> That may be cleaner than returning the created MemoryContext as\\n> returned result from tokenize_auth_file().\\n\\nI originally used two functions as it's a common pattern (e.g. ReadBuffer /\\nReadBufferExtended or all the *_internal versions of functions) and to avoid\\nunnecessary branch, but I agree that here having an extra branch is unlikely to\\nmake any measurable difference. It would only matter on -DEXEC_BACKEND /\\nwin32, and in that case the extra overhead (over an already expensive new\\nbackend mechanism) is way more than that, so +1 to keep things simple here.\\n\\n> + /* Cumulate errors if any. */\\n> + if (err_msg)\\n> + {\\n> + if (err_buf.len > 0)\\n> + appendStringInfoChar(&err_buf, '\\\\n');\\n> + appendStringInfoString(&err_buf, err_msg);\\n> + }\\n>\\n> This aggregates all the error messages for all the files included in a\\n> given repository. As the patch stands, it seems to me that we would\\n> get errors related to an include_dir clause for two cases:\\n> - The specified path does not exist, in which case we have only one\\n> err_msg to consume and report back.\\n> - Multiple failures in opening and/or reading included files.\\n> In the second case, aggregating the reports would provide a full set\\n> of information, but that's not something a user would be able to act\\n> on directly as this is system-related. Or there is a case to know a\\n> full list of files in the case of multiple files that cannot be read\\n> because of permission issues? We may be fine with just the first\\n> system error here. Note that in the case where files can be read and\\n> opened, these would have their own TokenizedAuthLines for each line\\n> parsed, meaning one line in the SQL views once translated to an\\n> HbaLine or an IdentLine.\\n\\nIf you have an include_dir directive and multiple files have wrong permission\\n(or maybe broken symlink or something like that), you will get multiple errors\\nwhen trying to process that single directive. I think it's friendlier to\\nreport as much detail as we can, so users can make sure they fix everything\\nrather than iterating over the first error. That's especially helpful if the\\nfix is done in some external tooling (puppet or whatever) rather than directly\\non the server.\\n>\\n> This line of thoughts brings an interesting point, actually: there is\\n> an inconsistency between \\\"include_if_exists\\\" and \\\"include\\\" compared to\\n> the GUC processing. As of the patch, if we use \\\"include\\\" on a file\\n> that does not exist, the tokenization logic would jump over it and\\n> continue processing the follow-up entries anyway. This is a different\\n> policy than the GUCs, where we would immediately stop looking at\\n> parameters after an \\\"include\\\" if it fails because its file does not\\n> exist, working as a immediate stop in the processing. The difference\\n> that the patch brings between \\\"include_if_exists\\\" and \\\"include\\\" is\\n> that we report an error in one case but not the other, still skip the\\n> files in both cases and move on with the rest. Hence my question,\\n> shouldn't we do like the GUC processing for the hba and ident files,\\n> aka stop immediately when we fail to find a file on an \\\"include\\\"\\n> clause? This would be equivalent to doing a \\\"break\\\" in\\n> tokenize_file_with_context() after failing an include file.\\n\\nI think that the problem is that we have the same interface for processing the\\nfiles on a startup/reload and for filling the views, so if we want the views to\\nbe helpful and report all errors we have to also allow a bogus \\\"include\\\" to\\ncontinue in the reload case too. The same problem doesn't exists for GUCs, so\\na slightly different behavior there might be acceptable.\\n\\n\\n\",\n \"msg_date\": \"Mon, 14 Nov 2022 15:47:27 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 14, 2022 at 03:47:27PM +0800, Julien Rouhaud wrote:\\n> On Mon, Nov 14, 2022 at 02:40:37PM +0900, Michael Paquier wrote:\\n>> If the caller passes NULL for *linectx as the initial line context,\\n>> just create it as we do now. If *linectx is not NULL, just reuse it.\\n>> That may be cleaner than returning the created MemoryContext as\\n>> returned result from tokenize_auth_file().\\n> \\n> I originally used two functions as it's a common pattern (e.g. ReadBuffer /\\n> ReadBufferExtended or all the *_internal versions of functions) and to avoid\\n> unnecessary branch, but I agree that here having an extra branch is unlikely to\\n> make any measurable difference. It would only matter on -DEXEC_BACKEND /\\n> win32, and in that case the extra overhead (over an already expensive new\\n> backend mechanism) is way more than that, so +1 to keep things simple here.\\n\\nOh, Okay. So that was your intention here.\\n\\n>> This aggregates all the error messages for all the files included in a\\n>> given repository. As the patch stands, it seems to me that we would\\n>> get errors related to an include_dir clause for two cases:\\n>> - The specified path does not exist, in which case we have only one\\n>> err_msg to consume and report back.\\n>> - Multiple failures in opening and/or reading included files.\\n>> In the second case, aggregating the reports would provide a full set\\n>> of information, but that's not something a user would be able to act\\n>> on directly as this is system-related. Or there is a case to know a\\n>> full list of files in the case of multiple files that cannot be read\\n>> because of permission issues? We may be fine with just the first\\n>> system error here. Note that in the case where files can be read and\\n>> opened, these would have their own TokenizedAuthLines for each line\\n>> parsed, meaning one line in the SQL views once translated to an\\n>> HbaLine or an IdentLine.\\n> \\n> If you have an include_dir directive and multiple files have wrong permission\\n> (or maybe broken symlink or something like that), you will get multiple errors\\n> when trying to process that single directive. I think it's friendlier to\\n> report as much detail as we can, so users can make sure they fix everything\\n> rather than iterating over the first error. That's especially helpful if the\\n> fix is done in some external tooling (puppet or whatever) rather than directly\\n> on the server.\\n\\nHmm, Okay. Well, this would include only errors on I/O or permission\\nproblems for existing files.. I have not seen deployments that have\\ndozens of sub-files for GUCs with an included dir, but perhaps I lack\\nof experience on user histories.\\n\\n> I think that the problem is that we have the same interface for processing the\\n> files on a startup/reload and for filling the views, so if we want the views to\\n> be helpful and report all errors we have to also allow a bogus \\\"include\\\" to\\n> continue in the reload case too. The same problem doesn't exists for GUCs, so\\n> a slightly different behavior there might be acceptable.\\n\\nWell, there is as well the point that we don't have yet a view for\\nGUCs that does the equivalent of HBA and ident, but that does not\\nmean, it seems to me, that there should be an inconsistency in the way\\nwe process those commands because one has implemented a feature but\\nnot the other. On the contrary, I'd rather try to make them\\nconsistent. As things are in the patch, the only difference between\\n\\\"include_if_exists\\\" and \\\"include\\\" is that the latter would report some\\ninformation if the file goes missing, the former generates a LOG entry\\nabout the file being skipped without something in the system view.\\nNow, wouldn't it be useful for the end-user to report that a file is\\nskipped as an effect of \\\"include_if_exists\\\" in the system views? If\\nnot, then what's the point of having this clause to begin with? My\\nopinion is that both clauses are useful, still on the ground of\\nconsistency both clauses should work the same as for GUCs. Both\\nshould report something in err_msg even if that's just about a file\\nbeing skipped, though I agree that this could be considered confusing\\nas well for an if_exists clause (does not look like a big deal to me\\nbased on the debuggability gain).\\n--\\nMichael\",\n \"msg_date\": \"Tue, 15 Nov 2022 08:46:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 15, 2022 at 08:46:55AM +0900, Michael Paquier wrote:\\n> On Mon, Nov 14, 2022 at 03:47:27PM +0800, Julien Rouhaud wrote:\\n> >\\n> > If you have an include_dir directive and multiple files have wrong permission\\n> > (or maybe broken symlink or something like that), you will get multiple errors\\n> > when trying to process that single directive. I think it's friendlier to\\n> > report as much detail as we can, so users can make sure they fix everything\\n> > rather than iterating over the first error. That's especially helpful if the\\n> > fix is done in some external tooling (puppet or whatever) rather than directly\\n> > on the server.\\n>\\n> Hmm, Okay. Well, this would include only errors on I/O or permission\\n> problems for existing files.. I have not seen deployments that have\\n> dozens of sub-files for GUCs with an included dir, but perhaps I lack\\n> of experience on user histories.\\n\\nWhile being the same inclusion infrastructure, it's likely that people will\\nhave different usage. I'm assuming that for GUCs the main usage is to have\\nyour automation tool put one of your template conf for instance (small_vm.conf,\\nbig_vm.conf or something like that) in an included directory, so you don't\\nreally need a lot of them. You can also rely on ALTER SYSTEM to avoid manually\\nhandling configuration files entirely.\\n\\nFor authentication there are probably very different pattern. The use case I\\nhad when writing this patch is some complex application that relies on many\\nservices, each service having dedicated role and authentication, and services\\ncan be enabled or disabled dynamically pretty much anytime. I had to write\\ncode to merge possibly new entries with existing pg_hba/pg_ident configuration\\nfiles. With this feature it would be much easier (and robust) to simply have a\\nmain pg_hba/pg_ident that includes some directory, and have the service\\nenable/disable simply creates/removes a dedicated file for each service, and we\\nthen would usually have at least a dozen files there.\\n\\nI'm assuming people doing multi-tenant can also have similar usage.\\n\\n> > I think that the problem is that we have the same interface for processing the\\n> > files on a startup/reload and for filling the views, so if we want the views to\\n> > be helpful and report all errors we have to also allow a bogus \\\"include\\\" to\\n> > continue in the reload case too. The same problem doesn't exists for GUCs, so\\n> > a slightly different behavior there might be acceptable.\\n>\\n> Well, there is as well the point that we don't have yet a view for\\n> GUCs that does the equivalent of HBA and ident,\\n\\nYes that's what I meant by \\\"the same problem doesn't exists for GUCs\\\".\\n\\n> but that does not\\n> mean, it seems to me, that there should be an inconsistency in the way\\n> we process those commands because one has implemented a feature but\\n> not the other. On the contrary, I'd rather try to make them\\n> consistent.\\n\\nYou mean stopping at the first error, even if it's only for the view reporting?\\nThat will make the reload consistent, but the view will be a bit useless then.\\n\\n> As things are in the patch, the only difference between\\n> \\\"include_if_exists\\\" and \\\"include\\\" is that the latter would report some\\n> information if the file goes missing, the former generates a LOG entry\\n> about the file being skipped without something in the system view.\\n> Now, wouldn't it be useful for the end-user to report that a file is\\n> skipped as an effect of \\\"include_if_exists\\\" in the system views? If\\n> not, then what's the point of having this clause to begin with?\\n\\nI don't really get the argument \\\"the proposed monitoring view doesn't give this\\ninformation, so the feature isn't needed\\\". Also, unless I'm missing something\\nthe other difference between \\\"include\\\" and \\\"include_if_exists\\\" is that the\\nfirst one will refuse to reload the conf (or start the server) if the file is\\nmissing, while the other one will?\\n\\n> My\\n> opinion is that both clauses are useful, still on the ground of\\n> consistency both clauses should work the same as for GUCs. Both\\n> should report something in err_msg even if that's just about a file\\n> being skipped, though I agree that this could be considered confusing\\n> as well for an if_exists clause (does not look like a big deal to me\\n> based on the debuggability gain).\\n\\nIt would be nice to have the information that an \\\"include_if_exists\\\" file\\ndidn't exist, but having a log-level message in the \\\"error\\\" column is a clear\\nPOLA violation. People will probably just do something like\\n\\nSELECT file_name, line_number, error FROM pg_hba_file_rules WHERE error IS NOT\\nNULL;\\n\\nand report an error if any row is found. Having to parse the error field to\\nknow if that's really an error or not is going to be a huge foot-gun. Maybe we\\ncould indeed report the problem in err_msg but for include_if_exists display it\\nin some other column of the view?\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Nov 2022 10:53:02 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"2022年11月14日(月) 14:41 Michael Paquier <michael@paquier.xyz>:\\n>\\n> On Sat, Nov 12, 2022 at 04:13:53PM +0800, Julien Rouhaud wrote:\\n> > It's looks good to me. I agree that file name and line number should be enough\\n> > to diagnose any unexpected error.\\n>\\n> Thanks for checking. I have looked at 0001 and 0002 again with a\\n> fresh mind, and applied both of them this morning.\\n\\nHi\\n\\nJust a quick note to mention that 0003 adds a new test (\\\"004_file_inclusion.pl\\\")\\nbut doesn't seem to include it in the \\\"meson.build\\\" file.\\n\\nRegards\\n\\nIan Barwick\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Nov 2022 14:01:06 +0900\",\n \"msg_from\": \"Ian Lawrence Barwick <barwick@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Le mer. 16 nov. 2022 à 13:01, Ian Lawrence Barwick <barwick@gmail.com> a\\nécrit :\\n\\n> 2022年11月14日(月) 14:41 Michael Paquier <michael@paquier.xyz>:\\n> >\\n> > On Sat, Nov 12, 2022 at 04:13:53PM +0800, Julien Rouhaud wrote:\\n> > > It's looks good to me. I agree that file name and line number should\\n> be enough\\n> > > to diagnose any unexpected error.\\n> >\\n> > Thanks for checking. I have looked at 0001 and 0002 again with a\\n> > fresh mind, and applied both of them this morning.\\n>\\n> Hi\\n>\\n> Just a quick note to mention that 0003 adds a new test (\\\"\\n> 004_file_inclusion.pl\\\")\\n> but doesn't seem to include it in the \\\"meson.build\\\" file.\\n>\\n\\nah thanks! Michael also mentioned that previously but as we were focusing\\non preliminary refactoring patches I didn't check if if was fixed in the\\npatches sent recently, but I will definitely take care of it for the next\\nround\\n\\n>\\n\\nLe mer. 16 nov. 2022 à 13:01, Ian Lawrence Barwick <barwick@gmail.com> a écrit :2022年11月14日(月) 14:41 Michael Paquier <michael@paquier.xyz>:\\n>\\n> On Sat, Nov 12, 2022 at 04:13:53PM +0800, Julien Rouhaud wrote:\\n> > It's looks good to me. I agree that file name and line number should be enough\\n> > to diagnose any unexpected error.\\n>\\n> Thanks for checking. I have looked at 0001 and 0002 again with a\\n> fresh mind, and applied both of them this morning.\\n\\nHi\\n\\nJust a quick note to mention that 0003 adds a new test (\\\"004_file_inclusion.pl\\\")\\nbut doesn't seem to include it in the \\\"meson.build\\\" file.ah thanks! Michael also mentioned that previously but as we were focusing on preliminary refactoring patches I didn't check if if was fixed in the patches sent recently, but I will definitely take care of it for the next round\",\n \"msg_date\": \"Wed, 16 Nov 2022 13:30:39 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 16, 2022 at 10:53:02AM +0800, Julien Rouhaud wrote:\\n> While being the same inclusion infrastructure, it's likely that people will\\n> have different usage. I'm assuming that for GUCs the main usage is to have\\n> your automation tool put one of your template conf for instance (small_vm.conf,\\n> big_vm.conf or something like that) in an included directory, so you don't\\n> really need a lot of them. You can also rely on ALTER SYSTEM to avoid manually\\n> handling configuration files entirely.\\n> \\n> For authentication there are probably very different pattern. The use case I\\n> had when writing this patch is some complex application that relies on many\\n> services, each service having dedicated role and authentication, and services\\n> can be enabled or disabled dynamically pretty much anytime. I had to write\\n> code to merge possibly new entries with existing pg_hba/pg_ident configuration\\n> files. With this feature it would be much easier (and robust) to simply have a\\n> main pg_hba/pg_ident that includes some directory, and have the service\\n> enable/disable simply creates/removes a dedicated file for each service, and we\\n> then would usually have at least a dozen files there.\\n> \\n> I'm assuming people doing multi-tenant can also have similar usage.\\n\\nSo you main application for HBA/ident is include_dir/.. For GUCs, we\\nwould just skip the directory if it has no files, but the directory\\nhas to exist. Your patch is behaving the same.\\n\\n>> but that does not\\n>> mean, it seems to me, that there should be an inconsistency in the way\\n>> we process those commands because one has implemented a feature but\\n>> not the other. On the contrary, I'd rather try to make them\\n>> consistent.\\n> \\n> You mean stopping at the first error, even if it's only for the view reporting?\\n> That will make the reload consistent, but the view will be a bit useless then.\\n\\nYep, I meant to stop the generation of the TokenizedAuthLines at the\\nfirst inclusion error by letting tokenize_auth_file() return a status\\nto be able to stop the recursions. But after some second-thoughts\\npondering about this, I see why I am wrong and why you are right. As\\nyou say, stopping the generation of the TokenizedAuthLines would just\\nlimit the amount of data reported at once in the view, the golden rule\\nfor HBA/ident being that we would reload nothing as long as there is\\nat least one error reported when parsing things. So giving more\\ncontrol to tokenize_auth_file() to stop a recursion, say by making it\\nreturn a boolean status, makes little sense.\\n\\n> It would be nice to have the information that an \\\"include_if_exists\\\" file\\n> didn't exist, but having a log-level message in the \\\"error\\\" column is a clear\\n> POLA violation. People will probably just do something like\\n> \\n> SELECT file_name, line_number, error FROM pg_hba_file_rules WHERE error IS NOT\\n> NULL;\\n> \\n> and report an error if any row is found. Having to parse the error field to\\n> know if that's really an error or not is going to be a huge foot-gun. Maybe we\\n> could indeed report the problem in err_msg but for include_if_exists display it\\n> in some other column of the view?\\n\\nHmm. One possibility would be to add a direct mention to the\\n\\\"include\\\", \\\"include_dir\\\" and \\\"include_if_exists\\\" clauses through\\npg_hba_file_rules.type? I don't see how to do that without making the\\npatch much more invasive as per the existing separation between\\ntokenization and Hba/IdentLine filling, though, and perhaps the error \\nprovided with the full path to the file would provide enough context\\nfor one to know if the failure is happening on an included file for\\ndatabase/user lists or a full file from an \\\"include\\\" clause. It also\\nmeans that include_if_exists remains transparent all the time.\\nSimpler may be for the best here, at the end. \\n\\nBy the way, I am wondering whether process_included_authfile() is\\nthe most intuitive interface here. The only thing that prevents a\\ncommon routine to process the include commands is the failure on\\nGetConfFilesInDir(), where we need to build a TokenizedAuthLine when\\nthe others have already done so tokenize_file_with_context(). Could\\nit be cleaner to have a small routine like makeTokenizedAuthLine()\\nthat gets reused when we fail scanning a directory to build a\\nTokenizedAuthLine, in combination with a small-ish routine working on\\na directory like ParseConfigDirectory() but for HBA/ident? Or we\\ncould just drop process_included_authfile() entirely? On failure,\\nthis would make the code do a next_line all the time for all the\\ninclude clauses.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 17 Nov 2022 11:33:05 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 17, 2022 at 11:33:05AM +0900, Michael Paquier wrote:\\n> By the way, I am wondering whether process_included_authfile() is\\n> the most intuitive interface here. The only thing that prevents a\\n> common routine to process the include commands is the failure on\\n> GetConfFilesInDir(), where we need to build a TokenizedAuthLine when\\n> the others have already done so tokenize_file_with_context(). Could\\n> it be cleaner to have a small routine like makeTokenizedAuthLine()\\n> that gets reused when we fail scanning a directory to build a\\n> TokenizedAuthLine, in combination with a small-ish routine working on\\n> a directory like ParseConfigDirectory() but for HBA/ident? Or we\\n> could just drop process_included_authfile() entirely? On failure,\\n> this would make the code do a next_line all the time for all the\\n> include clauses.\\n\\nI have been waiting for your reply for some time, so I have taken some\\nto look at this patch by myself and hacked on it.\\n\\nAt the end, the thing I was not really happy about is the\\nMemoryContext used to store the set of TokenizedAuthLines. I have\\nlooked at a couple of approaches, like passing around the context as\\nyou do, but at the end there is something I found annoying: we may\\ntokenize a file in the line context of a different file, storing in\\nmuch more data than just the TokenizedAuthLines. Then I got a few\\nsteps back and began using a static memory context that only stored\\nthe TokenizedAuthLines, switching to it in one place as of the end of\\ntokenize_auth_file() when inserting an item. This makes hbafuncs.c a\\nbit less aware of the memory context, but I think that we could live\\nwith that with a cleaner tokenization interface. That's close to what\\nGUCs do, in some way. Note that this may be better as an independent\\npatch, actually, as it impacts the current @ inclusions.\\n\\nI have noticed a bug in the logic of include_if_exists: we should\\nignore only files on ENOENT, but complain for other errnos after\\nopening the file.\\n\\nThe docs and the sample files have been tweaked a bit, giving a\\ncleaner separation between the main record types and the inclusion\\nones.\\n\\n+ /* XXX: this should stick to elevel for some cases? */\\n+ ereport(LOG,\\n+ (errmsg(\\\"skipping missing authentication file \\\\\\\"%s\\\\\\\"\\\",\\n+ inc_fullname)));\\nShould we always issue a LOG here? In some cases we use an elevel of\\nDEBUG3.\\n\\nSo, what do you think about something like the attached? I have begun\\na lookup at the tests, but I don't have enough material for an actual\\nreview of this part yet. Note that I have removed temporarily\\nprocess_included_authfile(), as I was looking at all the code branches\\nin details. The final result ought to include AbsoluteConfigLocation(),\\nopen_auth_file() and tokenize_auth_file() with an extra missing_ok\\nargument, I guess (\\\"strict\\\" as proposed originally is not the usual\\nPG-way for ENOENT-ish problems). process_line should be used only\\nwhen we have no err_msg, meaning that these have been consumed in some\\nTokenizedAuthLines already.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 22 Nov 2022 17:20:01 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 22, 2022 at 05:20:01PM +0900, Michael Paquier wrote:\\n> + /* XXX: this should stick to elevel for some cases? */\\n> + ereport(LOG,\\n> + (errmsg(\\\"skipping missing authentication file \\\\\\\"%s\\\\\\\"\\\",\\n> + inc_fullname)));\\n> Should we always issue a LOG here? In some cases we use an elevel of\\n> DEBUG3.\\n\\nAnd here I think that we need to use elevel. In hba.c, this would\\ngenerate a LOG while hbafuncs.c uses DEBUG3, leading to less noise in\\nthe log files.\\n\\n> So, what do you think about something like the attached? I have begun\\n> a lookup at the tests, but I don't have enough material for an actual\\n> review of this part yet. Note that I have removed temporarily\\n> process_included_authfile(), as I was looking at all the code branches\\n> in details. The final result ought to include AbsoluteConfigLocation(),\\n> open_auth_file() and tokenize_auth_file() with an extra missing_ok\\n> argument, I guess (\\\"strict\\\" as proposed originally is not the usual\\n> PG-way for ENOENT-ish problems). process_line should be used only\\n> when we have no err_msg, meaning that these have been consumed in some\\n> TokenizedAuthLines already.\\n\\nThis, however, was still brittle in terms of memory handling.\\nReloading the server a few hundred times proved that this was leaking\\nmemory in the tokenization. This becomes quite simple once you switch\\nto an approach where tokenize_auth_file() uses its own memory context,\\nwhile we store all the TokenizedAuthLines in the static context. It\\nalso occurred to me that we can create and drop the static context\\nholding the tokens when opening the top-root auth file, aka with a\\ndepth at 0. It may be a bit simpler to switch to a single-context\\napproach for all the allocations of tokenize_auth_file(), though I\\nfind the use of a static context much cleaner for the inclusions with\\n@ files when we expand an existing list.\\n\\nThe patch can be split, again, into more pieces. Attached 0001\\nreworks the memory allocation contexts for the tokenization and 0002\\nis the main feature. As of things are, I am quite happy with the\\nshapes of 0001 and 0002. I have tested quite a bit its robustness,\\nwith valgrind for example, to make sure that there are no leaks in the\\npostmaster (with[out] -DEXEC_BACKEND). The inclusion logic is\\nrefactored to be in a state close to your previous patches, see\\ntokenize_inclusion_file().\\n\\nNote that the tests are failing for some of the Windows CIs, actually,\\ndue to a difference in some of the paths generated vs the file paths\\n(aka c:\\\\cirrus vs c:/cirrus, as far as I saw).\\n--\\nMichael\",\n \"msg_date\": \"Wed, 23 Nov 2022 15:05:18 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nSorry for the very late answer, I had quite a lot of other things going on\\nrecently. And thanks for taking care of the patchset!\\n\\nOn Wed, Nov 23, 2022 at 03:05:18PM +0900, Michael Paquier wrote:\\n> On Tue, Nov 22, 2022 at 05:20:01PM +0900, Michael Paquier wrote:\\n> > + /* XXX: this should stick to elevel for some cases? */\\n> > + ereport(LOG,\\n> > + (errmsg(\\\"skipping missing authentication file \\\\\\\"%s\\\\\\\"\\\",\\n> > + inc_fullname)));\\n> > Should we always issue a LOG here? In some cases we use an elevel of\\n> > DEBUG3.\\n>\\n> And here I think that we need to use elevel. In hba.c, this would\\n> generate a LOG while hbafuncs.c uses DEBUG3, leading to less noise in\\n> the log files.\\n\\nYeah I agree the LOG message is only interesting if you're reloading the conf.\\nI actually thought that this is what I did sorry about that.\\n\\n> > So, what do you think about something like the attached? I have begun\\n> > a lookup at the tests, but I don't have enough material for an actual\\n> > review of this part yet. Note that I have removed temporarily\\n> > process_included_authfile(), as I was looking at all the code branches\\n> > in details. The final result ought to include AbsoluteConfigLocation(),\\n> > open_auth_file() and tokenize_auth_file() with an extra missing_ok\\n> > argument, I guess (\\\"strict\\\" as proposed originally is not the usual\\n> > PG-way for ENOENT-ish problems). process_line should be used only\\n> > when we have no err_msg, meaning that these have been consumed in some\\n> > TokenizedAuthLines already.\\n>\\n> This, however, was still brittle in terms of memory handling.\\n> Reloading the server a few hundred times proved that this was leaking\\n> memory in the tokenization.\\n\\nOh, nice catch.\\n\\n> This becomes quite simple once you switch\\n> to an approach where tokenize_auth_file() uses its own memory context,\\n> while we store all the TokenizedAuthLines in the static context. It\\n> also occurred to me that we can create and drop the static context\\n> holding the tokens when opening the top-root auth file, aka with a\\n> depth at 0. It may be a bit simpler to switch to a single-context\\n> approach for all the allocations of tokenize_auth_file(), though I\\n> find the use of a static context much cleaner for the inclusions with\\n> @ files when we expand an existing list.\\n\\nAgreed.\\n\\nThe depth 0 is getting used quite a lot now, maybe we should have a define for\\nit to make it easier to grep, like TOP_LEVEL_AUTH_FILE or something like that?\\nAnd also add a define for the magical 10 for the max inclusion depth, for both\\nauth files and GUC files while at it?\\n\\n> The patch can be split, again, into more pieces. Attached 0001\\n> reworks the memory allocation contexts for the tokenization and 0002\\n> is the main feature. As of things are, I am quite happy with the\\n> shapes of 0001 and 0002. I have tested quite a bit its robustness,\\n> with valgrind for example, to make sure that there are no leaks in the\\n> postmaster (with[out] -DEXEC_BACKEND). The inclusion logic is\\n> refactored to be in a state close to your previous patches, see\\n> tokenize_inclusion_file().\\n\\nYep I saw that. I don't have much to say about the patch, it looks good to me.\\nThe only nitpicking I have is:\\n\\n+/*\\n+ * Memory context holding the list of TokenizedAuthLines when parsing\\n+ * HBA or ident config files. This is created when loading the top HBA\\n+ + or ident file (depth of 0).\\n+ */\\n+static MemoryContext tokenize_context = NULL;\\n\\nThe comment seems a bit ambiguous as with \\\"loading the top...\\\" you probably\\nmeant something like \\\"loading the file in memory\\\" rather than \\\"(re)loading the\\nconfiguration\\\". Maybe s/loading/opening/?\\n>\\n> Note that the tests are failing for some of the Windows CIs, actually,\\n> due to a difference in some of the paths generated vs the file paths\\n> (aka c:\\\\cirrus vs c:/cirrus, as far as I saw).\\n\\nMmm, I haven't looked deeply so I'm not sure if the perl podules are aware of\\nit or not, but maybe we could somehow detect the used delimiter at the\\nbeginning after normalizing the directory, and use a $DELIM rather than a plain\\n\\\"/\\\"?\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Nov 2022 15:56:50 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 23, 2022 at 03:56:50PM +0800, Julien Rouhaud wrote:\\n> The depth 0 is getting used quite a lot now, maybe we should have a define for\\n> it to make it easier to grep, like TOP_LEVEL_AUTH_FILE or something like that?\\n> And also add a define for the magical 10 for the max inclusion depth, for both\\n> auth files and GUC files while at it?\\n\\nSounds like a good idea to me, and it seems to me that this had better\\nbe unified between the GUCs (see ParseConfigFp() that hardcodes a\\ndepth of 0) and hba.c. It looks like they could be added to\\nconffiles.h, as of CONF_FILE_START_{LEVEL,DEPTH} and\\nCONF_FILE_MAX_{LEVEL,DEPTH}. Would you like to send a patch?\\n\\n> The comment seems a bit ambiguous as with \\\"loading the top...\\\" you probably\\n> meant something like \\\"loading the file in memory\\\" rather than \\\"(re)loading the\\n> configuration\\\". Maybe s/loading/opening/?\\n\\nRight. I have used \\\"opening\\\" at the end.\\n\\nI have worked on all that, did more polishing to the docs, the\\ncomments and some tiny bits of the logic, and applied both 0001 and\\n0002. Now, to the tests..\\n\\n> Mmm, I haven't looked deeply so I'm not sure if the perl podules are aware of\\n> it or not, but maybe we could somehow detect the used delimiter at the\\n> beginning after normalizing the directory, and use a $DELIM rather than a plain\\n> \\\"/\\\"?\\n\\nI am not sure. Could you have a look and see if you can get the CI\\nback to green? The first thing I would test is to switch the error\\npatterns to be regexps based on the basenames rather than the full\\npaths (tweaking the queries on the system views to do htat), so as we\\navoid all this business with slash and backslash transformations.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 24 Nov 2022 14:07:21 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 24, 2022 at 02:07:21PM +0900, Michael Paquier wrote:\\n> On Wed, Nov 23, 2022 at 03:56:50PM +0800, Julien Rouhaud wrote:\\n> > The depth 0 is getting used quite a lot now, maybe we should have a define for\\n> > it to make it easier to grep, like TOP_LEVEL_AUTH_FILE or something like that?\\n> > And also add a define for the magical 10 for the max inclusion depth, for both\\n> > auth files and GUC files while at it?\\n> \\n> Sounds like a good idea to me, and it seems to me that this had better\\n> be unified between the GUCs (see ParseConfigFp() that hardcodes a\\n> depth of 0) and hba.c. It looks like they could be added to\\n> conffiles.h, as of CONF_FILE_START_{LEVEL,DEPTH} and\\n> CONF_FILE_MAX_{LEVEL,DEPTH}. Would you like to send a patch?\\n\\nAgreed, and yes I will take care of that shortly.\\n\\n> > The comment seems a bit ambiguous as with \\\"loading the top...\\\" you probably\\n> > meant something like \\\"loading the file in memory\\\" rather than \\\"(re)loading the\\n> > configuration\\\". Maybe s/loading/opening/?\\n> \\n> Right. I have used \\\"opening\\\" at the end.\\n> \\n> I have worked on all that, did more polishing to the docs, the\\n> comments and some tiny bits of the logic, and applied both 0001 and\\n> 0002.\\n\\nThanks a lot!\\n\\n> Now, to the tests..\\n> \\n> > Mmm, I haven't looked deeply so I'm not sure if the perl podules are aware of\\n> > it or not, but maybe we could somehow detect the used delimiter at the\\n> > beginning after normalizing the directory, and use a $DELIM rather than a plain\\n> > \\\"/\\\"?\\n> \\n> I am not sure. Could you have a look and see if you can get the CI\\n> back to green? The first thing I would test is to switch the error\\n> patterns to be regexps based on the basenames rather than the full\\n> paths (tweaking the queries on the system views to do htat), so as we\\n> avoid all this business with slash and backslash transformations.\\n\\nI'm also working on it! Hopefully I should be able to come up with a fix soon.\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Nov 2022 14:37:23 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Thu, Nov 24, 2022 at 02:37:23PM +0800, Julien Rouhaud wrote:\\n> On Thu, Nov 24, 2022 at 02:07:21PM +0900, Michael Paquier wrote:\\n> > On Wed, Nov 23, 2022 at 03:56:50PM +0800, Julien Rouhaud wrote:\\n> > > The depth 0 is getting used quite a lot now, maybe we should have a define for\\n> > > it to make it easier to grep, like TOP_LEVEL_AUTH_FILE or something like that?\\n> > > And also add a define for the magical 10 for the max inclusion depth, for both\\n> > > auth files and GUC files while at it?\\n> > \\n> > Sounds like a good idea to me, and it seems to me that this had better\\n> > be unified between the GUCs (see ParseConfigFp() that hardcodes a\\n> > depth of 0) and hba.c. It looks like they could be added to\\n> > conffiles.h, as of CONF_FILE_START_{LEVEL,DEPTH} and\\n> > CONF_FILE_MAX_{LEVEL,DEPTH}. Would you like to send a patch?\\n\\nSo I went with CONF_FILE_START_DEPTH and CONF_FILE_MAX_DEPTH. Attached v22\\nthat fixes it in all the places I found.\\n\\n> > Now, to the tests..\\n> > \\n> > > Mmm, I haven't looked deeply so I'm not sure if the perl podules are aware of\\n> > > it or not, but maybe we could somehow detect the used delimiter at the\\n> > > beginning after normalizing the directory, and use a $DELIM rather than a plain\\n> > > \\\"/\\\"?\\n> > \\n> > I am not sure. Could you have a look and see if you can get the CI\\n> > back to green? The first thing I would test is to switch the error\\n> > patterns to be regexps based on the basenames rather than the full\\n> > paths (tweaking the queries on the system views to do htat), so as we\\n> > avoid all this business with slash and backslash transformations.\\n\\nApparently just making sure that the $node->data_dir consistently uses forward\\nslashes is enough to make the CI happy, for VS 2019 [1] and MinGW64 [2], so\\ndone this way with an extra normalization step.\\n\\n[1] https://cirrus-ci.com/task/4944203946917888\\n[2] https://cirrus-ci.com/task/6070103853760512\",\n \"msg_date\": \"Thu, 24 Nov 2022 17:07:24 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 24, 2022 at 05:07:24PM +0800, Julien Rouhaud wrote:\\n> So I went with CONF_FILE_START_DEPTH and CONF_FILE_MAX_DEPTH. Attached v22\\n> that fixes it in all the places I found.\\n\\nSounds fine. Added one comment, fixed one comment, and applied.\\nThanks!\\n--\\nMichael\",\n \"msg_date\": \"Fri, 25 Nov 2022 07:41:59 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Fri, Nov 25, 2022 at 07:41:59AM +0900, Michael Paquier wrote:\\n> On Thu, Nov 24, 2022 at 05:07:24PM +0800, Julien Rouhaud wrote:\\n> > So I went with CONF_FILE_START_DEPTH and CONF_FILE_MAX_DEPTH. Attached v22\\n> > that fixes it in all the places I found.\\n> \\n> Sounds fine. Added one comment, fixed one comment, and applied.\\n> Thanks!\\n\\nThanks a lot!\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Nov 2022 10:25:30 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"2022年11月25日(金) 11:25 Julien Rouhaud <rjuju123@gmail.com>:\\n>\\n> On Fri, Nov 25, 2022 at 07:41:59AM +0900, Michael Paquier wrote:\\n> > On Thu, Nov 24, 2022 at 05:07:24PM +0800, Julien Rouhaud wrote:\\n> > > So I went with CONF_FILE_START_DEPTH and CONF_FILE_MAX_DEPTH. Attached v22\\n> > > that fixes it in all the places I found.\\n> >\\n> > Sounds fine. Added one comment, fixed one comment, and applied.\\n> > Thanks!\\n>\\n> Thanks a lot!\\n\\nHi\\n\\nI'm trying to reconcile open CommitFest entries with their actual\\nstatus; the entry for this:\\n\\n https://commitfest.postgresql.org/40/3558/\\n\\nshows as \\\"Waiting on Author\\\", but looks like it's all been committed;\\nis there anything\\nleft to do on this?\\n\\nThanks\\n\\nIan Barwick\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Nov 2022 16:31:38 +0900\",\n \"msg_from\": \"Ian Lawrence Barwick <barwick@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"Le dim. 27 nov. 2022 à 15:31, Ian Lawrence Barwick\\n\\n>\\n> I'm trying to reconcile open CommitFest entries with their actual\\n> status; the entry for this:\\n>\\n> https://commitfest.postgresql.org/40/3558/\\n>\\n> shows as \\\"Waiting on Author\\\", but looks like it's all been committed;\\n> is there anything\\n> left to do on this?\\n>\\n\\nright the CF entry is out of date. there's still the additional tap test\\nthat needs to be taken care of. I sent a new version recently that works on\\nwindows CI, so I guess the correct status should now be needs review,\\nalthough the latest patchset probably doesn't apply anymore since the first\\npatch has been committed. I'm traveling today I'll try to send a rebased\\nversion in the evening\\n\\n>\\n\\nLe dim. 27 nov. 2022 à 15:31, Ian Lawrence Barwick \\n\\nI'm trying to reconcile open CommitFest entries with their actual\\nstatus; the entry for this:\\n\\n https://commitfest.postgresql.org/40/3558/\\n\\nshows as \\\"Waiting on Author\\\", but looks like it's all been committed;\\nis there anything\\nleft to do on this?right the CF entry is out of date. there's still the additional tap test that needs to be taken care of. I sent a new version recently that works on windows CI, so I guess the correct status should now be needs review, although the latest patchset probably doesn't apply anymore since the first patch has been committed. I'm traveling today I'll try to send a rebased version in the evening\",\n \"msg_date\": \"Sun, 27 Nov 2022 15:39:44 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 27, 2022 at 03:39:44PM +0800, Julien Rouhaud wrote:\\n> Le dim. 27 nov. 2022 � 15:31, Ian Lawrence Barwick\\n> \\n> >\\n> > I'm trying to reconcile open CommitFest entries with their actual\\n> > status; the entry for this:\\n> >\\n> > https://commitfest.postgresql.org/40/3558/\\n> >\\n> > shows as \\\"Waiting on Author\\\", but looks like it's all been committed;\\n> > is there anything\\n> > left to do on this?\\n> >\\n> \\n> right the CF entry is out of date. there's still the additional tap test\\n> that needs to be taken care of. I sent a new version recently that works on\\n> windows CI, so I guess the correct status should now be needs review,\\n> although the latest patchset probably doesn't apply anymore since the first\\n> patch has been committed. I'm traveling today I'll try to send a rebased\\n> version in the evening\\n\\nAnd here's the rebased patch for the TAP tests. I will switch the CF entry to\\nNeeds Review.\",\n \"msg_date\": \"Sun, 27 Nov 2022 19:04:46 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"2022年11月27日(日) 20:04 Julien Rouhaud <rjuju123@gmail.com>:\\n>\\n> On Sun, Nov 27, 2022 at 03:39:44PM +0800, Julien Rouhaud wrote:\\n> > Le dim. 27 nov. 2022 à 15:31, Ian Lawrence Barwick\\n> >\\n> > >\\n> > > I'm trying to reconcile open CommitFest entries with their actual\\n> > > status; the entry for this:\\n> > >\\n> > > https://commitfest.postgresql.org/40/3558/\\n> > >\\n> > > shows as \\\"Waiting on Author\\\", but looks like it's all been committed;\\n> > > is there anything\\n> > > left to do on this?\\n> > >\\n> >\\n> > right the CF entry is out of date. there's still the additional tap test\\n> > that needs to be taken care of. I sent a new version recently that works on\\n> > windows CI, so I guess the correct status should now be needs review,\\n> > although the latest patchset probably doesn't apply anymore since the first\\n> > patch has been committed. I'm traveling today I'll try to send a rebased\\n> > version in the evening\\n>\\n> And here's the rebased patch for the TAP tests. I will switch the CF entry to\\n> Needs Review.\\n\\nThanks for the quick update!\\n\\nRegards\\n\\nIan Barwick\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Nov 2022 21:49:31 +0900\",\n \"msg_from\": \"Ian Lawrence Barwick <barwick@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 27, 2022 at 09:49:31PM +0900, Ian Lawrence Barwick wrote:\\n> Thanks for the quick update!\\n\\nFWIW, I do intend to tackle this last part ASAP, as the last piece to\\ncommit to get the full picture in the tree.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 28 Nov 2022 10:08:36 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 27, 2022 at 07:04:46PM +0800, Julien Rouhaud wrote:\\n> And here's the rebased patch for the TAP tests. I will switch the CF entry to\\n> Needs Review.\\n\\nI have been looking at that, and applied the part 1 of the test for\\nthe positive tests to cover the basic ground I'd like to see covered\\nfor this feature. I have done quite a few changes to this part, like\\nreworking the way the incrementation of the counters is done for the\\nrule/map numbers and the line numbers of each file. The idea to use a\\nglobal dictionary to track the current number of lines was rather\\nclear, so I have kept that. add_{hba,ident}_line() rely on the\\nfiles included in a directory to be ordered in the same way as the\\nbackend, and I am fine to live with that.\\n\\n+# Normalize the data directory for Windows\\n+$data_dir =~ s/\\\\/\\\\.\\\\//\\\\//g; # reduce /./ to /\\n+$data_dir =~ s/\\\\/\\\\//\\\\//g; # reduce // to /\\n+$data_dir =~ s/\\\\/$//; # remove trailing /\\n+$data_dir =~ s/\\\\\\\\/\\\\//g; # change \\\\ to /\\nThe mysticism around the conversion of the data directory path is\\nsomething I'd like to avoid, or I suspect that we are going to have a\\nhard time debugging any issues in this area. For the positive cases\\ncommitted, I have gone through the approach of using the base name of\\nthe configuration file to avoid those compatibility issues. Wouldn't\\nit better to do the same for the expected error messages, switching to\\nan approach where we only check for the configuration file name in all\\nthese regexps?\\n\\n+# Generate the expected output for the auth file view error reporting (file\\n+# name, file line, error), for the given array of error conditions, as\\n+# generated generated by add_hba_line/add_ident_line with an err_str.\\n+sub generate_log_err_rows\\n+{\\ngenerate_log_err_rows() does not strike me as a good interface. The\\nissues I am pointed at here is it depends directly on the result\\nreturned by add_hba_line() or add_ident_line() when the caller passes\\ndown an error string. As far as I get it, this interface is done\\nas-is to correctly increment the line number of one file, and I'd like\\nto believe that add_{hba,ident}_line() should be kept designed so as\\nthey only return the expected matching entry in their catalog views,\\nwithout any code path returning something else (in your case the\\ntriplet [ $filename, $fileline, $err_str ]). That makes the whole\\nharder to understand.\\n\\nAttached is a rebased patch of the rest. With everything we have\\ndealt with in this CF, perhaps it would be better to mark this entry\\nas committed and switch to a new thread where the negative TAP tests\\ncould be discussed? It looks like the part for the error pattern\\nchecks compared with the logs could be the easiest portion of what's\\nremaining.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 28 Nov 2022 16:12:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 28, 2022 at 04:12:40PM +0900, Michael Paquier wrote:\\n> Attached is a rebased patch of the rest. With everything we have\\n> dealt with in this CF, perhaps it would be better to mark this entry\\n> as committed and switch to a new thread where the negative TAP tests\\n> could be discussed?\\n\\nFor now, I have marked the CF entry as committed.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 29 Nov 2022 08:48:44 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Allow file inclusion in pg_hba and pg_ident files\"\n }\n]","isConversational":false}}},{"rowIdx":53562,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nIn logical replication, currently Walsender sends the data that is\ngenerated locally and the data that are replicated from other\ninstances. This results in infinite recursion in circular logical\nreplication setup.\nHere the user is trying to have a 2-way replication setup with node 1\npublishing data to node2 and node2 publishing data to node1, so that\nthe user can perform dml operations from any node, it can act as a\n2-way multi master replication setup.\n\nThis problem can be reproduced with the following steps:\n-- Instance 1\ncreate publication pub1 for table t1;\ncreate table t1(c1 int);\n\n-- Instance 2\ncreate table t1(c1 int);\ncreate publication pub2 for table t1;\ncreate subscription sub1 CONNECTION 'dbname=postgres port=5432'\npublication pub1;\n\n-- Instance 1\ncreate subscription sub2 CONNECTION 'dbname=postgres port=5433'\npublication pub2; insert into t1 values(10);\n\nIn this scenario, the Walsender in publisher pub1 sends data to the\napply worker in subscriber sub1, the apply worker in sub1 maps the\ndata to local tables and applies the individual changes as they are\nreceived. Then the Walsender in publisher pub2 sends data to the apply\nworker in subscriber sub2, the apply worker in sub2 maps the data to\nlocal tables and applies the individual changes as they are received.\nThis process repeats infinitely.\n\nCurrently we do not differentiate if the data is locally generated\ndata, or a replicated data and we send both the data which causes\ninfinite recursion.\n\nWe could see that the record count has increased significantly within sometime:\nselect count(*) from t1;\n count\n--------------\n 4000000\n(1 row)\n\nIf the table had primary key constraint, we could notice that the\nfirst insert is successful and when the same insert is sent back, the\ninsert fails because of constraint error:\n2022-02-23 09:28:43.592 IST [14743] ERROR: duplicate key value\nviolates unique constraint \"t1_pkey\"\n2022-02-23 09:28:43.592 IST [14743] DETAIL: Key (c1)=(10) already exists.\n2022-02-23 09:28:43.592 IST [14743] CONTEXT: processing remote data\nduring \"INSERT\" for replication target relation \"public.t1\" in\ntransaction 727 at 2022-02-23 09:28:43.406738+05:30\n2022-02-23 09:28:43.593 IST [14678] LOG: background worker \"logical\nreplication worker\" (PID 14743) exited with exit code 1\n2022-02-23 09:28:48.608 IST [14745] LOG: logical replication apply\nworker for subscription \"sub2\" has started\n2022-02-23 09:28:48.624 IST [14745] ERROR: duplicate key value\nviolates unique constraint \"t1_pkey\"\n2022-02-23 09:28:48.624 IST [14745] DETAIL: Key (c1)=(10) already exists.\n2022-02-23 09:28:48.624 IST [14745] CONTEXT: processing remote data\nduring \"INSERT\" for replication target relation \"public.t1\" in\ntransaction 727 at 2022-02-23 09:28:43.406738+05:30\n2022-02-23 09:28:48.626 IST [14678] LOG: background worker \"logical\nreplication worker\" (PID 14745) exited with exit code 1\n\nThe same problem can occur in any circular node setup like 3 nodes,\n4node etc like: a) node1 publishing to node2 b) node2 publishing to\nnode3 c) node3 publishing back to node1.\n\nHere there are two problems for the user: a) incremental\nsynchronization of table sending both local data and replicated data\nby walsender b) Table synchronization of table using copy command\nsending both local data and replicated data\n\nFor the first problem \"Incremental synchronization of table by\nWalsender\" can be solved by:\nCurrently the locally generated data does not have replication origin\nassociated and the data that has originated from another instance will\nhave a replication origin associated. We could use this information to\ndifferentiate locally generated data and replicated data and send only\nthe locally generated data. This \"only_local\" could be provided as an\noption while subscription is created:\nex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\nPUBLICATION pub1 with (only_local = on);\n\nI have attached a basic patch for this, if the idea is accepted, I\nwill work further to test more scenarios, add documentation, and test\nand post an updated patch.\nFor the second problem, Table synchronization of table including local\ndata and replicated data using copy command.\n\nLet us consider the following scenario:\na) node1 publishing to node2 b) node2 publishing to node1. Here in\nthis case node1 will have replicated data from node2 and vice versa.\n\nIn the above if user wants to include node3 to subscribe data from\nnode2. Users will have to create a subscription in node3 to get the\ndata from node2. During table synchronization we send the complete\ntable data from node2 to node3. Node2 will have local data from node2\nand also replicated data from node1. Currently we don't have an option\nto differentiate between the locally generated data and replicated\ndata in the heap which will cause infinite recursion as described\nabove.\n\nTo handle this if user has specified only_local option, we could throw\na warning or error out while creating subscription in this case, we\ncould have a column srreplicateddata in pg_subscription_rel which\ncould indicate if the table has any replicated data or not:\npostgres=# select * from pg_subscription_rel;\n srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\n---------+---------+------------+-----------+------------------\n 16389 | 16384 | r | 0/14A4640 | t\n 16389 | 16385 | r | 0/14A4690 | f\n(1 row)\n\nIn the above example, srreplicateddata with true indicates, tabel t1\nwhose relid is 16384 has replicated data and the other row having\nsrreplicateddata as false indicates table t2 whose relid is 16385\ndoes not have replicated data.\nWhen creating a new subscription, the subscriber will connect to the\npublisher and check if the relation has replicated data by checking\nsrreplicateddata in pg_subscription_rel table.\nIf the table has any replicated data, log a warning or error for this.\n\nAlso, we could document the steps on how to handle the initial sync like:\na) Complete the ongoing transactions on this table in the replication\nsetup nodes i.e. node1 and node2 in the above case, so that the table\ndata is consistent, b) Once there are no ongoing transaction, Copy the\ntable data using copy command from any one of the nodes, c) create\nsubscription with copy_data option as off d) Perform further\ntransactions on the table e) All the further transactions performed\nwill be handled by the walsender which will take care of skipping\nreplicated data and sending only the local data. i.e. node2 will send\nthe locally generated data to node3.\n\nI'm not sure if there is any other better way to handle this. If there\nis a better way, we could handle it accordingly.\nThoughts?\n\nRegards,\nVignesh","msg_date":"Wed, 23 Feb 2022 11:59:02 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi Vignesh,\r\n\r\n> In logical replication, currently Walsender sends the data that is\r\n> generated locally and the data that are replicated from other\r\n> instances. This results in infinite recursion in circular logical\r\n> replication setup.\r\n\r\nThank you for good explanation. I understand that this fix can be used\r\nfor a bidirectional replication.\r\n\r\n> Here there are two problems for the user: a) incremental\r\n> synchronization of table sending both local data and replicated data\r\n> by walsender b) Table synchronization of table using copy command\r\n> sending both local data and replicated data\r\n\r\nSo you wanted to solve these two problem and currently focused on\r\nthe first one, right? We can check one by one.\r\n\r\n> For the first problem \"Incremental synchronization of table by\r\n> Walsender\" can be solved by:\r\n> Currently the locally generated data does not have replication origin\r\n> associated and the data that has originated from another instance will\r\n> have a replication origin associated. We could use this information to\r\n> differentiate locally generated data and replicated data and send only\r\n> the locally generated data. This \"only_local\" could be provided as an\r\n> option while subscription is created:\r\n> ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\r\n> PUBLICATION pub1 with (only_local = on);\r\n\r\nSounds good, but I cannot distinguish whether the assumption will keep.\r\n\r\nI played with your patch, but it could not be applied to current master.\r\nI tested from bd74c40 and I confirmed infinite loop was not appeared.\r\n\r\nlocal_only could not be set from ALTER SUBSCRIPTION command.\r\nIs it expected?\r\n\r\nBest Regards,\r\nHayato Kuroda\r\nFUJITSU LIMITED\r\n\r\n","msg_date":"Tue, 1 Mar 2022 10:42:27 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Mar 1, 2022 at 4:12 PM kuroda.hayato@fujitsu.com\n<kuroda.hayato@fujitsu.com> wrote:\n>\n> Hi Vignesh,\n>\n> > In logical replication, currently Walsender sends the data that is\n> > generated locally and the data that are replicated from other\n> > instances. This results in infinite recursion in circular logical\n> > replication setup.\n>\n> Thank you for good explanation. I understand that this fix can be used\n> for a bidirectional replication.\n\nOnce these issues are resolved, it can be used for bi-directional\nlogical replication.\n\n> > Here there are two problems for the user: a) incremental\n> > synchronization of table sending both local data and replicated data\n> > by walsender b) Table synchronization of table using copy command\n> > sending both local data and replicated data\n>\n> So you wanted to solve these two problem and currently focused on\n> the first one, right? We can check one by one.\n\nYes.\n\n> > For the first problem \"Incremental synchronization of table by\n> > Walsender\" can be solved by:\n> > Currently the locally generated data does not have replication origin\n> > associated and the data that has originated from another instance will\n> > have a replication origin associated. We could use this information to\n> > differentiate locally generated data and replicated data and send only\n> > the locally generated data. This \"only_local\" could be provided as an\n> > option while subscription is created:\n> > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\n> > PUBLICATION pub1 with (only_local = on);\n>\n> Sounds good, but I cannot distinguish whether the assumption will keep.\n\nReplication origin is created by the apply worker and it will be used\nfor all the transactions received from the walsender. I feel the\nreplication origin will be present always.\n\n> I played with your patch, but it could not be applied to current master.\n> I tested from bd74c40 and I confirmed infinite loop was not appeared.\n\nI will post an updated version for this soon.\n\n> local_only could not be set from ALTER SUBSCRIPTION command.\n> Is it expected?\n\nI wanted to get the opinion from others too just to make sure the\napproach is right. I will fix this including the documentation, test,\netc in the later versions.\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 2 Mar 2022 08:56:57 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\n>\n...\n...\n> I have attached a basic patch for this, if the idea is accepted, I\n> will work further to test more scenarios, add documentation, and test\n> and post an updated patch.\n> For the second problem, Table synchronization of table including local\n> data and replicated data using copy command.\n>\n> Let us consider the following scenario:\n> a) node1 publishing to node2 b) node2 publishing to node1. Here in\n> this case node1 will have replicated data from node2 and vice versa.\n>\n> In the above if user wants to include node3 to subscribe data from\n> node2. Users will have to create a subscription in node3 to get the\n> data from node2. During table synchronization we send the complete\n> table data from node2 to node3. Node2 will have local data from node2\n> and also replicated data from node1. Currently we don't have an option\n> to differentiate between the locally generated data and replicated\n> data in the heap which will cause infinite recursion as described\n> above.\n>\n> To handle this if user has specified only_local option, we could throw\n> a warning or error out while creating subscription in this case, we\n> could have a column srreplicateddata in pg_subscription_rel which\n> could indicate if the table has any replicated data or not:\n> postgres=# select * from pg_subscription_rel;\n> srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\n> ---------+---------+------------+-----------+------------------\n> 16389 | 16384 | r | 0/14A4640 | t\n> 16389 | 16385 | r | 0/14A4690 | f\n> (1 row)\n>\n> In the above example, srreplicateddata with true indicates, tabel t1\n> whose relid is 16384 has replicated data and the other row having\n> srreplicateddata as false indicates table t2 whose relid is 16385\n> does not have replicated data.\n> When creating a new subscription, the subscriber will connect to the\n> publisher and check if the relation has replicated data by checking\n> srreplicateddata in pg_subscription_rel table.\n> If the table has any replicated data, log a warning or error for this.\n>\n\nIf you want to give the error in this case, then I think we need to\nprovide an option to the user to allow copy. One possibility could be\nto extend existing copy_data option as 'false', 'true', 'force'. For\n'false', there shouldn't be any change, for 'true', if 'only_local'\noption is also set and the new column indicates replicated data then\ngive an error, for 'force', we won't give an error even if the\nconditions as mentioned for 'true' case are met, rather we will allow\ncopy in this case.\n\n> Also, we could document the steps on how to handle the initial sync like:\n> a) Complete the ongoing transactions on this table in the replication\n> setup nodes i.e. node1 and node2 in the above case, so that the table\n> data is consistent, b) Once there are no ongoing transaction, Copy the\n> table data using copy command from any one of the nodes, c) create\n> subscription with copy_data option as off d) Perform further\n> transactions on the table e) All the further transactions performed\n> will be handled by the walsender which will take care of skipping\n> replicated data and sending only the local data. i.e. node2 will send\n> the locally generated data to node3.\n>\n> I'm not sure if there is any other better way to handle this.\n>\n\nI could think of the below options for users to set up bi-directional\nreplication for the same table.\n\nOption-1:\nThere is no pre-existing data in the tables that are going to\nparticipate in bi-directional replication. In such a case, Users can\ncreate pub/sub (with only_local option as proposed by you) on both\nnodes before starting any writes on tables. This will allow\nbi-directional replication for the required tables. Now, if the user\nwants one of the nodes to join at a later point, then the strategy in\nOption-2/3 could be used.\n\nOption-2:\nOne of the nodes (say node-1) has some pre-existing data and another\nnode (say node-2) doesn't have any pre-existing data. In this case,\nthe user can set up pub/sub (with only_local and copy_data as 'false'\noptions) for node-1 first before any of the operations on node-2.\nThen, it can set up pub/sub on node-2. This will allow bi-directional\nreplication for the required tables.\n\nOption-3:\nBoth the nodes have some pre-existing data. I think the easiest option\ncould be to truncate data on one of the nodes and set up pub/sub on\nboth nodes. See, one way to achieve it among two nodes as below:\n\nNode-1:\nTable t1 has data\n1, 2, 3, 4\n\nPublication for t1, pub1: Create Publication pub1 For Table t1;\n\nNode-2:\nTable t1 has data\n5, 6, 7, 8\nPublication for t1, pub1_2: Create Publication pub1_2 For Table t1;\n\nNow, Create Subscription for pub1 on node1: Create Subscription sub1_2\nConnection '<node-1 details>' Publication pub1 WITH (only_local =\ntrue);\n\nNode-1:\nBegin;\n# Disallow truncates to be published\nAlter Publication pub1 Set (publish = 'insert, update, delete');\nTruncate t1;\nCreate Subscription sub1 Connection '<node-2 details>' Publication\npub1 WITH (only_local = true, copy_data = 'force');\nAlter Publication pub1 Set (publish = 'insert, update, delete, truncate');\nCommit;\n\nI think this will allow the bi-directional replication between two\nnodes. In this scheme, the user needs to manually perform some steps\nincluding truncate of the table on one of the nodes which she might or\nmight not like but at least there will be a way to set up a\nbi-directional replication on two nodes for same table operations\nwhich is not possible now.\n\nI think one can even imagine using and extending this functionality so\nthat users don't need to perform TRUNCATE on one of the nodes. Say, in\nthe above case for tablesync phase, we make both nodes to start a\ntransaction, create a slot on another node (with USE_SNAPSHOT option),\nand then allow copy from another node. I think it will be important to\nallow copy on each node once the slots are created and the initial\nsnapshot is established.\n\nFor more than two nodes, I think we can suggest having either of the\noption-1 or 2 for setup. But, there could be other ways as well\ndepending on how the user wants to do the setup.\n\nThoughts?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 2 Mar 2022 15:58:55 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Mar 1, 2022 at 4:12 PM kuroda.hayato@fujitsu.com\n<kuroda.hayato@fujitsu.com> wrote:\n>\n> Hi Vignesh,\n>\n> > In logical replication, currently Walsender sends the data that is\n> > generated locally and the data that are replicated from other\n> > instances. This results in infinite recursion in circular logical\n> > replication setup.\n>\n> Thank you for good explanation. I understand that this fix can be used\n> for a bidirectional replication.\n>\n> > Here there are two problems for the user: a) incremental\n> > synchronization of table sending both local data and replicated data\n> > by walsender b) Table synchronization of table using copy command\n> > sending both local data and replicated data\n>\n> So you wanted to solve these two problem and currently focused on\n> the first one, right? We can check one by one.\n>\n> > For the first problem \"Incremental synchronization of table by\n> > Walsender\" can be solved by:\n> > Currently the locally generated data does not have replication origin\n> > associated and the data that has originated from another instance will\n> > have a replication origin associated. We could use this information to\n> > differentiate locally generated data and replicated data and send only\n> > the locally generated data. This \"only_local\" could be provided as an\n> > option while subscription is created:\n> > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\n> > PUBLICATION pub1 with (only_local = on);\n>\n> Sounds good, but I cannot distinguish whether the assumption will keep.\n>\n> I played with your patch, but it could not be applied to current master.\n> I tested from bd74c40 and I confirmed infinite loop was not appeared.\nRebased the patch on top of head\n\n> local_only could not be set from ALTER SUBSCRIPTION command.\n> Is it expected?\nModified\n\nThanks for the comments, the attached patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Wed, 2 Mar 2022 21:42:11 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"FYI, the v2 patch did not apply to HEAD\n\n[postgres@CentOS7-x64 oss_postgres_misc]$ git apply\n../patches_misc/v2-0001-Skip-replication-of-non-local-data.patch\n--verbose\n...\nerror: patch failed: src/backend/replication/slotfuncs.c:231\nerror: src/backend/replication/slotfuncs.c: patch does not apply\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia.\n\n\n","msg_date":"Mon, 7 Mar 2022 14:37:02 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\n> Here there are two problems for the user: a) incremental\n> synchronization of table sending both local data and replicated data\n> by walsender b) Table synchronization of table using copy command\n> sending both local data and replicated data\n>\n> For the first problem \"Incremental synchronization of table by\n> Walsender\" can be solved by:\n> Currently the locally generated data does not have replication origin\n> associated and the data that has originated from another instance will\n> have a replication origin associated. We could use this information to\n> differentiate locally generated data and replicated data and send only\n> the locally generated data. This \"only_local\" could be provided as an\n> option while subscription is created:\n> ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\n> PUBLICATION pub1 with (only_local = on);\n\nI haven't yet gone through the patch, but I have a question about the\nidea. Suppose I want to set up a logical replication like,\nnode1->node2->node3->node1. So how would I create the subscriber at\nnode1? only_local=on or off?. I mean on node1, I want the changes\nfrom node3 which are generated on node3 or which are replicated from\nnode2 but I do not want changes that are replicated from node1 itself?\n So if I set only_local=on then node1 will not get the changes\nreplicated from node2, is that right? and If I set only_local=off then\nit will create the infinite loop again? So how are we protecting\nagainst this case?\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 7 Mar 2022 09:25:33 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 9:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>\n> On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\n> > Here there are two problems for the user: a) incremental\n> > synchronization of table sending both local data and replicated data\n> > by walsender b) Table synchronization of table using copy command\n> > sending both local data and replicated data\n> >\n> > For the first problem \"Incremental synchronization of table by\n> > Walsender\" can be solved by:\n> > Currently the locally generated data does not have replication origin\n> > associated and the data that has originated from another instance will\n> > have a replication origin associated. We could use this information to\n> > differentiate locally generated data and replicated data and send only\n> > the locally generated data. This \"only_local\" could be provided as an\n> > option while subscription is created:\n> > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\n> > PUBLICATION pub1 with (only_local = on);\n>\n> I haven't yet gone through the patch, but I have a question about the\n> idea. Suppose I want to set up a logical replication like,\n> node1->node2->node3->node1. So how would I create the subscriber at\n> node1? only_local=on or off?. I mean on node1, I want the changes\n> from node3 which are generated on node3 or which are replicated from\n> node2 but I do not want changes that are replicated from node1 itself?\n> So if I set only_local=on then node1 will not get the changes\n> replicated from node2, is that right? and If I set only_local=off then\n> it will create the infinite loop again? So how are we protecting\n> against this case?\n>\n\nIn the above topology if you want local changes from both node3 and\nnode2 then I think the way to get that would be you have to create two\nsubscriptions on node1. The first one points to node2 (with\nonly_local=off) and the second one points to node3 (with only_local\n=off). Will that address your case or am I missing something?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 7 Mar 2022 10:05:09 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 10:05 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 9:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >\n> > On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\n> > > Here there are two problems for the user: a) incremental\n> > > synchronization of table sending both local data and replicated data\n> > > by walsender b) Table synchronization of table using copy command\n> > > sending both local data and replicated data\n> > >\n> > > For the first problem \"Incremental synchronization of table by\n> > > Walsender\" can be solved by:\n> > > Currently the locally generated data does not have replication origin\n> > > associated and the data that has originated from another instance will\n> > > have a replication origin associated. We could use this information to\n> > > differentiate locally generated data and replicated data and send only\n> > > the locally generated data. This \"only_local\" could be provided as an\n> > > option while subscription is created:\n> > > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\n> > > PUBLICATION pub1 with (only_local = on);\n> >\n> > I haven't yet gone through the patch, but I have a question about the\n> > idea. Suppose I want to set up a logical replication like,\n> > node1->node2->node3->node1. So how would I create the subscriber at\n> > node1? only_local=on or off?. I mean on node1, I want the changes\n> > from node3 which are generated on node3 or which are replicated from\n> > node2 but I do not want changes that are replicated from node1 itself?\n> > So if I set only_local=on then node1 will not get the changes\n> > replicated from node2, is that right? and If I set only_local=off then\n> > it will create the infinite loop again? So how are we protecting\n> > against this case?\n> >\n>\n> In the above topology if you want local changes from both node3 and\n> node2 then I think the way to get that would be you have to create two\n> subscriptions on node1. The first one points to node2 (with\n> only_local=off) and the second one points to node3 (with only_local\n> =off).\n>\n\nSorry, I intend to say 'only_local=on' at both places in my previous email.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 7 Mar 2022 10:15:46 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi Vignesh, I also have not looked at the patch yet, but I have what\nseems like a very fundamental (and possibly dumb) question...\n\nBasically, I do not understand the choice of syntax for setting things up.\n\nIMO that \"only-local\" option sounds very similar to the other\nPUBLICATION (\"publish\") options which decide the kinds of things that\nwill be published. So it feels more natural for me to think of the\npublisher as being the one to decide what will be published.\n\ne.g.\n\noption 1:\nCREATE PUBLICATION p1 FOR TABLE t1;\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n\noption 2:\nCREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n\n~~\n\nIIUC the patch is using option 1. My first impression was it feels\nback-to-front for the SUBSCRIPTION telling the PUBLICATION what to\npublish.\n\nSo, why does the patch use syntax option 1?\n\n------\nKind Regards,\nPeter Smith\nFujitsu Australia\n\n\n","msg_date":"Mon, 7 Mar 2022 15:56:25 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 3:56 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Hi Vignesh, I also have not looked at the patch yet, but I have what\n> seems like a very fundamental (and possibly dumb) question...\n>\n> Basically, I do not understand the choice of syntax for setting things up.\n>\n> IMO that \"only-local\" option sounds very similar to the other\n> PUBLICATION (\"publish\") options which decide the kinds of things that\n> will be published. So it feels more natural for me to think of the\n> publisher as being the one to decide what will be published.\n>\n> e.g.\n>\n> option 1:\n> CREATE PUBLICATION p1 FOR TABLE t1;\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n>\n> option 2:\n> CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n>\n\nSorry, I mean to write WITH.\n\noption 2:\nCREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n\n> ~~\n>\n> IIUC the patch is using option 1. My first impression was it feels\n> back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\n> publish.\n>\n> So, why does the patch use syntax option 1?\n>\n> ------\n> Kind Regards,\n> Peter Smith\n> Fujitsu Australia\n\n\n","msg_date":"Mon, 7 Mar 2022 16:06:53 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Hi Vignesh, I also have not looked at the patch yet, but I have what\n> seems like a very fundamental (and possibly dumb) question...\n>\n> Basically, I do not understand the choice of syntax for setting things up.\n>\n> IMO that \"only-local\" option sounds very similar to the other\n> PUBLICATION (\"publish\") options which decide the kinds of things that\n> will be published. So it feels more natural for me to think of the\n> publisher as being the one to decide what will be published.\n>\n> e.g.\n>\n> option 1:\n> CREATE PUBLICATION p1 FOR TABLE t1;\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n>\n> option 2:\n> CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n>\n> ~~\n>\n> IIUC the patch is using option 1. My first impression was it feels\n> back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\n> publish.\n>\n> So, why does the patch use syntax option 1?\n\nI felt the advantage with keeping it at the subscription side is that,\nthe subscriber from one node can subscribe with only_local option on\nand a different subscriber from a different node can subscribe with\nonly_local option as off. This might not be possible with having the\noption at publisher side. Having it at the subscriber side might give\nmore flexibility for the user.\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 7 Mar 2022 10:50:35 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Mar 1, 2022 at 4:12 PM kuroda.hayato@fujitsu.com\n<kuroda.hayato@fujitsu.com> wrote:\n>\n> Hi Vignesh,\n>\n> > In logical replication, currently Walsender sends the data that is\n> > generated locally and the data that are replicated from other\n> > instances. This results in infinite recursion in circular logical\n> > replication setup.\n>\n> Thank you for good explanation. I understand that this fix can be used\n> for a bidirectional replication.\n>\n> > Here there are two problems for the user: a) incremental\n> > synchronization of table sending both local data and replicated data\n> > by walsender b) Table synchronization of table using copy command\n> > sending both local data and replicated data\n>\n> So you wanted to solve these two problem and currently focused on\n> the first one, right? We can check one by one.\n>\n> > For the first problem \"Incremental synchronization of table by\n> > Walsender\" can be solved by:\n> > Currently the locally generated data does not have replication origin\n> > associated and the data that has originated from another instance will\n> > have a replication origin associated. We could use this information to\n> > differentiate locally generated data and replicated data and send only\n> > the locally generated data. This \"only_local\" could be provided as an\n> > option while subscription is created:\n> > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\n> > PUBLICATION pub1 with (only_local = on);\n>\n> Sounds good, but I cannot distinguish whether the assumption will keep.\n>\n> I played with your patch, but it could not be applied to current master.\n> I tested from bd74c40 and I confirmed infinite loop was not appeared.\n>\n> local_only could not be set from ALTER SUBSCRIPTION command.\n> Is it expected?\n\nI felt changing only_local option might be useful for the user while\nmodifying the subscription like setting it with a different set of\npublications. Changes for this are included in the v2 patch attached\nat [1].\n[2] - https://www.postgresql.org/message-id/CALDaNm0WSo5369pr2eN1obTGBeiJU9cQdF6Ju1sC4hMQNy5BfQ%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 7 Mar 2022 11:14:31 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Dear Peter,\r\n\r\n> > So, why does the patch use syntax option 1?\r\n\r\nIMU it might be useful for the following case.\r\n\r\nAssuming that multi-master configuration with node1, node2.\r\nNode1 has a publication pub1 and a subscription sub2, node2 has pub2 and sub1.\r\n\r\nFrom that situation, please consider that new node node3 is added\r\nthat subscribe some changes from node2.\r\n\r\nIf the feature is introduced as option1, new publication must be defined in node2.\r\nIf that is introduced as option2, however, maybe pub2 can be reused.\r\ni.e. multiple declaration of publications can be avoided.\r\n\r\nBest Regards,\r\nHayato Kuroda\r\nFUJITSU LIMITED\r\n\r\n","msg_date":"Mon, 7 Mar 2022 06:12:50 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Hi Vignesh, I also have not looked at the patch yet, but I have what\n> > seems like a very fundamental (and possibly dumb) question...\n> >\n> > Basically, I do not understand the choice of syntax for setting things up.\n> >\n> > IMO that \"only-local\" option sounds very similar to the other\n> > PUBLICATION (\"publish\") options which decide the kinds of things that\n> > will be published. So it feels more natural for me to think of the\n> > publisher as being the one to decide what will be published.\n> >\n> > e.g.\n> >\n> > option 1:\n> > CREATE PUBLICATION p1 FOR TABLE t1;\n> > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n> >\n> > option 2:\n> > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\n> > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n> >\n> > ~~\n> >\n> > IIUC the patch is using option 1. My first impression was it feels\n> > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\n> > publish.\n> >\n> > So, why does the patch use syntax option 1?\n>\n> I felt the advantage with keeping it at the subscription side is that,\n> the subscriber from one node can subscribe with only_local option on\n> and a different subscriber from a different node can subscribe with\n> only_local option as off. This might not be possible with having the\n> option at publisher side. Having it at the subscriber side might give\n> more flexibility for the user.\n>\n\nOK. Option 2 needs two publications for that scenario. IMO it's more\nintuitive this way, but maybe you wanted to avoid the extra\npublications?\n\nnode0:\nCREATE PUBLICATION p1 FOR TABLE t1;\nCREATE PUBLICATION p1_local FOR TABLE t1 WITH (publish = 'only_local');\n\nnode1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1_local;\nnode2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1;\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 7 Mar 2022 17:14:48 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 5:12 PM kuroda.hayato@fujitsu.com\n<kuroda.hayato@fujitsu.com> wrote:\n>\n> Dear Peter,\n>\n> > > So, why does the patch use syntax option 1?\n>\n> IMU it might be useful for the following case.\n>\n> Assuming that multi-master configuration with node1, node2.\n> Node1 has a publication pub1 and a subscription sub2, node2 has pub2 and sub1.\n>\n> From that situation, please consider that new node node3 is added\n> that subscribe some changes from node2.\n>\n> If the feature is introduced as option1, new publication must be defined in node2.\n> If that is introduced as option2, however, maybe pub2 can be reused.\n> i.e. multiple declaration of publications can be avoided.\n>\n\nYes. Thanks for the example. I had the same observation in my last post [1]\n\n------\n[1] https://www.postgresql.org/message-id/CAHut%2BPtRxiQR_4UFLNThg-NNRV447FvwtcR-BvqMzjyMJXKwfw%40mail.gmail.com\n\nKind Regards,\nPeter Smith\nFujitsu Australia.\n\n\n","msg_date":"Mon, 7 Mar 2022 17:20:03 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 11:45 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > >\n> > > Hi Vignesh, I also have not looked at the patch yet, but I have what\n> > > seems like a very fundamental (and possibly dumb) question...\n> > >\n> > > Basically, I do not understand the choice of syntax for setting things up.\n> > >\n> > > IMO that \"only-local\" option sounds very similar to the other\n> > > PUBLICATION (\"publish\") options which decide the kinds of things that\n> > > will be published. So it feels more natural for me to think of the\n> > > publisher as being the one to decide what will be published.\n> > >\n> > > e.g.\n> > >\n> > > option 1:\n> > > CREATE PUBLICATION p1 FOR TABLE t1;\n> > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n> > >\n> > > option 2:\n> > > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\n> > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n> > >\n> > > ~~\n> > >\n> > > IIUC the patch is using option 1. My first impression was it feels\n> > > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\n> > > publish.\n> > >\n> > > So, why does the patch use syntax option 1?\n> >\n> > I felt the advantage with keeping it at the subscription side is that,\n> > the subscriber from one node can subscribe with only_local option on\n> > and a different subscriber from a different node can subscribe with\n> > only_local option as off. This might not be possible with having the\n> > option at publisher side. Having it at the subscriber side might give\n> > more flexibility for the user.\n> >\n>\n> OK. Option 2 needs two publications for that scenario. IMO it's more\n> intuitive this way, but maybe you wanted to avoid the extra\n> publications?\n\nYes, I wanted to avoid the extra publication creation that you pointed\nout. Option 1 can handle this scenario without creating the extra\npublications:\nnode0: CREATE PUBLICATION p1 FOR TABLE t1;\nnode1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = on);\nnode2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = off);\n\nI'm ok with both the approaches, now that this scenario can be handled\nby using both the options. i.e providing only_local option as an\noption while creating publication or providing only_local option as an\noption while creating subscription as Peter has pointed out at [1].\noption 1:\nCREATE PUBLICATION p1 FOR TABLE t1;\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n\noption 2:\nCREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n\nShall we get a few opinions on this and take it in that direction?\n\n[1] - https://www.postgresql.org/message-id/CAHut%2BPsAWaETh9VMymbBfMrqiE1KuqMq%2BwpBg0s7eMzwLATr%2Bw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 7 Mar 2022 12:47:43 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Dear Vignesh,\r\n\r\n> I felt changing only_local option might be useful for the user while\r\n> modifying the subscription like setting it with a different set of\r\n> publications. Changes for this are included in the v2 patch attached\r\n> at [1].\r\n\r\n+1, thanks. I'll post if I notice something to say.\r\n\r\n> Shall we get a few opinions on this and take it in that direction?\r\n\r\nI prefer subscriber-option, but I also think both are reasonable.\r\n+1 about asking other reviewers.\r\n\r\nBest Regards,\r\nHayato Kuroda\r\nFUJITSU LIMITED\r\n\r\n","msg_date":"Mon, 7 Mar 2022 07:37:46 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 10:15 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > I haven't yet gone through the patch, but I have a question about the\n> > > idea. Suppose I want to set up a logical replication like,\n> > > node1->node2->node3->node1. So how would I create the subscriber at\n> > > node1? only_local=on or off?. I mean on node1, I want the changes\n> > > from node3 which are generated on node3 or which are replicated from\n> > > node2 but I do not want changes that are replicated from node1 itself?\n> > > So if I set only_local=on then node1 will not get the changes\n> > > replicated from node2, is that right? and If I set only_local=off then\n> > > it will create the infinite loop again? So how are we protecting\n> > > against this case?\n> > >\n> >\n> > In the above topology if you want local changes from both node3 and\n> > node2 then I think the way to get that would be you have to create two\n> > subscriptions on node1. The first one points to node2 (with\n> > only_local=off) and the second one points to node3 (with only_local\n> > =off).\n> >\n>\n> Sorry, I intend to say 'only_local=on' at both places in my previous email.\n\nHmm okay, so for this topology we will have to connect node1 directly\nto node2 as well as to node3 but can not cascade the changes. I was\nwondering can it be done without using the extra connection between\nnode2 to node1? I mean instead of making this a boolean flag that\nwhether we want local change or remote change, can't we control the\nchanges based on the origin id? Such that node1 will get the local\nchanges of node3 but with using the same subscription it will get\nchanges from node3 which are originated from node2 but it will not\nreceive the changes which are originated from node1.\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 7 Mar 2022 13:10:44 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 6:17 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 11:45 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > > >\n> > > > Hi Vignesh, I also have not looked at the patch yet, but I have what\n> > > > seems like a very fundamental (and possibly dumb) question...\n> > > >\n> > > > Basically, I do not understand the choice of syntax for setting things up.\n> > > >\n> > > > IMO that \"only-local\" option sounds very similar to the other\n> > > > PUBLICATION (\"publish\") options which decide the kinds of things that\n> > > > will be published. So it feels more natural for me to think of the\n> > > > publisher as being the one to decide what will be published.\n> > > >\n> > > > e.g.\n> > > >\n> > > > option 1:\n> > > > CREATE PUBLICATION p1 FOR TABLE t1;\n> > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n> > > >\n> > > > option 2:\n> > > > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\n> > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n> > > >\n> > > > ~~\n> > > >\n> > > > IIUC the patch is using option 1. My first impression was it feels\n> > > > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\n> > > > publish.\n> > > >\n> > > > So, why does the patch use syntax option 1?\n> > >\n> > > I felt the advantage with keeping it at the subscription side is that,\n> > > the subscriber from one node can subscribe with only_local option on\n> > > and a different subscriber from a different node can subscribe with\n> > > only_local option as off. This might not be possible with having the\n> > > option at publisher side. Having it at the subscriber side might give\n> > > more flexibility for the user.\n> > >\n> >\n> > OK. Option 2 needs two publications for that scenario. IMO it's more\n> > intuitive this way, but maybe you wanted to avoid the extra\n> > publications?\n>\n> Yes, I wanted to avoid the extra publication creation that you pointed\n> out. Option 1 can handle this scenario without creating the extra\n> publications:\n> node0: CREATE PUBLICATION p1 FOR TABLE t1;\n> node1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = on);\n> node2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = off);\n>\n> I'm ok with both the approaches, now that this scenario can be handled\n> by using both the options. i.e providing only_local option as an\n> option while creating publication or providing only_local option as an\n> option while creating subscription as Peter has pointed out at [1].\n> option 1:\n> CREATE PUBLICATION p1 FOR TABLE t1;\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n>\n> option 2:\n> CREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n>\n> Shall we get a few opinions on this and take it in that direction?\n>\n> [1] - https://www.postgresql.org/message-id/CAHut%2BPsAWaETh9VMymbBfMrqiE1KuqMq%2BwpBg0s7eMzwLATr%2Bw%40mail.gmail.com\n>\n> Regards,\n> Vignesh\n\nBTW here is a counter-example to your scenario from earlier.\n\nLet's say I have a publication p1 and p2 and want to subscribe to p1\nwith only_local=true, and p2 with only_local = false;\n\nUsing the current OPtion 1 syntax you cannot do this with a single\nsubscription because the option is tied to the subscription.\nBut using syntax Option 2 you may be able to do it.\n\nOption 1:\nCREATE PUBLICATION p1 FOR TABLE t1;\nCREATE PUBLICATION p2 FOR TABLE t2;\nCREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 WITH (local_only = true);\nCREATE SUBSCRIPTION s2 ... FOR PUBLICATION p1 WITH (local_only = false);\n\nOption 2:\nCREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'local_only');\nCREATE PUBLICATION p2 FOR TABLE t2;\nCREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1, p2;\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 7 Mar 2022 19:14:39 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 12:48 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 11:45 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > > >\n> > > > Hi Vignesh, I also have not looked at the patch yet, but I have what\n> > > > seems like a very fundamental (and possibly dumb) question...\n> > > >\n> > > > Basically, I do not understand the choice of syntax for setting things up.\n> > > >\n> > > > IMO that \"only-local\" option sounds very similar to the other\n> > > > PUBLICATION (\"publish\") options which decide the kinds of things that\n> > > > will be published. So it feels more natural for me to think of the\n> > > > publisher as being the one to decide what will be published.\n> > > >\n> > > > e.g.\n> > > >\n> > > > option 1:\n> > > > CREATE PUBLICATION p1 FOR TABLE t1;\n> > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n> > > >\n> > > > option 2:\n> > > > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\n> > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n> > > >\n> > > > ~~\n> > > >\n> > > > IIUC the patch is using option 1. My first impression was it feels\n> > > > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\n> > > > publish.\n> > > >\n> > > > So, why does the patch use syntax option 1?\n> > >\n> > > I felt the advantage with keeping it at the subscription side is that,\n> > > the subscriber from one node can subscribe with only_local option on\n> > > and a different subscriber from a different node can subscribe with\n> > > only_local option as off. This might not be possible with having the\n> > > option at publisher side. Having it at the subscriber side might give\n> > > more flexibility for the user.\n> > >\n> >\n> > OK. Option 2 needs two publications for that scenario. IMO it's more\n> > intuitive this way, but maybe you wanted to avoid the extra\n> > publications?\n>\n> Yes, I wanted to avoid the extra publication creation that you pointed\n> out. Option 1 can handle this scenario without creating the extra\n> publications:\n> node0: CREATE PUBLICATION p1 FOR TABLE t1;\n> node1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = on);\n> node2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = off);\n>\n> I'm ok with both the approaches, now that this scenario can be handled\n> by using both the options. i.e providing only_local option as an\n> option while creating publication or providing only_local option as an\n> option while creating subscription as Peter has pointed out at [1].\n> option 1:\n> CREATE PUBLICATION p1 FOR TABLE t1;\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n>\n> option 2:\n> CREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n>\n> Shall we get a few opinions on this and take it in that direction?\n>\n\nAllowing multiple publications will lead to a lot of duplicate data in\ncatalogs like pg_publication_rel, so, I don't see how option-2 can be\nbeneficial?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 7 Mar 2022 14:24:43 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi Vignesh,\n\nHere are some review comments for patch v2.\n\n======\n\n1. Question about syntax\n\nI already posted some questions about why the syntax is on the CREATE\nSUBSCRCRIBER side.\nIMO \"local_only\" is a publisher option, so it seemed more natural to\nme for it to be specified as a \"publish\" option.\n\nRef [1] my original question + suggestion for Option 2\nRef [2] some other examples of subscribing to multiple-publishers\n\nAnyway, +1 to see what other people think.\n\n~~~\n\n2. ALTER\n\n(related also to the question about syntax)\n\nIf subscribing to multiple publications then ALTER is going to change\nthe 'local_only' for all of them, which might not be what you want\n(??)\n\n~~~\n\n3. subscription_parameter\n\n(related also to the question about syntax)\n\nCREATE SUBSCRIPTION subscription_name\n CONNECTION 'conninfo'\n PUBLICATION publication_name [, ...]\n [ WITH ( subscription_parameter [= value] [, ... ] ) ]\n\n~\n\nThat WITH is for *subscription* options, not the publication options.\n\nSo IMO 'local_only' intuitively seems like \"local\" means local where\nthe subscriber is.\n\nSo, if the Option 1 syntax is chosen (see comment #1) then I think the\noption name maybe should change to be something more like\n'publish_local_only' or something similar to be more clear what local\nactually means.\n\n~~~\n\n4. contrib/test_decoding/test_decoding.c\n\n@@ -484,6 +487,16 @@ pg_decode_filter(LogicalDecodingContext *ctx,\n return false;\n }\n\n+static bool\n+pg_decode_filter_remotedata(LogicalDecodingContext *ctx,\n+ RepOriginId origin_id)\n+{\n+ TestDecodingData *data = ctx->output_plugin_private;\n+\n+ if (data->only_local && origin_id != InvalidRepOriginId)\n+ return true;\n+ return false;\n+}\n\n4a. Maybe needs function comment.\n\n4b. Missing blank line following this function\n\n~~~\n\n5. General - please check all of the patch.\n\nThere seems inconsistency with the member names, local variable names,\nparameter names etc. There are all variations of:\n\n- only_local\n- onlylocaldata\n- onlylocal_data\n- etc\n\nPlease try using the same name everywhere for everything if possible.\n\n~~~\n\n6. src/backend/replication/logical/decode.c - FilterRemoteOriginData\n\n@@ -585,7 +594,8 @@ logicalmsg_decode(LogicalDecodingContext *ctx,\nXLogRecordBuffer *buf)\n message = (xl_logical_message *) XLogRecGetData(r);\n\n if (message->dbId != ctx->slot->data.database ||\n- FilterByOrigin(ctx, origin_id))\n+ FilterByOrigin(ctx, origin_id) ||\n+ FilterRemoteOriginData(ctx, origin_id))\n return;\n\nI noticed that every call to FilterRemoteOriginData has an associated\npreceding call to FilterByOrigin. It might be worth just combining the\nlogic into FilterByOrigin. Then none of that calling code (9 x places)\nwould need to change at all.\n\n~~~\n\n7. src/backend/replication/logical/logical.c - CreateInitDecodingContext\n\n@@ -451,6 +453,8 @@ CreateInitDecodingContext(const char *plugin,\n */\n ctx->twophase &= slot->data.two_phase;\n\n+ ctx->onlylocal_data &= slot->data.onlylocal_data;\n\nThe equivalent 'twophase' option had a big comment. Probably this new\noption should also have a similar comment?\n\n~~~\n\n8. src/backend/replication/logical/logical.c - filter_remotedata_cb_wrapper\n\n+bool\n+filter_remotedata_cb_wrapper(LogicalDecodingContext *ctx,\n+ RepOriginId origin_id)\n+{\n+ LogicalErrorCallbackState state;\n+ ErrorContextCallback errcallback;\n+ bool ret;\n+\n+ Assert(!ctx->fast_forward);\n+\n+ /* Push callback + info on the error context stack */\n+ state.ctx = ctx;\n+ state.callback_name = \"filter_remoteorigin\";\n\nThere is no consistency between the function and the name:\n\n\"filter_remoteorigin\" versus filter_remotedata_cb.\n\nA similar inconsistency for this is elsewhere. See review comment #9\n\n~~~\n\n9. src/backend/replication/pgoutput/pgoutput.c\n\n@@ -215,6 +217,7 @@ _PG_output_plugin_init(OutputPluginCallbacks *cb)\n cb->commit_prepared_cb = pgoutput_commit_prepared_txn;\n cb->rollback_prepared_cb = pgoutput_rollback_prepared_txn;\n cb->filter_by_origin_cb = pgoutput_origin_filter;\n+ cb->filter_remotedata_cb = pgoutput_remoteorigin_filter;\n\nInconsistent names for the member and function.\n\nfilter_remotedata_cb VS pgoutput_remoteorigin_filter.\n\n~~~\n\n10. src/backend/replication/pgoutput/pgoutput.c\n\n@@ -1450,6 +1465,16 @@ pgoutput_origin_filter(LogicalDecodingContext *ctx,\n return false;\n }\n\n+static bool\n+pgoutput_remoteorigin_filter(LogicalDecodingContext *ctx,\n+ RepOriginId origin_id)\n+{\n+ PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\n+\n+ if (data->onlylocal_data && origin_id != InvalidRepOriginId)\n+ return true;\n+ return false;\n+}\n /*\n * Shutdown the output plugin.\n *\n\n10a. Add a function comment.\n\n10b. Missing blank line after the function\n\n~~~\n\n11. src/backend/replication/slotfuncs.c - pg_create_logical_replication_slot\n\n@@ -171,6 +174,7 @@ pg_create_logical_replication_slot(PG_FUNCTION_ARGS)\n Name plugin = PG_GETARG_NAME(1);\n bool temporary = PG_GETARG_BOOL(2);\n bool two_phase = PG_GETARG_BOOL(3);\n+ bool onlylocal_data = PG_GETARG_BOOL(4);\n Datum result;\n TupleDesc tupdesc;\n HeapTuple tuple;\n\n\nWon't there be some PG Docs needing to be updated now there is another\nparameter?\n\n~~~\n\n12. src/include/catalog/pg_proc.dat - pg_get_replication_slots\n\nI did not see any update for pg_get_replication_slots, but you added\nthe 4th parameter elsewhere. Is something missing here?\n\n~~~\n\n13. src/include/replication/logical.h\n\n@@ -99,6 +99,8 @@ typedef struct LogicalDecodingContext\n */\n bool twophase_opt_given;\n\n+ bool onlylocal_data;\n+\n\nI think the new member needs some comment.\n\n~~~\n\n14. src/include/replication/walreceiver.h\n\n@@ -183,6 +183,7 @@ typedef struct\n bool streaming; /* Streaming of large transactions */\n bool twophase; /* Streaming of two-phase transactions at\n * prepare time */\n+ bool onlylocal_data;\n } logical;\n } proto;\n } WalRcvStreamOptions;\n\nI think the new member needs some comment.\n\n~~~\n\n15. src/test/regress/sql/subscription.sql\n\nALTER SUBSCRIPTION test missing?\n\n------\n[1] https://www.postgresql.org/message-id/CAHut%2BPsAWaETh9VMymbBfMrqiE1KuqMq%2BwpBg0s7eMzwLATr%2Bw%40mail.gmail.com\n[2] https://www.postgresql.org/message-id/CAHut%2BPvQonJd5epJBM0Yfh1499mL9kTL9a%3DGrMhvnL6Ok05zqw%40mail.gmail.com\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 7 Mar 2022 19:57:47 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 1:11 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 10:15 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > > I haven't yet gone through the patch, but I have a question about the\n> > > > idea. Suppose I want to set up a logical replication like,\n> > > > node1->node2->node3->node1. So how would I create the subscriber at\n> > > > node1? only_local=on or off?. I mean on node1, I want the changes\n> > > > from node3 which are generated on node3 or which are replicated from\n> > > > node2 but I do not want changes that are replicated from node1 itself?\n> > > > So if I set only_local=on then node1 will not get the changes\n> > > > replicated from node2, is that right? and If I set only_local=off then\n> > > > it will create the infinite loop again? So how are we protecting\n> > > > against this case?\n> > > >\n> > >\n> > > In the above topology if you want local changes from both node3 and\n> > > node2 then I think the way to get that would be you have to create two\n> > > subscriptions on node1. The first one points to node2 (with\n> > > only_local=off) and the second one points to node3 (with only_local\n> > > =off).\n> > >\n> >\n> > Sorry, I intend to say 'only_local=on' at both places in my previous email.\n>\n> Hmm okay, so for this topology we will have to connect node1 directly\n> to node2 as well as to node3 but can not cascade the changes. I was\n> wondering can it be done without using the extra connection between\n> node2 to node1? I mean instead of making this a boolean flag that\n> whether we want local change or remote change, can't we control the\n> changes based on the origin id? Such that node1 will get the local\n> changes of node3 but with using the same subscription it will get\n> changes from node3 which are originated from node2 but it will not\n> receive the changes which are originated from node1.\n>\n\nGood point. I think we can provide that as an additional option to\ngive more flexibility but we won't be able to use it for initial sync\nwhere we can't differentiate between data from different origins.\nAlso, I think as origins are internally generated, we may need some\nbetter way to expose it to users so that they can specify it as an\noption. Isn't it better to provide first some simple way like a\nboolean option so that users have some way to replicate the same table\ndata among different nodes without causing an infinite loop and then\nextend it as you are suggesting or may be in some other ways as well?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 7 Mar 2022 15:01:15 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 1:11 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >\n> > On Mon, Mar 7, 2022 at 10:15 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > > > I haven't yet gone through the patch, but I have a question about the\n> > > > > idea. Suppose I want to set up a logical replication like,\n> > > > > node1->node2->node3->node1. So how would I create the subscriber at\n> > > > > node1? only_local=on or off?. I mean on node1, I want the changes\n> > > > > from node3 which are generated on node3 or which are replicated from\n> > > > > node2 but I do not want changes that are replicated from node1 itself?\n> > > > > So if I set only_local=on then node1 will not get the changes\n> > > > > replicated from node2, is that right? and If I set only_local=off then\n> > > > > it will create the infinite loop again? So how are we protecting\n> > > > > against this case?\n> > > > >\n> > > >\n> > > > In the above topology if you want local changes from both node3 and\n> > > > node2 then I think the way to get that would be you have to create two\n> > > > subscriptions on node1. The first one points to node2 (with\n> > > > only_local=off) and the second one points to node3 (with only_local\n> > > > =off).\n> > > >\n> > >\n> > > Sorry, I intend to say 'only_local=on' at both places in my previous email.\n> >\n> > Hmm okay, so for this topology we will have to connect node1 directly\n> > to node2 as well as to node3 but can not cascade the changes. I was\n> > wondering can it be done without using the extra connection between\n> > node2 to node1? I mean instead of making this a boolean flag that\n> > whether we want local change or remote change, can't we control the\n> > changes based on the origin id? Such that node1 will get the local\n> > changes of node3 but with using the same subscription it will get\n> > changes from node3 which are originated from node2 but it will not\n> > receive the changes which are originated from node1.\n> >\n>\n> Good point. I think we can provide that as an additional option to\n> give more flexibility but we won't be able to use it for initial sync\n> where we can't differentiate between data from different origins.\n> Also, I think as origins are internally generated, we may need some\n> better way to expose it to users so that they can specify it as an\n> option. Isn't it better to provide first some simple way like a\n> boolean option so that users have some way to replicate the same table\n> data among different nodes without causing an infinite loop and then\n> extend it as you are suggesting or may be in some other ways as well?\n\nYeah, that makes sense that first we provide some simple mechanism to\nenable it and we can extend it later.\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 7 Mar 2022 15:25:18 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 1:45 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 6:17 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Mon, Mar 7, 2022 at 11:45 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > >\n> > > On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\n> > > >\n> > > > On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > > > >\n> > > > > Hi Vignesh, I also have not looked at the patch yet, but I have what\n> > > > > seems like a very fundamental (and possibly dumb) question...\n> > > > >\n> > > > > Basically, I do not understand the choice of syntax for setting things up.\n> > > > >\n> > > > > IMO that \"only-local\" option sounds very similar to the other\n> > > > > PUBLICATION (\"publish\") options which decide the kinds of things that\n> > > > > will be published. So it feels more natural for me to think of the\n> > > > > publisher as being the one to decide what will be published.\n> > > > >\n> > > > > e.g.\n> > > > >\n> > > > > option 1:\n> > > > > CREATE PUBLICATION p1 FOR TABLE t1;\n> > > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n> > > > >\n> > > > > option 2:\n> > > > > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\n> > > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n> > > > >\n> > > > > ~~\n> > > > >\n> > > > > IIUC the patch is using option 1. My first impression was it feels\n> > > > > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\n> > > > > publish.\n> > > > >\n> > > > > So, why does the patch use syntax option 1?\n> > > >\n> > > > I felt the advantage with keeping it at the subscription side is that,\n> > > > the subscriber from one node can subscribe with only_local option on\n> > > > and a different subscriber from a different node can subscribe with\n> > > > only_local option as off. This might not be possible with having the\n> > > > option at publisher side. Having it at the subscriber side might give\n> > > > more flexibility for the user.\n> > > >\n> > >\n> > > OK. Option 2 needs two publications for that scenario. IMO it's more\n> > > intuitive this way, but maybe you wanted to avoid the extra\n> > > publications?\n> >\n> > Yes, I wanted to avoid the extra publication creation that you pointed\n> > out. Option 1 can handle this scenario without creating the extra\n> > publications:\n> > node0: CREATE PUBLICATION p1 FOR TABLE t1;\n> > node1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = on);\n> > node2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = off);\n> >\n> > I'm ok with both the approaches, now that this scenario can be handled\n> > by using both the options. i.e providing only_local option as an\n> > option while creating publication or providing only_local option as an\n> > option while creating subscription as Peter has pointed out at [1].\n> > option 1:\n> > CREATE PUBLICATION p1 FOR TABLE t1;\n> > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\n> >\n> > option 2:\n> > CREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\n> > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\n> >\n> > Shall we get a few opinions on this and take it in that direction?\n> >\n> > [1] - https://www.postgresql.org/message-id/CAHut%2BPsAWaETh9VMymbBfMrqiE1KuqMq%2BwpBg0s7eMzwLATr%2Bw%40mail.gmail.com\n> >\n> > Regards,\n> > Vignesh\n>\n> BTW here is a counter-example to your scenario from earlier.\n>\n> Let's say I have a publication p1 and p2 and want to subscribe to p1\n> with only_local=true, and p2 with only_local = false;\n>\n> Using the current OPtion 1 syntax you cannot do this with a single\n> subscription because the option is tied to the subscription.\n> But using syntax Option 2 you may be able to do it.\n>\n> Option 1:\n> CREATE PUBLICATION p1 FOR TABLE t1;\n> CREATE PUBLICATION p2 FOR TABLE t2;\n> CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 WITH (local_only = true);\n> CREATE SUBSCRIPTION s2 ... FOR PUBLICATION p1 WITH (local_only = false);\n>\n> Option 2:\n> CREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'local_only');\n> CREATE PUBLICATION p2 FOR TABLE t2;\n> CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1, p2;\n\nI felt having multiple publications will create duplicate entries in\nthe system table, Amit also has pointed this at [1]. Also enhancing\nthis approach to support filtering based on replication origin which\nis suggested by dilip at [2] is also on the client side and also the\ninitial check to handle the copy_data specified by Amit at [3] will be\ndone by the client side. Based on the above I feel the existing\napproach is better. I might be missing something here.\n\n[1] - https://www.postgresql.org/message-id/CAA4eK1LgCVv8u-fOsMPbGC96sWXhT3EKOBAeFW3g84otjStztw%40mail.gmail.com\n[2] - https://www.postgresql.org/message-id/CAFiTN-tKbjHDjAFNnqRoR8u1B%2Bfs0wunGz%3D3wp0iU-sUaxZJTQ%40mail.gmail.com\n[3] - https://www.postgresql.org/message-id/CAA4eK1%2Bco2cd8a6okgUD_pcFEHcc7mVc0k_RE2%3D6ahyv3WPRMg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 7 Mar 2022 16:23:17 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi Vignesh,\nI agree with Peter's comment that the changes to\nFilterRemoteOriginData() should be part of FilterByOrigin()\n\nFurther, I wonder why \"onlylocal_data\" is a replication slot's\nproperty. A replication slot tracks the progress of replication and it\nmay be used by different receivers with different options. I could\nstart one receiver which wants only local data, say using\n\"pg_logical_slot_get_changes\" and later start another receiver which\nfetches all the data starting from where the first receiver left. This\noption prevents such flexibility.\n\nAs discussed earlier in the thread, local_only can be property of\npublication or subscription, depending upon the use case, but I can't\nsee any reason that it should be tied to a replication slot.\n\nI have a similar question for \"two_phase\" but the ship has sailed and\nprobably it makes some sense there which I don't know.\n\nAs for publication vs subscription, I think both are useful cases.\n1. It will be a publication's property, if we want the node to not\npublish any data that it receives from other nodes for a given set of\ntables.\n2. It will be the subscription's property, if we want the subscription\nto decide whether it wants to fetch the data changed on only upstream\nor other nodes as well.\n\nMaybe we want to add it to both and use the stricter of both. If a\npublication has local_only, it publishes only local changes. If the\nsubscription has local only then WAL sender sends only local changes.\n\n--\nBest Wishes,\nAshutosh\n\n\n","msg_date":"Mon, 7 Mar 2022 17:00:56 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 5:01 PM Ashutosh Bapat\n<ashutosh.bapat.oss@gmail.com> wrote:\n>\n> Hi Vignesh,\n> I agree with Peter's comment that the changes to\n> FilterRemoteOriginData() should be part of FilterByOrigin()\n>\n> Further, I wonder why \"onlylocal_data\" is a replication slot's\n> property. A replication slot tracks the progress of replication and it\n> may be used by different receivers with different options. I could\n> start one receiver which wants only local data, say using\n> \"pg_logical_slot_get_changes\" and later start another receiver which\n> fetches all the data starting from where the first receiver left. This\n> option prevents such flexibility.\n>\n> As discussed earlier in the thread, local_only can be property of\n> publication or subscription, depending upon the use case, but I can't\n> see any reason that it should be tied to a replication slot.\n>\n\nI thought it should be similar to 'streaming' option of subscription\nbut may be Vignesh has some other reason which makes it different.\n\n> I have a similar question for \"two_phase\" but the ship has sailed and\n> probably it makes some sense there which I don't know.\n>\n\ntwo_phase is different from some of the other subscription options\nlike 'streaming' such that it can be enabled only at the time of slot\nand subscription creation, we can't change/specify it via\npg_logical_slot_get_changes. This is to avoid the case where we won't\nknow at the time of the commit prepared whether the prepare for the\ntransaction has already been sent. For the same reason, we need to\nalso know the 'two_phase_at' information.\n\n> As for publication vs subscription, I think both are useful cases.\n> 1. It will be a publication's property, if we want the node to not\n> publish any data that it receives from other nodes for a given set of\n> tables.\n> 2. It will be the subscription's property, if we want the subscription\n> to decide whether it wants to fetch the data changed on only upstream\n> or other nodes as well.\n>\n\nI think it could be useful to allow it via both publication and\nsubscription but I guess it is better to provide it via one way\ninitially just to keep things simple and give users some way to deal\nwith such cases. I would prefer to allow it via subscription initially\nfor the reasons specified by Vignesh in his previous email [1]. Now,\nif we think those all are ignorable things and it is more important to\nallow this option first by publication or we must allow it via both\npublication and subscription then it makes sense to change it.\n\n\n[1] - https://www.postgresql.org/message-id/CALDaNm3jkotRhKfCqu5CXOf36_yiiW_cYE5%3DbG%3Dj6N3gOWJkqw%40mail.gmail.com\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 7 Mar 2022 17:50:59 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 5:51 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 5:01 PM Ashutosh Bapat\n> <ashutosh.bapat.oss@gmail.com> wrote:\n> >\n> > Hi Vignesh,\n> > I agree with Peter's comment that the changes to\n> > FilterRemoteOriginData() should be part of FilterByOrigin()\n> >\n> > Further, I wonder why \"onlylocal_data\" is a replication slot's\n> > property. A replication slot tracks the progress of replication and it\n> > may be used by different receivers with different options. I could\n> > start one receiver which wants only local data, say using\n> > \"pg_logical_slot_get_changes\" and later start another receiver which\n> > fetches all the data starting from where the first receiver left. This\n> > option prevents such flexibility.\n> >\n> > As discussed earlier in the thread, local_only can be property of\n> > publication or subscription, depending upon the use case, but I can't\n> > see any reason that it should be tied to a replication slot.\n> >\n>\n> I thought it should be similar to 'streaming' option of subscription\n> but may be Vignesh has some other reason which makes it different.\n\nYes, this can be removed from the replication slot. It is my mistake\nthat I have made while making the code similar to two-phase, I'm\nworking on making the changes for this. I will fix and post an updated\nversion for this.\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 7 Mar 2022 21:57:03 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"IIUC the new option may be implemented subscriber-side and/or\npublisher-side and/or both, and the subscriber-side option may be\n\"enhanced\" in future to prevent cycles. And probably there are more\nfeatures I don't know about or that have not yet been thought of.\n\n~~\n\nEven if the plan is only to implement just one part now and then add\nmore later, I think there still should be some consideration for what\nyou expect all possible future options to look like, because that may\naffect current implementation choices.\n\nThe point is:\n\n- we should take care so don't accidentally end up with an option that\nturned out to be inconsistent looking on the subscriber-side /\npublisher-side.\n\n- we should try to avoid accidentally painting ourselves into a corner\n(e.g. stuck with a boolean option that cannot be enhanced later on)\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 8 Mar 2022 16:00:45 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Mar 8, 2022 at 10:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> IIUC the new option may be implemented subscriber-side and/or\n> publisher-side and/or both, and the subscriber-side option may be\n> \"enhanced\" in future to prevent cycles. And probably there are more\n> features I don't know about or that have not yet been thought of.\n>\n> ~~\n>\n> Even if the plan is only to implement just one part now and then add\n> more later, I think there still should be some consideration for what\n> you expect all possible future options to look like, because that may\n> affect current implementation choices.\n>\n> The point is:\n>\n> - we should take care so don't accidentally end up with an option that\n> turned out to be inconsistent looking on the subscriber-side /\n> publisher-side.\n>\n> - we should try to avoid accidentally painting ourselves into a corner\n> (e.g. stuck with a boolean option that cannot be enhanced later on)\n>\n\nAgreed. I think it is important to see we shouldn't do something which\nis not extendable in future or paint us in the corner. But, as of now,\nwith the current proposal, the main thing we should consider is\nwhether exposing the boolean option is okay or shall we do something\ndifferent so that we can extend it later to specific origin ids?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 8 Mar 2022 10:51:08 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Mar 8, 2022 at 4:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Mar 8, 2022 at 10:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > IIUC the new option may be implemented subscriber-side and/or\n> > publisher-side and/or both, and the subscriber-side option may be\n> > \"enhanced\" in future to prevent cycles. And probably there are more\n> > features I don't know about or that have not yet been thought of.\n> >\n> > ~~\n> >\n> > Even if the plan is only to implement just one part now and then add\n> > more later, I think there still should be some consideration for what\n> > you expect all possible future options to look like, because that may\n> > affect current implementation choices.\n> >\n> > The point is:\n> >\n> > - we should take care so don't accidentally end up with an option that\n> > turned out to be inconsistent looking on the subscriber-side /\n> > publisher-side.\n> >\n> > - we should try to avoid accidentally painting ourselves into a corner\n> > (e.g. stuck with a boolean option that cannot be enhanced later on)\n> >\n>\n> Agreed. I think it is important to see we shouldn't do something which\n> is not extendable in future or paint us in the corner. But, as of now,\n> with the current proposal, the main thing we should consider is\n> whether exposing the boolean option is okay or shall we do something\n> different so that we can extend it later to specific origin ids?\n>\n\nI was wondering, assuming later there is an enhancement to detect and\nprevent cycles using origin ids, then what really is the purpose of\nthe option?\n\nI can imagine if a cycle happens should probably log some one-time\nWARNING (just to bring it to the user's attention in case it was\ncaused by some accidental misconfiguration) but apart from that why\nwould this need to be an option at all - i.e. is there a use-case\nwhere a user would NOT want to prevent a recursive error?\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 8 Mar 2022 17:46:37 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Mar 8, 2022 at 12:17 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Tue, Mar 8, 2022 at 4:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Tue, Mar 8, 2022 at 10:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > >\n> > > IIUC the new option may be implemented subscriber-side and/or\n> > > publisher-side and/or both, and the subscriber-side option may be\n> > > \"enhanced\" in future to prevent cycles. And probably there are more\n> > > features I don't know about or that have not yet been thought of.\n> > >\n> > > ~~\n> > >\n> > > Even if the plan is only to implement just one part now and then add\n> > > more later, I think there still should be some consideration for what\n> > > you expect all possible future options to look like, because that may\n> > > affect current implementation choices.\n> > >\n> > > The point is:\n> > >\n> > > - we should take care so don't accidentally end up with an option that\n> > > turned out to be inconsistent looking on the subscriber-side /\n> > > publisher-side.\n> > >\n> > > - we should try to avoid accidentally painting ourselves into a corner\n> > > (e.g. stuck with a boolean option that cannot be enhanced later on)\n> > >\n> >\n> > Agreed. I think it is important to see we shouldn't do something which\n> > is not extendable in future or paint us in the corner. But, as of now,\n> > with the current proposal, the main thing we should consider is\n> > whether exposing the boolean option is okay or shall we do something\n> > different so that we can extend it later to specific origin ids?\n> >\n>\n> I was wondering, assuming later there is an enhancement to detect and\n> prevent cycles using origin ids, then what really is the purpose of\n> the option?\n>\n\nWe can't use origin ids during the initial sync as we can't\ndifferentiate between data from local or remote node. Also, how will\nwe distinguish between different origins? AFAIU, all the changes\napplied on any particular node use the same origin.\n\n> I can imagine if a cycle happens should probably log some one-time\n> WARNING (just to bring it to the user's attention in case it was\n> caused by some accidental misconfiguration) but apart from that why\n> would this need to be an option at all - i.e. is there a use-case\n> where a user would NOT want to prevent a recursive error?\n>\n\nI think there will be plenty of such setups where there won't be a\nneed for any such option like where users won't need bi-directional\nreplication for the same table.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 8 Mar 2022 13:24:47 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 2:28 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Hi Vignesh,\n>\n> Here are some review comments for patch v2.\n>\n> ======\n>\n> 1. Question about syntax\n>\n> I already posted some questions about why the syntax is on the CREATE\n> SUBSCRCRIBER side.\n> IMO \"local_only\" is a publisher option, so it seemed more natural to\n> me for it to be specified as a \"publish\" option.\n>\n> Ref [1] my original question + suggestion for Option 2\n> Ref [2] some other examples of subscribing to multiple-publishers\n>\n> Anyway, +1 to see what other people think.\n>\n\nI feel we can support it in the subscriber side first and then extend\nit to the publisher side as being discussed in [1]. I have retained it\nas it is.\n\n> ~~~\n>\n> 2. ALTER\n>\n> (related also to the question about syntax)\n>\n> If subscribing to multiple publications then ALTER is going to change\n> the 'local_only' for all of them, which might not be what you want\n> (??)\n>\n\n I feel we can support it in the subscriber side first and then extend\nit to the publisher side as being discussed in [1]. When it is\nextended to the publisher side, it will get handled. I have retained\nit as it is.\n\n> ~~~\n>\n> 3. subscription_parameter\n>\n> (related also to the question about syntax)\n>\n> CREATE SUBSCRIPTION subscription_name\n> CONNECTION 'conninfo'\n> PUBLICATION publication_name [, ...]\n> [ WITH ( subscription_parameter [= value] [, ... ] ) ]\n>\n> ~\n>\n> That WITH is for *subscription* options, not the publication options.\n>\n> So IMO 'local_only' intuitively seems like \"local\" means local where\n> the subscriber is.\n>\n> So, if the Option 1 syntax is chosen (see comment #1) then I think the\n> option name maybe should change to be something more like\n> 'publish_local_only' or something similar to be more clear what local\n> actually means.\n>\n\nChanged it to publish_local_only\n\n> ~~~\n>\n> 4. contrib/test_decoding/test_decoding.c\n>\n> @@ -484,6 +487,16 @@ pg_decode_filter(LogicalDecodingContext *ctx,\n> return false;\n> }\n>\n> +static bool\n> +pg_decode_filter_remotedata(LogicalDecodingContext *ctx,\n> + RepOriginId origin_id)\n> +{\n> + TestDecodingData *data = ctx->output_plugin_private;\n> +\n> + if (data->only_local && origin_id != InvalidRepOriginId)\n> + return true;\n> + return false;\n> +}\n>\n> 4a. Maybe needs function comment.\n\nModified\n\n> 4b. Missing blank line following this function\n>\n\nModified\n\n> ~~~\n>\n> 5. General - please check all of the patch.\n>\n> There seems inconsistency with the member names, local variable names,\n> parameter names etc. There are all variations of:\n>\n> - only_local\n> - onlylocaldata\n> - onlylocal_data\n> - etc\n>\n> Please try using the same name everywhere for everything if possible.\n>\n\nI have changed it to only_local wherever possible.\n\n> ~~~\n>\n> 6. src/backend/replication/logical/decode.c - FilterRemoteOriginData\n>\n> @@ -585,7 +594,8 @@ logicalmsg_decode(LogicalDecodingContext *ctx,\n> XLogRecordBuffer *buf)\n> message = (xl_logical_message *) XLogRecGetData(r);\n>\n> if (message->dbId != ctx->slot->data.database ||\n> - FilterByOrigin(ctx, origin_id))\n> + FilterByOrigin(ctx, origin_id) ||\n> + FilterRemoteOriginData(ctx, origin_id))\n> return;\n>\n> I noticed that every call to FilterRemoteOriginData has an associated\n> preceding call to FilterByOrigin. It might be worth just combining the\n> logic into FilterByOrigin. Then none of that calling code (9 x places)\n> would need to change at all.\n\nModified\n\n> ~~~\n>\n> 7. src/backend/replication/logical/logical.c - CreateInitDecodingContext\n>\n> @@ -451,6 +453,8 @@ CreateInitDecodingContext(const char *plugin,\n> */\n> ctx->twophase &= slot->data.two_phase;\n>\n> + ctx->onlylocal_data &= slot->data.onlylocal_data;\n>\n> The equivalent 'twophase' option had a big comment. Probably this new\n> option should also have a similar comment?\n\nThese change is not required anymore, the comment no more applies. I\nhave not made any change for this.\n\n> ~~~\n>\n> 8. src/backend/replication/logical/logical.c - filter_remotedata_cb_wrapper\n>\n> +bool\n> +filter_remotedata_cb_wrapper(LogicalDecodingContext *ctx,\n> + RepOriginId origin_id)\n> +{\n> + LogicalErrorCallbackState state;\n> + ErrorContextCallback errcallback;\n> + bool ret;\n> +\n> + Assert(!ctx->fast_forward);\n> +\n> + /* Push callback + info on the error context stack */\n> + state.ctx = ctx;\n> + state.callback_name = \"filter_remoteorigin\";\n>\n> There is no consistency between the function and the name:\n>\n> \"filter_remoteorigin\" versus filter_remotedata_cb.\n>\n> A similar inconsistency for this is elsewhere. See review comment #9\n\nModified it to filter_remote_origin_cb_wrapper and changed to\n*_remotedata_* to *_remote_origin_*\n\n> ~~~\n>\n> 9. src/backend/replication/pgoutput/pgoutput.c\n>\n> @@ -215,6 +217,7 @@ _PG_output_plugin_init(OutputPluginCallbacks *cb)\n> cb->commit_prepared_cb = pgoutput_commit_prepared_txn;\n> cb->rollback_prepared_cb = pgoutput_rollback_prepared_txn;\n> cb->filter_by_origin_cb = pgoutput_origin_filter;\n> + cb->filter_remotedata_cb = pgoutput_remoteorigin_filter;\n>\n> Inconsistent names for the member and function.\n>\n> filter_remotedata_cb VS pgoutput_remoteorigin_filter.\n\nModified it to filter_remote_origin_cb and changed to *_remotedata_*\nto *_remote_origin_*\n\n> ~~~\n>\n> 10. src/backend/replication/pgoutput/pgoutput.c\n>\n> @@ -1450,6 +1465,16 @@ pgoutput_origin_filter(LogicalDecodingContext *ctx,\n> return false;\n> }\n>\n> +static bool\n> +pgoutput_remoteorigin_filter(LogicalDecodingContext *ctx,\n> + RepOriginId origin_id)\n> +{\n> + PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\n> +\n> + if (data->onlylocal_data && origin_id != InvalidRepOriginId)\n> + return true;\n> + return false;\n> +}\n> /*\n> * Shutdown the output plugin.\n> *\n>\n> 10a. Add a function comment.\n\nModified\n\n> 10b. Missing blank line after the function\n\nModified\n\n> ~~~\n>\n> 11. src/backend/replication/slotfuncs.c - pg_create_logical_replication_slot\n>\n> @@ -171,6 +174,7 @@ pg_create_logical_replication_slot(PG_FUNCTION_ARGS)\n> Name plugin = PG_GETARG_NAME(1);\n> bool temporary = PG_GETARG_BOOL(2);\n> bool two_phase = PG_GETARG_BOOL(3);\n> + bool onlylocal_data = PG_GETARG_BOOL(4);\n> Datum result;\n> TupleDesc tupdesc;\n> HeapTuple tuple;\n>\n>\n> Won't there be some PG Docs needing to be updated now there is another\n> parameter?\n\nThis change is not required anymore, the comment no longer applies. I\nhave not made any changes for this.\n\n> ~~~\n>\n> 12. src/include/catalog/pg_proc.dat - pg_get_replication_slots\n>\n> I did not see any update for pg_get_replication_slots, but you added\n> the 4th parameter elsewhere. Is something missing here?\n\nThis change is not required anymore, the comment no longer applies. I\nhave not made any changes for this.\n\n> ~~~\n>\n> 13. src/include/replication/logical.h\n>\n> @@ -99,6 +99,8 @@ typedef struct LogicalDecodingContext\n> */\n> bool twophase_opt_given;\n>\n> + bool onlylocal_data;\n> +\n>\n> I think the new member needs some comment.\n\nModified\n\n> ~~~\n>\n> 14. src/include/replication/walreceiver.h\n>\n> @@ -183,6 +183,7 @@ typedef struct\n> bool streaming; /* Streaming of large transactions */\n> bool twophase; /* Streaming of two-phase transactions at\n> * prepare time */\n> + bool onlylocal_data;\n> } logical;\n> } proto;\n> } WalRcvStreamOptions;\n>\n> I think the new member needs some comment.\n\nModified\n\n> ~~~\n>\n> 15. src/test/regress/sql/subscription.sql\n>\n> ALTER SUBSCRIPTION test missing?\n\nAdded\n\nThanks for the comments, the attached v3 patch has the fixes for the same.\n[1] - https://www.postgresql.org/message-id/CAA4eK1%2BMtz%2BStvNNtTg9%3D9BTq8%3DpMu-V5i4yWqs%3DKJUh0Z_L4g%40mail.gmail.com\n\nRegards,\nVignesh","msg_date":"Tue, 8 Mar 2022 20:08:00 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Mar 7, 2022 at 9:57 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 5:51 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Mon, Mar 7, 2022 at 5:01 PM Ashutosh Bapat\n> > <ashutosh.bapat.oss@gmail.com> wrote:\n> > >\n> > > Hi Vignesh,\n> > > I agree with Peter's comment that the changes to\n> > > FilterRemoteOriginData() should be part of FilterByOrigin()\n> > >\n> > > Further, I wonder why \"onlylocal_data\" is a replication slot's\n> > > property. A replication slot tracks the progress of replication and it\n> > > may be used by different receivers with different options. I could\n> > > start one receiver which wants only local data, say using\n> > > \"pg_logical_slot_get_changes\" and later start another receiver which\n> > > fetches all the data starting from where the first receiver left. This\n> > > option prevents such flexibility.\n> > >\n> > > As discussed earlier in the thread, local_only can be property of\n> > > publication or subscription, depending upon the use case, but I can't\n> > > see any reason that it should be tied to a replication slot.\n> > >\n> >\n> > I thought it should be similar to 'streaming' option of subscription\n> > but may be Vignesh has some other reason which makes it different.\n>\n> Yes, this can be removed from the replication slot. It is my mistake\n> that I have made while making the code similar to two-phase, I'm\n> working on making the changes for this. I will fix and post an updated\n> version for this.\n\nI have made the changes for this, the changes for this are available\nin the v3 patch attached at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm0JcV-7iQZhyy3kehnWTy6x%3Dz%2BsX6u6Df%2B%2By8z33pz%2BBw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 8 Mar 2022 20:16:24 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi Vegnesh,\r\n\r\nWhile considering about second problem, I was very confusing about it.\r\nI'm happy if you answer my question.\r\n\r\n> To handle this if user has specified only_local option, we could throw\r\n> a warning or error out while creating subscription in this case, we\r\n> could have a column srreplicateddata in pg_subscription_rel which\r\n> could indicate if the table has any replicated data or not:\r\n> postgres=# select * from pg_subscription_rel;\r\n> srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\r\n> ---------+---------+------------+-----------+------------------\r\n> 16389 | 16384 | r | 0/14A4640 | t\r\n> 16389 | 16385 | r | 0/14A4690 | f\r\n> (1 row)\r\n> In the above example, srreplicateddata with true indicates, tabel t1\r\n> whose relid is 16384 has replicated data and the other row having\r\n> srreplicateddata as false indicates table t2 whose relid is 16385\r\n> does not have replicated data.\r\n> When creating a new subscription, the subscriber will connect to the\r\n> publisher and check if the relation has replicated data by checking\r\n> srreplicateddata in pg_subscription_rel table.\r\n> If the table has any replicated data, log a warning or error for this.\r\n\r\nIIUC srreplicateddata represents whether the subscribed data is not\r\ngenerated from the publisher, but another node.\r\nMy first impression was that the name 'srreplicateddata' is not friendly\r\nbecause all subscribed data is replicated from publisher.\r\nAlso I was not sure how value of the column was set.\r\nIIUC a filtering by replication origins is done in publisher node\r\nand subscriber node cannot know\r\nwhether some data are really filtered or not.\r\nIf we distinguish by subscriber option publish_local_only,\r\nit cannot reproduce your example because same subscriber have different 'srreplicateddata'.\r\n\r\nBest Regards,\r\nHayato Kuroda\r\nFUJITSU LIMITED\r\n\r\n","msg_date":"Fri, 11 Mar 2022 10:58:25 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Mar 2, 2022 at 3:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> ...\n> ...\n> > I have attached a basic patch for this, if the idea is accepted, I\n> > will work further to test more scenarios, add documentation, and test\n> > and post an updated patch.\n> > For the second problem, Table synchronization of table including local\n> > data and replicated data using copy command.\n> >\n> > Let us consider the following scenario:\n> > a) node1 publishing to node2 b) node2 publishing to node1. Here in\n> > this case node1 will have replicated data from node2 and vice versa.\n> >\n> > In the above if user wants to include node3 to subscribe data from\n> > node2. Users will have to create a subscription in node3 to get the\n> > data from node2. During table synchronization we send the complete\n> > table data from node2 to node3. Node2 will have local data from node2\n> > and also replicated data from node1. Currently we don't have an option\n> > to differentiate between the locally generated data and replicated\n> > data in the heap which will cause infinite recursion as described\n> > above.\n> >\n> > To handle this if user has specified only_local option, we could throw\n> > a warning or error out while creating subscription in this case, we\n> > could have a column srreplicateddata in pg_subscription_rel which\n> > could indicate if the table has any replicated data or not:\n> > postgres=# select * from pg_subscription_rel;\n> > srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\n> > ---------+---------+------------+-----------+------------------\n> > 16389 | 16384 | r | 0/14A4640 | t\n> > 16389 | 16385 | r | 0/14A4690 | f\n> > (1 row)\n> >\n> > In the above example, srreplicateddata with true indicates, tabel t1\n> > whose relid is 16384 has replicated data and the other row having\n> > srreplicateddata as false indicates table t2 whose relid is 16385\n> > does not have replicated data.\n> > When creating a new subscription, the subscriber will connect to the\n> > publisher and check if the relation has replicated data by checking\n> > srreplicateddata in pg_subscription_rel table.\n> > If the table has any replicated data, log a warning or error for this.\n> >\n>\n> If you want to give the error in this case, then I think we need to\n> provide an option to the user to allow copy. One possibility could be\n> to extend existing copy_data option as 'false', 'true', 'force'. For\n> 'false', there shouldn't be any change, for 'true', if 'only_local'\n> option is also set and the new column indicates replicated data then\n> give an error, for 'force', we won't give an error even if the\n> conditions as mentioned for 'true' case are met, rather we will allow\n> copy in this case.\n\nWhen a subscription is created with publish_local_only and copy_data,\nit will connect to the publisher and check if the published tables\nhave also been subscribed from other nodes by checking if the entry is\npresent in pg_subscription_rel and throw an error if present. The\nattached v4-0002-Support-force-option-for-copy_data-check-and-thro.patch\nhas the implementation for the same.\nThoughts?\n\nRegards,\nVignesh","msg_date":"Sat, 12 Mar 2022 10:46:19 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Mar 11, 2022 at 4:28 PM kuroda.hayato@fujitsu.com\n<kuroda.hayato@fujitsu.com> wrote:\n>\n> Hi Vegnesh,\n>\n> While considering about second problem, I was very confusing about it.\n> I'm happy if you answer my question.\n>\n> > To handle this if user has specified only_local option, we could throw\n> > a warning or error out while creating subscription in this case, we\n> > could have a column srreplicateddata in pg_subscription_rel which\n> > could indicate if the table has any replicated data or not:\n> > postgres=# select * from pg_subscription_rel;\n> > srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\n> > ---------+---------+------------+-----------+------------------\n> > 16389 | 16384 | r | 0/14A4640 | t\n> > 16389 | 16385 | r | 0/14A4690 | f\n> > (1 row)\n> > In the above example, srreplicateddata with true indicates, tabel t1\n> > whose relid is 16384 has replicated data and the other row having\n> > srreplicateddata as false indicates table t2 whose relid is 16385\n> > does not have replicated data.\n> > When creating a new subscription, the subscriber will connect to the\n> > publisher and check if the relation has replicated data by checking\n> > srreplicateddata in pg_subscription_rel table.\n> > If the table has any replicated data, log a warning or error for this.\n>\n> IIUC srreplicateddata represents whether the subscribed data is not\n> generated from the publisher, but another node.\n> My first impression was that the name 'srreplicateddata' is not friendly\n> because all subscribed data is replicated from publisher.\n> Also I was not sure how value of the column was set.\n> IIUC a filtering by replication origins is done in publisher node\n> and subscriber node cannot know\n> whether some data are really filtered or not.\n> If we distinguish by subscriber option publish_local_only,\n> it cannot reproduce your example because same subscriber have different 'srreplicateddata'.\n\nLet's consider an existing Multi master logical replication setup\nbetween Node1 and Node2 that is created using the following steps:\na) Node1 - Publication publishing employee table - pub1\nb) Node2 - Subscription subscribing from publication pub1 with\npublish_local_only - sub1_pub1_node1\nc) Node2 - Publication publishing employee table - pub2\nd) Node1 - Subscription subscribing from publication pub2 with\npublish_local_only - sub2_pub2_node2\n\nTo create a subscription in node3, we will be using the following steps:\na) Node2 - Publication publishing employee table. - pub3\nb) Node3 - Subscription subscribing from publication in Node2 with\npublish_local_only - sub3_pub3_node2\n\nWhen we create a subscription in Node3, Node3 will connect to\nNode2(this will not be done in Node3) and check if the employee table\nis present in pg_subscription_rel, in our case Node2 will have\nemployee table present in pg_subscription_rel (sub1_pub1_node1\nsubscribing to employee table from pub1 in Node1). As employee table\nis being subscribed in node2 from node1, we will throw an error like\nbelow:\npostgres=# create subscription sub2 CONNECTION 'dbname =postgres port\n= 9999' publication pub2 with (publish_local_only=on);\nERROR: CREATE/ALTER SUBSCRIPTION with publish_local_only and\ncopy_data as true is not allowed when the publisher might have\nreplicated data, table:public.t1 might have replicated data in the\npublisher\nHINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\n\nI was initially planning to add srreplicateddata field but I have\nchanged it slightly to keep the design simple. Now we just check if\nthe relation is present in pg_subscription_rel and throw an error if\ncopy_data and publish_local_only option is specified. The changes for\nthe same are available at [1].\n\n[1] - https://www.postgresql.org/message-id/CALDaNm0V%2B%3Db%3DCeZJNAAUO2PmSXH5QzNX3jADXb-0hGO_jVj0vA%40mail.gmail.com\nThoughts?\n\nRegards,\nVignesh\n\n\n","msg_date":"Sat, 12 Mar 2022 12:25:27 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Dear Vignesh,\r\n\r\nThank you for updating your patch!\r\n\r\n> Let's consider an existing Multi master logical replication setup\r\n> between Node1 and Node2 that is created using the following steps:\r\n> a) Node1 - Publication publishing employee table - pub1\r\n> b) Node2 - Subscription subscribing from publication pub1 with\r\n> publish_local_only - sub1_pub1_node1\r\n> c) Node2 - Publication publishing employee table - pub2\r\n> d) Node1 - Subscription subscribing from publication pub2 with\r\n> publish_local_only - sub2_pub2_node2\r\n> \r\n> To create a subscription in node3, we will be using the following steps:\r\n> a) Node2 - Publication publishing employee table. - pub3\r\n> b) Node3 - Subscription subscribing from publication in Node2 with\r\n> publish_local_only - sub3_pub3_node2\r\n> \r\n> When we create a subscription in Node3, Node3 will connect to\r\n> Node2(this will not be done in Node3) and check if the employee table\r\n> is present in pg_subscription_rel, in our case Node2 will have\r\n> employee table present in pg_subscription_rel (sub1_pub1_node1\r\n> subscribing to employee table from pub1 in Node1). As employee table\r\n> is being subscribed in node2 from node1, we will throw an error like\r\n> below:\r\n> postgres=# create subscription sub2 CONNECTION 'dbname =postgres port\r\n> = 9999' publication pub2 with (publish_local_only=on);\r\n> ERROR: CREATE/ALTER SUBSCRIPTION with publish_local_only and\r\n> copy_data as true is not allowed when the publisher might have\r\n> replicated data, table:public.t1 might have replicated data in the\r\n> publisher\r\n> HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\r\n\r\nThanks for kind explanation. \r\nI read above and your doc in 0002, and I put some comments.\r\n\r\n1. alter_subscription.sgml\r\n\r\n```\r\n- <term><literal>copy_data</literal> (<type>boolean</type>)</term>\r\n+ <term><literal>copy_data</literal> (<type>boolean</type> | <literal>force</literal>)</term>\r\n```\r\n\r\nI thought that it should be written as enum. For example, huge_pages GUC parameter\r\ncan accept {on, off, try}, and it has been written as enum.\r\n\r\n2. create_subscription.sgml\r\n\r\n```\r\n- <term><literal>copy_data</literal> (<type>boolean</type>)</term>\r\n+ <term><literal>copy_data</literal> (<type>boolean</type> | <literal>force</literal>)</term>\r\n```\r\n\r\nSame as above.\r\n\r\n3. create_subscription.sgml\r\n\r\n```\r\n+\r\n+ <para>\r\n+ If the publication tables were also subscribing data in the publisher\r\n+ from other publishers, it will affect the\r\n+ <command>CREATE SUBSCRIPTION</command> based on the value specified\r\n+ for <literal>publish_local_only</literal> option. Refer to the\r\n+ <xref linkend=\"sql-createsubscription-notes\" /> for details.\r\n+ </para>\r\n```\r\n\r\nI seeked docs, but the words \" publication tables \" have not seen.\r\nHow about \"tables in the publication\"?\r\n\r\n4. create_subscription.sgml - about your example \r\n\r\nIn the first section, we should describe about 2-nodes case more detail\r\nlike Amit mentioned in [1]. I thought that Option-3 can be resolved by defining\r\nsubscriptions in both nodes with publish_local_only = true and copy_data = force.\r\n\r\n> I was initially planning to add srreplicateddata field but I have\r\n> changed it slightly to keep the design simple. Now we just check if\r\n> the relation is present in pg_subscription_rel and throw an error if\r\n> copy_data and publish_local_only option is specified. The changes for\r\n> the same are available at [1].\r\n> \r\n> [1] -\r\n> https://www.postgresql.org/message-id/CALDaNm0V%2B%3Db%3DCeZJNAAU\r\n> O2PmSXH5QzNX3jADXb-0hGO_jVj0vA%40mail.gmail.com\r\n> Thoughts?\r\n\r\nActually I doubted that the column is really needed, so it is good.\r\n\r\nBest Regards,\r\nHayato Kuroda\r\nFUJITSU LIMITED\r\n\r\n","msg_date":"Tue, 15 Mar 2022 01:39:49 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"Dear Vignesh,\r\n\r\n> Thanks for kind explanation.\r\n> I read above and your doc in 0002, and I put some comments.\r\n\r\nI forgot a comment about 0002 doc.\r\n\r\n5. create_subscription.sgml - about your example\r\n\r\nThree possibilities were listed in the doc,\r\nbut I was not sure about b) case.\r\nIn the situation Node1 and Node2 have already become multi-master,\r\nand data has already synced at that time.\r\nIf so, how do we realize that \"there is data present only in one Node\"?\r\nCase a) and c) seem reasonable.\r\n\r\nBest Regards,\r\nHayato Kuroda\r\nFUJITSU LIMITED\r\n\r\n","msg_date":"Tue, 15 Mar 2022 04:25:21 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Mar 15, 2022 at 7:09 AM kuroda.hayato@fujitsu.com\n<kuroda.hayato@fujitsu.com> wrote:\n>\n> Dear Vignesh,\n>\n> Thank you for updating your patch!\n>\n> > Let's consider an existing Multi master logical replication setup\n> > between Node1 and Node2 that is created using the following steps:\n> > a) Node1 - Publication publishing employee table - pub1\n> > b) Node2 - Subscription subscribing from publication pub1 with\n> > publish_local_only - sub1_pub1_node1\n> > c) Node2 - Publication publishing employee table - pub2\n> > d) Node1 - Subscription subscribing from publication pub2 with\n> > publish_local_only - sub2_pub2_node2\n> >\n> > To create a subscription in node3, we will be using the following steps:\n> > a) Node2 - Publication publishing employee table. - pub3\n> > b) Node3 - Subscription subscribing from publication in Node2 with\n> > publish_local_only - sub3_pub3_node2\n> >\n> > When we create a subscription in Node3, Node3 will connect to\n> > Node2(this will not be done in Node3) and check if the employee table\n> > is present in pg_subscription_rel, in our case Node2 will have\n> > employee table present in pg_subscription_rel (sub1_pub1_node1\n> > subscribing to employee table from pub1 in Node1). As employee table\n> > is being subscribed in node2 from node1, we will throw an error like\n> > below:\n> > postgres=# create subscription sub2 CONNECTION 'dbname =postgres port\n> > = 9999' publication pub2 with (publish_local_only=on);\n> > ERROR: CREATE/ALTER SUBSCRIPTION with publish_local_only and\n> > copy_data as true is not allowed when the publisher might have\n> > replicated data, table:public.t1 might have replicated data in the\n> > publisher\n> > HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\n>\n> Thanks for kind explanation.\n> I read above and your doc in 0002, and I put some comments.\n>\n> 1. alter_subscription.sgml\n>\n> ```\n> - <term><literal>copy_data</literal> (<type>boolean</type>)</term>\n> + <term><literal>copy_data</literal> (<type>boolean</type> | <literal>force</literal>)</term>\n> ```\n>\n> I thought that it should be written as enum. For example, huge_pages GUC parameter\n> can accept {on, off, try}, and it has been written as enum.\n\nModified\n\n> 2. create_subscription.sgml\n>\n> ```\n> - <term><literal>copy_data</literal> (<type>boolean</type>)</term>\n> + <term><literal>copy_data</literal> (<type>boolean</type> | <literal>force</literal>)</term>\n> ```\n>\n> Same as above.\n\nModified\n\n> 3. create_subscription.sgml\n>\n> ```\n> +\n> + <para>\n> + If the publication tables were also subscribing data in the publisher\n> + from other publishers, it will affect the\n> + <command>CREATE SUBSCRIPTION</command> based on the value specified\n> + for <literal>publish_local_only</literal> option. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for details.\n> + </para>\n> ```\n>\n> I seeked docs, but the words \" publication tables \" have not seen.\n> How about \"tables in the publication\"?\n\nModified\n\n> 4. create_subscription.sgml - about your example\n>\n> In the first section, we should describe about 2-nodes case more detail\n> like Amit mentioned in [1]. I thought that Option-3 can be resolved by defining\n> subscriptions in both nodes with publish_local_only = true and copy_data = force.\n\nI thought existing information is enough because we have mentioned\nthat node1 and node2 have bidirectional replication setup done and\nboth the table data will be replicated and synchronized as and when\nthe DML operations are happening. In option-3 we need to create a\nsubscription with copy_data as force to one node and copy_data as\nfalse to another node because both nodes will be having the same data,\ncopying the data just from one of the nodes should be enough.\n\nThanks for the comments, the attached v5 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Wed, 16 Mar 2022 11:15:09 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Mar 15, 2022 at 9:55 AM kuroda.hayato@fujitsu.com\n<kuroda.hayato@fujitsu.com> wrote:\n>\n> Dear Vignesh,\n>\n> > Thanks for kind explanation.\n> > I read above and your doc in 0002, and I put some comments.\n>\n> I forgot a comment about 0002 doc.\n>\n> 5. create_subscription.sgml - about your example\n>\n> Three possibilities were listed in the doc,\n> but I was not sure about b) case.\n> In the situation Node1 and Node2 have already become multi-master,\n> and data has already synced at that time.\n> If so, how do we realize that \"there is data present only in one Node\"?\n> Case a) and c) seem reasonable.\n\nYour point is valid, modified.\n\nThe changes for the same are available int the v5 patch available at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 16 Mar 2022 11:17:16 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi Vignesh,\nSome review comments on 0001\n\nOn Wed, Mar 16, 2022 at 11:17 AM vignesh C <vignesh21@gmail.com> wrote:\n\n>\n> The changes for the same are available int the v5 patch available at [1].\n> [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\n>\n\n cb->truncate_cb = pg_decode_truncate;\n cb->commit_cb = pg_decode_commit_txn;\n cb->filter_by_origin_cb = pg_decode_filter;\n+ cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\n\nWhy do we need a new hook? Can we use filter_by_origin_cb? Also it looks like\nimplementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\nunless my eyes are deceiving me.\n\n- <literal>binary</literal>, <literal>streaming</literal>, and\n- <literal>disable_on_error</literal>.\n+ <literal>binary</literal>, <literal>streaming</literal>,\n+ <literal>disable_on_error</literal> and\n+ <literal>publish_local_only</literal>.\n\n\"publish_local_only\" as a \"subscription\" option looks odd. Should it be\n\"subscribe_local_only\"?\n\n\n+ <varlistentry>\n+ <term><literal>publish_local_only</literal>\n(<type>boolean</type>)</term>\n+ <listitem>\n+ <para>\n+ Specifies whether the subscription should subscribe only to the\n+ locally generated changes or subscribe to both the locally generated\n+ changes and the replicated changes that was generated from other\n+ nodes in the publisher. The default is <literal>false</literal>.\n\nThis description to uses verb \"subscribe\" instead of \"publish\".\n\n@@ -936,6 +951,13 @@ AlterSubscription(ParseState *pstate,\nAlterSubscriptionStmt *stmt,\n = true;\n }\n\n+ if (IsSet(opts.specified_opts, SUBOPT_PUBLISH_LOCAL_ONLY))\n+ {\n+ values[Anum_pg_subscription_sublocal - 1] =\n+ BoolGetDatum(opts.streaming);\n\nshould this be opts.sublocal?\n\n cb->commit_prepared_cb = pgoutput_commit_prepared_txn;\n cb->rollback_prepared_cb = pgoutput_rollback_prepared_txn;\n cb->filter_by_origin_cb = pgoutput_origin_filter;\n+ cb->filter_remote_origin_cb = pgoutput_remote_origin_filter;\n\npgoutput_origin_filter just returns false now. I think we should just enhance\nthat function and reuse the callback, instead of adding a new callback.\n\n--- a/src/include/replication/logical.h\n+++ b/src/include/replication/logical.h\n@@ -99,6 +99,8 @@ typedef struct LogicalDecodingContext\n */\n bool twophase_opt_given;\n\n+ bool only_local; /* publish only locally generated data */\n+\n\nIf we get rid of the new callback, we won't need this new member as well.\nAnyway the filtering happens within the output plugin. There is nothing that\ncore is required to do here.\n\n--- a/src/include/replication/walreceiver.h\n+++ b/src/include/replication/walreceiver.h\n@@ -183,6 +183,7 @@ typedef struct\n bool streaming; /* Streaming of large transactions */\n bool twophase; /* Streaming of two-phase transactions at\n * prepare time */\n+ bool only_local; /* publish only locally generated data */\n\nAre we using this anywhere. I couldn't spot any usage of this member. I might\nbe missing something though.\n\n\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Wed, 30 Mar 2022 19:22:12 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Mar 30, 2022 at 7:40 PM Ashutosh Bapat\n<ashutosh.bapat.oss@gmail.com> wrote:\n> >\n> > The changes for the same are available int the v5 patch available at [1].\n> > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\n> >\n>\n> cb->truncate_cb = pg_decode_truncate;\n> cb->commit_cb = pg_decode_commit_txn;\n> cb->filter_by_origin_cb = pg_decode_filter;\n> + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\n>\n> Why do we need a new hook? Can we use filter_by_origin_cb?\n>\n\nI also think there is no need for a new hook in this case but I might\nalso be missing something that Vignesh has in his mind.\n\n> Also it looks like\n> implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\n> unless my eyes are deceiving me.\n>\n> - <literal>binary</literal>, <literal>streaming</literal>, and\n> - <literal>disable_on_error</literal>.\n> + <literal>binary</literal>, <literal>streaming</literal>,\n> + <literal>disable_on_error</literal> and\n> + <literal>publish_local_only</literal>.\n>\n> \"publish_local_only\" as a \"subscription\" option looks odd. Should it be\n> \"subscribe_local_only\"?\n>\n\nI also think \"subscribe_local_only\" fits better here.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 31 Mar 2022 09:14:53 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Mar 31, 2022 at 9:14 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Mar 30, 2022 at 7:40 PM Ashutosh Bapat\n> <ashutosh.bapat.oss@gmail.com> wrote:\n> > >\n> > > The changes for the same are available int the v5 patch available at [1].\n> > > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\n> > >\n> >\n> > cb->truncate_cb = pg_decode_truncate;\n> > cb->commit_cb = pg_decode_commit_txn;\n> > cb->filter_by_origin_cb = pg_decode_filter;\n> > + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\n> >\n> > Why do we need a new hook? Can we use filter_by_origin_cb?\n> >\n>\n> I also think there is no need for a new hook in this case but I might\n> also be missing something that Vignesh has in his mind.\n>\n> > Also it looks like\n> > implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\n> > unless my eyes are deceiving me.\n> >\n> > - <literal>binary</literal>, <literal>streaming</literal>, and\n> > - <literal>disable_on_error</literal>.\n> > + <literal>binary</literal>, <literal>streaming</literal>,\n> > + <literal>disable_on_error</literal> and\n> > + <literal>publish_local_only</literal>.\n> >\n> > \"publish_local_only\" as a \"subscription\" option looks odd. Should it be\n> > \"subscribe_local_only\"?\n> >\n>\n> I also think \"subscribe_local_only\" fits better here.\n>\n\nAbout 0002 patch,\nCommit message:\n------\nIf a subscription is created to Node1 from Node3 with publish_local_only\nand copy_data as ON, then throw an error so that user can handle\ncreation of subscription with table having consistent data.\n------\n\nDo you want to refer to Node2 instead of Node3 here as Node3 doesn't\nmake sense in the description?\n\nAlso, you haven't explained anywhere in the patch why\n'publish_local_only' (or whatever we call it) won't work for initial\nsync? IIUC, it is because we can identify origin changes only based on\nWAL and initial sync directly copies data from the heap. Am, I missing\nsomething here?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 31 Mar 2022 14:51:54 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Mar 30, 2022 at 7:22 PM Ashutosh Bapat\n<ashutosh.bapat.oss@gmail.com> wrote:\n>\n> Hi Vignesh,\n> Some review comments on 0001\n>\n> On Wed, Mar 16, 2022 at 11:17 AM vignesh C <vignesh21@gmail.com> wrote:\n>\n> >\n> > The changes for the same are available int the v5 patch available at [1].\n> > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\n> >\n>\n> cb->truncate_cb = pg_decode_truncate;\n> cb->commit_cb = pg_decode_commit_txn;\n> cb->filter_by_origin_cb = pg_decode_filter;\n> + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\n>\n> Why do we need a new hook? Can we use filter_by_origin_cb? Also it looks like\n> implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\n> unless my eyes are deceiving me.\n\nI have used filter_by_origin_cb for the implementation, removed\nfilter_remote_origin_cb\n\n> - <literal>binary</literal>, <literal>streaming</literal>, and\n> - <literal>disable_on_error</literal>.\n> + <literal>binary</literal>, <literal>streaming</literal>,\n> + <literal>disable_on_error</literal> and\n> + <literal>publish_local_only</literal>.\n>\n> \"publish_local_only\" as a \"subscription\" option looks odd. Should it be\n> \"subscribe_local_only\"?\n\nModified\n\n>\n> + <varlistentry>\n> + <term><literal>publish_local_only</literal>\n> (<type>boolean</type>)</term>\n> + <listitem>\n> + <para>\n> + Specifies whether the subscription should subscribe only to the\n> + locally generated changes or subscribe to both the locally generated\n> + changes and the replicated changes that was generated from other\n> + nodes in the publisher. The default is <literal>false</literal>.\n>\n> This description to uses verb \"subscribe\" instead of \"publish\".\n\nModified\n\n> @@ -936,6 +951,13 @@ AlterSubscription(ParseState *pstate,\n> AlterSubscriptionStmt *stmt,\n> = true;\n> }\n>\n> + if (IsSet(opts.specified_opts, SUBOPT_PUBLISH_LOCAL_ONLY))\n> + {\n> + values[Anum_pg_subscription_sublocal - 1] =\n> + BoolGetDatum(opts.streaming);\n>\n> should this be opts.sublocal?\n\nYes you are right, Modified\n\n> cb->commit_prepared_cb = pgoutput_commit_prepared_txn;\n> cb->rollback_prepared_cb = pgoutput_rollback_prepared_txn;\n> cb->filter_by_origin_cb = pgoutput_origin_filter;\n> + cb->filter_remote_origin_cb = pgoutput_remote_origin_filter;\n>\n> pgoutput_origin_filter just returns false now. I think we should just enhance\n> that function and reuse the callback, instead of adding a new callback.\n\nModified\n\n> --- a/src/include/replication/logical.h\n> +++ b/src/include/replication/logical.h\n> @@ -99,6 +99,8 @@ typedef struct LogicalDecodingContext\n> */\n> bool twophase_opt_given;\n>\n> + bool only_local; /* publish only locally generated data */\n> +\n>\n> If we get rid of the new callback, we won't need this new member as well.\n> Anyway the filtering happens within the output plugin. There is nothing that\n> core is required to do here.\n\nModified\n\n> --- a/src/include/replication/walreceiver.h\n> +++ b/src/include/replication/walreceiver.h\n> @@ -183,6 +183,7 @@ typedef struct\n> bool streaming; /* Streaming of large transactions */\n> bool twophase; /* Streaming of two-phase transactions at\n> * prepare time */\n> + bool only_local; /* publish only locally generated data */\n>\n> Are we using this anywhere. I couldn't spot any usage of this member. I might\n> be missing something though.\n\nThis is set in ApplyWorkerMain before calling libpqrcv_startstreaming.\nThis will be used in building the START_REPLICATION command.\nThanks for the comments, the attached v6 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Fri, 1 Apr 2022 12:34:21 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Mar 31, 2022 at 2:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Thu, Mar 31, 2022 at 9:14 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Wed, Mar 30, 2022 at 7:40 PM Ashutosh Bapat\n> > <ashutosh.bapat.oss@gmail.com> wrote:\n> > > >\n> > > > The changes for the same are available int the v5 patch available at [1].\n> > > > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\n> > > >\n> > >\n> > > cb->truncate_cb = pg_decode_truncate;\n> > > cb->commit_cb = pg_decode_commit_txn;\n> > > cb->filter_by_origin_cb = pg_decode_filter;\n> > > + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\n> > >\n> > > Why do we need a new hook? Can we use filter_by_origin_cb?\n> > >\n> >\n> > I also think there is no need for a new hook in this case but I might\n> > also be missing something that Vignesh has in his mind.\n> >\n> > > Also it looks like\n> > > implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\n> > > unless my eyes are deceiving me.\n> > >\n> > > - <literal>binary</literal>, <literal>streaming</literal>, and\n> > > - <literal>disable_on_error</literal>.\n> > > + <literal>binary</literal>, <literal>streaming</literal>,\n> > > + <literal>disable_on_error</literal> and\n> > > + <literal>publish_local_only</literal>.\n> > >\n> > > \"publish_local_only\" as a \"subscription\" option looks odd. Should it be\n> > > \"subscribe_local_only\"?\n> > >\n> >\n> > I also think \"subscribe_local_only\" fits better here.\n> >\n>\n> About 0002 patch,\n> Commit message:\n> ------\n> If a subscription is created to Node1 from Node3 with publish_local_only\n> and copy_data as ON, then throw an error so that user can handle\n> creation of subscription with table having consistent data.\n> ------\n> Do you want to refer to Node2 instead of Node3 here as Node3 doesn't\n> make sense in the description?\n\nIt should be Node3, I have added more details in the commit message\nmentioning about the scenario.\n\n> Also, you haven't explained anywhere in the patch why\n> 'publish_local_only' (or whatever we call it) won't work for initial\n> sync? IIUC, it is because we can identify origin changes only based on\n> WAL and initial sync directly copies data from the heap. Am, I missing\n> something here?\n\nI have added the explanation where we are throwing the error.\n\nThe v6 patch at [1] has the changes for the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm16eBx2BjLFjfFHSU4pb25HmgV--M692OPgH_91Dwn%3D2g%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 1 Apr 2022 12:45:50 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"I didn't find this in https://commitfest.postgresql.org/37/. Is this\nsomewhere in the commitfest?\n\nOn Fri, Apr 1, 2022 at 12:46 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Thu, Mar 31, 2022 at 2:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Thu, Mar 31, 2022 at 9:14 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > >\n> > > On Wed, Mar 30, 2022 at 7:40 PM Ashutosh Bapat\n> > > <ashutosh.bapat.oss@gmail.com> wrote:\n> > > > >\n> > > > > The changes for the same are available int the v5 patch available at [1].\n> > > > > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\n> > > > >\n> > > >\n> > > > cb->truncate_cb = pg_decode_truncate;\n> > > > cb->commit_cb = pg_decode_commit_txn;\n> > > > cb->filter_by_origin_cb = pg_decode_filter;\n> > > > + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\n> > > >\n> > > > Why do we need a new hook? Can we use filter_by_origin_cb?\n> > > >\n> > >\n> > > I also think there is no need for a new hook in this case but I might\n> > > also be missing something that Vignesh has in his mind.\n> > >\n> > > > Also it looks like\n> > > > implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\n> > > > unless my eyes are deceiving me.\n> > > >\n> > > > - <literal>binary</literal>, <literal>streaming</literal>, and\n> > > > - <literal>disable_on_error</literal>.\n> > > > + <literal>binary</literal>, <literal>streaming</literal>,\n> > > > + <literal>disable_on_error</literal> and\n> > > > + <literal>publish_local_only</literal>.\n> > > >\n> > > > \"publish_local_only\" as a \"subscription\" option looks odd. Should it be\n> > > > \"subscribe_local_only\"?\n> > > >\n> > >\n> > > I also think \"subscribe_local_only\" fits better here.\n> > >\n> >\n> > About 0002 patch,\n> > Commit message:\n> > ------\n> > If a subscription is created to Node1 from Node3 with publish_local_only\n> > and copy_data as ON, then throw an error so that user can handle\n> > creation of subscription with table having consistent data.\n> > ------\n> > Do you want to refer to Node2 instead of Node3 here as Node3 doesn't\n> > make sense in the description?\n>\n> It should be Node3, I have added more details in the commit message\n> mentioning about the scenario.\n>\n> > Also, you haven't explained anywhere in the patch why\n> > 'publish_local_only' (or whatever we call it) won't work for initial\n> > sync? IIUC, it is because we can identify origin changes only based on\n> > WAL and initial sync directly copies data from the heap. Am, I missing\n> > something here?\n>\n> I have added the explanation where we are throwing the error.\n>\n> The v6 patch at [1] has the changes for the same.\n> [1] - https://www.postgresql.org/message-id/CALDaNm16eBx2BjLFjfFHSU4pb25HmgV--M692OPgH_91Dwn%3D2g%40mail.gmail.com\n>\n> Regards,\n> Vignesh\n\n\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Mon, 4 Apr 2022 18:50:08 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my comments for the latest patch v6-0001.\n\n(I will post my v6-0002 review comments separately)\n\nPATCH v6-0001 comments\n======================\n\n1.1 General - Option name\n\nI still feel like the option name is not ideal. Unfortunately, this is\nimportant because any name change would impact lots of these patch\nfiles and docs, struct members etc.\n\nIt was originally called \"local_only\", but I thought that as a\nSUBSCRIPTION option that was confusing because \"local\" means local to\nwhat? Really it is local to the publisher, not local to the\nsubscriber, so that name seemed misleading.\n\nThen I suggested \"publish_local_only\". Although that resolved the\nambiguity problem, other people thought it seemed odd to have the\n\"publish\" prefix for a subscription-side option.\n\nSo now it is changed again to \"subscribe_local_only\" -- It's getting\nbetter but still, it is implied that the \"local\" means local to the\npublisher except there is nothing in the option name really to convey\nthat meaning. IMO we here all understand the meaning of this option\nmostly by familiarity with this discussion thread, but I think a user\ncoming to this for the first time will still be confused by the name.\n\nBelow are some other name ideas. Maybe they are not improvements, but\nit might help other people to come up with something better.\n\nsubscribe_publocal_only = true/false\norigin = pub_only/any\nadjacent_data_only = true/false\ndirect_subscriptions_only = true/false\n...\n\n\n(FYI, the remainder of these review comments will assume the option is\nstill called \"subscribe_local_only\")\n\n~~~\n\n1.2 General - inconsistent members and args\n\nIMO the struct members and args should also be named for close\nconsistency with whatever the option name is.\n\nCurrently the option is called \"subscription_local_only\". So I think\nthe members/args would be better to be called \"local_only\" instead of\n\"only_local\".\n\n~~~\n\n1.3 Commit message - wrong option name\n\nThe commit message refers to the option name as \"publish_local_only\"\ninstead of the option name that is currently implemented.\n\n~~~\n\n1.4 Commit message - wording\n\nThe wording seems a bit off. Below is suggested simpler wording which\nI AFAIK conveys the same information.\n\nBEFORE\nAdd an option publish_local_only which will subscribe only to the locally\ngenerated data in the publisher node. If subscriber is created with this\noption, publisher will skip publishing the data that was subscribed\nfrom other nodes. It can be created using following syntax:\nex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\nPUBLICATION pub1 with (publish_local_only = on);\n\nSUGGESTION\nThis patch adds a new SUBSCRIPTION boolean option\n\"subscribe_local_only\". The default is false. When a SUBSCRIPTION is\ncreated with this option enabled, the publisher will only publish data\nthat originated at the publisher node.\nUsage:\nCREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\nPUBLICATION pub1 with (subscribe_local_only = true);\n\n~~~\n\n1.5 doc/src/sgml/ref/create_subscription.sgml - \"generated\" changes.\n\n+ <para>\n+ Specifies whether the subscription will request the publisher to send\n+ locally generated changes or both the locally generated changes and\n+ the replicated changes that was generated from other nodes. The\n+ default is <literal>false</literal>.\n+ </para>\n\nFor some reason, it seemed a bit strange to me to use the term\n\"generated\" changes. Maybe better to refer to the origin of changes?\n\nSUGGESTION\nSpecifies whether the publisher should send only changes that\noriginated locally at the publisher node, or send any publisher node\nchanges regardless of their origin. The default is false.\n\n~~~\n\n1.6 src/backend/replication/pgoutput/pgoutput.c -\nLOGICALREP_PROTO_TWOPHASE_VERSION_NUM\n\n@@ -496,6 +509,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\nOutputPluginOptions *opt,\n else\n ctx->twophase_opt_given = true;\n\n+ if (data->only_local && data->protocol_version <\nLOGICALREP_PROTO_TWOPHASE_VERSION_NUM)\n+ ereport(ERROR,\n+ (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\n+ errmsg(\"requested proto_version=%d does not support\nsubscribe_local_only, need %d or higher\",\n+ data->protocol_version, LOGICALREP_PROTO_TWOPHASE_VERSION_NUM)));\n\nI thought this code should not be using\nLOGICALREP_PROTO_TWOPHASE_VERSION_NUM. Shouldn't there be some newly\nintroduced constant like LOGICALREP_PROTO_LOCALONLY_VERSION_NUM which\nyou will use here?\n\n~~~\n\n1.7 src/bin/pg_dump/pg_dump.c - 150000\n\n@@ -4451,11 +4452,13 @@ getSubscriptions(Archive *fout)\n if (fout->remoteVersion >= 150000)\n appendPQExpBufferStr(query,\n \" s.subtwophasestate,\\n\"\n- \" s.subdisableonerr\\n\");\n+ \" s.subdisableonerr,\\n\"\n+ \" s.sublocal\\n\");\n else\n appendPQExpBuffer(query,\n \" '%c' AS subtwophasestate,\\n\"\n- \" false AS subdisableonerr\\n\",\n+ \" false AS subdisableonerr,\\n\"\n+ \" false AS s.sublocal\\n\",\n LOGICALREP_TWOPHASE_STATE_DISABLED);\n\nI think this local_only feature is unlikely to get into the PG15\nrelease, so this code should be split out into a separate condition\nbecause later will need to change to say >= 160000.\n\n~~~\n\n1.8 src/bin/pg_dump/pg_dump.c - dumpSubscription\n\n@@ -4585,6 +4591,9 @@ dumpSubscription(Archive *fout, const\nSubscriptionInfo *subinfo)\n if (strcmp(subinfo->subdisableonerr, \"t\") == 0)\n appendPQExpBufferStr(query, \", disable_on_error = true\");\n\n+ if (strcmp(subinfo->sublocal, \"f\") != 0)\n+ appendPQExpBufferStr(query, \", subscribe_local_only = on\");\n+\n\nI felt it is more natural to say \"if it is true set to true\", instead\nof \"if it is not false set to on\".\n\nSUGGESTION\nif (strcmp(subinfo->sublocal, \"t\") == 0)\nappendPQExpBufferStr(query, \", subscribe_local_only = true\");\n\n~~~\n\n1.9 src/bin/psql/describe.c - 150000\n\n@@ -6318,9 +6318,11 @@ describeSubscriptions(const char *pattern, bool verbose)\n if (pset.sversion >= 150000)\n appendPQExpBuffer(&buf,\n \", subtwophasestate AS \\\"%s\\\"\\n\"\n- \", subdisableonerr AS \\\"%s\\\"\\n\",\n+ \", subdisableonerr AS \\\"%s\\\"\\n\"\n+ \", sublocal AS \\\"%s\\\"\\n\",\n gettext_noop(\"Two phase commit\"),\n- gettext_noop(\"Disable on error\"));\n+ gettext_noop(\"Disable on error\"),\n+ gettext_noop(\"Only local\"));\n\nI think this local_only feature is unlikely to get into the PG15\nrelease, so this code should be split out into a separate condition\nbecause later will need to change to say >= 160000.\n\n~~~\n\n1.10 src/bin/psql/describe.c - describeSubscriptions column\n\n@@ -6318,9 +6318,11 @@ describeSubscriptions(const char *pattern, bool verbose)\n if (pset.sversion >= 150000)\n appendPQExpBuffer(&buf,\n \", subtwophasestate AS \\\"%s\\\"\\n\"\n- \", subdisableonerr AS \\\"%s\\\"\\n\",\n+ \", subdisableonerr AS \\\"%s\\\"\\n\"\n+ \", sublocal AS \\\"%s\\\"\\n\",\n gettext_noop(\"Two phase commit\"),\n- gettext_noop(\"Disable on error\"));\n+ gettext_noop(\"Disable on error\"),\n+ gettext_noop(\"Only local\"));\n\nI think the column name here should be more consistent with the option\nname. e.g. it should be \"Local only\", not \"Only local\".\n\n~~~\n\n1.11 src/bin/psql/tab-complete.c - whitespace\n\n@@ -3167,7 +3167,7 @@ psql_completion(const char *text, int start, int end)\n /* Complete \"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\" */\n else if (HeadMatches(\"CREATE\", \"SUBSCRIPTION\") && TailMatches(\"WITH\", \"(\"))\n COMPLETE_WITH(\"binary\", \"connect\", \"copy_data\", \"create_slot\",\n- \"enabled\", \"slot_name\", \"streaming\",\n+ \"enabled\", \"slot_name\", \"streaming\", \"subscribe_local_only\",\n\nThe patch accidentally added a space char before the \"slot_name\".\n\n~~~\n\n1.12 src/include/replication/walreceiver.h - \"generated\"\n\n@@ -183,6 +183,7 @@ typedef struct\n bool streaming; /* Streaming of large transactions */\n bool twophase; /* Streaming of two-phase transactions at\n * prepare time */\n+ bool only_local; /* publish only locally generated data */\n\nThis is a similar review comment as #1.5 about saying the word \"generated\".\nMaybe there is another way to word this?\n\n~~~\n\n1.13 src/test/regress/sql/subscription.sql - missing test case\n\nIsn't there a missing test case for ensuring that the new option is boolean?\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 5 Apr 2022 10:51:20 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Apr 4, 2022 at 6:50 PM Ashutosh Bapat\n<ashutosh.bapat.oss@gmail.com> wrote:\n>\n> I didn't find this in https://commitfest.postgresql.org/37/. Is this\n> somewhere in the commitfest?\n\nI missed adding it earlier, I have added it to commitfest today:\nhttps://commitfest.postgresql.org/38/3610/\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 5 Apr 2022 07:24:37 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my comments for the latest patch v6-0001.\n>\n> (I will post my v6-0002 review comments separately)\n>\n> PATCH v6-0001 comments\n> ======================\n>\n> 1.1 General - Option name\n>\n> I still feel like the option name is not ideal. Unfortunately, this is\n> important because any name change would impact lots of these patch\n> files and docs, struct members etc.\n>\n> It was originally called \"local_only\", but I thought that as a\n> SUBSCRIPTION option that was confusing because \"local\" means local to\n> what? Really it is local to the publisher, not local to the\n> subscriber, so that name seemed misleading.\n>\n> Then I suggested \"publish_local_only\". Although that resolved the\n> ambiguity problem, other people thought it seemed odd to have the\n> \"publish\" prefix for a subscription-side option.\n>\n> So now it is changed again to \"subscribe_local_only\" -- It's getting\n> better but still, it is implied that the \"local\" means local to the\n> publisher except there is nothing in the option name really to convey\n> that meaning. IMO we here all understand the meaning of this option\n> mostly by familiarity with this discussion thread, but I think a user\n> coming to this for the first time will still be confused by the name.\n>\n> Below are some other name ideas. Maybe they are not improvements, but\n> it might help other people to come up with something better.\n>\n> subscribe_publocal_only = true/false\n> origin = pub_only/any\n> adjacent_data_only = true/false\n> direct_subscriptions_only = true/false\n> ...\n>\n\nI think local_only with a description should be okay considering other\ncurrent options. Some of the options like slot_name, binary, etc. are\npublisher-specific which becomes clear only after reading the\ndescription. I am not sure adding pub*/sub* to it is much helpful. So,\n+1 to a simple boolean like 'local_only', 'data_local', or something\nlike that.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 5 Apr 2022 08:06:18 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my comments for the latest patch v6-0002.\n\nPATCH v6-0002 comments\n======================\n\n2.1 General - should this be an independent patch?\n\nIn many ways, I think most of this patch is unrelated to the other\n\"local_only\" patch (v6-0001).\n\nFor example, IIUC even in the current HEAD, we could consider it to be\na user error if multiple SUBSCRIPTIONS or multiple PUBLICATIONS of the\nsame SUBSCRIPTION are replicating to the same TABLE on the same node\nand using \"copy_data = on\".\n\nSo I think it would be ok to throw an ERROR if such a copy_data clash\nis detected, and then the user will have to change to use \"copy_data =\noff\" for some/all of them to avoid data duplication.\n\nThe \"local_only\" option only adds some small logic to this new ERROR,\nbut it's not really a prerequisite at all.\n\ne.g. this whole ERROR part of the patch can be a separate thread.\n\n~~~\n\n2.2 General - can we remove the \"force\" enum?\n\nNow, because I consider the clashing \"copy_data = on\" ERROR to be a\nuser error, I think that is something that the user can already take\ncare of themselves just using the copy_data = off.\n\nI did not really like the modifying of the \"copy_data\" option from\njust boolean to some kind of hybrid boolean + \"force\".\n\na) It smells a bit off to me. IMO replication is supposed to end up\nwith the same (replicated) data on the standby machine but this\n\"force\" mode seems to be just helping the user to break that concept\nand say - \"I know what I'm doing, and I don't care if I get lots of\nduplicated data in the replica table - just let me do it\"...\n\nb) It also somehow feels like the new \"force\" was introduced mostly to\nmake the code ERROR handling implementation simpler, rather than to\nmake the life of the end-user better. Yes, if force is removed maybe\nthe copy-clash-detection-code will need to be internally quite more\ncomplex than it is now, but that is how it should be, instead of\nputting extra burden on the user (e.g. by complicating the PG docs and\ngiving them yet more alternatives to configure). I think any clashing\ncopy_data options really is a user error, but also I think the current\nboolean copy_data true/false already gives the user a way to fix it.\n\nc) Actually, your new error hint messages are similar to my\nperspective: They already say \"Use CREATE/ALTER SUBSCRIPTION with\ncopy_data = off or force\". All I am saying is remove the \"force\", and\nthe user can still fix the problem just by using \"copy_data = off\"\nappropriately.\n\n======\n\nSo (from above) I am not much in favour of the copy_data becoming a\nhybrid enum and using \"force\", yet that is what most of this patch is\nimplementing. Anyway, the remainder of my review comments here are for\nthe code in its current form. Maybe if \"force\" can be removed most of\nthe following comments end up being redundant.\n\n======\n\n2.3 Commit message - wording\n\nThis message is difficult to understand.\n\nI think that the long sentence \"Now when user is trying...\" can be\nbroken into more manageable parts.\n\nThis part \"and throw an error so that user can handle the initial copy\ndata.\" also seemed a bit vague.\n\n~~~\n\n2.4 Commit message - more functions\n\n\"This patch does couple of things:\"\n\nIIUC, there seems a third thing implemented by this patch but not\ndescribed by the comment. I think it also adds support for ALTER\nSUBSCRIPTION SET PUBLICATION WITH (subscribe_local_only)\n\n~~~\n\n2.5 doc/src/sgml/ref/create_subscription.sgml - wording\n\n@@ -161,6 +161,13 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n the replicated changes that was generated from other nodes. The\n default is <literal>false</literal>.\n </para>\n+ <para>\n+ If the tables in the publication were also subscribing to the data in\n+ the publisher from other publishers, it will affect the\n+ <command>CREATE SUBSCRIPTION</command> based on the value specified\n+ for <literal>copy_data</literal> option. Refer to the\n+ <xref linkend=\"sql-createsubscription-notes\" /> for details.\n+ </para>\n\nIs there is a simpler way to express all that?\n\nSUGGESTION\nThere is some interation between the option \"subscribe_local_only\" and\noption \"copy_data\". Refer to the <xref\nlinkend=\"sql-createsubscription-notes\" /> for details.\n\n~~~\n\n2.6 doc/src/sgml/ref/create_subscription.sgml - whitespace\n\n@@ -213,18 +220,28 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n </varlistentry>\n\n <varlistentry>\n- <term><literal>copy_data</literal> (<type>boolean</type>)</term>\n+ <term><literal>copy_data</literal> (<type>enum</type>)</term>\n <listitem>\n <para>\n Specifies whether to copy pre-existing data in the publications\n- that are being subscribed to when the replication starts.\n- The default is <literal>true</literal>.\n+ that are being subscribed to when the replication starts. This\n+ parameter may be either <literal>true</literal>,\n+ <literal>false</literal> or <literal>force</literal>. The default is\n+ <literal>true</literal>.\n\nThat last line has trailing whitespace.\n\n~~~\n\n2.7 doc/src/sgml/ref/create_subscription.sgml - wording\n\n+ <para>\n+ If the tables in the publication were also subscribing to the data in\n+ the publisher from other publishers, it will affect the\n+ <command>CREATE SUBSCRIPTION</command> based on the value specified\n+ for <literal>subscribe_local_only</literal> option. Refer to the\n+ <xref linkend=\"sql-createsubscription-notes\" /> for details.\n+ </para>\n\nThis is similar to review comment #2.5 which I thought could be\nwritten in a simpler way.\n\n~~~\n\n2.8 doc/src/sgml/ref/create_subscription.sgml\n\n@@ -375,6 +392,42 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n can have non-existent publications.\n </para>\n\n+ <para>\n+ Let's consider an existing Multi master logical replication setup between\n+ Node1 and Node2 that is created using the following steps:\n+ a) Node1 - Publication publishing employee table.\n+ b) Node2 - Subscription subscribing from publication pub1 with\n+ <literal>subscribe_local_only</literal>.\n+ c) Node2 - Publication publishing employee table.\n+ d) Node1 - Subscription subscribing from publication pub2 with\n+ <literal>subscribe_local_only</literal>.\n+ Now when user is trying to add another node Node3 to the above Multi master\n+ logical replication setup, user will have to create one subscription\n+ subscribing from Node1 and another subscription subscribing from Node2 in\n+ Node3 using <literal>subscribe_local_only</literal> option and\n+ <literal>copy_data</literal> as <literal>true</literal>, while\n+ the subscription is created, server will identify that Node2 is subscribing\n+ from Node1 and Node1 is subscribing from Node2 and throw an error like:\n+ <programlisting>\n+postgres=# CREATE SUBSCRIPTION mysub CONNECTION 'host=192.168.1.50\nport=5432 user=foo dbname=foodb'\n+ PUBLICATION mypublication, insert_only with\n(subscribe_local_only=on);\n+ERROR: CREATE/ALTER SUBSCRIPTION with subscribe_local_only and\ncopy_data as true is not allowed when the publisher might have\nreplicated data, table:public.t1 might have replicated data in the\npublisher\n+HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\n+ </programlisting>\n+ In this scenario user can solve this based on one of the 2 possibilities,\n+ a) If there are no data present in Node1 and Node2, then the user can create\n+ the subscriptions to Node1 and Node2 with\n+ <literal>subscribe_local_only</literal> as <literal>true</literal> and\n+ <literal>copy_data</literal> as <literal>false</literal>. b) If the data is\n+ present, then the user can create subscription with\n+ <literal>copy_data</literal> as <literal>force</literal> on Node1 and\n+ <literal>copy_data</literal> as <literal>false</literal> on Node2, before\n+ allowing any operations on the respective tables of Node1 and Node2, in this\n+ case <literal>copy_data</literal> is <literal>false</literal> on Node2\n+ because the data will be replicated to each other and available on both the\n+ nodes.\n+ </para>\n+\n\nThat is a large slab of text in the Notes, so not very easy to digest it.\n\nI'm not sure what to suggest for this -\n- Perhaps the a,b,c,d should all be \"lists\" so it renders differently?\n- It almost seems like too much information to be included in this\n\"Notes\" section, Maybe it needs its own full page in PG Docs \"Logical\nReplication\" to discuss this topic.\n\n~~~\n\n2.9 src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_VALID\n\n@@ -69,6 +69,18 @@\n /* check if the 'val' has 'bits' set */\n #define IsSet(val, bits) (((val) & (bits)) == (bits))\n\n+#define IS_COPY_DATA_VALID(copy_data) (copy_data != COPY_DATA_OFF)\n\nThe macro seems misnamed because \"off\" is also \"valid\". It seems like\nit should be called something different like IS_COPY_DATA, or\nIS_COPY_DATA_ON_OR_FORCE, etc.\n\n~~~\n\n2.10 src/backend/commands/subscriptioncmds.c - DefGetCopyData\n\n+ /*\n+ * The set of strings accepted here should match up with\n+ * the grammar's opt_boolean_or_string production.\n+ */\n+ if (pg_strcasecmp(sval, \"true\") == 0)\n+ return COPY_DATA_ON;\n+ if (pg_strcasecmp(sval, \"false\") == 0)\n+ return COPY_DATA_OFF;\n+ if (pg_strcasecmp(sval, \"on\") == 0)\n+ return COPY_DATA_ON;\n+ if (pg_strcasecmp(sval, \"off\") == 0)\n+ return COPY_DATA_OFF;\n+ if (pg_strcasecmp(sval, \"force\") == 0)\n+ return COPY_DATA_FORCE;\n\nMaybe combine the return for true/on and false/off?\n\n~~~\n\n2.11 src/backend/commands/subscriptioncmds.c - fetch_table_list\n\n+ appendStringInfoString(&cmd,\n+ \"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\nPS.srsubstate as replicated\\n\"\n+ \"FROM pg_publication P,\\n\"\n+ \"LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n+ \"LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n+ \"pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n+ \"WHERE C.oid = GPT.relid AND P.pubname in (\");\n+\n\nThat blank line is not needed (it was not there previously) because\nthe next get_publications_str is a continuation of this SQL.\n\n~~~\n\n2.12 src/backend/commands/subscriptioncmds.c - fetch_table_list comment\n\n+ * It is quite possible that subscriber has not yet pulled data to\n+ * the tables, but in ideal cases the table data will be subscribed.\n+ * Too keep the code simple it is not checked if the subscriber table\n+ * has pulled the data or not.\n+ */\n\nTypo \"Too\" -> \"To\"\n\n~~~\n\n2.13 src/backend/commands/subscriptioncmds.c - force\n\n+ if (copydata == COPY_DATA_ON && only_local && !slot_attisnull(slot, 3))\n+ ereport(ERROR,\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n+ errmsg(\"CREATE/ALTER SUBSCRIPTION with subscribe_local_only and\ncopy_data as true is not allowed when the publisher might have\nreplicated data, table:%s.%s might have replicated data in the\npublisher\",\n+ nspname, relname),\n+ errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\"));\n\nAFAIK this is the only code of this patch that is distinguishing\nbetween \"force\" and the \"on\". As I wrote at the beginning of this post\nI feel you can keep this ERROR (maybe the way to detect it needs to be\nmore complex) and the user can fix it just by using \"copy_data = off\"\nappropriately. So then copy_data does not need to be changed.\n\n~~~\n\n2.14 src/test/regress/sql/subscription.sql - missing tests\n\nThe new copy_data is not really an enum true/false/force like the PG\ndocs claims.\n\nIt seems more like some kind of hybrid boolean+force. So if that is\nhow it is going to be then there are missing test cases to make sure\nthat values like \"on\"/\"off\"/\"0\"/\"1\" still are working.\n\n~~~\n\n2.15 src/test/subscription/t/032_circular.pl\n\n@@ -65,6 +69,25 @@ $node_A->safe_psql('postgres', \"\n PUBLICATION tap_pub_B\n WITH (subscribe_local_only = on, copy_data = off)\");\n\n+($result, $stdout, $stderr) = $node_A->psql('postgres', \"\n+ CREATE SUBSCRIPTION tap_sub_A1\n+ CONNECTION '$node_B_connstr application_name=$appname_A'\n+ PUBLICATION tap_pub_B\n+ WITH (subscribe_local_only = on, copy_data = on)\");\n+like(\n+ $stderr,\n+ qr/ERROR: CREATE\\/ALTER SUBSCRIPTION with\nsubscribe_local_only and copy_data as true is not allowed when the\npublisher might have replicated data/,\n+ \"Create subscription with subscribe_local_only and copy_data\nhaving replicated table in publisher\");\n+\n+$node_A->safe_psql('postgres', \"\n+ CREATE SUBSCRIPTION tap_sub_A2\n+ CONNECTION '$node_B_connstr application_name=$appname_A'\n+ PUBLICATION tap_pub_B\n+ WITH (subscribe_local_only = on, copy_data = force)\");\n+\n+$node_A->safe_psql('postgres', \"\n+ DROP SUBSCRIPTION tap_sub_A2\");\n+\n\nMaybe underneath it is the same, but from the outside, this looks like\na slightly different scenario from what is mentioned everywhere else\nin the patch.\n\nI think it would be better to create a new Node_C (aka Node3) so then\nthe TAP test can use the same example that you give in the commit\nmessage and the PG docs notes.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 5 Apr 2022 12:47:04 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Apr 5, 2022 at 12:36 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Here are my comments for the latest patch v6-0001.\n> >\n> > (I will post my v6-0002 review comments separately)\n> >\n> > PATCH v6-0001 comments\n> > ======================\n> >\n> > 1.1 General - Option name\n> >\n> > I still feel like the option name is not ideal. Unfortunately, this is\n> > important because any name change would impact lots of these patch\n> > files and docs, struct members etc.\n> >\n> > It was originally called \"local_only\", but I thought that as a\n> > SUBSCRIPTION option that was confusing because \"local\" means local to\n> > what? Really it is local to the publisher, not local to the\n> > subscriber, so that name seemed misleading.\n> >\n> > Then I suggested \"publish_local_only\". Although that resolved the\n> > ambiguity problem, other people thought it seemed odd to have the\n> > \"publish\" prefix for a subscription-side option.\n> >\n> > So now it is changed again to \"subscribe_local_only\" -- It's getting\n> > better but still, it is implied that the \"local\" means local to the\n> > publisher except there is nothing in the option name really to convey\n> > that meaning. IMO we here all understand the meaning of this option\n> > mostly by familiarity with this discussion thread, but I think a user\n> > coming to this for the first time will still be confused by the name.\n> >\n> > Below are some other name ideas. Maybe they are not improvements, but\n> > it might help other people to come up with something better.\n> >\n> > subscribe_publocal_only = true/false\n> > origin = pub_only/any\n> > adjacent_data_only = true/false\n> > direct_subscriptions_only = true/false\n> > ...\n> >\n>\n> I think local_only with a description should be okay considering other\n> current options. Some of the options like slot_name, binary, etc. are\n> publisher-specific which becomes clear only after reading the\n> description. I am not sure adding pub*/sub* to it is much helpful. So,\n> +1 to a simple boolean like 'local_only', 'data_local', or something\n> like that.\n\n+1 to use a simple boolean option.\n\nThanks for the examples of the other options which are also really\npublisher-specific. Seen in that light, and with a clear description,\nprobably the original \"local_only\" was fine after all.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 5 Apr 2022 13:09:01 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Apr 5, 2022 at 8:17 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my comments for the latest patch v6-0002.\n>\n> PATCH v6-0002 comments\n> ======================\n>\n> 2.1 General - should this be an independent patch?\n>\n\nI think both the patches are dependent and might get committed\ntogether if the concept proved to be useful and doesn't have flaws.\n\n> In many ways, I think most of this patch is unrelated to the other\n> \"local_only\" patch (v6-0001).\n>\n> For example, IIUC even in the current HEAD, we could consider it to be\n> a user error if multiple SUBSCRIPTIONS or multiple PUBLICATIONS of the\n> same SUBSCRIPTION are replicating to the same TABLE on the same node\n> and using \"copy_data = on\".\n>\n> So I think it would be ok to throw an ERROR if such a copy_data clash\n> is detected, and then the user will have to change to use \"copy_data =\n> off\" for some/all of them to avoid data duplication.\n>\n> The \"local_only\" option only adds some small logic to this new ERROR,\n> but it's not really a prerequisite at all.\n>\n> e.g. this whole ERROR part of the patch can be a separate thread.\n>\n> ~~~\n>\n> 2.2 General - can we remove the \"force\" enum?\n>\n> Now, because I consider the clashing \"copy_data = on\" ERROR to be a\n> user error, I think that is something that the user can already take\n> care of themselves just using the copy_data = off.\n>\n> I did not really like the modifying of the \"copy_data\" option from\n> just boolean to some kind of hybrid boolean + \"force\".\n>\n> a) It smells a bit off to me. IMO replication is supposed to end up\n> with the same (replicated) data on the standby machine but this\n> \"force\" mode seems to be just helping the user to break that concept\n> and say - \"I know what I'm doing, and I don't care if I get lots of\n> duplicated data in the replica table - just let me do it\"...\n>\n> b) It also somehow feels like the new \"force\" was introduced mostly to\n> make the code ERROR handling implementation simpler, rather than to\n> make the life of the end-user better. Yes, if force is removed maybe\n> the copy-clash-detection-code will need to be internally quite more\n> complex than it is now, but that is how it should be, instead of\n> putting extra burden on the user (e.g. by complicating the PG docs and\n> giving them yet more alternatives to configure). I think any clashing\n> copy_data options really is a user error, but also I think the current\n> boolean copy_data true/false already gives the user a way to fix it.\n>\n> c) Actually, your new error hint messages are similar to my\n> perspective: They already say \"Use CREATE/ALTER SUBSCRIPTION with\n> copy_data = off or force\". All I am saying is remove the \"force\", and\n> the user can still fix the problem just by using \"copy_data = off\"\n> appropriately.\n>\n\nHow will it work if there is more than one table? copy_data = \"off\"\nmeans it won't copy any of the other tables in the corresponding\npublication(s). I think it is primarily to give the user some option\nwhere she understands the logical replication setup better and\nunderstands that this won't be a problem.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 5 Apr 2022 11:45:50 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my comments for the latest patch v6-0001.\n>\n> (I will post my v6-0002 review comments separately)\n>\n> PATCH v6-0001 comments\n> ======================\n>\n> 1.1 General - Option name\n>\n> I still feel like the option name is not ideal. Unfortunately, this is\n> important because any name change would impact lots of these patch\n> files and docs, struct members etc.\n>\n> It was originally called \"local_only\", but I thought that as a\n> SUBSCRIPTION option that was confusing because \"local\" means local to\n> what? Really it is local to the publisher, not local to the\n> subscriber, so that name seemed misleading.\n>\n> Then I suggested \"publish_local_only\". Although that resolved the\n> ambiguity problem, other people thought it seemed odd to have the\n> \"publish\" prefix for a subscription-side option.\n>\n> So now it is changed again to \"subscribe_local_only\" -- It's getting\n> better but still, it is implied that the \"local\" means local to the\n> publisher except there is nothing in the option name really to convey\n> that meaning. IMO we here all understand the meaning of this option\n> mostly by familiarity with this discussion thread, but I think a user\n> coming to this for the first time will still be confused by the name.\n>\n> Below are some other name ideas. Maybe they are not improvements, but\n> it might help other people to come up with something better.\n>\n> subscribe_publocal_only = true/false\n> origin = pub_only/any\n> adjacent_data_only = true/false\n> direct_subscriptions_only = true/false\n> ...\n\nFWIW, The subscriber wants \"changes originated on publisher\". From\nthat angle origin = publisher/any looks attractive. It also leaves\nopen the possibility that the subscriber may ask changes from a set of\norigins or even non-publisher origins.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\n\n","msg_date":"Tue, 5 Apr 2022 19:06:18 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Apr 5, 2022 at 7:06 PM Ashutosh Bapat\n<ashutosh.bapat.oss@gmail.com> wrote:\n>\n> On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Below are some other name ideas. Maybe they are not improvements, but\n> > it might help other people to come up with something better.\n> >\n> > subscribe_publocal_only = true/false\n> > origin = pub_only/any\n> > adjacent_data_only = true/false\n> > direct_subscriptions_only = true/false\n> > ...\n>\n> FWIW, The subscriber wants \"changes originated on publisher\". From\n> that angle origin = publisher/any looks attractive. It also leaves\n> open the possibility that the subscriber may ask changes from a set of\n> origins or even non-publisher origins.\n>\n\nSo, how are you imagining extending it for multiple origins? I think\nwe can have multiple origins data on a node when there are multiple\nsubscriptions pointing to the different or same node. The origin names\nare internally generated names but are present in\npg_replication_origin. So, are we going to ask the user to check that\ntable and specify it as an option? But, note, that the user may not be\nable to immediately recognize which origin data is useful for her.\nSurely, with some help or advice in docs, she may be able to identify\nit but that doesn't seem so straightforward. The other point is that\nsuch an option won't be applicable for initial sync as we can't\ndifferentiate between data from a local or remote node.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 6 Apr 2022 09:36:52 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"FYI, here is a test script that is using the current patch (v6) to\ndemonstrate a way to share table data between different numbers of\nnodes (up to 5 of them here).\n\nThe script starts off with just 2-way sharing (nodes N1, N2),\nthen expands to 3-way sharing (nodes N1, N2, N3),\nthen 4-way sharing (nodes N1, N2, N3, N4),\nthen 5-way sharing (nodes N1, N2, N3, N4, N5).\n\nAs an extra complication, for this test, all 5 nodes have different\ninitial table data, which gets replicated to the others whenever each\nnew node joins the existing share group.\n\nPSA.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia","msg_date":"Thu, 7 Apr 2022 14:09:23 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Dear Peter,\r\n\r\n> FYI, here is a test script that is using the current patch (v6) to\r\n> demonstrate a way to share table data between different numbers of\r\n> nodes (up to 5 of them here).\r\n\r\nThanks for sharing your script! It's very helpful for us.\r\n\r\nWhile reading your script, however, I had a question about it.\r\nLine 121-122, you defined subscriptions for 2-nodes cluster:\r\n\r\npsql -p $port_N1 -c \"create subscription sub12 connection 'port=$port_N2' publication pub2 with ($copy_force);\"\r\npsql -p $port_N2 -c \"create subscription sub21 connection 'port=$port_N1' publication pub1 with ($copy_force);\"\r\n\r\nBut I was not sure it works well.\r\nN2 already have shared data from N1 when subscription sub21 is created.\r\nDid you assume that the initial copying is not so quick and\r\ndata synchronization will be not done when creating sub21?\r\n\r\nBest Regards,\r\nHayato Kuroda\r\nFUJITSU LIMITED\r\n\r\n","msg_date":"Thu, 7 Apr 2022 06:03:34 +0000","msg_from":"\"kuroda.hayato@fujitsu.com\" <kuroda.hayato@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Apr 7, 2022 at 4:03 PM kuroda.hayato@fujitsu.com\n<kuroda.hayato@fujitsu.com> wrote:\n>\n> Dear Peter,\n>\n> > FYI, here is a test script that is using the current patch (v6) to\n> > demonstrate a way to share table data between different numbers of\n> > nodes (up to 5 of them here).\n>\n> Thanks for sharing your script! It's very helpful for us.\n>\n> While reading your script, however, I had a question about it.\n> Line 121-122, you defined subscriptions for 2-nodes cluster:\n>\n> psql -p $port_N1 -c \"create subscription sub12 connection 'port=$port_N2' publication pub2 with ($copy_force);\"\n> psql -p $port_N2 -c \"create subscription sub21 connection 'port=$port_N1' publication pub1 with ($copy_force);\"\n>\n> But I was not sure it works well.\n> N2 already have shared data from N1 when subscription sub21 is created.\n> Did you assume that the initial copying is not so quick and\n> data synchronization will be not done when creating sub21?\n\nOops. Good catch.\n\nAlthough the 2-way test was working OK for me, I think that it worked\nonly because of lucky timing. e.g. When I put a delay between those 2\nsubscriptions then the 2nd one would cause the PK violation that\nprobably you were anticipating would happen.\n\nI have modified the 2-way example to use the same truncate pattern as others.\n\nPSA the fixed test.sh script and accompanying files.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia","msg_date":"Thu, 7 Apr 2022 16:13:56 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi Vignesh, FYI the patch is recently broken again and is failing on cfbot [1].\n\n------\n[1] http://cfbot.cputube.org/patch_38_3610.log\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 8 Apr 2022 09:07:45 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Apr 8, 2022 at 4:38 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Hi Vignesh, FYI the patch is recently broken again and is failing on cfbot [1].\n>\n> ------\n> [1] http://cfbot.cputube.org/patch_38_3610.log\n\nI'm working on fixing a few review comments, I will be posting an\nupdated version to handle this today.\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 8 Apr 2022 06:41:50 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my comments for the latest patch v6-0001.\n>\n> (I will post my v6-0002 review comments separately)\n>\n> PATCH v6-0001 comments\n> ======================\n>\n> 1.1 General - Option name\n>\n> I still feel like the option name is not ideal. Unfortunately, this is\n> important because any name change would impact lots of these patch\n> files and docs, struct members etc.\n>\n> It was originally called \"local_only\", but I thought that as a\n> SUBSCRIPTION option that was confusing because \"local\" means local to\n> what? Really it is local to the publisher, not local to the\n> subscriber, so that name seemed misleading.\n>\n> Then I suggested \"publish_local_only\". Although that resolved the\n> ambiguity problem, other people thought it seemed odd to have the\n> \"publish\" prefix for a subscription-side option.\n>\n> So now it is changed again to \"subscribe_local_only\" -- It's getting\n> better but still, it is implied that the \"local\" means local to the\n> publisher except there is nothing in the option name really to convey\n> that meaning. IMO we here all understand the meaning of this option\n> mostly by familiarity with this discussion thread, but I think a user\n> coming to this for the first time will still be confused by the name.\n>\n> Below are some other name ideas. Maybe they are not improvements, but\n> it might help other people to come up with something better.\n>\n> subscribe_publocal_only = true/false\n> origin = pub_only/any\n> adjacent_data_only = true/false\n> direct_subscriptions_only = true/false\n> ...\n>\n>\n> (FYI, the remainder of these review comments will assume the option is\n> still called \"subscribe_local_only\")\n\nModified to local_only\n\n> ~~~\n>\n> 1.2 General - inconsistent members and args\n>\n> IMO the struct members and args should also be named for close\n> consistency with whatever the option name is.\n>\n> Currently the option is called \"subscription_local_only\". So I think\n> the members/args would be better to be called \"local_only\" instead of\n> \"only_local\".\n\nModified\n\n> ~~~\n>\n> 1.3 Commit message - wrong option name\n>\n> The commit message refers to the option name as \"publish_local_only\"\n> instead of the option name that is currently implemented.\n\nModified\n\n> ~~~\n>\n> 1.4 Commit message - wording\n>\n> The wording seems a bit off. Below is suggested simpler wording which\n> I AFAIK conveys the same information.\n>\n> BEFORE\n> Add an option publish_local_only which will subscribe only to the locally\n> generated data in the publisher node. If subscriber is created with this\n> option, publisher will skip publishing the data that was subscribed\n> from other nodes. It can be created using following syntax:\n> ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\n> PUBLICATION pub1 with (publish_local_only = on);\n>\n> SUGGESTION\n> This patch adds a new SUBSCRIPTION boolean option\n> \"subscribe_local_only\". The default is false. When a SUBSCRIPTION is\n> created with this option enabled, the publisher will only publish data\n> that originated at the publisher node.\n> Usage:\n> CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\n> PUBLICATION pub1 with (subscribe_local_only = true);\n\nModified\n\n> ~~~\n>\n> 1.5 doc/src/sgml/ref/create_subscription.sgml - \"generated\" changes.\n>\n> + <para>\n> + Specifies whether the subscription will request the publisher to send\n> + locally generated changes or both the locally generated changes and\n> + the replicated changes that was generated from other nodes. The\n> + default is <literal>false</literal>.\n> + </para>\n>\n> For some reason, it seemed a bit strange to me to use the term\n> \"generated\" changes. Maybe better to refer to the origin of changes?\n>\n> SUGGESTION\n> Specifies whether the publisher should send only changes that\n> originated locally at the publisher node, or send any publisher node\n> changes regardless of their origin. The default is false.\n\nModified\n\n> ~~~\n>\n> 1.6 src/backend/replication/pgoutput/pgoutput.c -\n> LOGICALREP_PROTO_TWOPHASE_VERSION_NUM\n>\n> @@ -496,6 +509,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\n> OutputPluginOptions *opt,\n> else\n> ctx->twophase_opt_given = true;\n>\n> + if (data->only_local && data->protocol_version <\n> LOGICALREP_PROTO_TWOPHASE_VERSION_NUM)\n> + ereport(ERROR,\n> + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\n> + errmsg(\"requested proto_version=%d does not support\n> subscribe_local_only, need %d or higher\",\n> + data->protocol_version, LOGICALREP_PROTO_TWOPHASE_VERSION_NUM)));\n>\n> I thought this code should not be using\n> LOGICALREP_PROTO_TWOPHASE_VERSION_NUM. Shouldn't there be some newly\n> introduced constant like LOGICALREP_PROTO_LOCALONLY_VERSION_NUM which\n> you will use here?\n\nModified, I have set LOGICALREP_PROTO_LOCALONLY_VERSION_NUM to same\nvalue as LOGICALREP_PROTO_TWOPHASE_VERSION_NUM, will increment this\nonce server version is changed.\n\n> ~~~\n>\n> 1.7 src/bin/pg_dump/pg_dump.c - 150000\n>\n> @@ -4451,11 +4452,13 @@ getSubscriptions(Archive *fout)\n> if (fout->remoteVersion >= 150000)\n> appendPQExpBufferStr(query,\n> \" s.subtwophasestate,\\n\"\n> - \" s.subdisableonerr\\n\");\n> + \" s.subdisableonerr,\\n\"\n> + \" s.sublocal\\n\");\n> else\n> appendPQExpBuffer(query,\n> \" '%c' AS subtwophasestate,\\n\"\n> - \" false AS subdisableonerr\\n\",\n> + \" false AS subdisableonerr,\\n\"\n> + \" false AS s.sublocal\\n\",\n> LOGICALREP_TWOPHASE_STATE_DISABLED);\n>\n> I think this local_only feature is unlikely to get into the PG15\n> release, so this code should be split out into a separate condition\n> because later will need to change to say >= 160000.\n\n I have split the condition and checked it with 150000, this will be\nchanged later to 160000 after the branch is created.\n\n> ~~~\n>\n> 1.8 src/bin/pg_dump/pg_dump.c - dumpSubscription\n>\n> @@ -4585,6 +4591,9 @@ dumpSubscription(Archive *fout, const\n> SubscriptionInfo *subinfo)\n> if (strcmp(subinfo->subdisableonerr, \"t\") == 0)\n> appendPQExpBufferStr(query, \", disable_on_error = true\");\n>\n> + if (strcmp(subinfo->sublocal, \"f\") != 0)\n> + appendPQExpBufferStr(query, \", subscribe_local_only = on\");\n> +\n>\n> I felt it is more natural to say \"if it is true set to true\", instead\n> of \"if it is not false set to on\".\n>\n> SUGGESTION\n> if (strcmp(subinfo->sublocal, \"t\") == 0)\n> appendPQExpBufferStr(query, \", subscribe_local_only = true\");\n\nModified\n\n> ~~~\n>\n> 1.9 src/bin/psql/describe.c - 150000\n>\n> @@ -6318,9 +6318,11 @@ describeSubscriptions(const char *pattern, bool verbose)\n> if (pset.sversion >= 150000)\n> appendPQExpBuffer(&buf,\n> \", subtwophasestate AS \\\"%s\\\"\\n\"\n> - \", subdisableonerr AS \\\"%s\\\"\\n\",\n> + \", subdisableonerr AS \\\"%s\\\"\\n\"\n> + \", sublocal AS \\\"%s\\\"\\n\",\n> gettext_noop(\"Two phase commit\"),\n> - gettext_noop(\"Disable on error\"));\n> + gettext_noop(\"Disable on error\"),\n> + gettext_noop(\"Only local\"));\n>\n> I think this local_only feature is unlikely to get into the PG15\n> release, so this code should be split out into a separate condition\n> because later will need to change to say >= 160000.\n\nI have split the condition and checked it with 150000, this will be\nchanged later to 160000 after the branch is created.\n\n> ~~\n>\n> 1.10 src/bin/psql/describe.c - describeSubscriptions column\n>\n> @@ -6318,9 +6318,11 @@ describeSubscriptions(const char *pattern, bool verbose)\n> if (pset.sversion >= 150000)\n> appendPQExpBuffer(&buf,\n> \", subtwophasestate AS \\\"%s\\\"\\n\"\n> - \", subdisableonerr AS \\\"%s\\\"\\n\",\n> + \", subdisableonerr AS \\\"%s\\\"\\n\"\n> + \", sublocal AS \\\"%s\\\"\\n\",\n> gettext_noop(\"Two phase commit\"),\n> - gettext_noop(\"Disable on error\"));\n> + gettext_noop(\"Disable on error\"),\n> + gettext_noop(\"Only local\"));\n>\n> I think the column name here should be more consistent with the option\n> name. e.g. it should be \"Local only\", not \"Only local\".\n\nModified\n\n> ~~~\n>\n> 1.11 src/bin/psql/tab-complete.c - whitespace\n>\n> @@ -3167,7 +3167,7 @@ psql_completion(const char *text, int start, int end)\n> /* Complete \"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\" */\n> else if (HeadMatches(\"CREATE\", \"SUBSCRIPTION\") && TailMatches(\"WITH\", \"(\"))\n> COMPLETE_WITH(\"binary\", \"connect\", \"copy_data\", \"create_slot\",\n> - \"enabled\", \"slot_name\", \"streaming\",\n> + \"enabled\", \"slot_name\", \"streaming\", \"subscribe_local_only\",\n>\n> The patch accidentally added a space char before the \"slot_name\".\n\nModified\n\n> ~~~\n>\n> 1.12 src/include/replication/walreceiver.h - \"generated\"\n>\n> @@ -183,6 +183,7 @@ typedef struct\n> bool streaming; /* Streaming of large transactions */\n> bool twophase; /* Streaming of two-phase transactions at\n> * prepare time */\n> + bool only_local; /* publish only locally generated data */\n>\n> This is a similar review comment as #1.5 about saying the word \"generated\".\n> Maybe there is another way to word this?\n\nModified\n\n> ~~~\n>\n> 1.13 src/test/regress/sql/subscription.sql - missing test case\n>\n> Isn't there a missing test case for ensuring that the new option is boolean?\n\nAdded test\n\nThanks for the comments, the attached v7 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Fri, 8 Apr 2022 13:00:04 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Apr 5, 2022 at 8:17 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my comments for the latest patch v6-0002.\n>\n> PATCH v6-0002 comments\n> ======================\n>\n> 2.1 General - should this be an independent patch?\n>\n> In many ways, I think most of this patch is unrelated to the other\n> \"local_only\" patch (v6-0001).\n>\n> For example, IIUC even in the current HEAD, we could consider it to be\n> a user error if multiple SUBSCRIPTIONS or multiple PUBLICATIONS of the\n> same SUBSCRIPTION are replicating to the same TABLE on the same node\n> and using \"copy_data = on\".\n>\n> So I think it would be ok to throw an ERROR if such a copy_data clash\n> is detected, and then the user will have to change to use \"copy_data =\n> off\" for some/all of them to avoid data duplication.\n>\n> The \"local_only\" option only adds some small logic to this new ERROR,\n> but it's not really a prerequisite at all.\n>\n> e.g. this whole ERROR part of the patch can be a separate thread.\n\nAs Amit also pointed out, the patches have some dependency, I will\nkeep it as it is.\n\n> ~~~\n>\n> 2.2 General - can we remove the \"force\" enum?\n>\n> Now, because I consider the clashing \"copy_data = on\" ERROR to be a\n> user error, I think that is something that the user can already take\n> care of themselves just using the copy_data = off.\n>\n> I did not really like the modifying of the \"copy_data\" option from\n> just boolean to some kind of hybrid boolean + \"force\".\n>\n> a) It smells a bit off to me. IMO replication is supposed to end up\n> with the same (replicated) data on the standby machine but this\n> \"force\" mode seems to be just helping the user to break that concept\n> and say - \"I know what I'm doing, and I don't care if I get lots of\n> duplicated data in the replica table - just let me do it\"...\n>\n> b) It also somehow feels like the new \"force\" was introduced mostly to\n> make the code ERROR handling implementation simpler, rather than to\n> make the life of the end-user better. Yes, if force is removed maybe\n> the copy-clash-detection-code will need to be internally quite more\n> complex than it is now, but that is how it should be, instead of\n> putting extra burden on the user (e.g. by complicating the PG docs and\n> giving them yet more alternatives to configure). I think any clashing\n> copy_data options really is a user error, but also I think the current\n> boolean copy_data true/false already gives the user a way to fix it.\n>\n> c) Actually, your new error hint messages are similar to my\n> perspective: They already say \"Use CREATE/ALTER SUBSCRIPTION with\n> copy_data = off or force\". All I am saying is remove the \"force\", and\n> the user can still fix the problem just by using \"copy_data = off\"\n> appropriately.\n\nWhen a user is trying to add a node to bidirectional replication\nsetup, the user will use the force option to copy the data from one of\nthe nodes and use off to skip copying the data from other nodes. This\noption will be used while adding nodes to bidirectional replication,\nthe same has been documented with examples in the patch. I felt we\nshould retain this option.\n\n> ======\n>\n> So (from above) I am not much in favour of the copy_data becoming a\n> hybrid enum and using \"force\", yet that is what most of this patch is\n> implementing. Anyway, the remainder of my review comments here are for\n> the code in its current form. Maybe if \"force\" can be removed most of\n> the following comments end up being redundant.\n>\n> ======\n>\n> 2.3 Commit message - wording\n>\n> This message is difficult to understand.\n>\n> I think that the long sentence \"Now when user is trying...\" can be\n> broken into more manageable parts.\n\nSlightly modified\n\n> This part \"and throw an error so that user can handle the initial copy\n> data.\" also seemed a bit vague.\n\nI have given the reference to the documentation section of the patch\nfor initial data copy handling\n\n> ~~~\n>\n> 2.4 Commit message - more functions\n>\n> \"This patch does couple of things:\"\n>\n> IIUC, there seems a third thing implemented by this patch but not\n> described by the comment. I think it also adds support for ALTER\n> SUBSCRIPTION SET PUBLICATION WITH (subscribe_local_only)\n\nThis is part of 0001 patch, I felt this should not be part of 002 patch commit.\n\n> ~~~\n>\n> 2.5 doc/src/sgml/ref/create_subscription.sgml - wording\n>\n> @@ -161,6 +161,13 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> the replicated changes that was generated from other nodes. The\n> default is <literal>false</literal>.\n> </para>\n> + <para>\n> + If the tables in the publication were also subscribing to the data in\n> + the publisher from other publishers, it will affect the\n> + <command>CREATE SUBSCRIPTION</command> based on the value specified\n> + for <literal>copy_data</literal> option. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for details.\n> + </para>\n>\n> Is there is a simpler way to express all that?\n>\n> SUGGESTION\n> There is some interation between the option \"subscribe_local_only\" and\n> option \"copy_data\". Refer to the <xref\n> linkend=\"sql-createsubscription-notes\" /> for details.\n\nModified\n\n> ~~~\n>\n> 2.6 doc/src/sgml/ref/create_subscription.sgml - whitespace\n>\n> @@ -213,18 +220,28 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> </varlistentry>\n>\n> <varlistentry>\n> - <term><literal>copy_data</literal> (<type>boolean</type>)</term>\n> + <term><literal>copy_data</literal> (<type>enum</type>)</term>\n> <listitem>\n> <para>\n> Specifies whether to copy pre-existing data in the publications\n> - that are being subscribed to when the replication starts.\n> - The default is <literal>true</literal>.\n> + that are being subscribed to when the replication starts. This\n> + parameter may be either <literal>true</literal>,\n> + <literal>false</literal> or <literal>force</literal>. The default is\n> + <literal>true</literal>.\n>\n> That last line has trailing whitespace.\n\nModified\n\n> ~~~\n>\n> 2.7 doc/src/sgml/ref/create_subscription.sgml - wording\n>\n> + <para>\n> + If the tables in the publication were also subscribing to the data in\n> + the publisher from other publishers, it will affect the\n> + <command>CREATE SUBSCRIPTION</command> based on the value specified\n> + for <literal>subscribe_local_only</literal> option. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for details.\n> + </para>\n>\n> This is similar to review comment #2.5 which I thought could be\n> written in a simpler way.\n\nModified\n\n> ~~~\n>\n> 2.8 doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -375,6 +392,42 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> can have non-existent publications.\n> </para>\n>\n> + <para>\n> + Let's consider an existing Multi master logical replication setup between\n> + Node1 and Node2 that is created using the following steps:\n> + a) Node1 - Publication publishing employee table.\n> + b) Node2 - Subscription subscribing from publication pub1 with\n> + <literal>subscribe_local_only</literal>.\n> + c) Node2 - Publication publishing employee table.\n> + d) Node1 - Subscription subscribing from publication pub2 with\n> + <literal>subscribe_local_only</literal>.\n> + Now when user is trying to add another node Node3 to the above Multi master\n> + logical replication setup, user will have to create one subscription\n> + subscribing from Node1 and another subscription subscribing from Node2 in\n> + Node3 using <literal>subscribe_local_only</literal> option and\n> + <literal>copy_data</literal> as <literal>true</literal>, while\n> + the subscription is created, server will identify that Node2 is subscribing\n> + from Node1 and Node1 is subscribing from Node2 and throw an error like:\n> + <programlisting>\n> +postgres=# CREATE SUBSCRIPTION mysub CONNECTION 'host=192.168.1.50\n> port=5432 user=foo dbname=foodb'\n> + PUBLICATION mypublication, insert_only with\n> (subscribe_local_only=on);\n> +ERROR: CREATE/ALTER SUBSCRIPTION with subscribe_local_only and\n> copy_data as true is not allowed when the publisher might have\n> replicated data, table:public.t1 might have replicated data in the\n> publisher\n> +HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\n> + </programlisting>\n> + In this scenario user can solve this based on one of the 2 possibilities,\n> + a) If there are no data present in Node1 and Node2, then the user can create\n> + the subscriptions to Node1 and Node2 with\n> + <literal>subscribe_local_only</literal> as <literal>true</literal> and\n> + <literal>copy_data</literal> as <literal>false</literal>. b) If the data is\n> + present, then the user can create subscription with\n> + <literal>copy_data</literal> as <literal>force</literal> on Node1 and\n> + <literal>copy_data</literal> as <literal>false</literal> on Node2, before\n> + allowing any operations on the respective tables of Node1 and Node2, in this\n> + case <literal>copy_data</literal> is <literal>false</literal> on Node2\n> + because the data will be replicated to each other and available on both the\n> + nodes.\n> + </para>\n> +\n>\n> That is a large slab of text in the Notes, so not very easy to digest it.\n>\n> I'm not sure what to suggest for this -\n> - Perhaps the a,b,c,d should all be \"lists\" so it renders differently?\n> - It almost seems like too much information to be included in this\n> \"Notes\" section, Maybe it needs its own full page in PG Docs \"Logical\n> Replication\" to discuss this topic.\n\nThis is moved to the new page \"BiDirectional logical replication Quick\nSetup\" page now.\n\n> ~~~\n>\n> 2.9 src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_VALID\n>\n> @@ -69,6 +69,18 @@\n> /* check if the 'val' has 'bits' set */\n> #define IsSet(val, bits) (((val) & (bits)) == (bits))\n>\n> +#define IS_COPY_DATA_VALID(copy_data) (copy_data != COPY_DATA_OFF)\n>\n> The macro seems misnamed because \"off\" is also \"valid\". It seems like\n> it should be called something different like IS_COPY_DATA, or\n> IS_COPY_DATA_ON_OR_FORCE, etc.\n\nModified\n\n> ~~~\n>\n> 2.10 src/backend/commands/subscriptioncmds.c - DefGetCopyData\n>\n> + /*\n> + * The set of strings accepted here should match up with\n> + * the grammar's opt_boolean_or_string production.\n> + */\n> + if (pg_strcasecmp(sval, \"true\") == 0)\n> + return COPY_DATA_ON;\n> + if (pg_strcasecmp(sval, \"false\") == 0)\n> + return COPY_DATA_OFF;\n> + if (pg_strcasecmp(sval, \"on\") == 0)\n> + return COPY_DATA_ON;\n> + if (pg_strcasecmp(sval, \"off\") == 0)\n> + return COPY_DATA_OFF;\n> + if (pg_strcasecmp(sval, \"force\") == 0)\n> + return COPY_DATA_FORCE;\n>\n> Maybe combine the return for true/on and false/off?\n\nModified\n\n> ~~~\n>\n> 2.11 src/backend/commands/subscriptioncmds.c - fetch_table_list\n>\n> + appendStringInfoString(&cmd,\n> + \"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\n> PS.srsubstate as replicated\\n\"\n> + \"FROM pg_publication P,\\n\"\n> + \"LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n> + \"LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n> + \"pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n> + \"WHERE C.oid = GPT.relid AND P.pubname in (\");\n> +\n>\n> That blank line is not needed (it was not there previously) because\n> the next get_publications_str is a continuation of this SQL.\n\nModified\n\n> ~~~\n>\n> 2.12 src/backend/commands/subscriptioncmds.c - fetch_table_list comment\n>\n> + * It is quite possible that subscriber has not yet pulled data to\n> + * the tables, but in ideal cases the table data will be subscribed.\n> + * Too keep the code simple it is not checked if the subscriber table\n> + * has pulled the data or not.\n> + */\n>\n> Typo \"Too\" -> \"To\"\n\nModified\n\n> ~~~\n>\n> 2.13 src/backend/commands/subscriptioncmds.c - force\n>\n> + if (copydata == COPY_DATA_ON && only_local && !slot_attisnull(slot, 3))\n> + ereport(ERROR,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"CREATE/ALTER SUBSCRIPTION with subscribe_local_only and\n> copy_data as true is not allowed when the publisher might have\n> replicated data, table:%s.%s might have replicated data in the\n> publisher\",\n> + nspname, relname),\n> + errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\"));\n>\n> AFAIK this is the only code of this patch that is distinguishing\n> between \"force\" and the \"on\". As I wrote at the beginning of this post\n> I feel you can keep this ERROR (maybe the way to detect it needs to be\n> more complex) and the user can fix it just by using \"copy_data = off\"\n> appropriately. So then copy_data does not need to be changed.\n\nWhen a user is trying to add a node to bidirectional replication\nsetup, the user will use the force option to copy the data from one of\nthe nodes and use off to skip copying the data from other nodes. This\noption will be used while adding nodes to bidirectional replication,\nthe same has been documented with examples in the patch. I felt we\nshould retain this option.\n\n> ~~~\n>\n> 2.14 src/test/regress/sql/subscription.sql - missing tests\n>\n> The new copy_data is not really an enum true/false/force like the PG\n> docs claims.\n>\n> It seems more like some kind of hybrid boolean+force. So if that is\n> how it is going to be then there are missing test cases to make sure\n> that values like \"on\"/\"off\"/\"0\"/\"1\" still are working.\n\nAdded tests\n\n> ~~~\n>\n> 2.15 src/test/subscription/t/032_circular.pl\n>\n> @@ -65,6 +69,25 @@ $node_A->safe_psql('postgres', \"\n> PUBLICATION tap_pub_B\n> WITH (subscribe_local_only = on, copy_data = off)\");\n>\n> +($result, $stdout, $stderr) = $node_A->psql('postgres', \"\n> + CREATE SUBSCRIPTION tap_sub_A1\n> + CONNECTION '$node_B_connstr application_name=$appname_A'\n> + PUBLICATION tap_pub_B\n> + WITH (subscribe_local_only = on, copy_data = on)\");\n> +like(\n> + $stderr,\n> + qr/ERROR: CREATE\\/ALTER SUBSCRIPTION with\n> subscribe_local_only and copy_data as true is not allowed when the\n> publisher might have replicated data/,\n> + \"Create subscription with subscribe_local_only and copy_data\n> having replicated table in publisher\");\n> +\n> +$node_A->safe_psql('postgres', \"\n> + CREATE SUBSCRIPTION tap_sub_A2\n> + CONNECTION '$node_B_connstr application_name=$appname_A'\n> + PUBLICATION tap_pub_B\n> + WITH (subscribe_local_only = on, copy_data = force)\");\n> +\n> +$node_A->safe_psql('postgres', \"\n> + DROP SUBSCRIPTION tap_sub_A2\");\n> +\n>\n> Maybe underneath it is the same, but from the outside, this looks like\n> a slightly different scenario from what is mentioned everywhere else\n> in the patch.\n>\n> I think it would be better to create a new Node_C (aka Node3) so then\n> the TAP test can use the same example that you give in the commit\n> message and the PG docs notes.\n\nModified the tests to include Node_C.\n\nThanks for the comments, the v7 patch attached at [1] has the fixes for the same\n[1] - https://www.postgresql.org/message-id/CALDaNm1ei%3DrRwCBKWtUu8b5OsS6FFcvaxg9h0oXcjgFn8GoZnQ%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 8 Apr 2022 13:10:05 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Apr 7, 2022 at 2:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> FYI, here is a test script that is using the current patch (v6) to\n> demonstrate a way to share table data between different numbers of\n> nodes (up to 5 of them here).\n>\n> The script starts off with just 2-way sharing (nodes N1, N2),\n> then expands to 3-way sharing (nodes N1, N2, N3),\n> then 4-way sharing (nodes N1, N2, N3, N4),\n> then 5-way sharing (nodes N1, N2, N3, N4, N5).\n>\n> As an extra complication, for this test, all 5 nodes have different\n> initial table data, which gets replicated to the others whenever each\n> new node joins the existing share group.\n>\n> PSA.\n>\n\n\nHi Vignesh. I had some problems getting the above test script working.\nIt was OK up until I tried to join the 5th node (N5) to the existing 4\nnodes. The ERROR was manifesting itself strangely because it appeared\nthat there was an index violation in the pg_subscription_rel catalog\neven though AFAIK the N5 did not have any entries in it.\n\n\ne.g.\n2022-04-07 09:13:28.361 AEST [24237] ERROR: duplicate key value\nviolates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n2022-04-07 09:13:28.361 AEST [24237] DETAIL: Key (srrelid,\nsrsubid)=(16384, 16393) already exists.\n2022-04-07 09:13:28.361 AEST [24237] STATEMENT: create subscription\nsub51 connection 'port=7651' publication pub1 with\n(subscribe_local_only=true,copy_data=force);\n2022-04-07 09:13:28.380 AEST [24237] ERROR: duplicate key value\nviolates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n2022-04-07 09:13:28.380 AEST [24237] DETAIL: Key (srrelid,\nsrsubid)=(16384, 16394) already exists.\n2022-04-07 09:13:28.380 AEST [24237] STATEMENT: create subscription\nsub52 connection 'port=7652' publication pub2 with\n(subscribe_local_only=true,copy_data=false);\n2022-04-07 09:13:28.405 AEST [24237] ERROR: duplicate key value\nviolates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n2022-04-07 09:13:28.405 AEST [24237] DETAIL: Key (srrelid,\nsrsubid)=(16384, 16395) already exists.\n2022-04-07 09:13:28.405 AEST [24237] STATEMENT: create subscription\nsub53 connection 'port=7653' publication pub3 with\n(subscribe_local_only=true,copy_data=false);\n2022-04-07 09:13:28.425 AEST [24237] ERROR: duplicate key value\nviolates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n2022-04-07 09:13:28.425 AEST [24237] DETAIL: Key (srrelid,\nsrsubid)=(16384, 16396) already exists.\n2022-04-07 09:13:28.425 AEST [24237] STATEMENT: create subscription\nsub54 connection 'port=7654' publication pub4 with\n(subscribe_local_only=true,copy_data=false);\n2022-04-07 09:17:52.472 AEST [25852] ERROR: duplicate key value\nviolates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n2022-04-07 09:17:52.472 AEST [25852] DETAIL: Key (srrelid,\nsrsubid)=(16384, 16397) already exists.\n2022-04-07 09:17:52.472 AEST [25852] STATEMENT: create subscription\nsub51 connection 'port=7651' publication pub1;\n\n~~~\n\nWhen I debugged this it seemed like each of the CREAT SUBSCRIPTION was\ntrying to make a double-entry, because the fetch_tables (your patch\nv6-0002 modified SQL of this) was retuning the same table 2x.\n\n(gdb) bt\n#0 errfinish (filename=0xbc1057 \"nbtinsert.c\", lineno=671,\nfuncname=0xbc25e0 <__func__.15798> \"_bt_check_unique\") at elog.c:510\n#1 0x0000000000526d83 in _bt_check_unique (rel=0x7f654219c2a0,\ninsertstate=0x7ffd9629ddd0, heapRel=0x7f65421b0e28,\ncheckUnique=UNIQUE_CHECK_YES, is_unique=0x7ffd9629de01,\nspeculativeToken=0x7ffd9629ddcc) at nbtinsert.c:664\n#2 0x0000000000526157 in _bt_doinsert (rel=0x7f654219c2a0,\nitup=0x19ea8e8, checkUnique=UNIQUE_CHECK_YES, indexUnchanged=false,\nheapRel=0x7f65421b0e28) at nbtinsert.c:208\n#3 0x000000000053450e in btinsert (rel=0x7f654219c2a0,\nvalues=0x7ffd9629df10, isnull=0x7ffd9629def0, ht_ctid=0x19ea894,\nheapRel=0x7f65421b0e28, checkUnique=UNIQUE_CHECK_YES,\nindexUnchanged=false, indexInfo=0x19dea80) at nbtree.c:201\n#4 0x00000000005213b6 in index_insert (indexRelation=0x7f654219c2a0,\nvalues=0x7ffd9629df10, isnull=0x7ffd9629def0, heap_t_ctid=0x19ea894,\nheapRelation=0x7f65421b0e28, checkUnique=UNIQUE_CHECK_YES,\nindexUnchanged=false, indexInfo=0x19dea80) at indexam.c:193\n#5 0x00000000005c81d5 in CatalogIndexInsert (indstate=0x19de540,\nheapTuple=0x19ea890) at indexing.c:158\n#6 0x00000000005c8325 in CatalogTupleInsert (heapRel=0x7f65421b0e28,\ntup=0x19ea890) at indexing.c:231\n#7 0x00000000005f0170 in AddSubscriptionRelState (subid=16400,\nrelid=16384, state=105 'i', sublsn=0) at pg_subscription.c:315\n#8 0x00000000006d6fa5 in CreateSubscription (pstate=0x1942dc0,\nstmt=0x191f6a0, isTopLevel=true) at subscriptioncmds.c:767\n\n~~\n\nAside: All this was happening when I did not have enough logical\nreplication workers configured. (There were WARNINGS in the logfile\nthat I had not noticed).\nWhen I fix the configuration then all these other problems went away!\n\n~~\n\nSo to summarize, I'm not sure if the fetch_tables still has some\npotential problem lurking or not, but I feel that the SQL in that\nfunction maybe needs a closer look to ensure it is always impossible\nto return the same table multiple times.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia.\n\n\n","msg_date":"Tue, 12 Apr 2022 14:56:12 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Apr 12, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Thu, Apr 7, 2022 at 2:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > FYI, here is a test script that is using the current patch (v6) to\n> > demonstrate a way to share table data between different numbers of\n> > nodes (up to 5 of them here).\n> >\n> > The script starts off with just 2-way sharing (nodes N1, N2),\n> > then expands to 3-way sharing (nodes N1, N2, N3),\n> > then 4-way sharing (nodes N1, N2, N3, N4),\n> > then 5-way sharing (nodes N1, N2, N3, N4, N5).\n> >\n> > As an extra complication, for this test, all 5 nodes have different\n> > initial table data, which gets replicated to the others whenever each\n> > new node joins the existing share group.\n> >\n> > PSA.\n> >\n>\n>\n> Hi Vignesh. I had some problems getting the above test script working.\n> It was OK up until I tried to join the 5th node (N5) to the existing 4\n> nodes. The ERROR was manifesting itself strangely because it appeared\n> that there was an index violation in the pg_subscription_rel catalog\n> even though AFAIK the N5 did not have any entries in it.\n>\n>\n> e.g.\n> 2022-04-07 09:13:28.361 AEST [24237] ERROR: duplicate key value\n> violates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n> 2022-04-07 09:13:28.361 AEST [24237] DETAIL: Key (srrelid,\n> srsubid)=(16384, 16393) already exists.\n> 2022-04-07 09:13:28.361 AEST [24237] STATEMENT: create subscription\n> sub51 connection 'port=7651' publication pub1 with\n> (subscribe_local_only=true,copy_data=force);\n> 2022-04-07 09:13:28.380 AEST [24237] ERROR: duplicate key value\n> violates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n> 2022-04-07 09:13:28.380 AEST [24237] DETAIL: Key (srrelid,\n> srsubid)=(16384, 16394) already exists.\n> 2022-04-07 09:13:28.380 AEST [24237] STATEMENT: create subscription\n> sub52 connection 'port=7652' publication pub2 with\n> (subscribe_local_only=true,copy_data=false);\n> 2022-04-07 09:13:28.405 AEST [24237] ERROR: duplicate key value\n> violates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n> 2022-04-07 09:13:28.405 AEST [24237] DETAIL: Key (srrelid,\n> srsubid)=(16384, 16395) already exists.\n> 2022-04-07 09:13:28.405 AEST [24237] STATEMENT: create subscription\n> sub53 connection 'port=7653' publication pub3 with\n> (subscribe_local_only=true,copy_data=false);\n> 2022-04-07 09:13:28.425 AEST [24237] ERROR: duplicate key value\n> violates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n> 2022-04-07 09:13:28.425 AEST [24237] DETAIL: Key (srrelid,\n> srsubid)=(16384, 16396) already exists.\n> 2022-04-07 09:13:28.425 AEST [24237] STATEMENT: create subscription\n> sub54 connection 'port=7654' publication pub4 with\n> (subscribe_local_only=true,copy_data=false);\n> 2022-04-07 09:17:52.472 AEST [25852] ERROR: duplicate key value\n> violates unique constraint \"pg_subscription_rel_srrelid_srsubid_index\"\n> 2022-04-07 09:17:52.472 AEST [25852] DETAIL: Key (srrelid,\n> srsubid)=(16384, 16397) already exists.\n> 2022-04-07 09:17:52.472 AEST [25852] STATEMENT: create subscription\n> sub51 connection 'port=7651' publication pub1;\n>\n> ~~~\n>\n> When I debugged this it seemed like each of the CREAT SUBSCRIPTION was\n> trying to make a double-entry, because the fetch_tables (your patch\n> v6-0002 modified SQL of this) was retuning the same table 2x.\n>\n> (gdb) bt\n> #0 errfinish (filename=0xbc1057 \"nbtinsert.c\", lineno=671,\n> funcname=0xbc25e0 <__func__.15798> \"_bt_check_unique\") at elog.c:510\n> #1 0x0000000000526d83 in _bt_check_unique (rel=0x7f654219c2a0,\n> insertstate=0x7ffd9629ddd0, heapRel=0x7f65421b0e28,\n> checkUnique=UNIQUE_CHECK_YES, is_unique=0x7ffd9629de01,\n> speculativeToken=0x7ffd9629ddcc) at nbtinsert.c:664\n> #2 0x0000000000526157 in _bt_doinsert (rel=0x7f654219c2a0,\n> itup=0x19ea8e8, checkUnique=UNIQUE_CHECK_YES, indexUnchanged=false,\n> heapRel=0x7f65421b0e28) at nbtinsert.c:208\n> #3 0x000000000053450e in btinsert (rel=0x7f654219c2a0,\n> values=0x7ffd9629df10, isnull=0x7ffd9629def0, ht_ctid=0x19ea894,\n> heapRel=0x7f65421b0e28, checkUnique=UNIQUE_CHECK_YES,\n> indexUnchanged=false, indexInfo=0x19dea80) at nbtree.c:201\n> #4 0x00000000005213b6 in index_insert (indexRelation=0x7f654219c2a0,\n> values=0x7ffd9629df10, isnull=0x7ffd9629def0, heap_t_ctid=0x19ea894,\n> heapRelation=0x7f65421b0e28, checkUnique=UNIQUE_CHECK_YES,\n> indexUnchanged=false, indexInfo=0x19dea80) at indexam.c:193\n> #5 0x00000000005c81d5 in CatalogIndexInsert (indstate=0x19de540,\n> heapTuple=0x19ea890) at indexing.c:158\n> #6 0x00000000005c8325 in CatalogTupleInsert (heapRel=0x7f65421b0e28,\n> tup=0x19ea890) at indexing.c:231\n> #7 0x00000000005f0170 in AddSubscriptionRelState (subid=16400,\n> relid=16384, state=105 'i', sublsn=0) at pg_subscription.c:315\n> #8 0x00000000006d6fa5 in CreateSubscription (pstate=0x1942dc0,\n> stmt=0x191f6a0, isTopLevel=true) at subscriptioncmds.c:767\n>\n> ~~\n>\n> Aside: All this was happening when I did not have enough logical\n> replication workers configured. (There were WARNINGS in the logfile\n> that I had not noticed).\n> When I fix the configuration then all these other problems went away!\n>\n> ~~\n>\n> So to summarize, I'm not sure if the fetch_tables still has some\n> potential problem lurking or not, but I feel that the SQL in that\n> function maybe needs a closer look to ensure it is always impossible\n> to return the same table multiple times.\n\nI have earlier used a distinct of srsubstate, but there is a\npossibility that when multiple subscriptions are being created there\ncan be multiple entries because the srsubstate can be different like i\nand r. I have changed srsubstate to srrelid to get the unique values.\nThe attached v8 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Tue, 12 Apr 2022 12:22:22 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for v8-0001.\n\n======\n\n1. Commit message\n\nCREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\nPUBLICATION pub1 with (local_only = true);\n\nThe spaces in the CONNECTION string are a bit strange.\n\n~~~\n\n2. src/backend/catalog/pg_subscription.\n\n@@ -71,6 +71,7 @@ GetSubscription(Oid subid, bool missing_ok)\n sub->stream = subform->substream;\n sub->twophasestate = subform->subtwophasestate;\n sub->disableonerr = subform->subdisableonerr;\n+ sub->local_only = subform->sublocal;\n\nMaybe call it subform->sublocalonly;\n\n~~~\n\n3. src/backend/catalog/system_views.sql\n\n@@ -1290,8 +1290,8 @@ REVOKE ALL ON pg_replication_origin_status FROM public;\n -- All columns of pg_subscription except subconninfo are publicly readable.\n REVOKE ALL ON pg_subscription FROM public;\n GRANT SELECT (oid, subdbid, subskiplsn, subname, subowner, subenabled,\n- subbinary, substream, subtwophasestate,\nsubdisableonerr, subslotname,\n- subsynccommit, subpublications)\n+ subbinary, substream, subtwophasestate, subdisableonerr,\n+ sublocal, subslotname, subsynccommit, subpublications)\n\nMaybe call the column sublocalonly\n\n~~~\n\n4. .../libpqwalreceiver/libpqwalreceiver.c\n\n@@ -453,6 +453,10 @@ libpqrcv_startstreaming(WalReceiverConn *conn,\n PQserverVersion(conn->streamConn) >= 150000)\n appendStringInfoString(&cmd, \", two_phase 'on'\");\n\n+ if (options->proto.logical.local_only &&\n+ PQserverVersion(conn->streamConn) >= 150000)\n+ appendStringInfoString(&cmd, \", local_only 'on'\");\n\nAdd a FIXME comment here as a reminder that this condition needs to\nchange for PG16.\n\n~~~\n\n5. src/bin/pg_dump/pg_dump.c\n\n@@ -4361,6 +4361,7 @@ getSubscriptions(Archive *fout)\n int i_subsynccommit;\n int i_subpublications;\n int i_subbinary;\n+ int i_sublocal;\n int i,\n ntups;\n\nMaybe call this member i_sublocalonly;\n\n~~~\n\n6. src/bin/pg_dump/pg_dump.c\n+ if (fout->remoteVersion >= 150000)\n+ appendPQExpBufferStr(query, \" s.sublocal\\n\");\n+ else\n+ appendPQExpBufferStr(query, \" false AS sublocal\\n\");\n+\n\n6a. Add a FIXME comment as a reminder that this condition needs to\nchange for PG16.\n\n6b. Change the column name to sublocalonly.\n\n~~~\n\n7. src/bin/pg_dump/pg_dump.h\n\n@@ -661,6 +661,7 @@ typedef struct _SubscriptionInfo\n char *subdisableonerr;\n char *subsynccommit;\n char *subpublications;\n+ char *sublocal;\n } SubscriptionInfo;\n\nChange the member name to sublocalonly\n\n~~~\n\n8. src/bin/psql/describe.c\n\n@@ -6241,6 +6241,12 @@ describeSubscriptions(const char *pattern, bool verbose)\n gettext_noop(\"Two phase commit\"),\n gettext_noop(\"Disable on error\"));\n\n+ /* local_only is supported only in v15 and higher */\n+ if (pset.sversion >= 150000)\n+ appendPQExpBuffer(&buf,\n+ \", sublocal AS \\\"%s\\\"\\n\",\n+ gettext_noop(\"Local only\"));\n\n7a. The comment is wrong to mention v15.\n\n7b. Add a FIXME comment as a reminder that this condition needs to\nchange for PG16.\n\n7c. Change the column name to sublocalonly.\n\n~~~\n\n9. src/include/catalog/pg_subscription.h\n\n@@ -70,6 +70,8 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\nBKI_SHARED_RELATION BKI_ROW\n\n bool substream; /* Stream in-progress transactions. */\n\n+ bool sublocal; /* skip copying of remote origin data */\n\nChange the member name to sublocalonly\n\n~~~\n\n10. src/include/replication/logicalproto.h\n\n@@ -32,12 +32,17 @@\n *\n * LOGICALREP_PROTO_TWOPHASE_VERSION_NUM is the minimum protocol version with\n * support for two-phase commit decoding (at prepare time). Introduced in PG15.\n+ *\n+ * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\n+ * support for sending only locally originated data from the publisher.\n+ * Introduced in PG16.\n\nAdd a FIXME comment here as a reminder that the proto version number\nneeds to be bumped to 4 in PG16.\n\n~~~\n\n11. src/test/subscription/t/032_circular.pl\n\nPerhaps there should be another test using a third \"Node_C\" which\npublishes some data to Node_B. Then you can ensure that by using\nlocal_only (when Node_A is subscribing to Node_A) that nothing from\nthe Node_C can find its way onto Node_A. But then the test name\n\"circular\" is a bit misleading. Maybe it should be \"032_localonly\"?\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 14 Apr 2022 18:22:59 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Apr 14, 2022 at 1:53 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for v8-0001.\n>\n> ======\n>\n> 1. Commit message\n>\n> CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\n> PUBLICATION pub1 with (local_only = true);\n>\n> The spaces in the CONNECTION string are a bit strange.\n\nModified\n\n> ~~~\n>\n> 2. src/backend/catalog/pg_subscription.\n>\n> @@ -71,6 +71,7 @@ GetSubscription(Oid subid, bool missing_ok)\n> sub->stream = subform->substream;\n> sub->twophasestate = subform->subtwophasestate;\n> sub->disableonerr = subform->subdisableonerr;\n> + sub->local_only = subform->sublocal;\n>\n> Maybe call it subform->sublocalonly;\n>\n\nModified\n\n> ~~~\n>\n> 3. src/backend/catalog/system_views.sql\n>\n> @@ -1290,8 +1290,8 @@ REVOKE ALL ON pg_replication_origin_status FROM public;\n> -- All columns of pg_subscription except subconninfo are publicly readable.\n> REVOKE ALL ON pg_subscription FROM public;\n> GRANT SELECT (oid, subdbid, subskiplsn, subname, subowner, subenabled,\n> - subbinary, substream, subtwophasestate,\n> subdisableonerr, subslotname,\n> - subsynccommit, subpublications)\n> + subbinary, substream, subtwophasestate, subdisableonerr,\n> + sublocal, subslotname, subsynccommit, subpublications)\n>\n> Maybe call the column sublocalonly\n>\n\nModified\n\n> ~~~\n>\n> 4. .../libpqwalreceiver/libpqwalreceiver.c\n>\n> @@ -453,6 +453,10 @@ libpqrcv_startstreaming(WalReceiverConn *conn,\n> PQserverVersion(conn->streamConn) >= 150000)\n> appendStringInfoString(&cmd, \", two_phase 'on'\");\n>\n> + if (options->proto.logical.local_only &&\n> + PQserverVersion(conn->streamConn) >= 150000)\n> + appendStringInfoString(&cmd, \", local_only 'on'\");\n>\n> Add a FIXME comment here as a reminder that this condition needs to\n> change for PG16.\n>\n\nModified\n\n> ~~~\n>\n> 5. src/bin/pg_dump/pg_dump.c\n>\n> @@ -4361,6 +4361,7 @@ getSubscriptions(Archive *fout)\n> int i_subsynccommit;\n> int i_subpublications;\n> int i_subbinary;\n> + int i_sublocal;\n> int i,\n> ntups;\n>\n> Maybe call this member i_sublocalonly;\n>\n\nModified\n\n> ~~~\n>\n> 6. src/bin/pg_dump/pg_dump.c\n> + if (fout->remoteVersion >= 150000)\n> + appendPQExpBufferStr(query, \" s.sublocal\\n\");\n> + else\n> + appendPQExpBufferStr(query, \" false AS sublocal\\n\");\n> +\n>\n> 6a. Add a FIXME comment as a reminder that this condition needs to\n> change for PG16.\n>\n\nModified\n\n> 6b. Change the column name to sublocalonly.\n>\n\nModified\n\n> ~~~\n>\n> 7. src/bin/pg_dump/pg_dump.h\n>\n> @@ -661,6 +661,7 @@ typedef struct _SubscriptionInfo\n> char *subdisableonerr;\n> char *subsynccommit;\n> char *subpublications;\n> + char *sublocal;\n> } SubscriptionInfo;\n>\n> Change the member name to sublocalonly\n>\n\nModified\n\n> ~~~\n>\n> 8. src/bin/psql/describe.c\n>\n> @@ -6241,6 +6241,12 @@ describeSubscriptions(const char *pattern, bool verbose)\n> gettext_noop(\"Two phase commit\"),\n> gettext_noop(\"Disable on error\"));\n>\n> + /* local_only is supported only in v15 and higher */\n> + if (pset.sversion >= 150000)\n> + appendPQExpBuffer(&buf,\n> + \", sublocal AS \\\"%s\\\"\\n\",\n> + gettext_noop(\"Local only\"));\n>\n> 7a. The comment is wrong to mention v15.\n>\n\nI have removed this comment and added a FIXME. I will add it in once\nversion change is done to avoid confusion.\n\n> 7b. Add a FIXME comment as a reminder that this condition needs to\n> change for PG16.\n>\n\nModified\n\n> 7c. Change the column name to sublocalonly.\n>\n\nModified\n\n> ~~~\n>\n> 9. src/include/catalog/pg_subscription.h\n>\n> @@ -70,6 +70,8 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\n> BKI_SHARED_RELATION BKI_ROW\n>\n> bool substream; /* Stream in-progress transactions. */\n>\n> + bool sublocal; /* skip copying of remote origin data */\n>\n> Change the member name to sublocalonly\n>\n\nModified\n\n> ~~~\n>\n> 10. src/include/replication/logicalproto.h\n>\n> @@ -32,12 +32,17 @@\n> *\n> * LOGICALREP_PROTO_TWOPHASE_VERSION_NUM is the minimum protocol version with\n> * support for two-phase commit decoding (at prepare time). Introduced in PG15.\n> + *\n> + * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\n> + * support for sending only locally originated data from the publisher.\n> + * Introduced in PG16.\n>\n> Add a FIXME comment here as a reminder that the proto version number\n> needs to be bumped to 4 in PG16.\n>\n\nModified\n\n> ~~~\n>\n> 11. src/test/subscription/t/032_circular.pl\n>\n> Perhaps there should be another test using a third \"Node_C\" which\n> publishes some data to Node_B. Then you can ensure that by using\n> local_only (when Node_A is subscribing to Node_A) that nothing from\n> the Node_C can find its way onto Node_A. But then the test name\n> \"circular\" is a bit misleading. Maybe it should be \"032_localonly\"?\n\nAdded the test and changed the file.\n\nThanks for the comments, attached v9 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Sat, 16 Apr 2022 19:48:22 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"I checked the latest v9-0001 patch. Below are my review comments.\n\nOther than these few trivial comments this 0001 patch looks good to me.\n\n~~~\n\n1. src/backend/replication/pgoutput/pgoutput.c - whitespace\n\n@@ -1696,6 +1714,10 @@ static bool\n pgoutput_origin_filter(LogicalDecodingContext *ctx,\n RepOriginId origin_id)\n {\n+ PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\n+\n+ if (data->local_only && origin_id != InvalidRepOriginId)\n+ return true;\n return false;\n }\n\nSuggest to add a blank line after the return true;\n\n~~~\n\n2. src/bin/psql/tab-complete.c - not alphabetical\n\n@@ -1874,7 +1874,7 @@ psql_completion(const char *text, int start, int end)\n COMPLETE_WITH(\"(\", \"PUBLICATION\");\n /* ALTER SUBSCRIPTION <name> SET ( */\n else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\nTailMatches(\"SET\", \"(\"))\n- COMPLETE_WITH(\"binary\", \"slot_name\", \"streaming\",\n\"synchronous_commit\", \"disable_on_error\");\n+ COMPLETE_WITH(\"binary\", \"slot_name\", \"streaming\", \"local_only\",\n\"synchronous_commit\", \"disable_on_error\");\n\n2a. AFAIK the code intended that these options be listed in\nalphabetical order (I think the recent addition of disable_on_error is\nalso wrong here). So \"local_only\" should be moved.\n\n@@ -3156,7 +3156,7 @@ psql_completion(const char *text, int start, int end)\n /* Complete \"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\" */\n else if (HeadMatches(\"CREATE\", \"SUBSCRIPTION\") && TailMatches(\"WITH\", \"(\"))\n COMPLETE_WITH(\"binary\", \"connect\", \"copy_data\", \"create_slot\",\n- \"enabled\", \"slot_name\", \"streaming\",\n+ \"enabled\", \"slot_name\", \"streaming\", \"local_only\",\n \"synchronous_commit\", \"two_phase\", \"disable_on_error\");\n\n2b. ditto\n\n~~~\n\n3. src/test/subscription/t/032_localonly.pl - wrong message\n\n+$node_C->wait_for_catchup($appname_B2);\n+$node_B->wait_for_catchup($appname_A);\n+\n+$result = $node_B->safe_psql('postgres', \"SELECT * FROM tab_full;\");\n+is( $result, qq(11\n+12\n+13),\n+ 'Inserted successfully without leading to infinite recursion in\ncircular replication setup'\n+);\n+\n+# check that the data published from node_C to node_B is not sent to node_A\n+$result = $node_A->safe_psql('postgres', \"SELECT * FROM tab_full;\");\n+is( $result, qq(11\n+12),\n+ 'Inserted successfully without leading to infinite recursion in\ncircular replication setup'\n+);\n+\n\nThe new test looked good, but the cut/paste text message ('Inserted\nsuccessfully without leading to infinite recursion in circular\nreplication setup') maybe needs changing because there is nothing\nreally \"circular\" about this test case.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia.\n\n\n","msg_date":"Tue, 19 Apr 2022 12:59:16 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Below are my review comments for the v9-0002 patch (except I did not\nyet look at the TAP tests).\n\n~~~\n\n1. General comment - describe.c\n\nI wondered why the copy_data enum value is not displayed by the psql\n\\drs+ command. Should it be?\n\n~~~\n\n2. General comment - SUBOPT_LOCAL_ONLY\n\n@@ -1134,7 +1204,7 @@ AlterSubscription(ParseState *pstate,\nAlterSubscriptionStmt *stmt,\n\n case ALTER_SUBSCRIPTION_SET_PUBLICATION:\n {\n- supported_opts = SUBOPT_COPY_DATA | SUBOPT_REFRESH;\n+ supported_opts = SUBOPT_COPY_DATA | SUBOPT_REFRESH | SUBOPT_LOCAL_ONLY;\n parse_subscription_options(pstate, stmt->options,\n supported_opts, &opts);\n\n@@ -1236,7 +1308,8 @@ AlterSubscription(ParseState *pstate,\nAlterSubscriptionStmt *stmt,\n errmsg(\"ALTER SUBSCRIPTION ... REFRESH is not allowed for disabled\nsubscriptions\")));\n\n parse_subscription_options(pstate, stmt->options,\n- SUBOPT_COPY_DATA, &opts);\n+ SUBOPT_COPY_DATA | SUBOPT_LOCAL_ONLY,\n+ &opts);\n\nI noticed there is some new code that appears to be setting the\nSUBOT_LOCAL_ONLY as a supported option. Shouldn't those changes belong\nin the patch 0001 instead of in this patch 0002?\n\n~~~\n\n3. Commit message - wording\n\nCURRENT\na) Node1 - Publication publishing employee table.\nb) Node2 - Subscription subscribing from publication pub1 with\nlocal_only.\nc) Node2 - Publication publishing employee table.\nd) Node1 - Subscription subscribing from publication pub2 with\nlocal_only.\n\nSUGGESTION (I think below has the same meaning but is simpler)\na) Node1 - PUBLICATION pub1 for the employee table\nb) Node2 - SUBSCRIPTION from pub1 with local_only=true\nc) Node2 - PUBLICATION pub2 for the employee table\nd) Node1 - SUBSCRIPTION from pub2 with local_only=true\n\n~~~\n\n4. Commit message - meaning?\n\nCURRENT\nNow when user is trying to add another node Node3 to the\nabove Multi master logical replication setup:\na) user will have to create one subscription subscribing from Node1 to\nNode3\nb) user wil have to create another subscription subscribing from\nNode2 to Node3 using local_only option and copy_data as true.\n\nWhile the subscription is created, server will identify that Node2 is\nsubscribing from Node1 and Node1 is subscribing from Node2 and throw an\nerror so that user can handle the initial copy data.\n\n~\n\nThe wording above confused me. Can you clarify it?\ne.g.\na) What exactly was the user hoping to achieve here?\nb) Are the user steps doing something deliberately wrong just so you\ncan describe later that an error gets thrown?\n\n~~~\n\n5. doc/src/sgml/logical-replication.sgml - how to get here?\n\nI didn’t see any easy way to get to this page. (No cross refs from anywhere?)\n\n~~~\n\n6. doc/src/sgml/logical-replication.sgml - section levels\n\nI think the section levels may be a bit messed up. e.g. The HTML\nrendering of sections looks a bit confused. Maybe this is same as my\nreview comment #13.\n\n~~\n\n7. doc/src/sgml/logical-replication.sgml - headings\n\n<title>Setting Bidirection logical replication between two nodes:</title>\n\n7a. Maybe better just to have a simpler main heading like\n\"Bidirectional logical replication\".\n\n7b. Don't put \":\" in the title.\n\n~~~\n\n8. doc/src/sgml/logical-replication.sgml - missing intro\n\nIMO this page needs some sort of introduction/blurb instead of leaping\nstraight into examples without any preamble text to give context.\n\n~~~\n\n9. doc/src/sgml/logical-replication.sgml - bullets\n\nSuggest removing all the bullets from the example steps. (I had\nsomething similar a while ago for the \"Row Filter\" page but\neventually, they all had to be removed).\n\n~~~\n\n10. doc/src/sgml/logical-replication.sgml - SQL too long\n\nMany of the example commands are much too long, and for me, they are\ngiving scroll bars when rendered. It would be better if they can be\nwrapped in appropriate places so easier to read (and no resulting\nscroll bars).\n\n~~~\n\n11. doc/src/sgml/logical-replication.sgml - add the psql prompt\n\nIMO it would also be easier to understand the examples if you show the\npsql prompt. Then you can easily know the node context without having\nto describe it in the text so often.\n\ne.g.\n\n+ <para>\n+ Create the subscription in node2 to subscribe the changes from node1:\n+<programlisting>\n+CREATE SUBSCRIPTION sub_node1_node2 ...\n\nSUGGGESTION\n+ <para>\n+ Create the subscription in node2 to subscribe the changes from node1\n+<programlisting>\n+node_2=# CREATE SUBSCRIPTION sub_node1_node2 ...\n\n~~~\n\n12. doc/src/sgml/logical-replication.sgml - typo\n\n+ <para>\n+ Now the BiDirectional logical replication setup is complete between node1\n\ntypo \"BiDirectional\"\n\n~~~\n\n13. doc/src/sgml/logical-replication.sgml - deep levels\n\nThe section levels are very deep, but > 3 will not appear in the table\nof contents when rendered. Maybe you can rearrange to raise them all\nup one level, then IMO the TOC will work better and the whole page\nwill be easier to read.\n\n~~~\n\n14. doc/src/sgml/logical-replication.sgml - unnecessarily complex?\n\n+<programlisting>\n+# Truncate the table data but do not replicate the truncate.\n+ALTER PUBLICATION pub_node3 SET (publish='insert,update,delete');\n+TRUNCATE t1;\n+\n+CREATE SUBSCRIPTION sub_node3_node1 CONNECTION 'dbname=foo host=node1\nuser=repuser' PUBLICATION pub_node1 WITH (copy_data = force,\nlocal_only = on);\n+\n+CREATE SUBSCRIPTION sub_node3_node2 CONNECTION 'dbname=foo host=node2\nuser=repuser' PUBLICATION pub_node2 WITH (copy_data = off, local_only\n= on);\n+\n+# Include truncate operations from now\n+ALTER PUBLICATION pub_node3 SET (publish='insert,update,delete,truncate');\n+</programlisting>\n\nIs it really necessary for those CREATE SUBSCRIPTION to be placed\nwhere they are? I felt it would be less confusing if you just do the\nTRUNCATE between the 2x ALTER PUBLICATION... and then do those CREATE\nSUBSCRIPTION separately later.\n\n~~~\n\n14. doc/src/sgml/logical-replication.sgml - force?\n\nThere seems no documentation anywhere that says what is the\npurpose/meaning of the copy_data enum \"force\".\n\n~~~\n\n15. doc/src/sgml/ref/alter_subscription.sgml - force?\n\nMeanings of copy_data true/false are self-evident but should something\nhere be explaining what is the meaning of \"force\"? Or a reference to\nsome place that explains it?\n\n~~~\n\n16. doc/src/sgml/ref/create_subscription.sgml - local_only missing notes?\n\n@@ -161,6 +161,11 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n publisher node changes regardless of their origin. The default is\n <literal>false</literal>.\n </para>\n+ <para>\n+ There is some interation between the \"local_only\" option and\n+ \"copy_data\" option. Refer to the\n+ <xref linkend=\"sql-createsubscription-notes\" /> for details.\n+ </para>\n </listitem>\n </varlistentry>\n\n16a. Typo \"interation\"\n\n16b. But where in the referenced notes does it actually say anything about this?\n\n~~~\n\n17. doc/src/sgml/ref/create_subscription.sgml - force?\n\nThe notes about the copy_data do not say anything about what the\nvalues mean. Specifically, there seems nothing that describes what is\nthe meaning of \"force\".\n\n~~~\n\n18. doc/src/sgml/ref/create_subscription.sgml - copy_data missing notes?\n\n+\n+ <para>\n+ There is some interation between the \"local_only\" option and\n+ \"copy_data\" option. Refer to the\n+ <xref linkend=\"sql-createsubscription-notes\" /> for details.\n+ </para>\n\n18a. Typo \"interation\"\n\n18b. But where in the referenced notes does it actually say anything about this?\n\n~~~\n\n19. doc/src/sgml/ref/create_subscription.sgml - xref to the new page\n\nShouldn't there be an xref on this page somewhere to your other new\n\"Bidirectional\" docs page?\n\n~~~\n\n20. src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_ON_OR_FORCE\n\n@@ -69,6 +69,18 @@\n /* check if the 'val' has 'bits' set */\n #define IsSet(val, bits) (((val) & (bits)) == (bits))\n\n+#define IS_COPY_DATA_ON_OR_FORCE(copy_data) (copy_data != COPY_DATA_OFF)\n\nMaybe this would be better as a static inline function?\n\n~~~\n\n21. src/backend/commands/subscriptioncmds.c - comment\n\n+/*\n+ * Represents whether copy_data option is specified with on, off or force.\n+ */\n+typedef enum CopyData\n+{\n+ COPY_DATA_OFF,\n+ COPY_DATA_ON,\n+ COPY_DATA_FORCE\n+} CopyData;\n\nI felt it might be better if the comment described these enums in the\nsame order they are defined.\n\nE.g. \"Represents whether copy_data option is specified as off/on, or force.\"\n\n~~~\n\n22. src/backend/commands/subscriptioncmds.c - CreateSubscription bug?\n\n@@ -720,7 +788,8 @@ CreateSubscription(ParseState *pstate,\nCreateSubscriptionStmt *stmt,\n * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\n * PUBLICATION to work.\n */\n- if (opts.twophase && !opts.copy_data && tables != NIL)\n+ if (opts.twophase && IS_COPY_DATA_ON_OR_FORCE(opts.copy_data)\n+ && tables != NIL)\n twophase_enabled = true;\nIs that a bug? It does not seem to match the previous code. Should\nthat IS_COPY_DATA_ON_OR_FORCE be \"not\" ?\n\n~~~\n\n23. src/backend/commands/subscriptioncmds.c - long errmsg\n\n+ if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\n+ ereport(ERROR,\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n+ errmsg(\"CREATE/ALTER SUBSCRIPTION with local_only and copy_data as\ntrue is not allowed when the publisher might have replicated data,\ntable:%s.%s might have replicated data in the publisher\",\n+ nspname, relname),\n+ errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\"));\n+\n\nThe errmsg seems way too long for the source code. Can you use string\nconcatenation or continuation chars to wrap the message over multiple\nlines?\n\n~~~\n\n24. src/test/regress/sql/subscription.sql - typo\n\n@@ -66,6 +69,15 @@ ALTER SUBSCRIPTION regress_testsub4 SET (local_only = false);\n DROP SUBSCRIPTION regress_testsub3;\n DROP SUBSCRIPTION regress_testsub4;\n\n+-- ok - valid coy_data options\n\nTypo \"coy_data\". (it looks like this typo is not caused by this patch,\nbut I think this patch should fix it anyhow).\n\n~~~\n\n25. src/test/regress/sql/subscription.sql - test order\n\nThe new tests are OK but IMO they could be re-ordered so then they\nwill be more consistent for the positive and negative tests.\n\nCURRENT\n\"true/force/on/1\" and \"off/false/0\"\n\nSUGGEST\n\"true/on/1/force\" and \"false/off/0\"\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 20 Apr 2022 11:56:08 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Below are my review comments for the v9-0002 patch (TAP test part only).\n\n(The order of my comments is a bit muddled because I jumped around in\nthe file a bit while reviewing it).\n\n======\n\n1. create_subscription - missing comment.\n\n+sub create_subscription\n+{\n+ my ($node_subscriber, $node_publisher, $sub_name, $node_connstr,\n+ $application_name, $pub_name, $copy_data_val)\n+ = @_;\n\nMaybe add a comment for this subroutine and describe the expected parameters.\n\n~~\n\n2. create_subscription - hides the options\n\nIMO the \"create_subscription\" subroutine is hiding too many of the\ndetails, so now it is less clear (from the caller's POV) what the test\nis really doing. Perhaps the options could be passed more explicitly\nto this subroutine.\n\ne.g. Instead of\n\ncreate_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\n $appname_B1, 'tap_pub_A', 'on');\n\nperhaps explicitly set the WITH options like:\n\ncreate_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\n $appname_B1, 'tap_pub_A', 'local_only = on, copy_data = on');\n\n~~~\n\n3. the application names are confusing\n\n+my $appname_A1 = 'tap_sub_A_B';\n+my $appname_A2 = 'tap_sub_A_C';\n+my $appname_B1 = 'tap_sub_B_A';\n+my $appname_B2 = 'tap_sub_B_C';\n+my $appname_C1 = 'tap_sub_C_A';\n+my $appname_C2 = 'tap_sub_C_B';\n\nI found the application names using '1' and '2' to be a bit confusing.\ni.e it was unnecessarily hard to associate them (in my head) with\ntheir relevant subscriptions. IMO it would be easier to name them\nusing letters like below:\n\nSUGGESTED\nmy $appname_AB = 'tap_sub_A_B';\nmy $appname_AC = 'tap_sub_A_C';\nmy $appname_BA = 'tap_sub_B_A';\nmy $appname_BC = 'tap_sub_B_C';\nmy $appname_CA = 'tap_sub_C_A';\nmy $appname_CB = 'tap_sub_C_B';\n\n~~~\n\n4. create_subscription - passing the $appname 2x.\n\n+create_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\n+ $appname_B1, 'tap_pub_A', 'on');\n\n\nIt seemed confusing that the $app_name is passed twice.\n\nIMO should rename all those $appname_XX vars (see previous comment) to\nbe $subname_XX, and just pass that create_subscription instead.\n\nmy $subname_AB = 'tap_sub_A_B';\nmy $subname_AC = 'tap_sub_A_C';\nmy $subname_BA = 'tap_sub_B_A';\nmy $subname_BC = 'tap_sub_B_C';\nmy $subname_CA = 'tap_sub_C_A';\nmy $subname_CB = 'tap_sub_C_B';\n\nThen create_subscription subroutine should have one less argument.\nJust add a comment saying that the appname is always assigned the same\nvalue as the subname.\n\n~~~\n\n5. cleanup part - seems a bit misleading\n\n+# cleanup\n+$node_B->safe_psql(\n+ 'postgres', \"\n+ DROP SUBSCRIPTION $appname_B2\");\n+$node_C->safe_psql(\n+ 'postgres', \"\n+ DELETE FROM tab_full\");\n+$node_B->safe_psql(\n+ 'postgres', \"\n+ DELETE FROM tab_full where a = 13\");\n\nIs that comment misleading? IIUC this is not really cleaning up\neverything. It is just a cleanup of the previous Node_C test part\nisn't it?\n\n~~~\n\n6. Error case (when copy_data is true)\n\n+# Error when creating susbcription with local_only and copy_data as true when\n+# the publisher has replicated data\n\n6a. Typo \"susbcription\"\n\n6b. That comment maybe needs some more explanation - eg. say that\nsince Node_A is already subscribing to Node_B so when Node_B makes\nanother subscription to Node_A the copy doesn't really know if the\ndata really originated from Node_A or not...\n\n6c. Maybe the comment needed to be more like ############## style to\ndenote this (and the one that follows) is a separate test case.\n\n~~~\n\n7. Error case (when copy_data is force)\n\n+# Creating subscription with local_only and copy_data as force should be\n+# successful when the publisher has replicated data\n\n7a. Same typo \"subscription\"\n\n~~~\n\n8. Add 3rd node when the existing node has some data\n\n8a. Maybe that comment needs to be expanded more. This isn't just\n\"adding a node\" - you are joining it to the others as another\nbi-directional participant in a 3-way group. And the comment should\nclarify exactly which nodes have the initial data. Both the existing 2\nyou are joining to? The new one only? All of them?\n\n8b. IMO is would be much clearer to do SELECT from all the nodes at\nthe start of this test just to re-confirm what is all the initial data\non these nodes before joining the 3rd node.\n\nNOTE - These same review comments apply to the other test combinations too\n- Add 3rd node when the existing node has no data\n- Add 3rd node when the new node has some data\n\n~~~\n\n9. Inserted data\n\n+# insert some data in all the nodes\n+$node_A->safe_psql('postgres', \"INSERT INTO tab_full VALUES (13);\");\n+$node_B->safe_psql('postgres', \"INSERT INTO tab_full VALUES (21);\");\n+$node_C->safe_psql('postgres', \"INSERT INTO tab_full VALUES (31);\");\n\nThey seemed strange values (13, 21, 31) to add. I couldn't work out\nthe \"meaning\" of them.\n\nWouldn't values like 13, 23, 33 make more sense (e.g pattern is 10 x\nnode# + something)\n\n~~~\n\n10. verify_data($node_A, $node_B, $node_C);\n\nAll the expected values are too buried in this subroutine which makes\nit hard to read. I think it would be easier to read (from the caller's\nPOV) if you can pass the *expected* values as another arg into the\nverify_data subroutine.\n\ne.g. is something like this possible?\nverify_data($node_A, $node_B, $node_C, [11,12,13])\n\n~~~\n\n11. ALTER for ignoring 'truncate'\n\n+$node_C->safe_psql('postgres',\n+ \"ALTER PUBLICATION tap_pub_C SET (publish='insert,update,delete');\");\n+\n+$node_C->safe_psql('postgres', \"TRUNCATE tab_full\");\n+\n+create_subscription($node_C, $node_A, $appname_C1, $node_A_connstr,\n+ $appname_C1, 'tap_pub_A', 'force');\n+create_subscription($node_C, $node_B, $appname_C2, $node_B_connstr,\n+ $appname_C2, 'tap_pub_B', 'off');\n+\n+#include truncates now\n+$node_C->safe_psql('postgres',\n+ \"ALTER PUBLICATION tap_pub_C SET\n(publish='insert,update,delete,truncate');\"\n+);\n\nDo those create_subscription calls need to be encapsulated by the\nALTER like that? I did not think so. It looks a bit more complex if\ndone this way.\n\n~~~\n\n12. clean_subscriber_contents - misnamed?\n\nThis subroutine feels a bit misnamed. It seems to be doing lots of\nthings like detaching the Node_C and deleting all table data from all\nnodes. That all seems quite different from just \"clean subscriber\ncontents\".\n\n~~~\n\n13. table initial data?\n\n+clean_subscriber_contents($node_A, $node_B, $node_C);\n+\n+##########################################################################\n+# Add 3rd node when the new node has some data\n+##########################################################################\n\nBut does this test case *really* have some data? I am not so sure.\nDoesn't the preceding \"clean_subscriber_contents\" call remove all the\ndata that might have been there? That is why I think all the tests out\nto have SELECT (previous comment #8) so they can re-confirm what data\nis really in those tables before doing each test part.\n\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 20 Apr 2022 15:49:24 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Apr 19, 2022 at 8:29 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> I checked the latest v9-0001 patch. Below are my review comments.\n>\n> Other than these few trivial comments this 0001 patch looks good to me.\n>\n> ~~~\n>\n> 1. src/backend/replication/pgoutput/pgoutput.c - whitespace\n>\n> @@ -1696,6 +1714,10 @@ static bool\n> pgoutput_origin_filter(LogicalDecodingContext *ctx,\n> RepOriginId origin_id)\n> {\n> + PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\n> +\n> + if (data->local_only && origin_id != InvalidRepOriginId)\n> + return true;\n> return false;\n> }\n>\n> Suggest to add a blank line after the return true;\n\nModified\n\n> ~~~\n>\n> 2. src/bin/psql/tab-complete.c - not alphabetical\n>\n> @@ -1874,7 +1874,7 @@ psql_completion(const char *text, int start, int end)\n> COMPLETE_WITH(\"(\", \"PUBLICATION\");\n> /* ALTER SUBSCRIPTION <name> SET ( */\n> else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\n> TailMatches(\"SET\", \"(\"))\n> - COMPLETE_WITH(\"binary\", \"slot_name\", \"streaming\",\n> \"synchronous_commit\", \"disable_on_error\");\n> + COMPLETE_WITH(\"binary\", \"slot_name\", \"streaming\", \"local_only\",\n> \"synchronous_commit\", \"disable_on_error\");\n>\n> 2a. AFAIK the code intended that these options be listed in\n> alphabetical order (I think the recent addition of disable_on_error is\n> also wrong here). So \"local_only\" should be moved.\n\nModified\n\n> @@ -3156,7 +3156,7 @@ psql_completion(const char *text, int start, int end)\n> /* Complete \"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\" */\n> else if (HeadMatches(\"CREATE\", \"SUBSCRIPTION\") && TailMatches(\"WITH\", \"(\"))\n> COMPLETE_WITH(\"binary\", \"connect\", \"copy_data\", \"create_slot\",\n> - \"enabled\", \"slot_name\", \"streaming\",\n> + \"enabled\", \"slot_name\", \"streaming\", \"local_only\",\n> \"synchronous_commit\", \"two_phase\", \"disable_on_error\");\n>\n> 2b. ditto\n\nModified\n\n> ~~~\n>\n> 3. src/test/subscription/t/032_localonly.pl - wrong message\n>\n> +$node_C->wait_for_catchup($appname_B2);\n> +$node_B->wait_for_catchup($appname_A);\n> +\n> +$result = $node_B->safe_psql('postgres', \"SELECT * FROM tab_full;\");\n> +is( $result, qq(11\n> +12\n> +13),\n> + 'Inserted successfully without leading to infinite recursion in\n> circular replication setup'\n> +);\n> +\n> +# check that the data published from node_C to node_B is not sent to node_A\n> +$result = $node_A->safe_psql('postgres', \"SELECT * FROM tab_full;\");\n> +is( $result, qq(11\n> +12),\n> + 'Inserted successfully without leading to infinite recursion in\n> circular replication setup'\n> +);\n> +\n>\n> The new test looked good, but the cut/paste text message ('Inserted\n> successfully without leading to infinite recursion in circular\n> replication setup') maybe needs changing because there is nothing\n> really \"circular\" about this test case.\n\nModified\n\nAttached v10 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Fri, 22 Apr 2022 21:45:25 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Apr 20, 2022 at 7:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Below are my review comments for the v9-0002 patch (except I did not\n> yet look at the TAP tests).\n>\n> ~~~\n>\n> 1. General comment - describe.c\n>\n> I wondered why the copy_data enum value is not displayed by the psql\n> \\drs+ command. Should it be?\n>\n\nI noticed that we generally don't display the values for the options.\nI did not add it to keep it consistent with other options.\n\n> ~~~\n>\n> 2. General comment - SUBOPT_LOCAL_ONLY\n>\n> @@ -1134,7 +1204,7 @@ AlterSubscription(ParseState *pstate,\n> AlterSubscriptionStmt *stmt,\n>\n> case ALTER_SUBSCRIPTION_SET_PUBLICATION:\n> {\n> - supported_opts = SUBOPT_COPY_DATA | SUBOPT_REFRESH;\n> + supported_opts = SUBOPT_COPY_DATA | SUBOPT_REFRESH | SUBOPT_LOCAL_ONLY;\n> parse_subscription_options(pstate, stmt->options,\n> supported_opts, &opts);\n>\n> @@ -1236,7 +1308,8 @@ AlterSubscription(ParseState *pstate,\n> AlterSubscriptionStmt *stmt,\n> errmsg(\"ALTER SUBSCRIPTION ... REFRESH is not allowed for disabled\n> subscriptions\")));\n>\n> parse_subscription_options(pstate, stmt->options,\n> - SUBOPT_COPY_DATA, &opts);\n> + SUBOPT_COPY_DATA | SUBOPT_LOCAL_ONLY,\n> + &opts);\n>\n> I noticed there is some new code that appears to be setting the\n> SUBOT_LOCAL_ONLY as a supported option. Shouldn't those changes belong\n> in the patch 0001 instead of in this patch 0002?\n\nModified\n\n> ~~~\n>\n> 3. Commit message - wording\n>\n> CURRENT\n> a) Node1 - Publication publishing employee table.\n> b) Node2 - Subscription subscribing from publication pub1 with\n> local_only.\n> c) Node2 - Publication publishing employee table.\n> d) Node1 - Subscription subscribing from publication pub2 with\n> local_only.\n>\n> SUGGESTION (I think below has the same meaning but is simpler)\n> a) Node1 - PUBLICATION pub1 for the employee table\n> b) Node2 - SUBSCRIPTION from pub1 with local_only=true\n> c) Node2 - PUBLICATION pub2 for the employee table\n> d) Node1 - SUBSCRIPTION from pub2 with local_only=true\n\nModified\n\n> ~~~\n>\n> 4. Commit message - meaning?\n>\n> CURRENT\n> Now when user is trying to add another node Node3 to the\n> above Multi master logical replication setup:\n> a) user will have to create one subscription subscribing from Node1 to\n> Node3\n> b) user wil have to create another subscription subscribing from\n> Node2 to Node3 using local_only option and copy_data as true.\n>\n> While the subscription is created, server will identify that Node2 is\n> subscribing from Node1 and Node1 is subscribing from Node2 and throw an\n> error so that user can handle the initial copy data.\n>\n> ~\n>\n> The wording above confused me. Can you clarify it?\n> e.g.\n> a) What exactly was the user hoping to achieve here?\n> b) Are the user steps doing something deliberately wrong just so you\n> can describe later that an error gets thrown?\n\nModified\n\n> ~~~\n>\n> 5. doc/src/sgml/logical-replication.sgml - how to get here?\n>\n> I didn’t see any easy way to get to this page. (No cross refs from anywhere?)\n\nI have added a link from create subscription page\n\n> ~~~\n>\n> 6. doc/src/sgml/logical-replication.sgml - section levels\n>\n> I think the section levels may be a bit messed up. e.g. The HTML\n> rendering of sections looks a bit confused. Maybe this is same as my\n> review comment #13.\n\nModified\n\n> ~~\n>\n> 7. doc/src/sgml/logical-replication.sgml - headings\n>\n> <title>Setting Bidirection logical replication between two nodes:</title>\n>\n> 7a. Maybe better just to have a simpler main heading like\n> \"Bidirectional logical replication\".\n\nModified\n\n> 7b. Don't put \":\" in the title.\n\nModified\n\n> ~~~\n>\n> 8. doc/src/sgml/logical-replication.sgml - missing intro\n>\n> IMO this page needs some sort of introduction/blurb instead of leaping\n> straight into examples without any preamble text to give context.\n\nModified\n\n> ~~~\n>\n> 9. doc/src/sgml/logical-replication.sgml - bullets\n>\n> Suggest removing all the bullets from the example steps. (I had\n> something similar a while ago for the \"Row Filter\" page but\n> eventually, they all had to be removed).\n\nModified\n\n> ~~~\n>\n> 10. doc/src/sgml/logical-replication.sgml - SQL too long\n>\n> Many of the example commands are much too long, and for me, they are\n> giving scroll bars when rendered. It would be better if they can be\n> wrapped in appropriate places so easier to read (and no resulting\n> scroll bars).\n\nModified\n\n> ~~~\n>\n> 11. doc/src/sgml/logical-replication.sgml - add the psql prompt\n>\n> IMO it would also be easier to understand the examples if you show the\n> psql prompt. Then you can easily know the node context without having\n> to describe it in the text so often.\n>\n> e.g.\n>\n> + <para>\n> + Create the subscription in node2 to subscribe the changes from node1:\n> +<programlisting>\n> +CREATE SUBSCRIPTION sub_node1_node2 ...\n>\n> SUGGGESTION\n> + <para>\n> + Create the subscription in node2 to subscribe the changes from node1\n> +<programlisting>\n> +node_2=# CREATE SUBSCRIPTION sub_node1_node2 ...\n\nModified\n\n> ~~~\n>\n> 12. doc/src/sgml/logical-replication.sgml - typo\n>\n> + <para>\n> + Now the BiDirectional logical replication setup is complete between node1\n>\n> typo \"BiDirectional\"\n\nModified to bidirectional\n\n> ~~~\n>\n> 13. doc/src/sgml/logical-replication.sgml - deep levels\n>\n> The section levels are very deep, but > 3 will not appear in the table\n> of contents when rendered. Maybe you can rearrange to raise them all\n> up one level, then IMO the TOC will work better and the whole page\n> will be easier to read.\n\nModified\n\n> ~~~\n>\n> 14. doc/src/sgml/logical-replication.sgml - unnecessarily complex?\n>\n> +<programlisting>\n> +# Truncate the table data but do not replicate the truncate.\n> +ALTER PUBLICATION pub_node3 SET (publish='insert,update,delete');\n> +TRUNCATE t1;\n> +\n> +CREATE SUBSCRIPTION sub_node3_node1 CONNECTION 'dbname=foo host=node1\n> user=repuser' PUBLICATION pub_node1 WITH (copy_data = force,\n> local_only = on);\n> +\n> +CREATE SUBSCRIPTION sub_node3_node2 CONNECTION 'dbname=foo host=node2\n> user=repuser' PUBLICATION pub_node2 WITH (copy_data = off, local_only\n> = on);\n> +\n> +# Include truncate operations from now\n> +ALTER PUBLICATION pub_node3 SET (publish='insert,update,delete,truncate');\n> +</programlisting>\n>\n> Is it really necessary for those CREATE SUBSCRIPTION to be placed\n> where they are? I felt it would be less confusing if you just do the\n> TRUNCATE between the 2x ALTER PUBLICATION... and then do those CREATE\n> SUBSCRIPTION separately later.\n\nModified\n\n> ~~~\n>\n> 14. doc/src/sgml/logical-replication.sgml - force?\n>\n> There seems no documentation anywhere that says what is the\n> purpose/meaning of the copy_data enum \"force\".\n\nThis is added in create subscription notes\n\n> ~~~\n>\n> 15. doc/src/sgml/ref/alter_subscription.sgml - force?\n>\n> Meanings of copy_data true/false are self-evident but should something\n> here be explaining what is the meaning of \"force\"? Or a reference to\n> some place that explains it?\n\nThis is mentioned in create subscription notes, added a reference to\ncreate subscription notes.\n\n> ~~~\n>\n> 16. doc/src/sgml/ref/create_subscription.sgml - local_only missing notes?\n>\n> @@ -161,6 +161,11 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> publisher node changes regardless of their origin. The default is\n> <literal>false</literal>.\n> </para>\n> + <para>\n> + There is some interation between the \"local_only\" option and\n> + \"copy_data\" option. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for details.\n> + </para>\n> </listitem>\n> </varlistentry>\n>\n> 16a. Typo \"interation\"\n\nModified\n\n> 16b. But where in the referenced notes does it actually say anything about this?\n\nModified\n\n> ~~~\n>\n> 17. doc/src/sgml/ref/create_subscription.sgml - force?\n>\n> The notes about the copy_data do not say anything about what the\n> values mean. Specifically, there seems nothing that describes what is\n> the meaning of \"force\".\n\nI have mentioned this in the notes section.\n\n> ~~~\n>\n> 18. doc/src/sgml/ref/create_subscription.sgml - copy_data missing notes?\n>\n> +\n> + <para>\n> + There is some interation between the \"local_only\" option and\n> + \"copy_data\" option. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for details.\n> + </para>\n>\n> 18a. Typo \"interation\"\n\nModified\n\n> 18b. But where in the referenced notes does it actually say anything about this?\n\nAdded\n\n> ~~~\n>\n> 19. doc/src/sgml/ref/create_subscription.sgml - xref to the new page\n>\n> Shouldn't there be an xref on this page somewhere to your other new\n> \"Bidirectional\" docs page?\n\nAdded reference\n\n> ~~~\n>\n> 20. src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_ON_OR_FORCE\n>\n> @@ -69,6 +69,18 @@\n> /* check if the 'val' has 'bits' set */\n> #define IsSet(val, bits) (((val) & (bits)) == (bits))\n>\n> +#define IS_COPY_DATA_ON_OR_FORCE(copy_data) (copy_data != COPY_DATA_OFF)\n>\n> Maybe this would be better as a static inline function?\n\n What is the advantage of doing this change? I have not changed this\nas the macro usage is fine.\nThoughts?\n\n> ~~~\n>\n> 21. src/backend/commands/subscriptioncmds.c - comment\n>\n> +/*\n> + * Represents whether copy_data option is specified with on, off or force.\n> + */\n> +typedef enum CopyData\n> +{\n> + COPY_DATA_OFF,\n> + COPY_DATA_ON,\n> + COPY_DATA_FORCE\n> +} CopyData;\n>\n> I felt it might be better if the comment described these enums in the\n> same order they are defined.\n>\n> E.g. \"Represents whether copy_data option is specified as off/on, or force.\"\n\nModified\n\n> ~~~\n>\n> 22. src/backend/commands/subscriptioncmds.c - CreateSubscription bug?\n>\n> @@ -720,7 +788,8 @@ CreateSubscription(ParseState *pstate,\n> CreateSubscriptionStmt *stmt,\n> * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\n> * PUBLICATION to work.\n> */\n> - if (opts.twophase && !opts.copy_data && tables != NIL)\n> + if (opts.twophase && IS_COPY_DATA_ON_OR_FORCE(opts.copy_data)\n> + && tables != NIL)\n> twophase_enabled = true;\n> Is that a bug? It does not seem to match the previous code. Should\n> that IS_COPY_DATA_ON_OR_FORCE be \"not\" ?\n\nModified\n\n> ~~~\n>\n> 23. src/backend/commands/subscriptioncmds.c - long errmsg\n>\n> + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\n> + ereport(ERROR,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"CREATE/ALTER SUBSCRIPTION with local_only and copy_data as\n> true is not allowed when the publisher might have replicated data,\n> table:%s.%s might have replicated data in the publisher\",\n> + nspname, relname),\n> + errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\"));\n> +\n>\n> The errmsg seems way too long for the source code. Can you use string\n> concatenation or continuation chars to wrap the message over multiple\n> lines?\n\nI had seen that the long error message elsewhere also is in a single\nline. I think we should keep it as it is to maintain the coding\nstandard.\nThoughts?\n\n> ~~~\n>\n> 24. src/test/regress/sql/subscription.sql - typo\n>\n> @@ -66,6 +69,15 @@ ALTER SUBSCRIPTION regress_testsub4 SET (local_only = false);\n> DROP SUBSCRIPTION regress_testsub3;\n> DROP SUBSCRIPTION regress_testsub4;\n>\n> +-- ok - valid coy_data options\n>\n> Typo \"coy_data\". (it looks like this typo is not caused by this patch,\n> but I think this patch should fix it anyhow).\n\nModified\n\n> ~~~\n>\n> 25. src/test/regress/sql/subscription.sql - test order\n>\n> The new tests are OK but IMO they could be re-ordered so then they\n> will be more consistent for the positive and negative tests.\n>\n> CURRENT\n> \"true/force/on/1\" and \"off/false/0\"\n>\n> SUGGEST\n> \"true/on/1/force\" and \"false/off/0\"\n\nModified\n\nThanks for the comments, the v10 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm0PmOz71O6ofhZkB0rts5Ak2HUhMuuMQoViH_LAXTBeBw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 22 Apr 2022 21:53:48 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Apr 20, 2022 at 11:19 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Below are my review comments for the v9-0002 patch (TAP test part only).\n>\n> (The order of my comments is a bit muddled because I jumped around in\n> the file a bit while reviewing it).\n>\n> ======\n>\n> 1. create_subscription - missing comment.\n>\n> +sub create_subscription\n> +{\n> + my ($node_subscriber, $node_publisher, $sub_name, $node_connstr,\n> + $application_name, $pub_name, $copy_data_val)\n> + = @_;\n>\n> Maybe add a comment for this subroutine and describe the expected parameters.\n\nAdded\n\n> ~~\n>\n> 2. create_subscription - hides the options\n>\n> IMO the \"create_subscription\" subroutine is hiding too many of the\n> details, so now it is less clear (from the caller's POV) what the test\n> is really doing. Perhaps the options could be passed more explicitly\n> to this subroutine.\n>\n> e.g. Instead of\n>\n> create_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\n> $appname_B1, 'tap_pub_A', 'on');\n>\n> perhaps explicitly set the WITH options like:\n>\n> create_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\n> $appname_B1, 'tap_pub_A', 'local_only = on, copy_data = on');\n\nModified\n\n> ~~~\n>\n> 3. the application names are confusing\n>\n> +my $appname_A1 = 'tap_sub_A_B';\n> +my $appname_A2 = 'tap_sub_A_C';\n> +my $appname_B1 = 'tap_sub_B_A';\n> +my $appname_B2 = 'tap_sub_B_C';\n> +my $appname_C1 = 'tap_sub_C_A';\n> +my $appname_C2 = 'tap_sub_C_B';\n>\n> I found the application names using '1' and '2' to be a bit confusing.\n> i.e it was unnecessarily hard to associate them (in my head) with\n> their relevant subscriptions. IMO it would be easier to name them\n> using letters like below:\n>\n> SUGGESTED\n> my $appname_AB = 'tap_sub_A_B';\n> my $appname_AC = 'tap_sub_A_C';\n> my $appname_BA = 'tap_sub_B_A';\n> my $appname_BC = 'tap_sub_B_C';\n> my $appname_CA = 'tap_sub_C_A';\n> my $appname_CB = 'tap_sub_C_B';\n\nRemoved appname and used subscription names for application name\n\n> ~~~\n>\n> 4. create_subscription - passing the $appname 2x.\n>\n> +create_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\n> + $appname_B1, 'tap_pub_A', 'on');\n>\n>\n> It seemed confusing that the $app_name is passed twice.\n>\n> IMO should rename all those $appname_XX vars (see previous comment) to\n> be $subname_XX, and just pass that create_subscription instead.\n>\n> my $subname_AB = 'tap_sub_A_B';\n> my $subname_AC = 'tap_sub_A_C';\n> my $subname_BA = 'tap_sub_B_A';\n> my $subname_BC = 'tap_sub_B_C';\n> my $subname_CA = 'tap_sub_C_A';\n> my $subname_CB = 'tap_sub_C_B';\n>\n> Then create_subscription subroutine should have one less argument.\n> Just add a comment saying that the appname is always assigned the same\n> value as the subname.\n\nModifeid\n\n> ~~~\n>\n> 5. cleanup part - seems a bit misleading\n>\n> +# cleanup\n> +$node_B->safe_psql(\n> + 'postgres', \"\n> + DROP SUBSCRIPTION $appname_B2\");\n> +$node_C->safe_psql(\n> + 'postgres', \"\n> + DELETE FROM tab_full\");\n> +$node_B->safe_psql(\n> + 'postgres', \"\n> + DELETE FROM tab_full where a = 13\");\n>\n> Is that comment misleading? IIUC this is not really cleaning up\n> everything. It is just a cleanup of the previous Node_C test part\n> isn't it?\n\nModified to clear the operations done by this test\n\n> ~~~\n>\n> 6. Error case (when copy_data is true)\n>\n> +# Error when creating susbcription with local_only and copy_data as true when\n> +# the publisher has replicated data\n>\n> 6a. Typo \"susbcription\"\n\nModified\n\n> 6b. That comment maybe needs some more explanation - eg. say that\n> since Node_A is already subscribing to Node_B so when Node_B makes\n> another subscription to Node_A the copy doesn't really know if the\n> data really originated from Node_A or not...\n\nSlightly reworded and modified\n\n> 6c. Maybe the comment needed to be more like ############## style to\n> denote this (and the one that follows) is a separate test case.\n\nModified\n\n> ~~~\n>\n> 7. Error case (when copy_data is force)\n>\n> +# Creating subscription with local_only and copy_data as force should be\n> +# successful when the publisher has replicated data\n>\n> 7a. Same typo \"subscription\"\n\nI felt subscription is correct in this case, made no change for this\n\n> ~~~\n>\n> 8. Add 3rd node when the existing node has some data\n>\n> 8a. Maybe that comment needs to be expanded more. This isn't just\n> \"adding a node\" - you are joining it to the others as another\n> bi-directional participant in a 3-way group. And the comment should\n> clarify exactly which nodes have the initial data. Both the existing 2\n> you are joining to? The new one only? All of them?\n\nModified\n\n> 8b. IMO is would be much clearer to do SELECT from all the nodes at\n> the start of this test just to re-confirm what is all the initial data\n> on these nodes before joining the 3rd node.\n\nModified\n\n> NOTE - These same review comments apply to the other test combinations too\n> - Add 3rd node when the existing node has no data\n> - Add 3rd node when the new node has some data\n\nModified\n\n> ~~~\n>\n> 9. Inserted data\n>\n> +# insert some data in all the nodes\n> +$node_A->safe_psql('postgres', \"INSERT INTO tab_full VALUES (13);\");\n> +$node_B->safe_psql('postgres', \"INSERT INTO tab_full VALUES (21);\");\n> +$node_C->safe_psql('postgres', \"INSERT INTO tab_full VALUES (31);\");\n>\n> They seemed strange values (13, 21, 31) to add. I couldn't work out\n> the \"meaning\" of them.\n>\n> Wouldn't values like 13, 23, 33 make more sense (e.g pattern is 10 x\n> node# + something)\n\nModified\n\n> ~~~\n>\n> 10. verify_data($node_A, $node_B, $node_C);\n>\n> All the expected values are too buried in this subroutine which makes\n> it hard to read. I think it would be easier to read (from the caller's\n> POV) if you can pass the *expected* values as another arg into the\n> verify_data subroutine.\n>\n> e.g. is something like this possible?\n> verify_data($node_A, $node_B, $node_C, [11,12,13])\n\nModified\n\n> ~~~\n>\n> 11. ALTER for ignoring 'truncate'\n>\n> +$node_C->safe_psql('postgres',\n> + \"ALTER PUBLICATION tap_pub_C SET (publish='insert,update,delete');\");\n> +\n> +$node_C->safe_psql('postgres', \"TRUNCATE tab_full\");\n> +\n> +create_subscription($node_C, $node_A, $appname_C1, $node_A_connstr,\n> + $appname_C1, 'tap_pub_A', 'force');\n> +create_subscription($node_C, $node_B, $appname_C2, $node_B_connstr,\n> + $appname_C2, 'tap_pub_B', 'off');\n> +\n> +#include truncates now\n> +$node_C->safe_psql('postgres',\n> + \"ALTER PUBLICATION tap_pub_C SET\n> (publish='insert,update,delete,truncate');\"\n> +);\n>\n> Do those create_subscription calls need to be encapsulated by the\n> ALTER like that? I did not think so. It looks a bit more complex if\n> done this way.\n\nModified\n\n> ~~~\n>\n> 12. clean_subscriber_contents - misnamed?\n>\n> This subroutine feels a bit misnamed. It seems to be doing lots of\n> things like detaching the Node_C and deleting all table data from all\n> nodes. That all seems quite different from just \"clean subscriber\n> contents\".\n\nchanged it to detach_node_clean_table_data\n\n> ~~~\n>\n> 13. table initial data?\n>\n> +clean_subscriber_contents($node_A, $node_B, $node_C);\n> +\n> +##########################################################################\n> +# Add 3rd node when the new node has some data\n> +##########################################################################\n>\n> But does this test case *really* have some data? I am not so sure.\n> Doesn't the preceding \"clean_subscriber_contents\" call remove all the\n> data that might have been there? That is why I think all the tests out\n> to have SELECT (previous comment #8) so they can re-confirm what data\n> is really in those tables before doing each test part.\n\nModified\n\nThanks for the comments, the v10 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm0PmOz71O6ofhZkB0rts5Ak2HUhMuuMQoViH_LAXTBeBw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 22 Apr 2022 21:58:22 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for v10-0001 and v10-0002.\n\n(I did not yet look at the v10-0002 TAP tests. I will check those\nseparately later).\n\n=================\nv10-0001 comments\n=================\n\n1.1 src/include/catalog/pg_subscription.h\n\n@@ -70,6 +70,8 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\nBKI_SHARED_RELATION BKI_ROW\n\n bool substream; /* Stream in-progress transactions. */\n\n+ bool sublocalonly; /* skip copying of remote origin data */\n+\n char subtwophasestate; /* Stream two-phase transactions */\n\n bool subdisableonerr; /* True if a worker error should cause the\n\n1.1.a Comment should start uppercase.\n\n1.1.b Perhaps it is better to say \"Skip replication of\" instead of\n\"Skip copying\" ?\n\n~~~\n\n1.2 src/include/catalog/pg_subscription.h\n\n@@ -110,6 +112,7 @@ typedef struct Subscription\n bool binary; /* Indicates if the subscription wants data in\n * binary format */\n bool stream; /* Allow streaming in-progress transactions. */\n+ bool local_only; /* Skip copying of remote origin data */\n char twophasestate; /* Allow streaming two-phase transactions */\n\nSame comment as #1.1\n\n=================\nv10-0002 comments\n=================\n\n2.1 Commit message\n\nb) user wil have to create another subscription subscribing from\nNode2 to Node3 using local_only option and copy_data as true.\n\nTypo: \"wil\"\n\n~~~\n\n2.2 Commit message\n\nHere when user has specified local_only 'on' which indicates that the\npublisher should only replicate the changes that are generated locally, but in\nthis case since the publisher node is also subscribing data from other nodes,\nthe publisher node can have remotely originated data, so throw an error in this\ncase to prevent remotely generated data being replicated to the subscriber. If\nuser still intends to continue with the operation user can specify copy_data\nas 'force' and proceed.\n\nSUGGESTED (minor rewording)\nIn the scenario above the user has specified local_only 'on' (which\nindicates that the publisher should only replicate the changes that\nare generated locally), but in this case, the publisher node is also\nsubscribing data from other nodes, so the publisher node may have\nremotely originated data. We throw an error, in this case, to draw\nattention to there being possible remote data. If the user still\nwishes to continue with the operation user can specify copy_data as\n'force' and proceed.\n\n~~~\n\n2.3 doc/src/sgml/logical-replication.sgml\n\n+ <para>\n+ Bidirectional replication is useful in creating multi master database\n+ which helps in performing read/write operations from any of the nodes.\n+ Setting up bidirectional logical replication between two nodes requires\n+ creation of the publication in all the nodes, creating subscription in\n+ each of the nodes that subcribes to data from all the nodes. The steps\n+ to create two node bidirectional replication is listed below:\n+ </para>\n\n2.3.a Typo: \"subcribes\"\n\n2.3.b Wording: \"creating subscription” -> \"and creating subscriptions...\"\n\n2.3.c Wording: \"The steps to create two node bidirectional replication\nis listed below:\" -> \"The steps to create a two-node bidirectional\nreplication are given below:\"\n\n~~~\n\n2.4 doc/src/sgml/logical-replication.sgml\n\n+<programlisting>\n+node2=# CREATE SUBSCRIPTION sub_node1_node2\n+node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\n+node2=# PUBLICATION pub_node1\n+node2=# WITH (copy_data = off, local_only = on);\n+CREATE SUBSCRIPTION\n+</programlisting>\n\nI am not sure if those psql continuation prompts are right. Shouldn't\nthey be \"node2-#\" instead of \"node2=#\"?\n\n~~~\n\n2.5 doc/src/sgml/logical-replication.sgml\n\n+<programlisting>\n+node2=# CREATE SUBSCRIPTION sub_node1_node2\n+node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\n+node2=# PUBLICATION pub_node1\n+node2=# WITH (copy_data = off, local_only = on);\n+CREATE SUBSCRIPTION\n+</programlisting>\n+ </para>\n\nIIUC the closing </para> should be on the same line as the\n</programlisting>. I recall there was some recent github push about\nthis sometime in the last month - maybe you can check to confirm it.\n\n~~~\n\n2.6 doc/src/sgml/logical-replication.sgml\n\n+ <para>\n+ Create the subscription in node2 to subscribe the changes from node1:\n+<programlisting>\n+node2=# CREATE SUBSCRIPTION sub_node1_node2\n\nIMO the naming convention here is backwards/confusing. E.g. The\n\"pub_node1\" is the publisher at node1. So similarly, I think the\nsubscriber at node2 should be called \"sub_node2_node1\" (not\n\"sub_node1_node2\")\n\n~~~\n\n2.7 doc/src/sgml/logical-replication.sgml\n\n+ <para>\n+ Adding a new node node3 to the existing node1 and node2 requires setting\n+ up subscription in node1 and node2 to replicate the data from node3 and\n+ setting up subscription in node3 to replicate data from node1 and node2.\n+ The steps for the same is listed below:\n+ </para>\n\nThere are several sections that say \"The steps for the same is listed\nbelow:\". The sentence for all of them seemed redundant to me.\n\n~~~\n\n2.8 doc/src/sgml/logical-replication.sgml\n\n+ <para>\n+ Adding a new node node3 to the existing node1 and node2 when data is\n+ present in existing nodes node1 and node2 needs similar steps, only change\n+ required here is that node3 should create subscription with copy_data as\n+ force to one of the existing nodes to receive the existing data during\n+ initial data synchronization. The steps for the same is listed below:\n+ </para>\n\nI thought it should be 2 sentences. So \"needs similar steps, only\nchange required here\" -> \"needs similar steps. The only change\nrequired here...\"\n\n~~~\n\n2.9 doc/src/sgml/logical-replication.sgml\n\nI think every time you say option names or values like \"copy_data\" in\nthe text they should be using <literal> font.\n\n~~~\n\n2.10 doc/src/sgml/logical-replication.sgml\n\n+ <para>\n+ Adding a new node node3 to the existing node1 and node2 when data is\n+ present in the new node node3 needs similar steps, few changes are\n+ required here to get the existing data from node3 to node1 and node2 and\n+ later cleaning up of data in node3 before synchronization of all the data\n+ from the existing nodes. The steps for the same is listed below:\n+ </para>\n\nI thought it should be 2 sentences. So \"needs similar steps, few\nchanges are required here\" -> \"needs similar steps. A few changes are\nrequired here...\"\n\n~~~\n\n2.11 doc/src/sgml/logical-replication.sgml\n\nAll the text that says \"create subscription\" and \"create publication\"\nmaybe should be change to \"create a subscription\" and \"create a\npublication\" etc.\n\n~~~\n\n2.12 doc/src/sgml/ref/alter_subscription.sgml - copy_data\n\n+ <para>\n+ There is some interaction between the \"local_only\" option and\n+ \"copy_data\" option. Refer to the\n+ <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n+ details and usage of force for copy_data option.\n </para>\n\n2.12.a Maybe options should not be in quotes like that. I think they\nshould be <literal> font.\n\n2.12.b It is a bit misleading because there is no \"Notes\" section here\non this page. Maybe it should say refer to the CREATE SUBSCRIPTION\nNotes.\n\n2.12.c The last copy_data should also be <literal> font.\n\n~~~\n\n2.13 doc/src/sgml/ref/create_subscription.sgml - local_only\n\n </para>\n+ <para>\n+ There is some interaction between the \"local_only\" option and\n+ \"copy_data\" option. Refer to the\n+ <xref linkend=\"sql-createsubscription-notes\" /> for details.\n+ </para>\n </listitem>\n\n2.13.a Maybe options should not be in quotes like that. I think they\nshould be <literal> font.\n\n2.13.b The last copy_data should also be <literal> font.\n\n~~~\n\n2.14 doc/src/sgml/ref/create_subscription.sgml - copy_data\n\n\n <para>\n Specifies whether to copy pre-existing data in the publications\n- that are being subscribed to when the replication starts.\n- The default is <literal>true</literal>.\n+ that are being subscribed to when the replication starts. This\n+ parameter may be either <literal>true</literal>,\n+ <literal>false</literal> or <literal>force</literal>. The default is\n+ <literal>true</literal>.\n </para>\n\n2.14.a Maybe options should not be in quotes like that. I think they\nshould be <literal> font.\n\n2.14.b The last copy_data should also be <literal> font.\n\n~~~\n\n2.15 doc/src/sgml/ref/create_subscription.sgml\n\n@@ -374,6 +388,16 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n can have non-existent publications.\n </para>\n\n+ <para>\n+ If subscription is created with local_only as 'on' and copy_data as 'on', it\n+ will check if the publisher tables are being subscribed to any other\n+ publisher and throw an error to prevent inconsistent data in the\n+ subscription. User can continue with the copy operation without throwing any\n+ error in this case by specifying copy_data as 'force'. Refer to the\n+ <xref linkend=\"bidirectional-logical-replication\"/> on how\n+ copy_data and local_only can be used in bidirectional replication.\n+ </para>\n\n2.15.a I think all those mentions of copy_data and local_only and\ntheir values should be in <literal> font instead of in quotes.\n\n2.15.b Wording: \"User can\" -> \"The user can\"\n\n~~~\n\n2.16 src/backend/commands/subscriptioncmds.c - macro\n\n> 20. src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_ON_OR_FORCE\n>\n> @@ -69,6 +69,18 @@\n> /* check if the 'val' has 'bits' set */\n> #define IsSet(val, bits) (((val) & (bits)) == (bits))\n>\n> +#define IS_COPY_DATA_ON_OR_FORCE(copy_data) (copy_data != COPY_DATA_OFF)\n>\n> Maybe this would be better as a static inline function?\n\nVignesh: What is the advantage of doing this change? I have not changed this\nas the macro usage is fine. Thoughts?\n\nOriginally I was going to suggest the macro should use extra parens\nlike ((copy_data) != COPY_DATA_OFF), but then I thought if it was a\nfunction then it would have enum type-checking which would be better.\nIf you want to keep the macro then please make the parens change.\n\n~~~\n\n2.17 src/backend/commands/subscriptioncmds.c - DefGetCopyData\n\n+ /*\n+ * The set of strings accepted here should match up with\n+ * the grammar's opt_boolean_or_string production.\n+ */\n+ if (pg_strcasecmp(sval, \"true\") == 0 ||\n+ pg_strcasecmp(sval, \"on\") == 0)\n+ return COPY_DATA_ON;\n+ if (pg_strcasecmp(sval, \"false\") == 0 ||\n+ pg_strcasecmp(sval, \"off\") == 0)\n+ return COPY_DATA_OFF;\n+ if (pg_strcasecmp(sval, \"force\") == 0)\n+ return COPY_DATA_FORCE;\n\nI think you can change the order of these to be off/on/force, so then\nthe order is consistent with the T_Integer case.\n\n~~~\n\n2.18 src/backend/commands/subscriptioncmds.c - DefGetCopyData\n\n+ ereport(ERROR,\n+ errcode(ERRCODE_SYNTAX_ERROR),\n+ errmsg(\"%s requires a boolean or \\\"force\\\"\", def->defname));\n+ return COPY_DATA_OFF; /*\nkeep compiler quiet */\n\nExcessive comment indent? Has pg_indent been run?\n\n~~~\n\n2.19 src/backend/commands/subscriptioncmds.c - AlterSubscription\n\n@@ -1236,7 +1306,8 @@ AlterSubscription(ParseState *pstate,\nAlterSubscriptionStmt *stmt,\n errmsg(\"ALTER SUBSCRIPTION ... REFRESH is not allowed for disabled\nsubscriptions\")));\n\n parse_subscription_options(pstate, stmt->options,\n- SUBOPT_COPY_DATA, &opts);\n+ SUBOPT_COPY_DATA,\n+ &opts);\n\nThis is a formatting change only. Maybe it does not belong in this\npatch unless it is the result of pg_indent.\n\n~~~\n\n2.20 src/backend/commands/subscriptioncmds.c - fetch_table_list\n\n> 23. src/backend/commands/subscriptioncmds.c - long errmsg\n>\n> + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\n> + ereport(ERROR,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"CREATE/ALTER SUBSCRIPTION with local_only and copy_data as\n> true is not allowed when the publisher might have replicated data,\n> table:%s.%s might have replicated data in the publisher\",\n> + nspname, relname),\n> + errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\"));\n> +\n>\n> The errmsg seems way too long for the source code. Can you use string\n> concatenation or continuation chars to wrap the message over multiple\n> lines?\n\nVignesh: I had seen that the long error message elsewhere also is in a\nsingle line. I think we should keep it as it is to maintain the coding\nstandard. Thoughts?\n\nOK, if you say it is already common practice then it's fine by me to\nleave it as-is.\n\n~~~\n\n2.21 src/test/regress/expected/subscription.out - make check\n\nmake check fails.\n 1 of 214 tests failed.\n\n2.21.a It looks like maybe you did not update the expected ordering of\nsome of the tests, after some minor adjustments in subscriprion.sql in\nv10. So the expected output needs to be fixed in the patch.\n\n2.21.b. Suggest adding this patch to CF so that the cfbot can pick up\nsuch test problems earlier.\n\n\n------\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 27 Apr 2022 15:45:28 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for v10-0002 (TAP tests part only)\n\nFIle: src/test/subscription/t/032_localonly.pl\n\n======\n\n1.\n\n+# Detach node C and clean the table contents.\n+sub detach_node_clean_table_data\n+{\n\n1a. Maybe say \"Detach node C from the node-group of (A, B, C) and\nclean the table contents from all nodes\"\n\n1b. Actually wondered do you need to TRUNCATE from both A and B (maybe\nonly truncate 1 is OK since those nodes are still using MMC). OTOH\nmaybe your explicit way makes the test simpler.\n\n~~~\n\n2.\n\n+# Subroutine for verify the data is replicated successfully.\n+sub verify_data\n+{\n\n2a. Typo: \"for verify\" -> \"to verify\"\n\n2b. The messages in this function maybe are not very appropriate. They\nsay 'Inserted successfully without leading to infinite recursion in\ncircular replication setup', but really the function is only testing\nall the data is the same as 'expected'. So it could be the result of\nany operation - not just Insert.\n\n~~~\n\n3. SELECT ORDER BY?\n\n$result = $node_B->safe_psql('postgres', \"SELECT * FROM tab_full;\");\nis($result, qq(11\n12\n13),\n 'Node_C data replicated to Node_B'\n);\n\nI am not sure are these OK like this or if *every* SELECT use ORDER BY\nto make sure the data is in the same qq expected order? There are\nmultiple cases like this.\n\n(BTW, I think this comment needs to be applied for the v10-0001 patch,\nmaybe not v10-0002).\n\n~~~\n\n4.\n\n+###############################################################################\n+# Specifying local_only 'on' which indicates that the publisher should only\n+# replicate the changes that are generated locally from node_B, but in\n+# this case since the node_B is also subscribing data from node_A, node_B can\n+# have data originated from node_A, so throw an error in this case to prevent\n+# node_A data being replicated to the node_C.\n+###############################################################################\n\nThere is something wrong with the description because there is no\n\"node_C\" in this test. You are just creating a 2nd subscription on\nnode A.\n\n~~\n\n5.\n\n+($result, $stdout, $stderr) = $node_A->psql(\n+ 'postgres', \"\n+ CREATE SUBSCRIPTION tap_sub_A3\n+ CONNECTION '$node_B_connstr application_name=$subname_AB'\n+ PUBLICATION tap_pub_B\n+ WITH (local_only = on, copy_data = on)\");\n+like(\n\nIt seemed strange to call this 2nd subscription \"tap_sub_A3\". Wouldn't\nit be better to call it \"tap_sub_A2\"?\n\n~~~\n\n6.\n\n+###############################################################################\n+# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\n+# replication setup when the existing nodes (node_A & node_B) has no data and\n+# the new node (node_C) some pre-existing data.\n+###############################################################################\n+$node_C->safe_psql('postgres', \"INSERT INTO tab_full VALUES (31);\");\n+\n+$result = $node_A->safe_psql('postgres', \"SELECT * FROM tab_full order by 1;\");\n+is( $result, qq(), 'Check existing data');\n+\n+$result = $node_B->safe_psql('postgres', \"SELECT * FROM tab_full order by 1;\");\n+is( $result, qq(), 'Check existing data');\n+\n+$result = $node_C->safe_psql('postgres', \"SELECT * FROM tab_full order by 1;\");\n+is($result, qq(31), 'Check existing data');\n+\n+create_subscription($node_A, $node_C, $subname_AC, $node_C_connstr,\n+ 'tap_pub_C', 'copy_data = force, local_only = on');\n+create_subscription($node_B, $node_C, $subname_BC, $node_C_connstr,\n+ 'tap_pub_C', 'copy_data = force, local_only = on');\n+\n\nBecause the Node_C does not yet have any subscriptions aren't these\ncases where you didn't really need to use \"force\"?\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 28 Apr 2022 12:27:15 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Apr 27, 2022 at 11:15 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for v10-0001 and v10-0002.\n>\n> (I did not yet look at the v10-0002 TAP tests. I will check those\n> separately later).\n>\n> =================\n> v10-0001 comments\n> =================\n>\n> 1.1 src/include/catalog/pg_subscription.h\n>\n> @@ -70,6 +70,8 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\n> BKI_SHARED_RELATION BKI_ROW\n>\n> bool substream; /* Stream in-progress transactions. */\n>\n> + bool sublocalonly; /* skip copying of remote origin data */\n> +\n> char subtwophasestate; /* Stream two-phase transactions */\n>\n> bool subdisableonerr; /* True if a worker error should cause the\n>\n> 1.1.a Comment should start uppercase.\n\nModified\n\n> 1.1.b Perhaps it is better to say \"Skip replication of\" instead of\n> \"Skip copying\" ?\n\nModified\n\n> ~~~\n>\n> 1.2 src/include/catalog/pg_subscription.h\n>\n> @@ -110,6 +112,7 @@ typedef struct Subscription\n> bool binary; /* Indicates if the subscription wants data in\n> * binary format */\n> bool stream; /* Allow streaming in-progress transactions. */\n> + bool local_only; /* Skip copying of remote origin data */\n> char twophasestate; /* Allow streaming two-phase transactions */\n>\n> Same comment as #1.1\n\nModified\n\n> =================\n> v10-0002 comments\n> =================\n>\n> 2.1 Commit message\n>\n> b) user wil have to create another subscription subscribing from\n> Node2 to Node3 using local_only option and copy_data as true.\n>\n> Typo: \"wil\"\n\nModified\n\n> ~~~\n>\n> 2.2 Commit message\n>\n> Here when user has specified local_only 'on' which indicates that the\n> publisher should only replicate the changes that are generated locally, but in\n> this case since the publisher node is also subscribing data from other nodes,\n> the publisher node can have remotely originated data, so throw an error in this\n> case to prevent remotely generated data being replicated to the subscriber. If\n> user still intends to continue with the operation user can specify copy_data\n> as 'force' and proceed.\n>\n> SUGGESTED (minor rewording)\n> In the scenario above the user has specified local_only 'on' (which\n> indicates that the publisher should only replicate the changes that\n> are generated locally), but in this case, the publisher node is also\n> subscribing data from other nodes, so the publisher node may have\n> remotely originated data. We throw an error, in this case, to draw\n> attention to there being possible remote data. If the user still\n> wishes to continue with the operation user can specify copy_data as\n> 'force' and proceed.\n\nModified\n\n> ~~~\n>\n> 2.3 doc/src/sgml/logical-replication.sgml\n>\n> + <para>\n> + Bidirectional replication is useful in creating multi master database\n> + which helps in performing read/write operations from any of the nodes.\n> + Setting up bidirectional logical replication between two nodes requires\n> + creation of the publication in all the nodes, creating subscription in\n> + each of the nodes that subcribes to data from all the nodes. The steps\n> + to create two node bidirectional replication is listed below:\n> + </para>\n>\n> 2.3.a Typo: \"subcribes\"\n\nModified\n\n> 2.3.b Wording: \"creating subscription” -> \"and creating subscriptions...\"\n\nModified\n\n> 2.3.c Wording: \"The steps to create two node bidirectional replication\n> is listed below:\" -> \"The steps to create a two-node bidirectional\n> replication are given below:\"\n\nModified\n\n> ~~~\n>\n> 2.4 doc/src/sgml/logical-replication.sgml\n>\n> +<programlisting>\n> +node2=# CREATE SUBSCRIPTION sub_node1_node2\n> +node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\n> +node2=# PUBLICATION pub_node1\n> +node2=# WITH (copy_data = off, local_only = on);\n> +CREATE SUBSCRIPTION\n> +</programlisting>\n>\n> I am not sure if those psql continuation prompts are right. Shouldn't\n> they be \"node2-#\" instead of \"node2=#\"?\n\nModified\n\n> ~~~\n>\n> 2.5 doc/src/sgml/logical-replication.sgml\n>\n> +<programlisting>\n> +node2=# CREATE SUBSCRIPTION sub_node1_node2\n> +node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\n> +node2=# PUBLICATION pub_node1\n> +node2=# WITH (copy_data = off, local_only = on);\n> +CREATE SUBSCRIPTION\n> +</programlisting>\n> + </para>\n>\n> IIUC the closing </para> should be on the same line as the\n> </programlisting>. I recall there was some recent github push about\n> this sometime in the last month - maybe you can check to confirm it.\n\n I have seen the programlisting across multiple places and noticed\nthat there is no such guideline being followed. I have not made any\nchange for this comment.\n\n> ~~~\n>\n> 2.6 doc/src/sgml/logical-replication.sgml\n>\n> + <para>\n> + Create the subscription in node2 to subscribe the changes from node1:\n> +<programlisting>\n> +node2=# CREATE SUBSCRIPTION sub_node1_node2\n>\n> IMO the naming convention here is backwards/confusing. E.g. The\n> \"pub_node1\" is the publisher at node1. So similarly, I think the\n> subscriber at node2 should be called \"sub_node2_node1\" (not\n> \"sub_node1_node2\")\n\nModified\n\n> ~~~\n>\n> 2.7 doc/src/sgml/logical-replication.sgml\n>\n> + <para>\n> + Adding a new node node3 to the existing node1 and node2 requires setting\n> + up subscription in node1 and node2 to replicate the data from node3 and\n> + setting up subscription in node3 to replicate data from node1 and node2.\n> + The steps for the same is listed below:\n> + </para>\n>\n> There are several sections that say \"The steps for the same is listed\n> below:\". The sentence for all of them seemed redundant to me.\n\nRemoved it.\n\n> ~~~\n>\n> 2.8 doc/src/sgml/logical-replication.sgml\n>\n> + <para>\n> + Adding a new node node3 to the existing node1 and node2 when data is\n> + present in existing nodes node1 and node2 needs similar steps, only change\n> + required here is that node3 should create subscription with copy_data as\n> + force to one of the existing nodes to receive the existing data during\n> + initial data synchronization. The steps for the same is listed below:\n> + </para>\n>\n> I thought it should be 2 sentences. So \"needs similar steps, only\n> change required here\" -> \"needs similar steps. The only change\n> required here...\"\n\nModified\n\n> ~~~\n>\n> 2.9 doc/src/sgml/logical-replication.sgml\n>\n> I think every time you say option names or values like \"copy_data\" in\n> the text they should be using <literal> font.\n\nModified\n\n> ~~~\n>\n> 2.10 doc/src/sgml/logical-replication.sgml\n>\n> + <para>\n> + Adding a new node node3 to the existing node1 and node2 when data is\n> + present in the new node node3 needs similar steps, few changes are\n> + required here to get the existing data from node3 to node1 and node2 and\n> + later cleaning up of data in node3 before synchronization of all the data\n> + from the existing nodes. The steps for the same is listed below:\n> + </para>\n>\n> I thought it should be 2 sentences. So \"needs similar steps, few\n> changes are required here\" -> \"needs similar steps. A few changes are\n> required here...\"\n\nModified\n\n> ~~~\n>\n> 2.11 doc/src/sgml/logical-replication.sgml\n>\n> All the text that says \"create subscription\" and \"create publication\"\n> maybe should be change to \"create a subscription\" and \"create a\n> publication\" etc.\n\nModified\n\n> ~~~\n>\n> 2.12 doc/src/sgml/ref/alter_subscription.sgml - copy_data\n>\n> + <para>\n> + There is some interaction between the \"local_only\" option and\n> + \"copy_data\" option. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n> + details and usage of force for copy_data option.\n> </para>\n>\n> 2.12.a Maybe options should not be in quotes like that. I think they\n> should be <literal> font.\n\nModified\n\n> 2.12.b It is a bit misleading because there is no \"Notes\" section here\n> on this page. Maybe it should say refer to the CREATE SUBSCRIPTION\n> Notes.\n\nModified\n\n> 2.12.c The last copy_data should also be <literal> font.\n\nModified\n\n> ~~~\n>\n> 2.13 doc/src/sgml/ref/create_subscription.sgml - local_only\n>\n> </para>\n> + <para>\n> + There is some interaction between the \"local_only\" option and\n> + \"copy_data\" option. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for details.\n> + </para>\n> </listitem>\n>\n> 2.13.a Maybe options should not be in quotes like that. I think they\n> should be <literal> font.\n\nModified\n\n> 2.13.b The last copy_data should also be <literal> font.\n\nModified\n\n> ~~~\n>\n> 2.14 doc/src/sgml/ref/create_subscription.sgml - copy_data\n>\n>\n> <para>\n> Specifies whether to copy pre-existing data in the publications\n> - that are being subscribed to when the replication starts.\n> - The default is <literal>true</literal>.\n> + that are being subscribed to when the replication starts. This\n> + parameter may be either <literal>true</literal>,\n> + <literal>false</literal> or <literal>force</literal>. The default is\n> + <literal>true</literal>.\n> </para>\n>\n> 2.14.a Maybe options should not be in quotes like that. I think they\n> should be <literal> font.\n\nModified\n\n> 2.14.b The last copy_data should also be <literal> font.\n\nModified\n\n> ~~~\n>\n> 2.15 doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -374,6 +388,16 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> can have non-existent publications.\n> </para>\n>\n> + <para>\n> + If subscription is created with local_only as 'on' and copy_data as 'on', it\n> + will check if the publisher tables are being subscribed to any other\n> + publisher and throw an error to prevent inconsistent data in the\n> + subscription. User can continue with the copy operation without throwing any\n> + error in this case by specifying copy_data as 'force'. Refer to the\n> + <xref linkend=\"bidirectional-logical-replication\"/> on how\n> + copy_data and local_only can be used in bidirectional replication.\n> + </para>\n>\n> 2.15.a I think all those mentions of copy_data and local_only and\n> their values should be in <literal> font instead of in quotes.\n\nModified\n\n> 2.15.b Wording: \"User can\" -> \"The user can\"\n\nModified\n\n> ~~~\n>\n> 2.16 src/backend/commands/subscriptioncmds.c - macro\n>\n> > 20. src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_ON_OR_FORCE\n> >\n> > @@ -69,6 +69,18 @@\n> > /* check if the 'val' has 'bits' set */\n> > #define IsSet(val, bits) (((val) & (bits)) == (bits))\n> >\n> > +#define IS_COPY_DATA_ON_OR_FORCE(copy_data) (copy_data != COPY_DATA_OFF)\n> >\n> > Maybe this would be better as a static inline function?\n>\n> Vignesh: What is the advantage of doing this change? I have not changed this\n> as the macro usage is fine. Thoughts?\n>\n> Originally I was going to suggest the macro should use extra parens\n> like ((copy_data) != COPY_DATA_OFF), but then I thought if it was a\n> function then it would have enum type-checking which would be better.\n> If you want to keep the macro then please make the parens change.\n\nModified to include parenthesis.\n\n> ~~~\n>\n> 2.17 src/backend/commands/subscriptioncmds.c - DefGetCopyData\n>\n> + /*\n> + * The set of strings accepted here should match up with\n> + * the grammar's opt_boolean_or_string production.\n> + */\n> + if (pg_strcasecmp(sval, \"true\") == 0 ||\n> + pg_strcasecmp(sval, \"on\") == 0)\n> + return COPY_DATA_ON;\n> + if (pg_strcasecmp(sval, \"false\") == 0 ||\n> + pg_strcasecmp(sval, \"off\") == 0)\n> + return COPY_DATA_OFF;\n> + if (pg_strcasecmp(sval, \"force\") == 0)\n> + return COPY_DATA_FORCE;\n>\n> I think you can change the order of these to be off/on/force, so then\n> the order is consistent with the T_Integer case.\n\nModified\n\n> ~~~\n>\n> 2.18 src/backend/commands/subscriptioncmds.c - DefGetCopyData\n>\n> + ereport(ERROR,\n> + errcode(ERRCODE_SYNTAX_ERROR),\n> + errmsg(\"%s requires a boolean or \\\"force\\\"\", def->defname));\n> + return COPY_DATA_OFF; /*\n> keep compiler quiet */\n>\n> Excessive comment indent? Has pg_indent been run?\n\nModified\n\n> ~~~\n>\n> 2.19 src/backend/commands/subscriptioncmds.c - AlterSubscription\n>\n> @@ -1236,7 +1306,8 @@ AlterSubscription(ParseState *pstate,\n> AlterSubscriptionStmt *stmt,\n> errmsg(\"ALTER SUBSCRIPTION ... REFRESH is not allowed for disabled\n> subscriptions\")));\n>\n> parse_subscription_options(pstate, stmt->options,\n> - SUBOPT_COPY_DATA, &opts);\n> + SUBOPT_COPY_DATA,\n> + &opts);\n>\n> This is a formatting change only. Maybe it does not belong in this\n> patch unless it is the result of pg_indent.\n\nModified\n\n> ~~~\n>\n> 2.20 src/backend/commands/subscriptioncmds.c - fetch_table_list\n>\n> > 23. src/backend/commands/subscriptioncmds.c - long errmsg\n> >\n> > + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\n> > + ereport(ERROR,\n> > + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> > + errmsg(\"CREATE/ALTER SUBSCRIPTION with local_only and copy_data as\n> > true is not allowed when the publisher might have replicated data,\n> > table:%s.%s might have replicated data in the publisher\",\n> > + nspname, relname),\n> > + errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\"));\n> > +\n> >\n> > The errmsg seems way too long for the source code. Can you use string\n> > concatenation or continuation chars to wrap the message over multiple\n> > lines?\n>\n> Vignesh: I had seen that the long error message elsewhere also is in a\n> single line. I think we should keep it as it is to maintain the coding\n> standard. Thoughts?\n>\n> OK, if you say it is already common practice then it's fine by me to\n> leave it as-is.\n\nOk, no change done.\n\n> ~~~\n>\n> 2.21 src/test/regress/expected/subscription.out - make check\n>\n> make check fails.\n> 1 of 214 tests failed.\n>\n> 2.21.a It looks like maybe you did not update the expected ordering of\n> some of the tests, after some minor adjustments in subscriprion.sql in\n> v10. So the expected output needs to be fixed in the patch.\n\nModified\n\n> 2.21.b. Suggest adding this patch to CF so that the cfbot can pick up\n> such test problems earlier.\n\nCF entry added\n\nThanks for the comments, the attached v11 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Thu, 28 Apr 2022 22:34:12 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Apr 28, 2022 at 7:57 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for v10-0002 (TAP tests part only)\n>\n> FIle: src/test/subscription/t/032_localonly.pl\n>\n> ======\n>\n> 1.\n>\n> +# Detach node C and clean the table contents.\n> +sub detach_node_clean_table_data\n> +{\n>\n> 1a. Maybe say \"Detach node C from the node-group of (A, B, C) and\n> clean the table contents from all nodes\"\n\nModified\n\n> 1b. Actually wondered do you need to TRUNCATE from both A and B (maybe\n> only truncate 1 is OK since those nodes are still using MMC). OTOH\n> maybe your explicit way makes the test simpler.\n\ncleaning the data in all nodes to keep the test simpler\n\n> ~~~\n>\n> 2.\n>\n> +# Subroutine for verify the data is replicated successfully.\n> +sub verify_data\n> +{\n>\n> 2a. Typo: \"for verify\" -> \"to verify\"\n\nModified\n\n> 2b. The messages in this function maybe are not very appropriate. They\n> say 'Inserted successfully without leading to infinite recursion in\n> circular replication setup', but really the function is only testing\n> all the data is the same as 'expected'. So it could be the result of\n> any operation - not just Insert.\n\nModified\n\n> ~~~\n>\n> 3. SELECT ORDER BY?\n>\n> $result = $node_B->safe_psql('postgres', \"SELECT * FROM tab_full;\");\n> is($result, qq(11\n> 12\n> 13),\n> 'Node_C data replicated to Node_B'\n> );\n>\n> I am not sure are these OK like this or if *every* SELECT use ORDER BY\n> to make sure the data is in the same qq expected order? There are\n> multiple cases like this.\n>\n> (BTW, I think this comment needs to be applied for the v10-0001 patch,\n> maybe not v10-0002).\n\nModified\n\n> ~~~\n>\n> 4.\n>\n> +###############################################################################\n> +# Specifying local_only 'on' which indicates that the publisher should only\n> +# replicate the changes that are generated locally from node_B, but in\n> +# this case since the node_B is also subscribing data from node_A, node_B can\n> +# have data originated from node_A, so throw an error in this case to prevent\n> +# node_A data being replicated to the node_C.\n> +###############################################################################\n>\n> There is something wrong with the description because there is no\n> \"node_C\" in this test. You are just creating a 2nd subscription on\n> node A.\n\nModified\n\n> ~~\n>\n> 5.\n>\n> +($result, $stdout, $stderr) = $node_A->psql(\n> + 'postgres', \"\n> + CREATE SUBSCRIPTION tap_sub_A3\n> + CONNECTION '$node_B_connstr application_name=$subname_AB'\n> + PUBLICATION tap_pub_B\n> + WITH (local_only = on, copy_data = on)\");\n> +like(\n>\n> It seemed strange to call this 2nd subscription \"tap_sub_A3\". Wouldn't\n> it be better to call it \"tap_sub_A2\"?\n\nModified\n\n> ~~~\n>\n> 6.\n>\n> +###############################################################################\n> +# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\n> +# replication setup when the existing nodes (node_A & node_B) has no data and\n> +# the new node (node_C) some pre-existing data.\n> +###############################################################################\n> +$node_C->safe_psql('postgres', \"INSERT INTO tab_full VALUES (31);\");\n> +\n> +$result = $node_A->safe_psql('postgres', \"SELECT * FROM tab_full order by 1;\");\n> +is( $result, qq(), 'Check existing data');\n> +\n> +$result = $node_B->safe_psql('postgres', \"SELECT * FROM tab_full order by 1;\");\n> +is( $result, qq(), 'Check existing data');\n> +\n> +$result = $node_C->safe_psql('postgres', \"SELECT * FROM tab_full order by 1;\");\n> +is($result, qq(31), 'Check existing data');\n> +\n> +create_subscription($node_A, $node_C, $subname_AC, $node_C_connstr,\n> + 'tap_pub_C', 'copy_data = force, local_only = on');\n> +create_subscription($node_B, $node_C, $subname_BC, $node_C_connstr,\n> + 'tap_pub_C', 'copy_data = force, local_only = on');\n> +\n>\n> Because the Node_C does not yet have any subscriptions aren't these\n> cases where you didn't really need to use \"force\"?\n\nModified\n\nThe v11 patch attached at [1] has the fixes for the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm2hRyXc4DKPLnBgem_LHBmy%3DWXFEsm9-xhckye1YXcpPA%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Thu, 28 Apr 2022 22:37:22 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for the v11* patches.\n\n======\n\nv11-0001 - no more comments. LGTM\n\n======\n\nV11-0002\n\n1. doc/src/sgml/logical-replication.sgml\n\n+ <para>\n+ Bidirectional replication is useful in creating multi master database\n+ which helps in performing read/write operations from any of the nodes.\n+ Setting up bidirectional logical replication between two nodes requires\n+ creation of the publication in all the nodes, creating subscriptions in\n+ each of the nodes that subscribes to data from all the nodes. The steps\n+ to create a two-node bidirectional replication are given below:\n+ </para>\n\nWording: \"creating multi master database\" -> \"creating a multi-master database\"\nWording: \"creation of the publication in all the nodes\" -> \"creation\nof a publication in all the nodes\".\n\n~~~\n\n2. doc/src/sgml/logical-replication.sgml\n\n> +<programlisting>\n> +node2=# CREATE SUBSCRIPTION sub_node1_node2\n> +node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\n> +node2=# PUBLICATION pub_node1\n> +node2=# WITH (copy_data = off, local_only = on);\n> +CREATE SUBSCRIPTION\n> +</programlisting>\n> + </para>\n>\n> IIUC the closing </para> should be on the same line as the\n> </programlisting>. I recall there was some recent github push about\n> this sometime in the last month - maybe you can check to confirm it.\n\nVignesh: I have seen the programlisting across multiple places and noticed\nthat there is no such guideline being followed. I have not made any\nchange for this comment.\n\nFYI – I found the push [1] by PeterE that I was referring to so. I\nthought we perhaps should follow this, even if not all other code is\ndoing so. But you can choose.\n\n~~~\n\n3. doc/src/sgml/logical-replication.sgml\n\n+ <para>\n+ Create a subscription in <literal>node3</literal> to subscribe the changes\n+ from <literal>node1</literal> and <literal>node2</literal>, here\n+ <literal>copy_data</literal> is specified as <literal>force</literal> when\n+ creating a subscription to <literal>node1</literal> so that the existing\n+ table data is copied during initial sync:\n+<programlisting>\n+node3=# CREATE SUBSCRIPTION\n+node3-# sub_node3_node1 CONNECTION 'dbname=foo host=node1 user=repuser'\n+node3-# PUBLICATION pub_node1\n+node3-# WITH (copy_data = force, local_only = on);\n+CREATE SUBSCRIPTION\n+node3=# CREATE SUBSCRIPTION\n+node3-# sub_node3_node2 CONNECTION 'dbname=foo host=node2 user=repuser'\n+node3-# PUBLICATION pub_node2\n+node3-# WITH (copy_data = off, local_only = on);\n+CREATE SUBSCRIPTION\n+</programlisting>\n\nI think this part should be split into 2 separate program listings\neach for the different subscriptions. Then those descriptions can be\ndescribed separately (e.g. why one is force and the other is not).\nThen it will also be more consistent with how you split/explained\nsomething similar in the previous section on this page.\n\n------\n[1] https://github.com/postgres/postgres/commit/d7ab2a9a3c0a2800ab36bb48d1cc97370067777e\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 29 Apr 2022 12:37:46 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Apr 29, 2022 at 8:08 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for the v11* patches.\n>\n> ======\n>\n> v11-0001 - no more comments. LGTM\n>\n> ======\n>\n> V11-0002\n>\n> 1. doc/src/sgml/logical-replication.sgml\n>\n> + <para>\n> + Bidirectional replication is useful in creating multi master database\n> + which helps in performing read/write operations from any of the nodes.\n> + Setting up bidirectional logical replication between two nodes requires\n> + creation of the publication in all the nodes, creating subscriptions in\n> + each of the nodes that subscribes to data from all the nodes. The steps\n> + to create a two-node bidirectional replication are given below:\n> + </para>\n>\n> Wording: \"creating multi master database\" -> \"creating a multi-master database\"\n> Wording: \"creation of the publication in all the nodes\" -> \"creation\n> of a publication in all the nodes\".\n\nModified\n\n> ~~~\n>\n> 2. doc/src/sgml/logical-replication.sgml\n>\n> > +<programlisting>\n> > +node2=# CREATE SUBSCRIPTION sub_node1_node2\n> > +node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\n> > +node2=# PUBLICATION pub_node1\n> > +node2=# WITH (copy_data = off, local_only = on);\n> > +CREATE SUBSCRIPTION\n> > +</programlisting>\n> > + </para>\n> >\n> > IIUC the closing </para> should be on the same line as the\n> > </programlisting>. I recall there was some recent github push about\n> > this sometime in the last month - maybe you can check to confirm it.\n>\n> Vignesh: I have seen the programlisting across multiple places and noticed\n> that there is no such guideline being followed. I have not made any\n> change for this comment.\n>\n> FYI – I found the push [1] by PeterE that I was referring to so. I\n> thought we perhaps should follow this, even if not all other code is\n> doing so. But you can choose.\n\nModified\n\n> ~~~\n>\n> 3. doc/src/sgml/logical-replication.sgml\n>\n> + <para>\n> + Create a subscription in <literal>node3</literal> to subscribe the changes\n> + from <literal>node1</literal> and <literal>node2</literal>, here\n> + <literal>copy_data</literal> is specified as <literal>force</literal> when\n> + creating a subscription to <literal>node1</literal> so that the existing\n> + table data is copied during initial sync:\n> +<programlisting>\n> +node3=# CREATE SUBSCRIPTION\n> +node3-# sub_node3_node1 CONNECTION 'dbname=foo host=node1 user=repuser'\n> +node3-# PUBLICATION pub_node1\n> +node3-# WITH (copy_data = force, local_only = on);\n> +CREATE SUBSCRIPTION\n> +node3=# CREATE SUBSCRIPTION\n> +node3-# sub_node3_node2 CONNECTION 'dbname=foo host=node2 user=repuser'\n> +node3-# PUBLICATION pub_node2\n> +node3-# WITH (copy_data = off, local_only = on);\n> +CREATE SUBSCRIPTION\n> +</programlisting>\n>\n> I think this part should be split into 2 separate program listings\n> each for the different subscriptions. Then those descriptions can be\n> described separately (e.g. why one is force and the other is not).\n> Then it will also be more consistent with how you split/explained\n> something similar in the previous section on this page.\n\nModified\nThe attached v12 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Fri, 29 Apr 2022 16:31:41 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Thanks for updating the patches for all my prior feedback.\n\nFor v12* I have only minor feedback for the docs.\n\n======\nV12-0001\n\nno comments\n\n======\nV12-0002\n\n1. Commit message\n\nIn another thread using the terminology \"multi-master\" seems to be\ncausing some problems. Maybe you should choose different wording in\nthis commit message to avoid having the same problems?\n\n~~~\n\n2. doc/src/sgml/logical-replication.sgml\n\nThese are all similar minor wording suggestions to split the sentences.\nWording: \", here copy_data is specified as XXX ...\" -> \". Use\ncopy_data specified as XXX ...\"\n\nAlso:\nWording: \"... to subscribe the changes from nodeXXX...\"\n-> \"... to subscribe to the changes from nodeXXX... \" (add \"to\")\n-> \"... to subscribe to nodeXXX.\" (or I preferred just remove the\nwhole \"changes\" part)\n\n2a.\nCreate a subscription in node3 to subscribe the changes from node1,\nhere copy_data is specified as force so that the existing table data\nis copied during initial sync:\n\nSUGGESTION\nCreate a subscription in node3 to subscribe to node1. Use copy_data\nspecified as force so that the existing table data is copied during\ninitial sync:\n\n2b.\nCreate a subscription in node1 to subscribe the changes fromnode3,\nhere copy_data is specified as force so that the existing table data\nis copied during initial sync:\n\nSUGGESTION\nCreate a subscription in node1 to subscribe to node3. Use copy_data\nspecified as force so that the existing table data is copied during\ninitial sync:\n\n2c.\nCreate a subscription in node2 to subscribe the changes from node3,\nhere copy_data is specified as force so that the existing table data\nis copied during initial sync:\n\nSUGGESTION\nCreate a subscription in node2 to subscribe to node3. Use copy_data\nspecified as force so that the existing table data is copied during\ninitial sync:\n\n2d.\nCreate a subscription in node3 to subscribe the changes from node1,\nhere copy_data is specified as force when creating a subscription to\nnode1 so that the existing table data is copied during initial sync:\n\nSUGGESTION\nCreate a subscription in node3 to subscribe to node1. Use copy_data\nspecified as force when creating a subscription to node1 so that the\nexisting table data is copied during initial sync:\n\n2e.\nCreate a subscription in node3 to subscribe the changes from node2,\nhere copy_data is specified as off as the initial table data would\nhave been copied in the earlier step:\n\nSUGGESTION (this one has a bit more re-wording than the others)\nCreate a subscription in node3 to subscribe to node2. Use copy_data\nspecified as off because the initial table data would have been\nalready copied in the previous step:\n\n~~~\n\n3. doc/src/sgml/logical-replication.sgml\n\n31.11.2. Adding new node when there is no data in any of the nodes\n31.11.3. Adding new node when data is present in the existing nodes\n31.11.4. Adding new node when data is present in the new node\n\nMinor change to the above heading?\n\nWording: \"Adding new node ...\" -> \"Adding a new node ...\"\n\n------\n[1] https://www.postgresql.org/message-id/flat/CAHut%2BPuwRAoWY9pz%3DEubps3ooQCOBFiYPU9Yi%3DVB-U%2ByORU7OA%40mail.gmail.com\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 2 May 2022 10:18:59 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, May 2, 2022 at 5:49 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Thanks for updating the patches for all my prior feedback.\n>\n> For v12* I have only minor feedback for the docs.\n>\n> ======\n> V12-0001\n>\n> no comments\n>\n> ======\n> V12-0002\n>\n> 1. Commit message\n>\n> In another thread using the terminology \"multi-master\" seems to be\n> causing some problems. Maybe you should choose different wording in\n> this commit message to avoid having the same problems?\n\nI felt, we can leave this to committer\n\n> ~~~\n>\n> 2. doc/src/sgml/logical-replication.sgml\n>\n> These are all similar minor wording suggestions to split the sentences.\n> Wording: \", here copy_data is specified as XXX ...\" -> \". Use\n> copy_data specified as XXX ...\"\n>\n> Also:\n> Wording: \"... to subscribe the changes from nodeXXX...\"\n> -> \"... to subscribe to the changes from nodeXXX... \" (add \"to\")\n> -> \"... to subscribe to nodeXXX.\" (or I preferred just remove the\n> whole \"changes\" part)\n>\n> 2a.\n> Create a subscription in node3 to subscribe the changes from node1,\n> here copy_data is specified as force so that the existing table data\n> is copied during initial sync:\n>\n> SUGGESTION\n> Create a subscription in node3 to subscribe to node1. Use copy_data\n> specified as force so that the existing table data is copied during\n> initial sync:\n\nModified\n\n> 2b.\n> Create a subscription in node1 to subscribe the changes fromnode3,\n> here copy_data is specified as force so that the existing table data\n> is copied during initial sync:\n>\n> SUGGESTION\n> Create a subscription in node1 to subscribe to node3. Use copy_data\n> specified as force so that the existing table data is copied during\n> initial sync:\n\nModified\n\n> 2c.\n> Create a subscription in node2 to subscribe the changes from node3,\n> here copy_data is specified as force so that the existing table data\n> is copied during initial sync:\n>\n> SUGGESTION\n> Create a subscription in node2 to subscribe to node3. Use copy_data\n> specified as force so that the existing table data is copied during\n> initial sync:\n\nModified\n\n> 2d.\n> Create a subscription in node3 to subscribe the changes from node1,\n> here copy_data is specified as force when creating a subscription to\n> node1 so that the existing table data is copied during initial sync:\n>\n> SUGGESTION\n> Create a subscription in node3 to subscribe to node1. Use copy_data\n> specified as force when creating a subscription to node1 so that the\n> existing table data is copied during initial sync:\n\nModified\n\n> 2e.\n> Create a subscription in node3 to subscribe the changes from node2,\n> here copy_data is specified as off as the initial table data would\n> have been copied in the earlier step:\n>\n> SUGGESTION (this one has a bit more re-wording than the others)\n> Create a subscription in node3 to subscribe to node2. Use copy_data\n> specified as off because the initial table data would have been\n> already copied in the previous step:\n\nModified\n\n> ~~~\n>\n> 3. doc/src/sgml/logical-replication.sgml\n>\n> 31.11.2. Adding new node when there is no data in any of the nodes\n> 31.11.3. Adding new node when data is present in the existing nodes\n> 31.11.4. Adding new node when data is present in the new node\n>\n> Minor change to the above heading?\n>\n> Wording: \"Adding new node ...\" -> \"Adding a new node ...\"\n\nModified\n\nThanks for the comments, the attached v13 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Wed, 4 May 2022 12:17:00 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Thanks for making changes from all my past reviews.\n\n1. I checked the cfbot result is OK\n\n2. Both patches apply cleanly in my VM\n\n3. Patch changes\nV13-0001 - No diffs; it is identical to v12-0001\nV13-0002 - only changed for pg docs. So I rebuilt v13 docs and checked\nthe HTML rendering. Looked OK.\n\nSo I have no more review comments. LGTM.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 5 May 2022 17:22:47 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On 2022-May-04, vignesh C wrote:\n\n> On Mon, May 2, 2022 at 5:49 AM Peter Smith <smithpb2250@gmail.com> wrote:\n\n> > 1. Commit message\n> >\n> > In another thread using the terminology \"multi-master\" seems to be\n> > causing some problems. Maybe you should choose different wording in\n> > this commit message to avoid having the same problems?\n> \n> I felt, we can leave this to committer\n\nPlease don't. See\nhttps://postgr.es/m/20200615182235.x7lch5n6kcjq4aue@alap3.anarazel.de\n\n-- \nÁlvaro Herrera Breisgau, Deutschland — https://www.EnterpriseDB.com/\n\n\n","msg_date":"Thu, 5 May 2022 11:11:05 +0200","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, May 5, 2022 at 2:41 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> On 2022-May-04, vignesh C wrote:\n>\n> > On Mon, May 2, 2022 at 5:49 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> > > 1. Commit message\n> > >\n> > > In another thread using the terminology \"multi-master\" seems to be\n> > > causing some problems. Maybe you should choose different wording in\n> > > this commit message to avoid having the same problems?\n> >\n> > I felt, we can leave this to committer\n>\n> Please don't. See\n> https://postgr.es/m/20200615182235.x7lch5n6kcjq4aue@alap3.anarazel.de\n\nIn this thread changing multimaster to multi-source did not get agreed\nand the changes did not get committed. The code still uses multimaster\nas in [1]. I think we should keep it as \"multimaster\" as that is\nalready being used in [1].\n\n[1] - https://www.postgresql.org/docs/current/different-replication-solutions.html\n\nRegards,\nVignesh\n\n\n","msg_date":"Thu, 5 May 2022 16:44:14 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, May 4, 2022 at 12:17 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> Thanks for the comments, the attached v13 patch has the changes for the same.\n>\n\nFew comments on v13-0001\n======================\n1.\n+ *\n+ * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\n+ * PG16.\n...\n@@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\nOutputPluginOptions *opt,\n else\n ctx->twophase_opt_given = true;\n\n+ if (data->local_only && data->protocol_version <\nLOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\n+ ereport(ERROR,\n+ (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\n+ errmsg(\"requested proto_version=%d does not support local_only, need\n%d or higher\",\n+ data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\n\nWhat is the need to change the protocol version for this parameter? As\nper my understanding, we only need a different protocol version when\nwe are sending some new message or some additional information in an\nexisting message as we have done for the streaming/two_phase options\nwhich doesn't seem to be the case here.\n\n2.\n@@ -29,6 +29,7 @@ typedef struct PGOutputData\n bool streaming;\n bool messages;\n bool two_phase;\n+ bool local_only;\n } PGOutputData;\n\nIt seems we already have a similar option named 'only_local' in\nTestDecodingData which has the exactly same functionality. So, isn't\nit better to name this as 'only_local' instead of 'local_only'? Also,\nlet's add a test case for 'only_local' option of test_decoding.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 18 May 2022 10:29:30 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, May 4, 2022 2:47 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> Thanks for the comments, the attached v13 patch has the changes for the same.\r\n> \r\n\r\nThanks for your patch. Here are some comments on v13-0001 patch.\r\n\r\n1) \r\n@@ -152,6 +152,18 @@ CREATE SUBSCRIPTION <replaceable class=\"parameter\">subscription_name</replaceabl\r\n </listitem>\r\n </varlistentry>\r\n \r\n+ <varlistentry>\r\n+ <term><literal>local_only</literal> (<type>boolean</type>)</term>\r\n+ <listitem>\r\n+ <para>\r\n+ Specifies whether the subscription will request the publisher to send\r\n+ locally originated changes at the publisher node, or send any\r\n+ publisher node changes regardless of their origin. The default is\r\n+ <literal>false</literal>.\r\n+ </para>\r\n+ </listitem>\r\n+ </varlistentry>\r\n+\r\n <varlistentry>\r\n <term><literal>slot_name</literal> (<type>string</type>)</term>\r\n <listitem>\r\n\r\nI think this change should be put after \"The following parameters control the\r\nsubscription's replication behavior after it has been created\", thoughts?\r\n\r\n2)\r\nA new column \"sublocalonly\" is added to pg_subscription, so maybe we need add it\r\nto pg_subscription document, too. (in doc/src/sgml/catalogs.sgml)\r\n\r\n3)\r\n/*\r\n * Currently we always forward.\r\n */\r\nstatic bool\r\npgoutput_origin_filter(LogicalDecodingContext *ctx,\r\n\t\t\t\t\t RepOriginId origin_id)\r\n\r\nShould we modify the comment of pgoutput_origin_filter()? It doesn't match the\r\ncode.\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Wed, 18 May 2022 08:10:42 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, May 4, 2022 at 12:17 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached v13 patch has the changes for the same.\n> >\n>\n> Few comments on v13-0001\n> ======================\n>\n\nFew comments on v13-0002\n===========================\n1.\nThe steps\n+ to create a two-node bidirectional replication are given below:\n+ </para>\n\nThe steps given after this will be successful only when there is no\ndata in any of the nodes and that is not clear by reading docs.\n\n2.\n+ <sect2 id=\"add-new-node-data-in-existing-node\">\n+ <title>Adding a new node when data is present in the existing nodes</title>\n+ <para>\n+ Adding a new node <literal>node3</literal> to the existing\n+ <literal>node1</literal> and <literal>node2</literal> when data is present\n+ in existing nodes <literal>node1</literal> and <literal>node2</literal>\n+ needs similar steps. The only change required here is that\n+ <literal>node3</literal> should create a subscription with\n+ <literal>copy_data = force</literal> to one of the existing nodes to\n+ receive the existing data during initial data synchronization.\n+ </para>\n\nI think the steps for these require the user to lock the required\ntables/database (or in some other way hold the operations on required\ntables) for node-2 till the time setup is complete, otherwise, node-3\nmight miss some data. It seems the same is also missing in the\nsection: \"Adding a new node when data is present in the new node\".\n\n3.\n+\n+ <sect2 id=\"add-node-data-present-in-new-node\">\n+ <title>Adding a new node when data is present in the new node</title>\n...\n...\n\n+ <para>\n+ Create a subscription in <literal>node1</literal> to subscribe to\n+ <literal>node3</literal>. Use <literal>copy_data</literal> specified as\n+ <literal>force</literal> so that the existing table data is copied during\n+ initial sync:\n+<programlisting>\n+node1=# CREATE SUBSCRIPTION sub_node1_node3\n+node1-# CONNECTION 'dbname=foo host=node3 user=repuser'\n+node1-# PUBLICATION pub_node3\n+node1-# WITH (copy_data = force, local_only = on);\n+CREATE SUBSCRIPTION\n+</programlisting></para>\n+\n+ <para>\n+ Create a subscription in <literal>node2</literal> to subscribe to\n+ <literal>node3</literal>. Use <literal>copy_data</literal> specified as\n+ <literal>force</literal> so that the existing table data is copied during\n+ initial sync:\n+<programlisting>\n+node2=# CREATE SUBSCRIPTION sub_node2_node3\n+node2-# CONNECTION 'dbname=foo host=node3 user=repuser'\n+node2-# PUBLICATION pub_node3\n+node2-# WITH (copy_data = force, local_only = on);\n+CREATE SUBSCRIPTION\n+</programlisting></para>\n\nWhy do we need to use \"copy_data = force\" here? AFAIU, unless, we\ncreate any subscription on node-3, we don't need the 'force' option.\n\n4. We should have a generic section to explain how users can add a new\nnode using the new options to the existing set of nodes in all cases.\nFor example, say when the existing set of nodes has some data and the\nnew node also has some pre-existing data. I think the basic steps are\nsomething like: a. create a required publication(s) on the new node.\n(b) create subscriptions on existing nodes pointing to publication on\nthe new node with the local_only option as true and copy_data = on.\n(c) wait for data to be copied from the new node to existing nodes.\n(d) Truncate the data on the new node. (e) create subscriptions\ncorresponding to each of the existing publisher nodes on the new node\nwith local_only as true and copy_data = force for one of the nodes.\n(f) One needs to ensure that there is no new activity on required\ntables/database (either by locking or in some other way) in all the\nnodes for which copy_data option is kept as false while creating\nsubscriptions in the previous step to avoid any data loss.\n\nWe also need to mention in Notes that as all operations are not\ntransactional, user is advised to take backup of existing data to\navoid any inconsistency.\n\n5.\n * It is quite possible that subscriber has not yet pulled data to\n+ * the tables, but in ideal cases the table data will be subscribed.\n+ * To keep the code simple it is not checked if the subscriber table\n+ * has pulled the data or not.\n+ */\n+ if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\n\nSorry, but I don't understand what you intend to say by the above\ncomment. Can you please explain?\n\n6. I feel we can explain the case with only two nodes in the commit\nmessage, why do we need to use a three-node case?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 18 May 2022 16:22:22 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, May 4, 2022 2:47 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> Thanks for the comments, the attached v13 patch has the changes for the same.\r\n>\r\n\r\nHere are some comments on v13-0002 patch.\r\n\r\n1)\r\n+\t\t * Throw an error so that the user can take care of the initial data\r\n+\t\t * copying and then create subscription with copy_data off.\r\n\r\nShould \"with copy_data off\" be changed to \"with copy_data off or force\"?\r\n\r\n2)\r\n\t\tcase ALTER_SUBSCRIPTION_ADD_PUBLICATION:\r\n\t\tcase ALTER_SUBSCRIPTION_DROP_PUBLICATION:\r\n...\r\n\t\t\t\t\t/*\r\n\t\t\t\t\t * See ALTER_SUBSCRIPTION_REFRESH for details why this is\r\n\t\t\t\t\t * not allowed.\r\n\t\t\t\t\t */\r\n\t\t\t\t\tif (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\r\n\r\nI think we need some changes here, too. Should it be modified to:\r\nif (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Thu, 19 May 2022 08:41:54 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, May 4, 2022 at 12:17 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached v13 patch has the changes for the same.\n> >\n>\n> Few comments on v13-0001\n> ======================\n> 1.\n> + *\n> + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\n> + * PG16.\n> ...\n> @@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\n> OutputPluginOptions *opt,\n> else\n> ctx->twophase_opt_given = true;\n>\n> + if (data->local_only && data->protocol_version <\n> LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\n> + ereport(ERROR,\n> + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\n> + errmsg(\"requested proto_version=%d does not support local_only, need\n> %d or higher\",\n> + data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\n>\n> What is the need to change the protocol version for this parameter? As\n> per my understanding, we only need a different protocol version when\n> we are sending some new message or some additional information in an\n> existing message as we have done for the streaming/two_phase options\n> which doesn't seem to be the case here.\n\nModified\n\n> 2.\n> @@ -29,6 +29,7 @@ typedef struct PGOutputData\n> bool streaming;\n> bool messages;\n> bool two_phase;\n> + bool local_only;\n> } PGOutputData;\n>\n> It seems we already have a similar option named 'only_local' in\n> TestDecodingData which has the exactly same functionality. So, isn't\n> it better to name this as 'only_local' instead of 'local_only'? Also,\n> let's add a test case for 'only_local' option of test_decoding.\n\nModified and added test for only_local option of test_decoding.\n\nThanks for the comments, the attached v14 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Fri, 20 May 2022 15:07:47 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, May 18, 2022 at 1:40 PM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Wed, May 4, 2022 2:47 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached v13 patch has the changes for the same.\n> >\n>\n> Thanks for your patch. Here are some comments on v13-0001 patch.\n>\n> 1)\n> @@ -152,6 +152,18 @@ CREATE SUBSCRIPTION <replaceable class=\"parameter\">subscription_name</replaceabl\n> </listitem>\n> </varlistentry>\n>\n> + <varlistentry>\n> + <term><literal>local_only</literal> (<type>boolean</type>)</term>\n> + <listitem>\n> + <para>\n> + Specifies whether the subscription will request the publisher to send\n> + locally originated changes at the publisher node, or send any\n> + publisher node changes regardless of their origin. The default is\n> + <literal>false</literal>.\n> + </para>\n> + </listitem>\n> + </varlistentry>\n> +\n> <varlistentry>\n> <term><literal>slot_name</literal> (<type>string</type>)</term>\n> <listitem>\n>\n> I think this change should be put after \"The following parameters control the\n> subscription's replication behavior after it has been created\", thoughts?\n\nModified\n\n> 2)\n> A new column \"sublocalonly\" is added to pg_subscription, so maybe we need add it\n> to pg_subscription document, too. (in doc/src/sgml/catalogs.sgml)\n\nModified\n\n> 3)\n> /*\n> * Currently we always forward.\n> */\n> static bool\n> pgoutput_origin_filter(LogicalDecodingContext *ctx,\n> RepOriginId origin_id)\n>\n> Should we modify the comment of pgoutput_origin_filter()? It doesn't match the\n> code.\n\nModified\n\nThanks for the comments, the v14 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm0xuYy35vOudVHBjov3fQ%3DjBRHJHKUUN9VarqO%3DYqtaxg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 20 May 2022 15:15:42 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, May 18, 2022 at 4:22 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Wed, May 4, 2022 at 12:17 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > Thanks for the comments, the attached v13 patch has the changes for the same.\n> > >\n> >\n> > Few comments on v13-0001\n> > ======================\n> >\n>\n> Few comments on v13-0002\n> ===========================\n> 1.\n> The steps\n> + to create a two-node bidirectional replication are given below:\n> + </para>\n>\n> The steps given after this will be successful only when there is no\n> data in any of the nodes and that is not clear by reading docs.\n\nModified\n\n> 2.\n> + <sect2 id=\"add-new-node-data-in-existing-node\">\n> + <title>Adding a new node when data is present in the existing nodes</title>\n> + <para>\n> + Adding a new node <literal>node3</literal> to the existing\n> + <literal>node1</literal> and <literal>node2</literal> when data is present\n> + in existing nodes <literal>node1</literal> and <literal>node2</literal>\n> + needs similar steps. The only change required here is that\n> + <literal>node3</literal> should create a subscription with\n> + <literal>copy_data = force</literal> to one of the existing nodes to\n> + receive the existing data during initial data synchronization.\n> + </para>\n>\n> I think the steps for these require the user to lock the required\n> tables/database (or in some other way hold the operations on required\n> tables) for node-2 till the time setup is complete, otherwise, node-3\n> might miss some data. It seems the same is also missing in the\n> section: \"Adding a new node when data is present in the new node\".\n\nModified. Mentioned that lock is required on new node-3 and node-2. I\nhave mentioned node-3 also should be locked so that no operation is\nhappened in node-3, else there can be some dml operation after\ntruncate which is sent to node-1 and the same data is synced when\ncreate subscription in node-3 subscribing to the publisher in node-1.\n\n> 3.\n> +\n> + <sect2 id=\"add-node-data-present-in-new-node\">\n> + <title>Adding a new node when data is present in the new node</title>\n> ...\n> ...\n>\n> + <para>\n> + Create a subscription in <literal>node1</literal> to subscribe to\n> + <literal>node3</literal>. Use <literal>copy_data</literal> specified as\n> + <literal>force</literal> so that the existing table data is copied during\n> + initial sync:\n> +<programlisting>\n> +node1=# CREATE SUBSCRIPTION sub_node1_node3\n> +node1-# CONNECTION 'dbname=foo host=node3 user=repuser'\n> +node1-# PUBLICATION pub_node3\n> +node1-# WITH (copy_data = force, local_only = on);\n> +CREATE SUBSCRIPTION\n> +</programlisting></para>\n> +\n> + <para>\n> + Create a subscription in <literal>node2</literal> to subscribe to\n> + <literal>node3</literal>. Use <literal>copy_data</literal> specified as\n> + <literal>force</literal> so that the existing table data is copied during\n> + initial sync:\n> +<programlisting>\n> +node2=# CREATE SUBSCRIPTION sub_node2_node3\n> +node2-# CONNECTION 'dbname=foo host=node3 user=repuser'\n> +node2-# PUBLICATION pub_node3\n> +node2-# WITH (copy_data = force, local_only = on);\n> +CREATE SUBSCRIPTION\n> +</programlisting></para>\n>\n> Why do we need to use \"copy_data = force\" here? AFAIU, unless, we\n> create any subscription on node-3, we don't need the 'force' option.\n\nModified to on.\n\n> 4. We should have a generic section to explain how users can add a new\n> node using the new options to the existing set of nodes in all cases.\n> For example, say when the existing set of nodes has some data and the\n> new node also has some pre-existing data. I think the basic steps are\n> something like: a. create a required publication(s) on the new node.\n> (b) create subscriptions on existing nodes pointing to publication on\n> the new node with the local_only option as true and copy_data = on.\n> (c) wait for data to be copied from the new node to existing nodes.\n> (d) Truncate the data on the new node. (e) create subscriptions\n> corresponding to each of the existing publisher nodes on the new node\n> with local_only as true and copy_data = force for one of the nodes.\n> (f) One needs to ensure that there is no new activity on required\n> tables/database (either by locking or in some other way) in all the\n> nodes for which copy_data option is kept as false while creating\n> subscriptions in the previous step to avoid any data loss.\n\nAdded\n\n> We also need to mention in Notes that as all operations are not\n> transactional, user is advised to take backup of existing data to\n> avoid any inconsistency.\n\nModified\n\n> 5.\n> * It is quite possible that subscriber has not yet pulled data to\n> + * the tables, but in ideal cases the table data will be subscribed.\n> + * To keep the code simple it is not checked if the subscriber table\n> + * has pulled the data or not.\n> + */\n> + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\n>\n> Sorry, but I don't understand what you intend to say by the above\n> comment. Can you please explain?\n\nWhen the user specifies copy_data as on, we should check if the\npublisher has the publication tables being subscribed from a remote\npublisher. If so throw an error as it remote origin data present.\nEx:\nNode1 - pub1 for table t1 -- no data\nNode2 - Sub1 subscribing to data from pub1\nNode2 - pub2 for table t1 -- no data\nNode3 - create subscription to Node2 with copy_data = ON\n\nIn this case even though the table does not have any remote origin\ndata, as Node2 is subscribing to data from Node1, throw an error.\nWe throw an error irrespective of data present in Node1 or not to keep\nthe code simple.\n\n> 6. I feel we can explain the case with only two nodes in the commit\n> message, why do we need to use a three-node case?\n\nModified\n\nThanks for the comments, the v14 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm0xuYy35vOudVHBjov3fQ%3DjBRHJHKUUN9VarqO%3DYqtaxg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 20 May 2022 15:30:54 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, May 19, 2022 at 2:12 PM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Wed, May 4, 2022 2:47 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached v13 patch has the changes for the same.\n> >\n>\n> Here are some comments on v13-0002 patch.\n>\n> 1)\n> + * Throw an error so that the user can take care of the initial data\n> + * copying and then create subscription with copy_data off.\n>\n> Should \"with copy_data off\" be changed to \"with copy_data off or force\"?\n\nModified\n\n> 2)\n> case ALTER_SUBSCRIPTION_ADD_PUBLICATION:\n> case ALTER_SUBSCRIPTION_DROP_PUBLICATION:\n> ...\n> /*\n> * See ALTER_SUBSCRIPTION_REFRESH for details why this is\n> * not allowed.\n> */\n> if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\n>\n> I think we need some changes here, too. Should it be modified to:\n> if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\n\nModified\n\nThanks for the comments, the v14 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm0xuYy35vOudVHBjov3fQ%3DjBRHJHKUUN9VarqO%3DYqtaxg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 20 May 2022 15:32:11 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, May 20, 2022 at 3:08 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> >\n> > Few comments on v13-0001\n> > ======================\n> > 1.\n> > + *\n> > + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\n> > + * PG16.\n> > ...\n> > @@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\n> > OutputPluginOptions *opt,\n> > else\n> > ctx->twophase_opt_given = true;\n> >\n> > + if (data->local_only && data->protocol_version <\n> > LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\n> > + ereport(ERROR,\n> > + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\n> > + errmsg(\"requested proto_version=%d does not support local_only, need\n> > %d or higher\",\n> > + data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\n> >\n> > What is the need to change the protocol version for this parameter? As\n> > per my understanding, we only need a different protocol version when\n> > we are sending some new message or some additional information in an\n> > existing message as we have done for the streaming/two_phase options\n> > which doesn't seem to be the case here.\n>\n> Modified\n>\n\nIt seems you forgot to remove the comments after removing the code\ncorresponding to the above. See below.\n+ *\n+ * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\n+ * support for sending only locally originated data from the publisher.\n+ * Introduced in PG16.\n+ *\n+ * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\n+ * PG16.\n */\n\nFew other comments on 0001\n========================\n1.\n+ * Return true if data has originated remotely when only_local option is\n+ * enabled, false otherwise.\n\nCan we slightly change the comment to:\"Return true if the data source\n(origin) is remote and user has requested only local data, false\notherwise.\"\n\n2.\n+SELECT pg_replication_origin_session_reset();\n+SELECT pg_drop_replication_slot('regression_slot_only_local');\n+SELECT pg_replication_origin_drop('regress_test_decoding:\nregression_slot_only_local');\n\\ No newline at end of file\n\nAt the end of the file, there should be a new line.\n\n3.\n+ <structfield>subonlylocal</structfield> <type>bool</type>\n+ </para>\n+ <para>\n+ If true, subscription will request the publisher to send locally\n+ originated changes at the publisher node, or send any publisher node\n+ changes regardless of their origin\n\nThe case where this option is not set is not clear from the\ndescription. Can we change the description to: \"If true, the\nsubscription will request that the publisher send locally originated\nchanges. False indicates that the publisher sends any changes\nregardless of their origin.\"\n\n4. This new option 'subonlylocal' is placed before 'substream' in docs\n(catalogs.sgml) and after it in structures in pg_subscription.h. I\nsuggest adding it after 'subdisableonerr' in docs and\npg_subscription.h. Also, adjust other places in subscriber-side code\nto place it consistently.\n\n5.\n@@ -4516,6 +4524,8 @@ getSubscriptions(Archive *fout)\n pg_strdup(PQgetvalue(res, i, i_subtwophasestate));\n subinfo[i].subdisableonerr =\n pg_strdup(PQgetvalue(res, i, i_subdisableonerr));\n+ subinfo[i].subonlylocal =\n+ pg_strdup(PQgetvalue(res, i, i_subonlylocal));\n\n /* Decide whether we want to dump it */\n selectDumpableObject(&(subinfo[i].dobj), fout);\n@@ -4589,6 +4599,9 @@ dumpSubscription(Archive *fout, const\nSubscriptionInfo *subinfo)\n if (strcmp(subinfo->subdisableonerr, \"t\") == 0)\n appendPQExpBufferStr(query, \", disable_on_error = true\");\n\n+ if (strcmp(subinfo->subonlylocal, \"t\") == 0)\n+ appendPQExpBufferStr(query, \", only_local = true\");\n+\n if (strcmp(subinfo->subsynccommit, \"off\") != 0)\n appendPQExpBuffer(query, \", synchronous_commit = %s\",\nfmtId(subinfo->subsynccommit));\n\ndiff --git a/src/bin/pg_dump/pg_dump.h b/src/bin/pg_dump/pg_dump.h\nindex 1d21c2906f..ddb855fd16 100644\n--- a/src/bin/pg_dump/pg_dump.h\n+++ b/src/bin/pg_dump/pg_dump.h\n@@ -661,6 +661,7 @@ typedef struct _SubscriptionInfo\n char *subdisableonerr;\n char *subsynccommit;\n char *subpublications;\n+ char *subonlylocal;\n } SubscriptionInfo;\n\nTo keep this part of the code consistent, I think it is better to\nplace 'subonlylocal' after 'subdisableonerr' in SubscriptionInfo.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 23 May 2022 10:01:12 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, May 23, 2022 at 10:01 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, May 20, 2022 at 3:08 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > >\n> > >\n> > > Few comments on v13-0001\n> > > ======================\n> > > 1.\n> > > + *\n> > > + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\n> > > + * PG16.\n> > > ...\n> > > @@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\n> > > OutputPluginOptions *opt,\n> > > else\n> > > ctx->twophase_opt_given = true;\n> > >\n> > > + if (data->local_only && data->protocol_version <\n> > > LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\n> > > + ereport(ERROR,\n> > > + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\n> > > + errmsg(\"requested proto_version=%d does not support local_only, need\n> > > %d or higher\",\n> > > + data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\n> > >\n> > > What is the need to change the protocol version for this parameter? As\n> > > per my understanding, we only need a different protocol version when\n> > > we are sending some new message or some additional information in an\n> > > existing message as we have done for the streaming/two_phase options\n> > > which doesn't seem to be the case here.\n> >\n> > Modified\n> >\n>\n> It seems you forgot to remove the comments after removing the code\n> corresponding to the above. See below.\n> + *\n> + * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\n> + * support for sending only locally originated data from the publisher.\n> + * Introduced in PG16.\n> + *\n> + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\n> + * PG16.\n> */\n\nRemoved it.\n\n> Few other comments on 0001\n> ========================\n> 1.\n> + * Return true if data has originated remotely when only_local option is\n> + * enabled, false otherwise.\n>\n> Can we slightly change the comment to:\"Return true if the data source\n> (origin) is remote and user has requested only local data, false\n> otherwise.\"\n\nModified\n\n> 2.\n> +SELECT pg_replication_origin_session_reset();\n> +SELECT pg_drop_replication_slot('regression_slot_only_local');\n> +SELECT pg_replication_origin_drop('regress_test_decoding:\n> regression_slot_only_local');\n> \\ No newline at end of file\n>\n> At the end of the file, there should be a new line.\n\nI checked the other regression sql files and found that a newline is\nnot added at the end of the file. I did not make any change for this.\nThoughts?\n\n> 3.\n> + <structfield>subonlylocal</structfield> <type>bool</type>\n> + </para>\n> + <para>\n> + If true, subscription will request the publisher to send locally\n> + originated changes at the publisher node, or send any publisher node\n> + changes regardless of their origin\n>\n> The case where this option is not set is not clear from the\n> description. Can we change the description to: \"If true, the\n> subscription will request that the publisher send locally originated\n> changes. False indicates that the publisher sends any changes\n> regardless of their origin.\"\n\nModified\n\n> 4. This new option 'subonlylocal' is placed before 'substream' in docs\n> (catalogs.sgml) and after it in structures in pg_subscription.h. I\n> suggest adding it after 'subdisableonerr' in docs and\n> pg_subscription.h. Also, adjust other places in subscriber-side code\n> to place it consistently.\n\nModified\n\n> 5.\n> @@ -4516,6 +4524,8 @@ getSubscriptions(Archive *fout)\n> pg_strdup(PQgetvalue(res, i, i_subtwophasestate));\n> subinfo[i].subdisableonerr =\n> pg_strdup(PQgetvalue(res, i, i_subdisableonerr));\n> + subinfo[i].subonlylocal =\n> + pg_strdup(PQgetvalue(res, i, i_subonlylocal));\n>\n> /* Decide whether we want to dump it */\n> selectDumpableObject(&(subinfo[i].dobj), fout);\n> @@ -4589,6 +4599,9 @@ dumpSubscription(Archive *fout, const\n> SubscriptionInfo *subinfo)\n> if (strcmp(subinfo->subdisableonerr, \"t\") == 0)\n> appendPQExpBufferStr(query, \", disable_on_error = true\");\n>\n> + if (strcmp(subinfo->subonlylocal, \"t\") == 0)\n> + appendPQExpBufferStr(query, \", only_local = true\");\n> +\n> if (strcmp(subinfo->subsynccommit, \"off\") != 0)\n> appendPQExpBuffer(query, \", synchronous_commit = %s\",\n> fmtId(subinfo->subsynccommit));\n>\n> diff --git a/src/bin/pg_dump/pg_dump.h b/src/bin/pg_dump/pg_dump.h\n> index 1d21c2906f..ddb855fd16 100644\n> --- a/src/bin/pg_dump/pg_dump.h\n> +++ b/src/bin/pg_dump/pg_dump.h\n> @@ -661,6 +661,7 @@ typedef struct _SubscriptionInfo\n> char *subdisableonerr;\n> char *subsynccommit;\n> char *subpublications;\n> + char *subonlylocal;\n> } SubscriptionInfo;\n>\n> To keep this part of the code consistent, I think it is better to\n> place 'subonlylocal' after 'subdisableonerr' in SubscriptionInfo.\n\nModified\n\nThanks for the comments, the attached v15 patch has the fixes for the same.\n\nRegards,\nVignesh","msg_date":"Mon, 23 May 2022 23:04:03 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, May 24, 2022 1:34 AM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> Thanks for the comments, the attached v15 patch has the fixes for the same.\r\n> \r\n\r\nThanks for updating the patch. I have a comment on the document in 0002 patch.\r\n\r\n@@ -300,6 +310,11 @@ CREATE SUBSCRIPTION <replaceable class=\"parameter\">subscription_name</replaceabl\r\n replication from the publisher. The default is\r\n <literal>false</literal>.\r\n </para>\r\n+ <para>\r\n+ There is some interaction between the <literal>only_local</literal>\r\n+ option and <literal>copy_data</literal> option. Refer to the\r\n+ <xref linkend=\"sql-createsubscription-notes\" /> for details.\r\n+ </para>\r\n </listitem>\r\n </varlistentry>\r\n\r\nThis change is related to \"only_local\" option and \"copy_data\" option, so I think\r\nit should be put together with \"only_local\", instead of \"disable_on_error\".\r\n\r\nBesides, I tested some cross version cases, pg_dump and describe command, the\r\nresults are all as expected.\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Tue, 24 May 2022 02:36:48 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, May 23, 2022 at 10:01 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, May 20, 2022 at 3:08 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > >\n> > >\n> > > Few comments on v13-0001\n> > > ======================\n> > > 1.\n> > > + *\n> > > + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\n> > > + * PG16.\n> > > ...\n> > > @@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\n> > > OutputPluginOptions *opt,\n> > > else\n> > > ctx->twophase_opt_given = true;\n> > >\n> > > + if (data->local_only && data->protocol_version <\n> > > LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\n> > > + ereport(ERROR,\n> > > + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\n> > > + errmsg(\"requested proto_version=%d does not support local_only, need\n> > > %d or higher\",\n> > > + data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\n> > >\n> > > What is the need to change the protocol version for this parameter? As\n> > > per my understanding, we only need a different protocol version when\n> > > we are sending some new message or some additional information in an\n> > > existing message as we have done for the streaming/two_phase options\n> > > which doesn't seem to be the case here.\n> >\n> > Modified\n> >\n>\n> It seems you forgot to remove the comments after removing the code\n> corresponding to the above. See below.\n> + *\n> + * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\n> + * support for sending only locally originated data from the publisher.\n> + * Introduced in PG16.\n> + *\n> + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\n> + * PG16.\n> */\n>\n> Few other comments on 0001\n> ========================\n> 1.\n> + * Return true if data has originated remotely when only_local option is\n> + * enabled, false otherwise.\n>\n> Can we slightly change the comment to:\"Return true if the data source\n> (origin) is remote and user has requested only local data, false\n> otherwise.\"\n>\n> 2.\n> +SELECT pg_replication_origin_session_reset();\n> +SELECT pg_drop_replication_slot('regression_slot_only_local');\n> +SELECT pg_replication_origin_drop('regress_test_decoding:\n> regression_slot_only_local');\n> \\ No newline at end of file\n>\n> At the end of the file, there should be a new line.\n>\n\nThanks for pointing this out, for some reason vim does not show these\nnew lines at the end of the file, VS Code shows the last new line. I\nhave added it.\n\nThe attached v16 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Wed, 25 May 2022 17:25:02 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, May 24, 2022 at 8:06 AM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Tue, May 24, 2022 1:34 AM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached v15 patch has the fixes for the same.\n> >\n>\n> Thanks for updating the patch. I have a comment on the document in 0002 patch.\n>\n> @@ -300,6 +310,11 @@ CREATE SUBSCRIPTION <replaceable class=\"parameter\">subscription_name</replaceabl\n> replication from the publisher. The default is\n> <literal>false</literal>.\n> </para>\n> + <para>\n> + There is some interaction between the <literal>only_local</literal>\n> + option and <literal>copy_data</literal> option. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for details.\n> + </para>\n> </listitem>\n> </varlistentry>\n>\n> This change is related to \"only_local\" option and \"copy_data\" option, so I think\n> it should be put together with \"only_local\", instead of \"disable_on_error\".\n\nModified.\n\nThe v16 patch attached at [1] has the changes for the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm2_Ytder-6C68ia%3Dm39cmknAhxf2KGkeNAtxt84MxMT3w%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 25 May 2022 17:26:32 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for patch v15-0001.\n\n======\n\n1. Commit message\n\nShould this also say new test cases were added for the test_decoding plugin?\n\n~~~\n\n2. contrib/test_decoding/sql/replorigin.sql\n\n@@ -119,3 +119,18 @@ SELECT data FROM\npg_logical_slot_get_changes('regression_slot_no_lsn', NULL, NUL\n SELECT pg_replication_origin_session_reset();\n SELECT pg_drop_replication_slot('regression_slot_no_lsn');\n SELECT pg_replication_origin_drop('regress_test_decoding:\nregression_slot_no_lsn');\n+\n+-- Verify that remote origin data is not returned with only-local option\n+SELECT 'init' FROM\npg_create_logical_replication_slot('regression_slot_only_local',\n'test_decoding');\n+SELECT pg_replication_origin_create('regress_test_decoding:\nregression_slot_only_local');\n+SELECT pg_replication_origin_session_setup('regress_test_decoding:\nregression_slot_only_local');\n+INSERT INTO origin_tbl(data) VALUES ('only_local, commit1');\n+-- remote origin data returned when only-local option is not set\n+SELECT data FROM\npg_logical_slot_get_changes('regression_slot_only_local', NULL, NULL,\n'skip-empty-xacts', '1', 'include-xids', '0', 'only-local', '0');\n+INSERT INTO origin_tbl(data) VALUES ('only_local, commit2');\n+-- remote origin data not returned when only-local option is set\n+SELECT data FROM\npg_logical_slot_get_changes('regression_slot_only_local', NULL, NULL,\n'skip-empty-xacts', '1', 'include-xids', '0', 'only-local', '1');\n+-- Clean up\n+SELECT pg_replication_origin_session_reset();\n+SELECT pg_drop_replication_slot('regression_slot_only_local');\n+SELECT pg_replication_origin_drop('regress_test_decoding:\nregression_slot_only_local');\n\nAll the new comments should consistently start with upper case.\n\n~~~\n\n3. doc/src/sgml/catalogs.sgml\n\n+ <para>\n+ If true, the subscription will request that the publisher send locally\n+ originated changes. False indicates that the publisher sends any changes\n+ regardless of their origin.\n+ </para></entry>\n\nSUGGESTION\nIf true, the subscription will request the publisher to only send\nchanges that originated locally.\n\n~~~\n\n4. doc/src/sgml/ref/create_subscription.sgml\n\n+ <para>\n+ Specifies whether the subscription will request the publisher to send\n+ locally originated changes at the publisher node, or send any\n+ publisher node changes regardless of their origin. The default is\n+ <literal>false</literal>.\n+ </para>\n\nThis wording should be more similar to the same information in catalogs.sgml\n\nSUGGESTION\nSpecifies whether the subscription will request the publisher to only\nsend changes that originated locally, or to send any changes\nregardless of origin.\n\n~~~\n\n5. src/include/replication/walreceiver.h\n\n@@ -183,6 +183,7 @@ typedef struct\n bool streaming; /* Streaming of large transactions */\n bool twophase; /* Streaming of two-phase transactions at\n * prepare time */\n+ bool only_local; /* publish only locally originated data */\n } logical;\n } proto;\n } WalRcvStreamOptions;\n\nSUGGESTION\n/* Publish only local origin data */\n\n~~~\n\n6. src/test/subscription/t/032_onlylocal.pl - cosmetic changes\n\n6a.\n+# Setup a bidirectional logical replication between Node_A & Node_B\n\nSUGGESTION\n\"... node_A and node_B\"\n\n6b.\n+is(1, 1, \"Circular replication setup is complete\");\n\nSUGGESTION (or maybe saying \"circular\" is also OK - I wasn't sure)\n\"Bidirectional replication setup is complete\"\n\n6c.\n+# check that bidirectional logical replication setup...\n\nStart comment sentence with upper case.\n\n6d.\n+###############################################################################\n+# check that remote data that is originated from node_C to node_B is not\n+# published to node_A\n+###############################################################################\n\nSUGGESTION\nCheck that remote data of node_B (that originated from node_C) is not\npublished to node_A\n\n6e.\n+is($result, qq(11\n+12\n+13), 'Node_C data replicated to Node_B'\n+);\n\nSUGGESTION for message\n'node_C data replicated to node_B'\n\n6f.\n+is($result, qq(11\n+12), 'Remote data originated from other node is not replicated when\nonly_local option is ON'\n+);\n\nSUGGESTION for message\n'Remote data originating from another node (not the publisher) is not\nreplicated when only_local option is ON'\n\n6g.\n\"Circular replication setup is complete\"\n'Inserted successfully without leading to infinite recursion in\nbidirectional replication setup'\n'Inserted successfully without leading to infinite recursion in\nbidirectional replication setup'\n'Node_C data replicated to Node_B'\n'Remote data originated from other node is not replicated when\nonly_local option is ON'\n\nWhy do some of the \"is\" messages have single quotes and others have\ndouble quotes? Should be consistent.\n\n~~~\n\n7. src/test/subscription/t/032_onlylocal.pl\n\n+my $appname_B2 = 'tap_sub_B2';\n+$node_B->safe_psql(\n+ 'postgres', \"\n+ CREATE SUBSCRIPTION tap_sub_B2\n+ CONNECTION '$node_C_connstr application_name=$appname_B2'\n+ PUBLICATION tap_pub_C\n+ WITH (only_local = on)\");\n+\n\nAFAIK the \"WITH (only_local = on)\" is unnecessary here. We don't care\nwhere the node_C data came from for this test case.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 26 May 2022 11:36:29 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for patch v15-0001.\n>\n> ======\n>\n> 1. Commit message\n>\n> Should this also say new test cases were added for the test_decoding plugin?\n>\n> ~~~\n\nShall we take this out as a separate test-case patch? I think it is a\ngood idea to do a missing test case for existing functionality.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 26 May 2022 09:50:13 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, May 26, 2022 at 2:20 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Here are my review comments for patch v15-0001.\n> >\n> > ======\n> >\n> > 1. Commit message\n> >\n> > Should this also say new test cases were added for the test_decoding plugin?\n> >\n> > ~~~\n>\n> Shall we take this out as a separate test-case patch? I think it is a\n> good idea to do a missing test case for existing functionality.\n>\n\n+1. I also think this new test could be posted separately.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 26 May 2022 14:59:41 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n>\n> 3. doc/src/sgml/catalogs.sgml\n>\n> + <para>\n> + If true, the subscription will request that the publisher send locally\n> + originated changes. False indicates that the publisher sends any changes\n> + regardless of their origin.\n> + </para></entry>\n>\n> SUGGESTION\n> If true, the subscription will request the publisher to only send\n> changes that originated locally.\n>\n> ~~~\n>\n\nI think it is a good idea to keep the description related to the\n'False' value as without that it is slightly unclear.\n\n> 4. doc/src/sgml/ref/create_subscription.sgml\n>\n> + <para>\n> + Specifies whether the subscription will request the publisher to send\n> + locally originated changes at the publisher node, or send any\n> + publisher node changes regardless of their origin. The default is\n> + <literal>false</literal>.\n> + </para>\n>\n> This wording should be more similar to the same information in catalogs.sgml\n>\n> SUGGESTION\n> Specifies whether the subscription will request the publisher to only\n> send changes that originated locally, or to send any changes\n> regardless of origin.\n>\n> ~~~\n>\n\nHere, also, let's keep the wording related to the default value.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 26 May 2022 10:38:42 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, May 26, 2022 at 3:08 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> >\n> > 3. doc/src/sgml/catalogs.sgml\n> >\n> > + <para>\n> > + If true, the subscription will request that the publisher send locally\n> > + originated changes. False indicates that the publisher sends any changes\n> > + regardless of their origin.\n> > + </para></entry>\n> >\n> > SUGGESTION\n> > If true, the subscription will request the publisher to only send\n> > changes that originated locally.\n> >\n> > ~~~\n> >\n>\n> I think it is a good idea to keep the description related to the\n> 'False' value as without that it is slightly unclear.\n>\n> > 4. doc/src/sgml/ref/create_subscription.sgml\n> >\n> > + <para>\n> > + Specifies whether the subscription will request the publisher to send\n> > + locally originated changes at the publisher node, or send any\n> > + publisher node changes regardless of their origin. The default is\n> > + <literal>false</literal>.\n> > + </para>\n> >\n> > This wording should be more similar to the same information in catalogs.sgml\n> >\n> > SUGGESTION\n> > Specifies whether the subscription will request the publisher to only\n> > send changes that originated locally, or to send any changes\n> > regardless of origin.\n> >\n> > ~~~\n> >\n>\n> Here, also, let's keep the wording related to the default value.\n>\n> --\n\nI agree. Probably I was not being clear enough with those review comments.\n\nMy suggestions were only meant as substitutes for the FIRST sentences\nof those patch fragments, not the entire paragraph.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 26 May 2022 15:38:20 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"This post completes my review of patch v15*.\n\n======\n\n1. A general comment affecting both patches\n\nI realised the things defined in the subscription WITH are referred to\nin the PG DOCS as \"parameters\" (not \"options\"). See CREATE\nSUBSCRIPTION [1]. This is already reflected in some of my review\ncomments for the v15-0002 below., but I'd already posted my v15-0001\nreview comments before noticing this.\n\nPlease also go through the v15-0001 patch to make all the necessary\nterminology changes (including in the v15-0001 commit message).\n\n=====\n\nPlease find below review comments for patch v15-0002.\n\n2. Commit message\n\nThis patch does couple of things:\nchange 1) Added force option for copy_data.\nchange 2) Check and throw an error if the publication tables were also\nsubscribing data in the publisher from other publishers.\n\n--\n\n2a.\n\"couple\" -> \"a couple\"\n\"Added force\" -> \"Adds force ...\"\n\"Check and throw\" -> \"Checks and throws ...\"\n\n2b.\nIsn't \"Change 2)\" only applicable when copy_data and only_local are\non? It should say so here.\n\n~~~\n\n3. Commit message\n\nnode1: Table t1 (c1 int) has data 1, 2, 3, 4\nnode2: Table t1 (c1 int) has data 5, 6, 7, 8\n\n--\n\nI feel these kinds of replication examples are always much easier to\nvisualize when the data somehow indicates where it came from. E.g.\nconsider using different data in your example like:\n\nnode1: Table t1 (c1 int) has data 11, 12, 13, 14\nnode2: Table t1 (c1 int) has data 21, 22, 23, 24\n\n~~~\n\n4. Commit message\n\nExamples in this message exceed 80 chars. Please wrap the SQL appropriately\n\n~~~\n\n5. Commit message\n\nAfter this the data will be something like this:\nnode1:\n1, 2, 3, 4, 5, 6, 7, 8\n\nnode2:\n1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\n\nSo, you can see that data on node2 (5, 6, 7, 8) is duplicated.\n\n--\n\nYeah, but is that entirely correct though? AFAIK the scenario\ndescribed by this example is going to recurse forever.\n\n~~~\n\n6. Commit message\n\nThis problem can be solved by using only_local and copy_data option as\ngiven below\n\n--\n\n\"option\" -> \"parameters\"\n\n~~~\n\n7. Commit message\n\nThere are other minor review comments but it's too messy to report in\ndetail here. I suggest checking all the \":\" (some were missing) and\nthe letter case of/within the sentences. I also suggest\ncutting/pasting all this text into MSWord or some grammar checker to\ncorrect anything else reported.\n\n~~~\n\n8. Commit message\n\nCREATE SUBSCRIPTION sub_node2_node1 CONNECTION '<node1 details>'\nPUBLICATION pub_node1 WITH (copy_data = force, only_local = on);\nERROR: CREATE/ALTER SUBSCRIPTION with only_local and copy_data as\ntrue is not allowed when the publisher might have replicated data\n\n--\n\nLooks like a typo: e.g. If copy_data is \"force\" then how is the error happening?\n\n~~~\n\n9. Commit message\n\nThe examples seem confused/muddled in some way IMO.\n\nIIUC the new ERROR is like a new integrity check for when the\nsubscription is using the only_local = on, along with copy_data = on.\n\nAnd so then the copy_data = force was added as a way to override that\nprevious error being reported.\n\nSo unless you know about the possible error then the \"force\" option\nmakes no sense. So I think you need to change the order of the\nexplanations in the commit message to describe the new error first.\n\n======\n\n10. doc/src/sgml/logical-replication.sgml\n\n+\n+ <sect1 id=\"bidirectional-logical-replication\">\n+ <title>Bidirectional logical replication</title>\n+\n+ <sect2 id=\"setting-bidirectional-replication-two-nodes\">\n+ <title>Setting bidirectional replication between two nodes</title>\n+ <para>\n+ Bidirectional replication is useful in creating a multi-master database\n+ which helps in performing read/write operations from any of the nodes.\n+ Setting up bidirectional logical replication between two nodes requires\n+ creation of a publication in all the nodes, creating subscriptions in\n+ each of the nodes that subscribes to data from all the nodes. The steps\n+ to create a two-node bidirectional replication when there is no data in\n+ both the nodes are given below:\n+ </para>\n\nThis description is confusing. It seems to be trying to say something\nabout a multi-master system (e.g. you say \"all the nodes\") yet this is\nall written inside the section about \"... two nodes\", so saying \"all\nthe nodes\" makes no sense here.\n\nI think you need to split this information and put some kind of blurb\ntext at the top level (section 31.11), and then this section (for \"two\nnodes\") should ONLY be talking about the case when there are just TWO\nnodes.\n\n~~~\n\n11. doc/src/sgml/logical-replication.sgml\n\n+ <para>\n+ Lock the required tables in <literal>node1</literal> and\n+ <literal>node2</literal> till the setup is completed.\n+ </para>\n\n\"till\" -> \"until\" (make this same change in multiple places)\n\n~~~\n\n12. doc/src/sgml/logical-replication.sgml\n\n+ <sect2 id=\"add-new-node-data-in-existing-node\">\n+ <title>Adding a new node when data is present in the existing nodes</title>\n+ <para>\n+ Adding a new node <literal>node3</literal> to the existing\n+ <literal>node1</literal> and <literal>node2</literal> when data is present\n+ in existing nodes <literal>node1</literal> and <literal>node2</literal>\n+ needs similar steps. The only change required here is that\n+ <literal>node3</literal> should create a subscription with\n+ <literal>copy_data = force</literal> to one of the existing nodes to\n+ receive the existing data during initial data synchronization.\n+ </para>\n\nI think perhaps this should clarify that there is NO data in node3 for\nthis example.\n\n~~~\n\n13. doc/src/sgml/logical-replication.sgml - section 31.11.3\n\n+ <para>\n+ Lock the required tables in <literal>node2</literal> and\n+ <literal>node3</literal> till the setup is completed.\n+ </para>\n\nShould you describe why you say no lock is required for tables on\nnode1? (Similar information is in section 31.11.4 also)\n\n~~~\n\n14. doc/src/sgml/logical-replication.sgml - section 31.11.3\n\n+ <para>\n+ Create a subscription in <literal>node3</literal> to subscribe to\n+ <literal>node2</literal>:\n+<programlisting>\n+node3=# CREATE SUBSCRIPTION sub_node3_node2\n+node3-# CONNECTION 'dbname=foo host=node2 user=repuser'\n+node3-# PUBLICATION pub_node2\n+node3-# WITH (copy_data = off, only_local = on);\n+CREATE SUBSCRIPTION\n+</programlisting></para>\n\nUsing a similar description as in section 31.11.4 this should probably\nalso say something like:\n\"Use copy_data specified as off because the initial table data would\nhave been already copied in the previous step\".\n\n~~~\n\n15. doc/src/sgml/logical-replication.sgml - section 31.11.4\n\n+ <sect2 id=\"add-node-data-present-in-new-node\">\n+ <title>Adding a new node when data is present in the new node</title>\n+ <para>\n+ Adding a new node <literal>node3</literal> to the existing\n+ <literal>node1</literal> and <literal>node2</literal> when data is present\n+ in the new node <literal>node3</literal> needs similar steps. A\nfew changes\n+ are required here to get the existing data from <literal>node3</literal>\n+ to <literal>node1</literal> and <literal>node2</literal> and later\n+ cleaning up of data in <literal>node3</literal> before synchronization of\n+ all the data from the existing nodes.\n+ </para>\n\nI think perhaps this should clarify that there is NO data in node1 or\nnode2 for this example.\n\n~~~\n\n16. doc/src/sgml/logical-replication.sgml\n\n+ <sect2 id=\"generic-steps-add-new-node\">\n+ <title>Generic steps to add a new node to the existing set of nodes</title>\n\nI think all the steps in this section should be numbered because the\norder is important.\n\n~~~\n\n17. doc/src/sgml/logical-replication.sgml\n\n+ <sect2 id=\"generic-steps-add-new-node\">\n+ <title>Generic steps to add a new node to the existing set of nodes</title>\n\nIIUC the word \"parameter\" should replace the word \"option\" in all the\ntext of this section.\n\n~~~\n\n18. doc/src/sgml/logical-replication.sgml - Notes\n\n+ <sect2>\n+ <title>Notes</title>\n+ <para>\n+ Setting up bidirectional logical replication across nodes requires multiple\n+ steps to be performed on various nodes, as all operations are not\n+ transactional, user is advised to take backup of existing data to avoid any\n+ inconsistency.\n+ </para>\n+ </sect2>\n\n18a.\nI thought this \"Notes\" section should be rendered more like an sgml\nnote or a warning. E.g. like the warning on this page [2].\n\n18b.\nSUGGESTION (split the para into multiple sentences). e.g.\nSetting up bidirectional logical replication across nodes requires\nmultiple steps to be performed on various nodes. Because not all\noperations are transactional, the user is advised to take backups.\n\n======\n\n19. doc/src/sgml/ref/alter_subscription.sgml\n\n+ <para>\n+ There is some interaction between the <literal>only_local</literal>\n+ option and <literal>copy_data</literal> option. Refer to the\n+ <command>CREATE SUBSCRIPTION</command>\n+ <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n+ details and usage of <literal>force</literal> for\n+ <literal>copy_data</literal> option.\n </para>\n\n\"option\" -> \"parameter\" (3x)\n\n======\n\n20. doc/src/sgml/ref/create_subscription.sgml\n\n+ <para>\n+ There is some interaction between the <literal>only_local</literal>\n+ option and <literal>copy_data</literal> option. Refer to the\n+ <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n+ details and usage of <literal>force</literal> for\n+ <literal>copy_data</literal> option.\n+ </para>\n\n\"option\" -> \"parameter\" (3x)\n\n======\n\n21. src/backend/commands/subscriptioncmds.c\n\n@@ -69,6 +69,18 @@\n /* check if the 'val' has 'bits' set */\n #define IsSet(val, bits) (((val) & (bits)) == (bits))\n\n+#define IS_COPY_DATA_ON_OR_FORCE(copy_data) ((copy_data) != COPY_DATA_OFF)\n+\n+/*\n+ * Represents whether copy_data option is specified with off, on or force.\n+ */\n+typedef enum CopyData\n+{\n+ COPY_DATA_OFF,\n+ COPY_DATA_ON,\n+ COPY_DATA_FORCE\n+} CopyData;\n+\n\n21a.\n\"option\" -> \"parameter\"\n\n21b.\nI think the new #define is redundant and can be removed (see the next\nreview comment)\n\n21c.\nI think you set the OFF enum value like:\ntypedef enum CopyData\n{\n COPY_DATA_OFF = 0,\n COPY_DATA_ON,\n COPY_DATA_FORCE\n} CopyData;\n\nthen it will greatly simplify things; many changes in this file will\nbe unnecessary (see subsequent review comments)\n\n~~~\n\n22. src/backend/commands/subscriptioncmds.c\n\n+/*\n+ * Validate the value specified for copy_data option.\n+ */\n+static CopyData\n+DefGetCopyData(DefElem *def)\n\n\"option\" -> \"parameter\"\n\n~~~\n\n23. src/backend/commands/subscriptioncmds.c\n\n+ /*\n+ * Allow 0, 1, \"true\", \"false\", \"on\", \"off\" or \"force\".\n+ */\n+ switch (nodeTag(def->arg))\n+ {\n+ case T_Integer:\n+ switch (intVal(def->arg))\n+ {\n+ case 0:\n+ return COPY_DATA_OFF;\n+ case 1:\n+ return COPY_DATA_ON;\n+ default:\n+ /* otherwise, error out below */\n+ break;\n+ }\n+ break;\n\nWhat about also allowing copy_data = 2, and making it equivalent to \"force\"?\n\n~~~\n\n24. src/backend/commands/subscriptioncmds.c\n\n@@ -333,17 +400,17 @@ parse_subscription_options(ParseState *pstate,\nList *stmt_options,\n errmsg(\"%s and %s are mutually exclusive options\",\n \"connect = false\", \"create_slot = true\")));\n\n- if (opts->copy_data &&\n+ if (IS_COPY_DATA_ON_OR_FORCE(opts->copy_data) &&\n IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\n\nI think this change would not be needed if you set the OFF enum = 0.\n\n~~~\n25. src/backend/commands/subscriptioncmds.c\n\n@@ -671,13 +738,14 @@ CreateSubscription(ParseState *pstate,\nCreateSubscriptionStmt *stmt,\n * Set sync state based on if we were asked to do data copy or\n * not.\n */\n- table_state = opts.copy_data ? SUBREL_STATE_INIT : SUBREL_STATE_READY;\n+ table_state = IS_COPY_DATA_ON_OR_FORCE(opts.copy_data) ?\nSUBREL_STATE_INIT : SUBREL_STATE_READY;\n\nI think this change would not be needed if you set the OFF enum = 0.\n\n~~~\n\n26. src/backend/commands/subscriptioncmds.c\n\n@@ -720,7 +788,8 @@ CreateSubscription(ParseState *pstate,\nCreateSubscriptionStmt *stmt,\n * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\n * PUBLICATION to work.\n */\n- if (opts.twophase && !opts.copy_data && tables != NIL)\n+ if (opts.twophase && opts.copy_data == COPY_DATA_OFF &&\n+ tables != NIL)\n twophase_enabled = true;\n\nI think this change would not be needed if you set the OFF enum = 0.\n\n~~~\n\n27. src/backend/commands/subscriptioncmds.c\n\n@@ -851,7 +921,7 @@ AlterSubscription_refresh(Subscription *sub, bool copy_data,\n list_length(subrel_states), sizeof(Oid), oid_cmp))\n {\n AddSubscriptionRelState(sub->oid, relid,\n- copy_data ? SUBREL_STATE_INIT : SUBREL_STATE_READY,\n+ IS_COPY_DATA_ON_OR_FORCE(copy_data) ? SUBREL_STATE_INIT : SUBREL_STATE_READY,\n InvalidXLogRecPtr);\n\nI think this change would not be needed if you set the OFF enum = 0.\n\n\n~~~\n28. src/backend/commands/subscriptioncmds.c\n\n@@ -1157,7 +1227,7 @@ AlterSubscription(ParseState *pstate,\nAlterSubscriptionStmt *stmt,\n * See ALTER_SUBSCRIPTION_REFRESH for details why this is\n * not allowed.\n */\n- if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\n+ if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\nIS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\n ereport(ERROR,\n (errcode(ERRCODE_SYNTAX_ERROR),\n errmsg(\"ALTER SUBSCRIPTION with refresh and copy_data is not allowed\nwhen two_phase is enabled\"),\n\nI think this change would not be needed if you set the OFF enum = 0.\n\n~~~\n\n29. src/backend/commands/subscriptioncmds.c\n\n@@ -1209,7 +1279,7 @@ AlterSubscription(ParseState *pstate,\nAlterSubscriptionStmt *stmt,\n * See ALTER_SUBSCRIPTION_REFRESH for details why this is\n * not allowed.\n */\n- if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\n+ if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\nIS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\n ereport(ERROR,\n (errcode(ERRCODE_SYNTAX_ERROR),\n errmsg(\"ALTER SUBSCRIPTION with refresh and copy_data is not allowed\nwhen two_phase is enabled\"),\n\nI think this change would not be needed if you set the OFF enum = 0.\n\n~~~\n\n30. src/backend/commands/subscriptioncmds.c\n\n@@ -1255,7 +1325,8 @@ AlterSubscription(ParseState *pstate,\nAlterSubscriptionStmt *stmt,\n *\n * For more details see comments atop worker.c.\n */\n- if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\n+ if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\n+ IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\n ereport(ERROR,\n (errcode(ERRCODE_SYNTAX_ERROR),\n errmsg(\"ALTER SUBSCRIPTION ... REFRESH with copy_data is not allowed\nwhen two_phase is enabled\"),\n\nI think this change would not be needed if you set the OFF enum = 0.\n\n~~~\n\n31. src/backend/commands/subscriptioncmds.c\n\n+ appendStringInfoString(&cmd,\n+ \"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\nPS.srrelid as replicated\\n\"\n+ \"FROM pg_publication P,\\n\"\n+ \"LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n+ \"LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n+ \"pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n+ \"WHERE C.oid = GPT.relid AND P.pubname in (\");\n\nuse upper case \"in\" -> \"IN\"\n\n======\n\n32. src/test/regress/sql/subscription.sql\n\n@@ -40,6 +40,9 @@ SET SESSION AUTHORIZATION 'regress_subscription_user';\n\n -- fail - invalid option combinations\n CREATE SUBSCRIPTION regress_testsub2 CONNECTION\n'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\nfalse, copy_data = true);\n+CREATE SUBSCRIPTION regress_testsub2 CONNECTION\n'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\nfalse, copy_data = on);\n+CREATE SUBSCRIPTION regress_testsub2 CONNECTION\n'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\nfalse, copy_data = 1);\n+CREATE SUBSCRIPTION regress_testsub2 CONNECTION\n'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\nfalse, copy_data = force);\n\n32a.\nThe test for \"copy_data = 1\" should confirm it is the same as \"on\".\n\n32b.\nShould also test copy_data = 2 (and confirm it is the same as \"force\").\n\n======\n\n33. src/test/subscription/t/032_localonly.pl - cosmetic changes\n\n33a.\n+# Detach node C from the node-group of (A, B, C) and clean the table contents\n+# from all nodes.\n\nSUGGESTION\nDetach node_C from the node-group of (node_A, node_B, node_C) and ...\n\n33b.\n+# Subroutine to create subscription and wait till the initial sync is\ncompleted.\n\n\"till\" -> \"until\"\n\n33c.\n+# Subroutine to create subscription and wait till the initial sync is\ncompleted.\n+# Subroutine expects subscriber node, publisher node, subscription name,\n+# destination connection string, publication name and the subscription with\n+# options to be passed as input parameters.\n+sub create_subscription\n+{\n+ my ($node_subscriber, $node_publisher, $sub_name, $node_connstr,\n+ $pub_name, $with_options)\n+ = @_;\n\n\"subscription with options\" => \"subscription parameters\"\n\"$with_options\" -> \"$sub_params\"\n\n33d.\n+# Specifying only_local 'on' which indicates that the publisher should only\n\n\"Specifying\" => \"Specify\"\n\n------\n[1] https://www.postgresql.org/docs/15/sql-createsubscription.html\n[2] https://www.postgresql.org/docs/current/bgworker.html\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 26 May 2022 18:36:41 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, May 20, 2022 at 3:31 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Wed, May 18, 2022 at 4:22 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > 5.\n> > * It is quite possible that subscriber has not yet pulled data to\n> > + * the tables, but in ideal cases the table data will be subscribed.\n> > + * To keep the code simple it is not checked if the subscriber table\n> > + * has pulled the data or not.\n> > + */\n> > + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\n> >\n> > Sorry, but I don't understand what you intend to say by the above\n> > comment. Can you please explain?\n>\n> When the user specifies copy_data as on, we should check if the\n> publisher has the publication tables being subscribed from a remote\n> publisher. If so throw an error as it remote origin data present.\n> Ex:\n> Node1 - pub1 for table t1 -- no data\n> Node2 - Sub1 subscribing to data from pub1\n> Node2 - pub2 for table t1 -- no data\n> Node3 - create subscription to Node2 with copy_data = ON\n>\n> In this case even though the table does not have any remote origin\n> data, as Node2 is subscribing to data from Node1, throw an error.\n> We throw an error irrespective of data present in Node1 or not to keep\n> the code simple.\n>\n\nI think we can change the contents of comments something like: \"Throw\nan error if the publisher has subscribed to the same table from some\nother publisher. We cannot differentiate between the local and\nnon-local data that is present in the HEAP during the initial sync.\nIdentification of local data can be done only from the WAL by using\nthe origin id. XXX: For simplicity, we don't check whether the table\nhas any data or not. If the table doesn't have any data then we don't\nneed to distinguish between local and non-local data so we can avoid\nthrowing error in that case.\"\n\nFew more comments:\n==================\nPatch 0002\n======\n1.\n+ if (copydata == COPY_DATA_ON && only_local && !slot_attisnull(slot, 3))\n+ ereport(ERROR,\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n+ errmsg(\"CREATE/ALTER SUBSCRIPTION with only_local and copy_data as\ntrue is not allowed when the publisher might have replicated data,\ntable:%s.%s might have replicated data in the publisher\",\n+ nspname, relname),\n+ errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\"));\n\nCan we split the error message? errmsg: table:%s.%s has replicated\ndata in the publisher; errdetail:CREATE/ALTER SUBSCRIPTION with\nonly_local and copy_data as true is not allowed when the publisher has\nreplicated data\n\n2.\n+ <para>\n+ Lock the required tables in the new node until the setup is complete.\n+ </para>\n+ <para>\n+ Create subscriptions on existing nodes pointing to publication on\n+ the new node with <literal>only_local</literal> option specified as\n+ <literal>on</literal> and <literal>copy_data</literal> specified as\n+ <literal>on</literal>.\n+ </para>\n+ <para>\n+ Wait for data to be copied from the new node to existing nodes.\n+ </para>\n+ <para>\n+ Alter the publication in new node so that the truncate operation is not\n+ replicated to the subscribers.\n+ </para>\n\nHere and at other places, we should specify that the lock mode should\nto acquire the lock on table should be EXCLUSIVE so that no concurrent\nDML is allowed on it. Also, it is better if somewhere we explain why\nand which nodes need locks?\n\nPatch 0001:\n==========\n1.\n+$node_A->append_conf(\n+ 'postgresql.conf', qq(\n+max_prepared_transactions = 10\n+logical_decoding_work_mem = 64kB\n+));\n\nI don't see why you need to set these parameters. There is no test\ncase that needs these parameters. Please remove these from here and\nall other similar places in 032_onlylocal.pl.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 26 May 2022 16:40:22 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Apr 6, 2022 at 9:36 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Apr 5, 2022 at 7:06 PM Ashutosh Bapat\n> <ashutosh.bapat.oss@gmail.com> wrote:\n> >\n> > On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > >\n> > > Below are some other name ideas. Maybe they are not improvements, but\n> > > it might help other people to come up with something better.\n> > >\n> > > subscribe_publocal_only = true/false\n> > > origin = pub_only/any\n> > > adjacent_data_only = true/false\n> > > direct_subscriptions_only = true/false\n> > > ...\n> >\n> > FWIW, The subscriber wants \"changes originated on publisher\". From\n> > that angle origin = publisher/any looks attractive. It also leaves\n> > open the possibility that the subscriber may ask changes from a set of\n> > origins or even non-publisher origins.\n> >\n>\n> So, how are you imagining extending it for multiple origins? I think\n> we can have multiple origins data on a node when there are multiple\n> subscriptions pointing to the different or same node. The origin names\n> are internally generated names but are present in\n> pg_replication_origin. So, are we going to ask the user to check that\n> table and specify it as an option? But, note, that the user may not be\n> able to immediately recognize which origin data is useful for her.\n>\n\nI still don't have a very clear answer for the usability aspect but it\nseems this was discussed in PGCon-Unconference [1] (Section: Origin\nfilter) and there also it was suggested to allow users to specify\nmultiple origin names. So, probably Ashutosh's idea makes sense and we\nshould use \"origin = publisher/any or origin=local/any\". Among these,\nI would prefer later.\n\nNote: I don't have email-id of Julian Markwort who has raised this\ntopic in Unconference. Euler, if you know, it might be better to add\nhim here, so that he is aware of what is going on and share his\nopinion if he wishes to.\n\n[1] - https://wiki.postgresql.org/wiki/PgCon_2022_Developer_Unconference\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 27 May 2022 07:54:02 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are some more review comments for v15-0002 I found while working\nthrough the documented examples. These are all for\ndoc/src/sgml/logical-replication.sgml. The section numbers (e.g.\n31.11.1) are the assigned numbers after HTML rendering.\n\nThe general comment is that the sections appeared at first to be\nindependent of each other, but they are not really - sometimes\nexisting data and existing pub/sub are assumed to remain. Therefore, I\nthink some more text needs to be added to clarify the initial state\nfor each of these sections.\n\n======\n\n1. Section id.\n\n<sect1 id=\"bidirectional-logical-replication\">\n\nIMO this section ID should be renamed\n\"logical-replication-bidirectional\" then it will be more consistent\nfilenames with all the other logical replication section names.\n\n~~~\n\n2. Section 31.11.2. Adding a new node when there is no data in any of the nodes\n\n2a.\nThis seems like a continuation of 31.11.1 because pub/sub of\nnode1,node2 is assumed to exist still.\n\n2b.\nThe bottom of this section should say after these steps any subsequent\nincremental data changes on any node will be replicated to all nodes\n(node1, node2, node3)\n\n~~~\n\n3. Section 31.11.3. Adding a new node when data is present in the existing nodes\n\n3a.\nThis seems to be a continuation of 31.11.1 because pub/sub (and\ninitial data) of node1/node2 is assumed to exist.\n\n3b.\n\nThe bottom of this section should say after these steps that any\ninitial data of node1/node2 will be seen in node3, and any subsequent\nincremental data changes on any node will be replicated to all nodes\n(node1, node2, node3)\n\n~~~\n\n4. Section 31.11.4. Adding a new node when data is present in the new node\n\n4a.\nThis seems to be a continuation of 31.11.1 because pub/sub (and\ninitial data) of node1/node2 is assumed to exist.\n\n4b.\nIt is not made very clear up-front if the tables on node1 and node2\nare empty or not. Apparently, they are considered NOT empty because\nlater in the example you are using \"force\" when you create the\nsubscription on node3.\n\n4c.\nThe SQL wrapping here is strangely different from others (put the\nsubscription name on 1st line)\nnode3=# CREATE SUBSCRIPTION\nnode3-# sub_node3_node1 CONNECTION 'dbname=foo host=node1 user=repuser'\nnode3-# PUBLICATION pub_node1\nnode3-# WITH (copy_data = force, only_local = on);\nCREATE SUBSCRIPTION\n\n4d.\nThe SQL wrapping here is strangely different from others (put the\nsubscription name on 1st line)\nnode3=# CREATE SUBSCRIPTION\nnode3-# sub_node3_node2 CONNECTION 'dbname=foo host=node2 user=repuser'\nnode3-# PUBLICATION pub_node2\nnode3-# WITH (copy_data = off, only_local = on);\nCREATE SUBSCRIPTION\n\n4e.\nThe bottom of this section should say after this that any initial data\nof node1/node2 will be seen in node3, and any subsequent incremental\ndata changes on any node will be replicated to all nodes (node1,\nnode2, node3)\n\n~~~\n\n5. Section 31.11.5. Generic steps to add a new node to the existing set of nodes\n\n5a\nCreate subscriptions on the new node pointing to publication on the\nfirst node with only_local option specified as on and copy_data option\nspecified as \"force\".\n\n-> that should say \"Create a subscription\" (singular)\n\n5b.\nCreate subscriptions on the new node pointing to publications on the\nremaining node with only_local option specified as on and copy_data\noption specified as off.\n\n-> that should say \"on the remaining node\" (plural)\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 27 May 2022 15:25:36 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, May 25, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> The attached v16 patch has the changes for the same.\r\n> \r\n\r\nThanks for updating the patch.\r\n\r\nSome comments for the document in 0002 patch.\r\n\r\n1.\r\n+ <para>\r\n+ Lock the required tables in <literal>node1</literal> and\r\n+ <literal>node2</literal> till the setup is completed.\r\n+ </para>\r\n+\r\n+ <para>\r\n+ Create a publication in <literal>node1</literal>:\r\n+<programlisting>\r\n+node1=# CREATE PUBLICATION pub_node1 FOR TABLE t1;\r\n+CREATE PUBLICATION\r\n+</programlisting></para>\r\n\r\nIf the table is locked in the very beginning, we will not be able to create the\r\npublication (because the locks have conflict). Maybe we should switch the order\r\nof creating publication and locking tables here.\r\n\r\n2.\r\nIn the case of \"Adding a new node when data is present in the new node\", we need\r\nto truncate table t1 in node3, but the truncate operation would be blocked\r\nbecause the table has be locked before. Maybe we need some changes for it.\r\n\r\nRegards,\r\nShi yu\r\n\r\n","msg_date":"Fri, 27 May 2022 07:04:18 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, May 27, 2022 at 5:04 PM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Wed, May 25, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > The attached v16 patch has the changes for the same.\n> >\n>\n> Thanks for updating the patch.\n>\n> Some comments for the document in 0002 patch.\n>\n> 1.\n> + <para>\n> + Lock the required tables in <literal>node1</literal> and\n> + <literal>node2</literal> till the setup is completed.\n> + </para>\n> +\n> + <para>\n> + Create a publication in <literal>node1</literal>:\n> +<programlisting>\n> +node1=# CREATE PUBLICATION pub_node1 FOR TABLE t1;\n> +CREATE PUBLICATION\n> +</programlisting></para>\n>\n> If the table is locked in the very beginning, we will not be able to create the\n> publication (because the locks have conflict). Maybe we should switch the order\n> of creating publication and locking tables here.\n>\n\nI agree. It seems some of the locking instructions in the earlier\nsections 31.11.1 - 31.11.4 are contradictory to the later \"generic\"\nsteps given in \"31.11.5. Generic steps to add a new node to the\nexisting set of nodes\". I'm assuming the generic steps are the\n\"correct\" steps\n\ne.g. generic steps say get the lock on new node tables AFTER the\npublication of new node.\ne.g. generic steps say do NOT get a table lock on the node one you are\n(first) joining to.\n\n~~~\n\nFurthermore, the generic steps are describing attaching a new N+1th\nnode to some (1 ... N) other nodes.\n\nSo I think in the TWO-node case (section 31.11.1) node2 should be\ntreated as the \"new\" node that you are attaching to the \"first\" node1.\nIMO the section 31.11.1 example should be reversed so that it obeys\nthe \"generic\" pattern.\ne.g. It should be doing the CREATE PUBLICATION pub_node2 first (since\nthat is the \"new\" node)\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 27 May 2022 18:38:25 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for patch v15-0001.\n>\n> ======\n>\n> 1. Commit message\n>\n> Should this also say new test cases were added for the test_decoding plugin?\n\nSeparated the test as a 0001 patch.\n\n> ~~~\n>\n> 2. contrib/test_decoding/sql/replorigin.sql\n>\n> @@ -119,3 +119,18 @@ SELECT data FROM\n> pg_logical_slot_get_changes('regression_slot_no_lsn', NULL, NUL\n> SELECT pg_replication_origin_session_reset();\n> SELECT pg_drop_replication_slot('regression_slot_no_lsn');\n> SELECT pg_replication_origin_drop('regress_test_decoding:\n> regression_slot_no_lsn');\n> +\n> +-- Verify that remote origin data is not returned with only-local option\n> +SELECT 'init' FROM\n> pg_create_logical_replication_slot('regression_slot_only_local',\n> 'test_decoding');\n> +SELECT pg_replication_origin_create('regress_test_decoding:\n> regression_slot_only_local');\n> +SELECT pg_replication_origin_session_setup('regress_test_decoding:\n> regression_slot_only_local');\n> +INSERT INTO origin_tbl(data) VALUES ('only_local, commit1');\n> +-- remote origin data returned when only-local option is not set\n> +SELECT data FROM\n> pg_logical_slot_get_changes('regression_slot_only_local', NULL, NULL,\n> 'skip-empty-xacts', '1', 'include-xids', '0', 'only-local', '0');\n> +INSERT INTO origin_tbl(data) VALUES ('only_local, commit2');\n> +-- remote origin data not returned when only-local option is set\n> +SELECT data FROM\n> pg_logical_slot_get_changes('regression_slot_only_local', NULL, NULL,\n> 'skip-empty-xacts', '1', 'include-xids', '0', 'only-local', '1');\n> +-- Clean up\n> +SELECT pg_replication_origin_session_reset();\n> +SELECT pg_drop_replication_slot('regression_slot_only_local');\n> +SELECT pg_replication_origin_drop('regress_test_decoding:\n> regression_slot_only_local');\n>\n> All the new comments should consistently start with upper case.\n\nModified\n\n> ~~~\n>\n> 3. doc/src/sgml/catalogs.sgml\n>\n> + <para>\n> + If true, the subscription will request that the publisher send locally\n> + originated changes. False indicates that the publisher sends any changes\n> + regardless of their origin.\n> + </para></entry>\n>\n> SUGGESTION\n> If true, the subscription will request the publisher to only send\n> changes that originated locally.\n\nModified\n\n> ~~~\n>\n> 4. doc/src/sgml/ref/create_subscription.sgml\n>\n> + <para>\n> + Specifies whether the subscription will request the publisher to send\n> + locally originated changes at the publisher node, or send any\n> + publisher node changes regardless of their origin. The default is\n> + <literal>false</literal>.\n> + </para>\n>\n> This wording should be more similar to the same information in catalogs.sgml\n>\n> SUGGESTION\n> Specifies whether the subscription will request the publisher to only\n> send changes that originated locally, or to send any changes\n> regardless of origin.\n\nModified\n\n> ~~~\n>\n> 5. src/include/replication/walreceiver.h\n>\n> @@ -183,6 +183,7 @@ typedef struct\n> bool streaming; /* Streaming of large transactions */\n> bool twophase; /* Streaming of two-phase transactions at\n> * prepare time */\n> + bool only_local; /* publish only locally originated data */\n> } logical;\n> } proto;\n> } WalRcvStreamOptions;\n>\n> SUGGESTION\n> /* Publish only local origin data */\n\nModified\n\n> ~~~\n>\n> 6. src/test/subscription/t/032_onlylocal.pl - cosmetic changes\n>\n> 6a.\n> +# Setup a bidirectional logical replication between Node_A & Node_B\n>\n> SUGGESTION\n> \"... node_A and node_B\"\n\nModified\n\n> 6b.\n> +is(1, 1, \"Circular replication setup is complete\");\n>\n> SUGGESTION (or maybe saying \"circular\" is also OK - I wasn't sure)\n> \"Bidirectional replication setup is complete\"\n\nModified\n\n> 6c.\n> +# check that bidirectional logical replication setup...\n>\n> Start comment sentence with upper case.\n\nModified\n\n> 6d.\n> +###############################################################################\n> +# check that remote data that is originated from node_C to node_B is not\n> +# published to node_A\n> +###############################################################################\n>\n> SUGGESTION\n> Check that remote data of node_B (that originated from node_C) is not\n> published to node_A\n\nModified\n\n> 6e.\n> +is($result, qq(11\n> +12\n> +13), 'Node_C data replicated to Node_B'\n> +);\n>\n> SUGGESTION for message\n> 'node_C data replicated to node_B'\n\nModified\n\n> 6f.\n> +is($result, qq(11\n> +12), 'Remote data originated from other node is not replicated when\n> only_local option is ON'\n> +);\n>\n> SUGGESTION for message\n> 'Remote data originating from another node (not the publisher) is not\n> replicated when only_local option is ON'\n\nModified\n\n> 6g.\n> \"Circular replication setup is complete\"\n> 'Inserted successfully without leading to infinite recursion in\n> bidirectional replication setup'\n> 'Inserted successfully without leading to infinite recursion in\n> bidirectional replication setup'\n> 'Node_C data replicated to Node_B'\n> 'Remote data originated from other node is not replicated when\n> only_local option is ON'\n>\n> Why do some of the \"is\" messages have single quotes and others have\n> double quotes? Should be consistent.\n\nModified\n\n> ~~~\n>\n> 7. src/test/subscription/t/032_onlylocal.pl\n>\n> +my $appname_B2 = 'tap_sub_B2';\n> +$node_B->safe_psql(\n> + 'postgres', \"\n> + CREATE SUBSCRIPTION tap_sub_B2\n> + CONNECTION '$node_C_connstr application_name=$appname_B2'\n> + PUBLICATION tap_pub_C\n> + WITH (only_local = on)\");\n> +\n>\n> AFAIK the \"WITH (only_local = on)\" is unnecessary here. We don't care\n> where the node_C data came from for this test case.\n\nThis test was added to handle the comment as in [1]\n\nThanks for the comments, the attached v17 patch has the fixes for the same.\nI have also separated the bidirectional logical replication steps,\nsince the steps are slightly complex and keeping it separate will help\nin easier review and modifying.\n[1] - https://www.postgresql.org/message-id/CAHut%2BPvwqwzuoQio-hJniarDUOTyxFrwrS9hucd47gEW-9wu-g%40mail.gmail.com\n\nRegards,\nVignesh","msg_date":"Wed, 1 Jun 2022 23:08:40 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, May 26, 2022 at 2:06 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> This post completes my review of patch v15*.\n>\n> ======\n>\n> 1. A general comment affecting both patches\n>\n> I realised the things defined in the subscription WITH are referred to\n> in the PG DOCS as \"parameters\" (not \"options\"). See CREATE\n> SUBSCRIPTION [1]. This is already reflected in some of my review\n> comments for the v15-0002 below., but I'd already posted my v15-0001\n> review comments before noticing this.\n>\n> Please also go through the v15-0001 patch to make all the necessary\n> terminology changes (including in the v15-0001 commit message).\n\nModified\n\n> =====\n>\n> Please find below review comments for patch v15-0002.\n>\n> 2. Commit message\n>\n> This patch does couple of things:\n> change 1) Added force option for copy_data.\n> change 2) Check and throw an error if the publication tables were also\n> subscribing data in the publisher from other publishers.\n>\n> --\n>\n> 2a.\n> \"couple\" -> \"a couple\"\n> \"Added force\" -> \"Adds force ...\"\n> \"Check and throw\" -> \"Checks and throws ...\"\n\nModified\n\n> 2b.\n> Isn't \"Change 2)\" only applicable when copy_data and only_local are\n> on? It should say so here.\n\nModified\n\n> ~~~\n>\n> 3. Commit message\n>\n> node1: Table t1 (c1 int) has data 1, 2, 3, 4\n> node2: Table t1 (c1 int) has data 5, 6, 7, 8\n>\n> --\n>\n> I feel these kinds of replication examples are always much easier to\n> visualize when the data somehow indicates where it came from. E.g.\n> consider using different data in your example like:\n>\n> node1: Table t1 (c1 int) has data 11, 12, 13, 14\n> node2: Table t1 (c1 int) has data 21, 22, 23, 24\n\nModified\n\n> ~~~\n>\n> 4. Commit message\n>\n> Examples in this message exceed 80 chars. Please wrap the SQL appropriately\n\nModified\n\n> ~~~\n>\n> 5. Commit message\n>\n> After this the data will be something like this:\n> node1:\n> 1, 2, 3, 4, 5, 6, 7, 8\n>\n> node2:\n> 1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\n>\n> So, you can see that data on node2 (5, 6, 7, 8) is duplicated.\n>\n> --\n>\n> Yeah, but is that entirely correct though? AFAIK the scenario\n> described by this example is going to recurse forever.\n\nModified\n\n> ~~~\n>\n> 6. Commit message\n>\n> This problem can be solved by using only_local and copy_data option as\n> given below\n>\n> --\n>\n> \"option\" -> \"parameters\"\n\nModified\n\n> ~~~\n>\n> 7. Commit message\n>\n> There are other minor review comments but it's too messy to report in\n> detail here. I suggest checking all the \":\" (some were missing) and\n> the letter case of/within the sentences. I also suggest\n> cutting/pasting all this text into MSWord or some grammar checker to\n> correct anything else reported.\n\nModified\n\n> ~~~\n>\n> 8. Commit message\n>\n> CREATE SUBSCRIPTION sub_node2_node1 CONNECTION '<node1 details>'\n> PUBLICATION pub_node1 WITH (copy_data = force, only_local = on);\n> ERROR: CREATE/ALTER SUBSCRIPTION with only_local and copy_data as\n> true is not allowed when the publisher might have replicated data\n>\n> --\n>\n> Looks like a typo: e.g. If copy_data is \"force\" then how is the error happening?\n\nModified\n\n> ~~~\n>\n> 9. Commit message\n>\n> The examples seem confused/muddled in some way IMO.\n>\n> IIUC the new ERROR is like a new integrity check for when the\n> subscription is using the only_local = on, along with copy_data = on.\n>\n> And so then the copy_data = force was added as a way to override that\n> previous error being reported.\n>\n> So unless you know about the possible error then the \"force\" option\n> makes no sense. So I think you need to change the order of the\n> explanations in the commit message to describe the new error first.\n\nModified\n\n> ======\n>\n> 10. doc/src/sgml/logical-replication.sgml\n>\n> +\n> + <sect1 id=\"bidirectional-logical-replication\">\n> + <title>Bidirectional logical replication</title>\n> +\n> + <sect2 id=\"setting-bidirectional-replication-two-nodes\">\n> + <title>Setting bidirectional replication between two nodes</title>\n> + <para>\n> + Bidirectional replication is useful in creating a multi-master database\n> + which helps in performing read/write operations from any of the nodes.\n> + Setting up bidirectional logical replication between two nodes requires\n> + creation of a publication in all the nodes, creating subscriptions in\n> + each of the nodes that subscribes to data from all the nodes. The steps\n> + to create a two-node bidirectional replication when there is no data in\n> + both the nodes are given below:\n> + </para>\n>\n> This description is confusing. It seems to be trying to say something\n> about a multi-master system (e.g. you say \"all the nodes\") yet this is\n> all written inside the section about \"... two nodes\", so saying \"all\n> the nodes\" makes no sense here.\n>\n> I think you need to split this information and put some kind of blurb\n> text at the top level (section 31.11), and then this section (for \"two\n> nodes\") should ONLY be talking about the case when there are just TWO\n> nodes.\n\nModified\n\n> ~~~\n>\n> 11. doc/src/sgml/logical-replication.sgml\n>\n> + <para>\n> + Lock the required tables in <literal>node1</literal> and\n> + <literal>node2</literal> till the setup is completed.\n> + </para>\n>\n> \"till\" -> \"until\" (make this same change in multiple places)\n\nModified\n\n> ~~~\n>\n> 12. doc/src/sgml/logical-replication.sgml\n>\n> + <sect2 id=\"add-new-node-data-in-existing-node\">\n> + <title>Adding a new node when data is present in the existing nodes</title>\n> + <para>\n> + Adding a new node <literal>node3</literal> to the existing\n> + <literal>node1</literal> and <literal>node2</literal> when data is present\n> + in existing nodes <literal>node1</literal> and <literal>node2</literal>\n> + needs similar steps. The only change required here is that\n> + <literal>node3</literal> should create a subscription with\n> + <literal>copy_data = force</literal> to one of the existing nodes to\n> + receive the existing data during initial data synchronization.\n> + </para>\n>\n> I think perhaps this should clarify that there is NO data in node3 for\n> this example.\n\nModified\n\n> ~~~\n>\n> 13. doc/src/sgml/logical-replication.sgml - section 31.11.3\n>\n> + <para>\n> + Lock the required tables in <literal>node2</literal> and\n> + <literal>node3</literal> till the setup is completed.\n> + </para>\n>\n> Should you describe why you say no lock is required for tables on\n> node1? (Similar information is in section 31.11.4 also)\n\nModified\n\n> ~~~\n>\n> 14. doc/src/sgml/logical-replication.sgml - section 31.11.3\n>\n> + <para>\n> + Create a subscription in <literal>node3</literal> to subscribe to\n> + <literal>node2</literal>:\n> +<programlisting>\n> +node3=# CREATE SUBSCRIPTION sub_node3_node2\n> +node3-# CONNECTION 'dbname=foo host=node2 user=repuser'\n> +node3-# PUBLICATION pub_node2\n> +node3-# WITH (copy_data = off, only_local = on);\n> +CREATE SUBSCRIPTION\n> +</programlisting></para>\n>\n> Using a similar description as in section 31.11.4 this should probably\n> also say something like:\n> \"Use copy_data specified as off because the initial table data would\n> have been already copied in the previous step\".\n\nModified\n\n> ~~~\n>\n> 15. doc/src/sgml/logical-replication.sgml - section 31.11.4\n>\n> + <sect2 id=\"add-node-data-present-in-new-node\">\n> + <title>Adding a new node when data is present in the new node</title>\n> + <para>\n> + Adding a new node <literal>node3</literal> to the existing\n> + <literal>node1</literal> and <literal>node2</literal> when data is present\n> + in the new node <literal>node3</literal> needs similar steps. A\n> few changes\n> + are required here to get the existing data from <literal>node3</literal>\n> + to <literal>node1</literal> and <literal>node2</literal> and later\n> + cleaning up of data in <literal>node3</literal> before synchronization of\n> + all the data from the existing nodes.\n> + </para>\n>\n> I think perhaps this should clarify that there is NO data in node1 or\n> node2 for this example.\n\nThis section contents have been removed because this is not feasible currently.\n\n> ~~~\n>\n> 16. doc/src/sgml/logical-replication.sgml\n>\n> + <sect2 id=\"generic-steps-add-new-node\">\n> + <title>Generic steps to add a new node to the existing set of nodes</title>\n>\n> I think all the steps in this section should be numbered because the\n> order is important.\n\nModified\n\n> ~~~\n>\n> 17. doc/src/sgml/logical-replication.sgml\n>\n> + <sect2 id=\"generic-steps-add-new-node\">\n> + <title>Generic steps to add a new node to the existing set of nodes</title>\n>\n> IIUC the word \"parameter\" should replace the word \"option\" in all the\n> text of this section.\n\nModified\n\n> ~~~\n>\n> 18. doc/src/sgml/logical-replication.sgml - Notes\n>\n> + <sect2>\n> + <title>Notes</title>\n> + <para>\n> + Setting up bidirectional logical replication across nodes requires multiple\n> + steps to be performed on various nodes, as all operations are not\n> + transactional, user is advised to take backup of existing data to avoid any\n> + inconsistency.\n> + </para>\n> + </sect2>\n>\n> 18a.\n> I thought this \"Notes\" section should be rendered more like an sgml\n> note or a warning. E.g. like the warning on this page [2].\n\nModified\n\n> 18b.\n> SUGGESTION (split the para into multiple sentences). e.g.\n> Setting up bidirectional logical replication across nodes requires\n> multiple steps to be performed on various nodes. Because not all\n> operations are transactional, the user is advised to take backups.\n\nModified\n\n> ======\n>\n> 19. doc/src/sgml/ref/alter_subscription.sgml\n>\n> + <para>\n> + There is some interaction between the <literal>only_local</literal>\n> + option and <literal>copy_data</literal> option. Refer to the\n> + <command>CREATE SUBSCRIPTION</command>\n> + <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n> + details and usage of <literal>force</literal> for\n> + <literal>copy_data</literal> option.\n> </para>\n>\n> \"option\" -> \"parameter\" (3x)\n\nModified\n\n> ======\n>\n> 20. doc/src/sgml/ref/create_subscription.sgml\n>\n> + <para>\n> + There is some interaction between the <literal>only_local</literal>\n> + option and <literal>copy_data</literal> option. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n> + details and usage of <literal>force</literal> for\n> + <literal>copy_data</literal> option.\n> + </para>\n>\n> \"option\" -> \"parameter\" (3x)\n\nModified\n\n> ======\n>\n> 21. src/backend/commands/subscriptioncmds.c\n>\n> @@ -69,6 +69,18 @@\n> /* check if the 'val' has 'bits' set */\n> #define IsSet(val, bits) (((val) & (bits)) == (bits))\n>\n> +#define IS_COPY_DATA_ON_OR_FORCE(copy_data) ((copy_data) != COPY_DATA_OFF)\n> +\n> +/*\n> + * Represents whether copy_data option is specified with off, on or force.\n> + */\n> +typedef enum CopyData\n> +{\n> + COPY_DATA_OFF,\n> + COPY_DATA_ON,\n> + COPY_DATA_FORCE\n> +} CopyData;\n> +\n>\n> 21a.\n> \"option\" -> \"parameter\"\n\nModified\n\n> 21b.\n> I think the new #define is redundant and can be removed (see the next\n> review comment)\n\nModified\n\n> 21c.\n> I think you set the OFF enum value like:\n> typedef enum CopyData\n> {\n> COPY_DATA_OFF = 0,\n> COPY_DATA_ON,\n> COPY_DATA_FORCE\n> } CopyData;\n>\n> then it will greatly simplify things; many changes in this file will\n> be unnecessary (see subsequent review comments)\n\nModified\n\n> ~~~\n>\n> 22. src/backend/commands/subscriptioncmds.c\n>\n> +/*\n> + * Validate the value specified for copy_data option.\n> + */\n> +static CopyData\n> +DefGetCopyData(DefElem *def)\n>\n> \"option\" -> \"parameter\"\n\nModified\n\n> ~~~\n>\n> 23. src/backend/commands/subscriptioncmds.c\n>\n> + /*\n> + * Allow 0, 1, \"true\", \"false\", \"on\", \"off\" or \"force\".\n> + */\n> + switch (nodeTag(def->arg))\n> + {\n> + case T_Integer:\n> + switch (intVal(def->arg))\n> + {\n> + case 0:\n> + return COPY_DATA_OFF;\n> + case 1:\n> + return COPY_DATA_ON;\n> + default:\n> + /* otherwise, error out below */\n> + break;\n> + }\n> + break;\n>\n> What about also allowing copy_data = 2, and making it equivalent to \"force\"?\n\nI felt the existing looks ok, no need to support 2. It might confuse the user.\n\n> ~~~\n>\n> 24. src/backend/commands/subscriptioncmds.c\n>\n> @@ -333,17 +400,17 @@ parse_subscription_options(ParseState *pstate,\n> List *stmt_options,\n> errmsg(\"%s and %s are mutually exclusive options\",\n> \"connect = false\", \"create_slot = true\")));\n>\n> - if (opts->copy_data &&\n> + if (IS_COPY_DATA_ON_OR_FORCE(opts->copy_data) &&\n> IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\n>\n> I think this change would not be needed if you set the OFF enum = 0.\n\nModified\n\n> ~~~\n> 25. src/backend/commands/subscriptioncmds.c\n>\n> @@ -671,13 +738,14 @@ CreateSubscription(ParseState *pstate,\n> CreateSubscriptionStmt *stmt,\n> * Set sync state based on if we were asked to do data copy or\n> * not.\n> */\n> - table_state = opts.copy_data ? SUBREL_STATE_INIT : SUBREL_STATE_READY;\n> + table_state = IS_COPY_DATA_ON_OR_FORCE(opts.copy_data) ?\n> SUBREL_STATE_INIT : SUBREL_STATE_READY;\n>\n> I think this change would not be needed if you set the OFF enum = 0.\n\nModified\n\n> ~~~\n>\n> 26. src/backend/commands/subscriptioncmds.c\n>\n> @@ -720,7 +788,8 @@ CreateSubscription(ParseState *pstate,\n> CreateSubscriptionStmt *stmt,\n> * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\n> * PUBLICATION to work.\n> */\n> - if (opts.twophase && !opts.copy_data && tables != NIL)\n> + if (opts.twophase && opts.copy_data == COPY_DATA_OFF &&\n> + tables != NIL)\n> twophase_enabled = true;\n>\n> I think this change would not be needed if you set the OFF enum = 0.\n\nModified\n\n> ~~~\n>\n> 27. src/backend/commands/subscriptioncmds.c\n>\n> @@ -851,7 +921,7 @@ AlterSubscription_refresh(Subscription *sub, bool copy_data,\n> list_length(subrel_states), sizeof(Oid), oid_cmp))\n> {\n> AddSubscriptionRelState(sub->oid, relid,\n> - copy_data ? SUBREL_STATE_INIT : SUBREL_STATE_READY,\n> + IS_COPY_DATA_ON_OR_FORCE(copy_data) ? SUBREL_STATE_INIT : SUBREL_STATE_READY,\n> InvalidXLogRecPtr);\n>\n> I think this change would not be needed if you set the OFF enum = 0.\n>\n\nModified\n\n> ~~~\n> 28. src/backend/commands/subscriptioncmds.c\n>\n> @@ -1157,7 +1227,7 @@ AlterSubscription(ParseState *pstate,\n> AlterSubscriptionStmt *stmt,\n> * See ALTER_SUBSCRIPTION_REFRESH for details why this is\n> * not allowed.\n> */\n> - if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\n> + if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\n> IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\n> ereport(ERROR,\n> (errcode(ERRCODE_SYNTAX_ERROR),\n> errmsg(\"ALTER SUBSCRIPTION with refresh and copy_data is not allowed\n> when two_phase is enabled\"),\n>\n> I think this change would not be needed if you set the OFF enum = 0.\n\nModified\n\n> ~~~\n>\n> 29. src/backend/commands/subscriptioncmds.c\n>\n> @@ -1209,7 +1279,7 @@ AlterSubscription(ParseState *pstate,\n> AlterSubscriptionStmt *stmt,\n> * See ALTER_SUBSCRIPTION_REFRESH for details why this is\n> * not allowed.\n> */\n> - if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\n> + if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\n> IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\n> ereport(ERROR,\n> (errcode(ERRCODE_SYNTAX_ERROR),\n> errmsg(\"ALTER SUBSCRIPTION with refresh and copy_data is not allowed\n> when two_phase is enabled\"),\n>\n> I think this change would not be needed if you set the OFF enum = 0.\n\nModified\n\n> ~~~\n>\n> 30. src/backend/commands/subscriptioncmds.c\n>\n> @@ -1255,7 +1325,8 @@ AlterSubscription(ParseState *pstate,\n> AlterSubscriptionStmt *stmt,\n> *\n> * For more details see comments atop worker.c.\n> */\n> - if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\n> + if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\n> + IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\n> ereport(ERROR,\n> (errcode(ERRCODE_SYNTAX_ERROR),\n> errmsg(\"ALTER SUBSCRIPTION ... REFRESH with copy_data is not allowed\n> when two_phase is enabled\"),\n>\n> I think this change would not be needed if you set the OFF enum = 0.\n\nModified\n\n> ~~~\n>\n> 31. src/backend/commands/subscriptioncmds.c\n>\n> + appendStringInfoString(&cmd,\n> + \"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\n> PS.srrelid as replicated\\n\"\n> + \"FROM pg_publication P,\\n\"\n> + \"LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n> + \"LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n> + \"pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n> + \"WHERE C.oid = GPT.relid AND P.pubname in (\");\n>\n> use upper case \"in\" -> \"IN\"\n\nModified\n\n> ======\n>\n> 32. src/test/regress/sql/subscription.sql\n>\n> @@ -40,6 +40,9 @@ SET SESSION AUTHORIZATION 'regress_subscription_user';\n>\n> -- fail - invalid option combinations\n> CREATE SUBSCRIPTION regress_testsub2 CONNECTION\n> 'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\n> false, copy_data = true);\n> +CREATE SUBSCRIPTION regress_testsub2 CONNECTION\n> 'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\n> false, copy_data = on);\n> +CREATE SUBSCRIPTION regress_testsub2 CONNECTION\n> 'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\n> false, copy_data = 1);\n> +CREATE SUBSCRIPTION regress_testsub2 CONNECTION\n> 'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\n> false, copy_data = force);\n>\n> 32a.\n> The test for \"copy_data = 1\" should confirm it is the same as \"on\".\n\nWe have the test for copy_data = on and copy_data = 1, both the\nresults are the same. Isn't that sufficient?\n\n> 32b.\n> Should also test copy_data = 2 (and confirm it is the same as \"force\").\n\nCurrently only on, off, true, false, 0 & 1 is supported. 2 is not\nsupported. Added a test for copy_data as 2\n\n> ======\n>\n> 33. src/test/subscription/t/032_localonly.pl - cosmetic changes\n>\n> 33a.\n> +# Detach node C from the node-group of (A, B, C) and clean the table contents\n> +# from all nodes.\n>\n> SUGGESTION\n> Detach node_C from the node-group of (node_A, node_B, node_C) and ...\n\nModified\n\n> 33b.\n> +# Subroutine to create subscription and wait till the initial sync is\n> completed.\n>\n> \"till\" -> \"until\"\n\nModified\n\n> 33c.\n> +# Subroutine to create subscription and wait till the initial sync is\n> completed.\n> +# Subroutine expects subscriber node, publisher node, subscription name,\n> +# destination connection string, publication name and the subscription with\n> +# options to be passed as input parameters.\n> +sub create_subscription\n> +{\n> + my ($node_subscriber, $node_publisher, $sub_name, $node_connstr,\n> + $pub_name, $with_options)\n> + = @_;\n>\n> \"subscription with options\" => \"subscription parameters\"\n> \"$with_options\" -> \"$sub_params\"\n\nModified\n\n> 33d.\n> +# Specifying only_local 'on' which indicates that the publisher should only\n>\n> \"Specifying\" => \"Specify\"\n\nModified\n\nThanks for the comments, the v17 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 1 Jun 2022 23:16:16 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, May 26, 2022 at 4:40 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, May 20, 2022 at 3:31 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Wed, May 18, 2022 at 4:22 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > >\n> > > 5.\n> > > * It is quite possible that subscriber has not yet pulled data to\n> > > + * the tables, but in ideal cases the table data will be subscribed.\n> > > + * To keep the code simple it is not checked if the subscriber table\n> > > + * has pulled the data or not.\n> > > + */\n> > > + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\n> > >\n> > > Sorry, but I don't understand what you intend to say by the above\n> > > comment. Can you please explain?\n> >\n> > When the user specifies copy_data as on, we should check if the\n> > publisher has the publication tables being subscribed from a remote\n> > publisher. If so throw an error as it remote origin data present.\n> > Ex:\n> > Node1 - pub1 for table t1 -- no data\n> > Node2 - Sub1 subscribing to data from pub1\n> > Node2 - pub2 for table t1 -- no data\n> > Node3 - create subscription to Node2 with copy_data = ON\n> >\n> > In this case even though the table does not have any remote origin\n> > data, as Node2 is subscribing to data from Node1, throw an error.\n> > We throw an error irrespective of data present in Node1 or not to keep\n> > the code simple.\n> >\n>\n> I think we can change the contents of comments something like: \"Throw\n> an error if the publisher has subscribed to the same table from some\n> other publisher. We cannot differentiate between the local and\n> non-local data that is present in the HEAP during the initial sync.\n> Identification of local data can be done only from the WAL by using\n> the origin id. XXX: For simplicity, we don't check whether the table\n> has any data or not. If the table doesn't have any data then we don't\n> need to distinguish between local and non-local data so we can avoid\n> throwing error in that case.\"\n\nModified\n\n> Few more comments:\n> ==================\n> Patch 0002\n> ======\n> 1.\n> + if (copydata == COPY_DATA_ON && only_local && !slot_attisnull(slot, 3))\n> + ereport(ERROR,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"CREATE/ALTER SUBSCRIPTION with only_local and copy_data as\n> true is not allowed when the publisher might have replicated data,\n> table:%s.%s might have replicated data in the publisher\",\n> + nspname, relname),\n> + errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\"));\n>\n> Can we split the error message? errmsg: table:%s.%s has replicated\n> data in the publisher; errdetail:CREATE/ALTER SUBSCRIPTION with\n> only_local and copy_data as true is not allowed when the publisher has\n> replicated data\n\nModified\n\n> 2.\n> + <para>\n> + Lock the required tables in the new node until the setup is complete.\n> + </para>\n> + <para>\n> + Create subscriptions on existing nodes pointing to publication on\n> + the new node with <literal>only_local</literal> option specified as\n> + <literal>on</literal> and <literal>copy_data</literal> specified as\n> + <literal>on</literal>.\n> + </para>\n> + <para>\n> + Wait for data to be copied from the new node to existing nodes.\n> + </para>\n> + <para>\n> + Alter the publication in new node so that the truncate operation is not\n> + replicated to the subscribers.\n> + </para>\n>\n> Here and at other places, we should specify that the lock mode should\n> to acquire the lock on table should be EXCLUSIVE so that no concurrent\n> DML is allowed on it. Also, it is better if somewhere we explain why\n> and which nodes need locks?\n\nModified\n\n> Patch 0001:\n> ==========\n> 1.\n> +$node_A->append_conf(\n> + 'postgresql.conf', qq(\n> +max_prepared_transactions = 10\n> +logical_decoding_work_mem = 64kB\n> +));\n>\n> I don't see why you need to set these parameters. There is no test\n> case that needs these parameters. Please remove these from here and\n> all other similar places in 032_onlylocal.pl.\n\nRemoved it.\n\nThanks for the comments, the v17 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 1 Jun 2022 23:18:07 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, May 27, 2022 at 7:54 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Apr 6, 2022 at 9:36 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Tue, Apr 5, 2022 at 7:06 PM Ashutosh Bapat\n> > <ashutosh.bapat.oss@gmail.com> wrote:\n> > >\n> > > On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > > >\n> > > > Below are some other name ideas. Maybe they are not improvements, but\n> > > > it might help other people to come up with something better.\n> > > >\n> > > > subscribe_publocal_only = true/false\n> > > > origin = pub_only/any\n> > > > adjacent_data_only = true/false\n> > > > direct_subscriptions_only = true/false\n> > > > ...\n> > >\n> > > FWIW, The subscriber wants \"changes originated on publisher\". From\n> > > that angle origin = publisher/any looks attractive. It also leaves\n> > > open the possibility that the subscriber may ask changes from a set of\n> > > origins or even non-publisher origins.\n> > >\n> >\n> > So, how are you imagining extending it for multiple origins? I think\n> > we can have multiple origins data on a node when there are multiple\n> > subscriptions pointing to the different or same node. The origin names\n> > are internally generated names but are present in\n> > pg_replication_origin. So, are we going to ask the user to check that\n> > table and specify it as an option? But, note, that the user may not be\n> > able to immediately recognize which origin data is useful for her.\n> >\n>\n> I still don't have a very clear answer for the usability aspect but it\n> seems this was discussed in PGCon-Unconference [1] (Section: Origin\n> filter) and there also it was suggested to allow users to specify\n> multiple origin names. So, probably Ashutosh's idea makes sense and we\n> should use \"origin = publisher/any or origin=local/any\". Among these,\n> I would prefer later.\n\nI have changed it to origin = local/any.\n\nThe v17 patch attached at [1] has the changes for the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 1 Jun 2022 23:19:03 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, May 27, 2022 at 10:56 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some more review comments for v15-0002 I found while working\n> through the documented examples. These are all for\n> doc/src/sgml/logical-replication.sgml. The section numbers (e.g.\n> 31.11.1) are the assigned numbers after HTML rendering.\n>\n> The general comment is that the sections appeared at first to be\n> independent of each other, but they are not really - sometimes\n> existing data and existing pub/sub are assumed to remain. Therefore, I\n> think some more text needs to be added to clarify the initial state\n> for each of these sections.\n>\n> ======\n>\n> 1. Section id.\n>\n> <sect1 id=\"bidirectional-logical-replication\">\n>\n> IMO this section ID should be renamed\n> \"logical-replication-bidirectional\" then it will be more consistent\n> filenames with all the other logical replication section names.\n\nModified\n\n> ~~~\n>\n> 2. Section 31.11.2. Adding a new node when there is no data in any of the nodes\n>\n> 2a.\n> This seems like a continuation of 31.11.1 because pub/sub of\n> node1,node2 is assumed to exist still.\n\nAdded a note\n\n> 2b.\n> The bottom of this section should say after these steps any subsequent\n> incremental data changes on any node will be replicated to all nodes\n> (node1, node2, node3)\n\nModified\n\n> ~~~\n>\n> 3. Section 31.11.3. Adding a new node when data is present in the existing nodes\n>\n> 3a.\n> This seems to be a continuation of 31.11.1 because pub/sub (and\n> initial data) of node1/node2 is assumed to exist.\n\nAdded a note\n\n> 3b.\n>\n> The bottom of this section should say after these steps that any\n> initial data of node1/node2 will be seen in node3, and any subsequent\n> incremental data changes on any node will be replicated to all nodes\n> (node1, node2, node3)\n\nModified\n\n> ~~~\n>\n> 4. Section 31.11.4. Adding a new node when data is present in the new node\n>\n> 4a.\n> This seems to be a continuation of 31.11.1 because pub/sub (and\n> initial data) of node1/node2 is assumed to exist.\n\nThis section contents have been removed because this is not feasible currently\n\n> 4b.\n> It is not made very clear up-front if the tables on node1 and node2\n> are empty or not. Apparently, they are considered NOT empty because\n> later in the example you are using \"force\" when you create the\n> subscription on node3.\n\nThis section contents have been removed because this is not feasible currently\n\n> 4c.\n> The SQL wrapping here is strangely different from others (put the\n> subscription name on 1st line)\n> node3=# CREATE SUBSCRIPTION\n> node3-# sub_node3_node1 CONNECTION 'dbname=foo host=node1 user=repuser'\n> node3-# PUBLICATION pub_node1\n> node3-# WITH (copy_data = force, only_local = on);\n> CREATE SUBSCRIPTION\n\nThis section contents have been removed because this is not feasible currently\n\n> 4d.\n> The SQL wrapping here is strangely different from others (put the\n> subscription name on 1st line)\n> node3=# CREATE SUBSCRIPTION\n> node3-# sub_node3_node2 CONNECTION 'dbname=foo host=node2 user=repuser'\n> node3-# PUBLICATION pub_node2\n> node3-# WITH (copy_data = off, only_local = on);\n> CREATE SUBSCRIPTION\n\nThis section contents have been removed because this is not feasible currently\n\n> 4e.\n> The bottom of this section should say after this that any initial data\n> of node1/node2 will be seen in node3, and any subsequent incremental\n> data changes on any node will be replicated to all nodes (node1,\n> node2, node3)\n\nThis section contents have been removed because this is not feasible currently\n\n> ~~~\n>\n> 5. Section 31.11.5. Generic steps to add a new node to the existing set of nodes\n>\n> 5a\n> Create subscriptions on the new node pointing to publication on the\n> first node with only_local option specified as on and copy_data option\n> specified as \"force\".\n>\n> -> that should say \"Create a subscription\" (singular)\n\nModified\n\n> 5b.\n> Create subscriptions on the new node pointing to publications on the\n> remaining node with only_local option specified as on and copy_data\n> option specified as off.\n>\n> -> that should say \"on the remaining node\" (plural)\n\nModified\n\nThanks for the comments, the v17 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 1 Jun 2022 23:23:15 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, May 27, 2022 at 12:34 PM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Wed, May 25, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > The attached v16 patch has the changes for the same.\n> >\n>\n> Thanks for updating the patch.\n>\n> Some comments for the document in 0002 patch.\n>\n> 1.\n> + <para>\n> + Lock the required tables in <literal>node1</literal> and\n> + <literal>node2</literal> till the setup is completed.\n> + </para>\n> +\n> + <para>\n> + Create a publication in <literal>node1</literal>:\n> +<programlisting>\n> +node1=# CREATE PUBLICATION pub_node1 FOR TABLE t1;\n> +CREATE PUBLICATION\n> +</programlisting></para>\n>\n> If the table is locked in the very beginning, we will not be able to create the\n> publication (because the locks have conflict). Maybe we should switch the order\n> of creating publication and locking tables here.\n\nModified\n\n> 2.\n> In the case of \"Adding a new node when data is present in the new node\", we need\n> to truncate table t1 in node3, but the truncate operation would be blocked\n> because the table has be locked before. Maybe we need some changes for it.\n\nThis section contents have been removed because this is not feasible currently\n\nThanks for the comments, the v17 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 1 Jun 2022 23:25:02 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, May 27, 2022 at 2:08 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Fri, May 27, 2022 at 5:04 PM shiy.fnst@fujitsu.com\n> <shiy.fnst@fujitsu.com> wrote:\n> >\n> > On Wed, May 25, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > The attached v16 patch has the changes for the same.\n> > >\n> >\n> > Thanks for updating the patch.\n> >\n> > Some comments for the document in 0002 patch.\n> >\n> > 1.\n> > + <para>\n> > + Lock the required tables in <literal>node1</literal> and\n> > + <literal>node2</literal> till the setup is completed.\n> > + </para>\n> > +\n> > + <para>\n> > + Create a publication in <literal>node1</literal>:\n> > +<programlisting>\n> > +node1=# CREATE PUBLICATION pub_node1 FOR TABLE t1;\n> > +CREATE PUBLICATION\n> > +</programlisting></para>\n> >\n> > If the table is locked in the very beginning, we will not be able to create the\n> > publication (because the locks have conflict). Maybe we should switch the order\n> > of creating publication and locking tables here.\n> >\n>\n> I agree. It seems some of the locking instructions in the earlier\n> sections 31.11.1 - 31.11.4 are contradictory to the later \"generic\"\n> steps given in \"31.11.5. Generic steps to add a new node to the\n> existing set of nodes\". I'm assuming the generic steps are the\n> \"correct\" steps\n>\n> e.g. generic steps say get the lock on new node tables AFTER the\n> publication of new node.\n> e.g. generic steps say do NOT get a table lock on the node one you are\n> (first) joining to.\n\nYes, the generic steps are correct. modified\n\n> ~~~\n>\n> Furthermore, the generic steps are describing attaching a new N+1th\n> node to some (1 ... N) other nodes.\n>\n> So I think in the TWO-node case (section 31.11.1) node2 should be\n> treated as the \"new\" node that you are attaching to the \"first\" node1.\n> IMO the section 31.11.1 example should be reversed so that it obeys\n> the \"generic\" pattern.\n> e.g. It should be doing the CREATE PUBLICATION pub_node2 first (since\n> that is the \"new\" node)\n\nIn generic steps we mention the publication must be created in a new\nnode and we don't say about the existing nodes, because the\npublication would be existing already. I feel the existing order of\npublication creation in the TWO-node case (section 31.11.1) is ok.\n\nThanks for the comments, the v17 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 1 Jun 2022 23:27:36 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Please find below my review comments for patch v17-0001:\n\n======\n\n1. Commit message\nAdd a missing test to verify only-local option in test_decoding plugin.\n\n\"option\" -> \"parameter\"\n\n======\n\n2. contrib/test_decoding/sql/replorigin.sql\n\n2a.\n+-- Verify that remote origin data is not returned with only-local option\n\nSUGGESTION\n-- Verify the behaviour of the only-local parameter\n\n2b.\n+-- Remote origin data returned when only-local option is not set\n\n\"option\" -> \"parameter\"\n\n2c.\n+-- Remote origin data not returned when only-local option is set\n\n\"option\" -> \"parameter\"\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 3 Jun 2022 15:26:29 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Please find below my review comments for patch v17-0002:\n\n======\n\n1. Commit message\n\nThis patch adds a new SUBSCRIPTION parameter \"origin\". It Specifies whether\nthe subscription will request the publisher to only send changes that\noriginated locally, or to send any changes regardless of origin.\n\n\"It Specifies\" -> \"It specifies\"\n\n~~~\n\n2. Commit message\n\nUsage:\nCREATE SUBSCRIPTION sub1 CONNECTION 'dbname=postgres port=9999'\nPUBLICATION pub1 with (origin = local);\n\n\"with\" -> \"WITH\"\n\n======\n\n3. doc/src/sgml/catalogs.sgml\n\n+ <para>\n+ Possible origin values are <literal>local</literal>,\n+ <literal>any</literal>, or <literal>NULL</literal> if none is specified.\n+ If <literal>local</literal>, the subscription will request the\n+ publisher to only send changes that originated locally. If\n+ <literal>any</literal>, the publisher sends any changes regardless of\n+ their origin.\n+ </para></entry>\n\nShould this also mention that NULL (default) is equivalent to 'any'?\n\nSUGGESTION\nIf <literal>any</literal> (or <literal>NULL</literal>), the publisher\nsends any changes regardless of their origin.\n\n======\n\n4. src/backend/catalog/pg_subscription.c\n\n@@ -72,6 +72,16 @@ GetSubscription(Oid subid, bool missing_ok)\n sub->twophasestate = subform->subtwophasestate;\n sub->disableonerr = subform->subdisableonerr;\n\n+ datum = SysCacheGetAttr(SUBSCRIPTIONOID,\n+ tup,\n+ Anum_pg_subscription_suborigin,\n+ &isnull);\n+\n+ if (!isnull)\n+ sub->origin = TextDatumGetCString(datum);\n+ else\n+ sub->origin = NULL;\n+\n /* Get conninfo */\n datum = SysCacheGetAttr(SUBSCRIPTIONOID,\n tup,\n\nMissing comment like the nearby code has:\n/* Get origin */\n\n======\n\n5. src/backend/replication/logical/worker.c\n\n+/* Macro for comparing string fields that might be NULL */\n+#define equalstr(a, b) \\\n+ (((a) != NULL && (b) != NULL) ? (strcmp(a, b) == 0) : (a) == (b))\n+\n\nShould that have some extra parens for the macro args?\ne.g. \"strcmp((a), (b))\"\n\n~~~\n\n6. src/backend/replication/logical/worker.c - maybe_reread_subscription\n\n@@ -3059,6 +3063,7 @@ maybe_reread_subscription(void)\n strcmp(newsub->slotname, MySubscription->slotname) != 0 ||\n newsub->binary != MySubscription->binary ||\n newsub->stream != MySubscription->stream ||\n+ equalstr(newsub->origin, MySubscription->origin) ||\n newsub->owner != MySubscription->owner ||\n !equal(newsub->publications, MySubscription->publications))\n {\n\nIs that right? Shouldn't it say !equalstr(...)?\n\n======\n\n7. src/backend/replication/pgoutput/pgoutput.c - parse_output_parameters\n\n@@ -380,6 +382,16 @@ parse_output_parameters(List *options, PGOutputData *data)\n\n data->two_phase = defGetBoolean(defel);\n }\n+ else if (strcmp(defel->defname, \"origin\") == 0)\n+ {\n+ if (origin_option_given)\n+ ereport(ERROR,\n+ (errcode(ERRCODE_SYNTAX_ERROR),\n+ errmsg(\"conflicting or redundant options\")));\n+ origin_option_given = true;\n+\n+ data->origin = defGetString(defel);\n+ }\n\nShould this function also be validating that the origin parameter\nvalue is only permitted to be one of \"local\" or \"any\"?\n\n~~~\n\n8. src/backend/replication/pgoutput/pgoutput.c - pgoutput_origin_filter\n\n@@ -1698,12 +1710,20 @@ pgoutput_message(LogicalDecodingContext *ctx,\nReorderBufferTXN *txn,\n }\n\n /*\n- * Currently we always forward.\n+ * Return true if the data source (origin) is remote and user has requested\n+ * only local data, false otherwise.\n */\n static bool\n pgoutput_origin_filter(LogicalDecodingContext *ctx,\n RepOriginId origin_id)\n {\n+ PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\n+\n+ if (data->origin &&\n+ (strcmp(data->origin, \"local\") == 0) &&\n+ origin_id != InvalidRepOriginId)\n+ return true;\n+\n return false;\n }\n\n8a.\nCould rewrite the condition by putting the strcmp last so you can\navoid doing unnecessary strcmp.\n\ne.g\n+ if (data->origin &&\n+ origin_id != InvalidRepOriginId &&\n+ strcmp(data->origin, \"local\" == 0)\n\n\n8b.\nI also wondered if it might be worth considering caching the origin\nparameter value when it was parsed so that you can avoid doing any\nstrcmp at all during this function. Because otherwise this might be\ncalled millions of times, right?\n\n======\n\n9. src/include/catalog/pg_subscription.h\n\n@@ -87,6 +87,9 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\nBKI_SHARED_RELATION BKI_ROW\n\n /* List of publications subscribed to */\n text subpublications[1] BKI_FORCE_NOT_NULL;\n+\n+ /* Publish the data originated from the specified origin */\n+ text suborigin;\n #endif\n } FormData_pg_subscription;\n\nSUGGESTION (for the comment text)\n/* Only publish data originating from the specified origin */\n\n~~~\n\n10 src/include/catalog/pg_subscription.h - Subscription\n\n@@ -118,6 +121,9 @@ typedef struct Subscription\n char *slotname; /* Name of the replication slot */\n char *synccommit; /* Synchronous commit setting for worker */\n List *publications; /* List of publication names to subscribe to */\n+ char *origin; /* Publish the data originated from the\n+ * specified origin */\n+\n } Subscription;\n\n10a.\nReword comment same as suggested in review comment #9\n\n10b.\nRemove spurious blank line\n\n======\n\n11. src/include/replication/walreceiver.h\n\n@@ -183,6 +183,8 @@ typedef struct\n bool streaming; /* Streaming of large transactions */\n bool twophase; /* Streaming of two-phase transactions at\n * prepare time */\n+ char *origin; /* Publish the data originated from the\n+ * specified origin */\n } logical;\n\nReword comment same as suggested in review comment #9\n\n======\n\n12. src/test/subscription/t/032_origin.pl\n\n+# Test logical replication using origin option.\n\n# Test the CREATE SUBSCRIPTION 'origin' parameter\n\n~~~\n\n13. src/test/subscription/t/032_origin.pl\n\n+# check that the data published from node_C to node_B is not sent to node_A\n+$result = $node_A->safe_psql('postgres', \"SELECT * FROM tab_full ORDER BY 1;\");\n+is($result, qq(11\n+12), 'Remote data originating from another node (not the publisher)\nis not replicated when origin option is local'\n+);\n\n\"option\" -> \"parameter\"\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 3 Jun 2022 15:30:49 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jun 3, 2022 at 10:56 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Please find below my review comments for patch v17-0001:\n>\n> ======\n>\n> 1. Commit message\n> Add a missing test to verify only-local option in test_decoding plugin.\n>\n> \"option\" -> \"parameter\"\n\nModified\n\n> ======\n>\n> 2. contrib/test_decoding/sql/replorigin.sql\n>\n> 2a.\n> +-- Verify that remote origin data is not returned with only-local option\n>\n> SUGGESTION\n> -- Verify the behaviour of the only-local parameter\n\nModified\n\n> 2b.\n> +-- Remote origin data returned when only-local option is not set\n>\n> \"option\" -> \"parameter\"\n\nModified\n\n> 2c.\n> +-- Remote origin data not returned when only-local option is set\n>\n> \"option\" -> \"parameter\"\n\nModified\n\nThe attached v18 patch has the fixes for the same.\n\nRegards,\nVignesh","msg_date":"Mon, 6 Jun 2022 10:44:08 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jun 3, 2022 at 11:01 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Please find below my review comments for patch v17-0002:\n>\n> ======\n>\n> 1. Commit message\n>\n> This patch adds a new SUBSCRIPTION parameter \"origin\". It Specifies whether\n> the subscription will request the publisher to only send changes that\n> originated locally, or to send any changes regardless of origin.\n>\n> \"It Specifies\" -> \"It specifies\"\n\nModified\n\n> ~~~\n>\n> 2. Commit message\n>\n> Usage:\n> CREATE SUBSCRIPTION sub1 CONNECTION 'dbname=postgres port=9999'\n> PUBLICATION pub1 with (origin = local);\n>\n> \"with\" -> \"WITH\"\n\nModified\n\n> ======\n>\n> 3. doc/src/sgml/catalogs.sgml\n>\n> + <para>\n> + Possible origin values are <literal>local</literal>,\n> + <literal>any</literal>, or <literal>NULL</literal> if none is specified.\n> + If <literal>local</literal>, the subscription will request the\n> + publisher to only send changes that originated locally. If\n> + <literal>any</literal>, the publisher sends any changes regardless of\n> + their origin.\n> + </para></entry>\n>\n> Should this also mention that NULL (default) is equivalent to 'any'?\n>\n> SUGGESTION\n> If <literal>any</literal> (or <literal>NULL</literal>), the publisher\n> sends any changes regardless of their origin.\n\nModified\n\n> ======\n>\n> 4. src/backend/catalog/pg_subscription.c\n>\n> @@ -72,6 +72,16 @@ GetSubscription(Oid subid, bool missing_ok)\n> sub->twophasestate = subform->subtwophasestate;\n> sub->disableonerr = subform->subdisableonerr;\n>\n> + datum = SysCacheGetAttr(SUBSCRIPTIONOID,\n> + tup,\n> + Anum_pg_subscription_suborigin,\n> + &isnull);\n> +\n> + if (!isnull)\n> + sub->origin = TextDatumGetCString(datum);\n> + else\n> + sub->origin = NULL;\n> +\n> /* Get conninfo */\n> datum = SysCacheGetAttr(SUBSCRIPTIONOID,\n> tup,\n>\n> Missing comment like the nearby code has:\n> /* Get origin */\n\nModified\n\n> ======\n>\n> 5. src/backend/replication/logical/worker.c\n>\n> +/* Macro for comparing string fields that might be NULL */\n> +#define equalstr(a, b) \\\n> + (((a) != NULL && (b) != NULL) ? (strcmp(a, b) == 0) : (a) == (b))\n> +\n>\n> Should that have some extra parens for the macro args?\n> e.g. \"strcmp((a), (b))\"\n\nModified\n\n> ~~~\n>\n> 6. src/backend/replication/logical/worker.c - maybe_reread_subscription\n>\n> @@ -3059,6 +3063,7 @@ maybe_reread_subscription(void)\n> strcmp(newsub->slotname, MySubscription->slotname) != 0 ||\n> newsub->binary != MySubscription->binary ||\n> newsub->stream != MySubscription->stream ||\n> + equalstr(newsub->origin, MySubscription->origin) ||\n> newsub->owner != MySubscription->owner ||\n> !equal(newsub->publications, MySubscription->publications))\n> {\n>\n> Is that right? Shouldn't it say !equalstr(...)?\n\nModified\n\n> ======\n>\n> 7. src/backend/replication/pgoutput/pgoutput.c - parse_output_parameters\n>\n> @@ -380,6 +382,16 @@ parse_output_parameters(List *options, PGOutputData *data)\n>\n> data->two_phase = defGetBoolean(defel);\n> }\n> + else if (strcmp(defel->defname, \"origin\") == 0)\n> + {\n> + if (origin_option_given)\n> + ereport(ERROR,\n> + (errcode(ERRCODE_SYNTAX_ERROR),\n> + errmsg(\"conflicting or redundant options\")));\n> + origin_option_given = true;\n> +\n> + data->origin = defGetString(defel);\n> + }\n>\n> Should this function also be validating that the origin parameter\n> value is only permitted to be one of \"local\" or \"any\"?\n\nModified\n\n> ~~~\n>\n> 8. src/backend/replication/pgoutput/pgoutput.c - pgoutput_origin_filter\n>\n> @@ -1698,12 +1710,20 @@ pgoutput_message(LogicalDecodingContext *ctx,\n> ReorderBufferTXN *txn,\n> }\n>\n> /*\n> - * Currently we always forward.\n> + * Return true if the data source (origin) is remote and user has requested\n> + * only local data, false otherwise.\n> */\n> static bool\n> pgoutput_origin_filter(LogicalDecodingContext *ctx,\n> RepOriginId origin_id)\n> {\n> + PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\n> +\n> + if (data->origin &&\n> + (strcmp(data->origin, \"local\") == 0) &&\n> + origin_id != InvalidRepOriginId)\n> + return true;\n> +\n> return false;\n> }\n>\n> 8a.\n> Could rewrite the condition by putting the strcmp last so you can\n> avoid doing unnecessary strcmp.\n>\n> e.g\n> + if (data->origin &&\n> + origin_id != InvalidRepOriginId &&\n> + strcmp(data->origin, \"local\" == 0)\n\nI have avoided the strcmp by caching locally as suggested in 8b. I\nhave not made any change for this.\n\n>\n> 8b.\n> I also wondered if it might be worth considering caching the origin\n> parameter value when it was parsed so that you can avoid doing any\n> strcmp at all during this function. Because otherwise this might be\n> called millions of times, right?\n\nModified\n\n> ======\n>\n> 9. src/include/catalog/pg_subscription.h\n>\n> @@ -87,6 +87,9 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\n> BKI_SHARED_RELATION BKI_ROW\n>\n> /* List of publications subscribed to */\n> text subpublications[1] BKI_FORCE_NOT_NULL;\n> +\n> + /* Publish the data originated from the specified origin */\n> + text suborigin;\n> #endif\n> } FormData_pg_subscription;\n>\n> SUGGESTION (for the comment text)\n> /* Only publish data originating from the specified origin */\n\nModified\n\n> ~~~\n>\n> 10 src/include/catalog/pg_subscription.h - Subscription\n>\n> @@ -118,6 +121,9 @@ typedef struct Subscription\n> char *slotname; /* Name of the replication slot */\n> char *synccommit; /* Synchronous commit setting for worker */\n> List *publications; /* List of publication names to subscribe to */\n> + char *origin; /* Publish the data originated from the\n> + * specified origin */\n> +\n> } Subscription;\n>\n> 10a.\n> Reword comment same as suggested in review comment #9\n\nModified\n\n> 10b.\n> Remove spurious blank line\n\nModified\n\n> ======\n>\n> 11. src/include/replication/walreceiver.h\n>\n> @@ -183,6 +183,8 @@ typedef struct\n> bool streaming; /* Streaming of large transactions */\n> bool twophase; /* Streaming of two-phase transactions at\n> * prepare time */\n> + char *origin; /* Publish the data originated from the\n> + * specified origin */\n> } logical;\n>\n> Reword comment same as suggested in review comment #9\n\nModified\n\n> ======\n>\n> 12. src/test/subscription/t/032_origin.pl\n>\n> +# Test logical replication using origin option.\n>\n> # Test the CREATE SUBSCRIPTION 'origin' parameter\n\nModified\n\n> ~~~\n>\n> 13. src/test/subscription/t/032_origin.pl\n>\n> +# check that the data published from node_C to node_B is not sent to node_A\n>+$result = $node_A->safe_psql('postgres', \"SELECT * FROM tab_full ORDER BY 1;\");\n> +is($result, qq(11\n> +12), 'Remote data originating from another node (not the publisher)\n> is not replicated when origin option is local'\n> +);\n>\n> \"option\" -> \"parameter\"\n\nModified\n\nThanks for the comments, the v18 patch attached at [1] has the fixes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm0Haovukx2q7Yd987Sm8fbQ0nsh8F0EWaO_qsw0uObGBQ%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 6 Jun 2022 10:47:09 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jun 6, 2022 1:14 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> The attached v18 patch has the fixes for the same.\r\n> \r\n\r\nThanks for updating the patch, here are some comments.\r\n\r\n0002 patch\r\n==============\r\n1. \r\n+ <varlistentry>\r\n+ <term><literal>origin</literal> (<type>string</type>)</term>\r\n+ <listitem>\r\n+ <para>\r\n\r\nIt maybe better if the type of \"origin\" parameter is enum, as it cannot be any\r\nstring and only has two valid values.\r\n\r\n2.\r\n@@ -607,6 +626,11 @@ CreateSubscription(ParseState *pstate, CreateSubscriptionStmt *stmt,\r\n \t\t\t\t\t LOGICALREP_TWOPHASE_STATE_PENDING :\r\n \t\t\t\t\t LOGICALREP_TWOPHASE_STATE_DISABLED);\r\n \tvalues[Anum_pg_subscription_subdisableonerr - 1] = BoolGetDatum(opts.disableonerr);\r\n+\tif (opts.origin)\r\n+\t\tvalues[Anum_pg_subscription_suborigin - 1] =\r\n+\t\t\tCStringGetTextDatum(opts.origin);\r\n+\telse\r\n+\t\tnulls[Anum_pg_subscription_suborigin - 1] = true;\r\n \tvalues[Anum_pg_subscription_subconninfo - 1] =\r\n \t\tCStringGetTextDatum(conninfo);\r\n \tif (opts.slot_name)\r\n\r\nDocument of \"CREATE SUBSCRIPTION\" says, the default value of \"origin\" is \"any\", so why not set\r\nsuborigin to \"any\" when user doesn't specify this parameter?\r\n\r\n\r\n0003 patch\r\n==============\r\n1.\r\n@@ -300,6 +310,11 @@ CREATE SUBSCRIPTION <replaceable class=\"parameter\">subscription_name</replaceabl\r\n replication from the publisher. The default is\r\n <literal>false</literal>.\r\n </para>\r\n+ <para>\r\n+ There is some interaction between the <literal>origin</literal>\r\n+ parameter and <literal>copy_data</literal> parameter. Refer to the\r\n+ <xref linkend=\"sql-createsubscription-notes\" /> for details.\r\n+ </para>\r\n </listitem>\r\n </varlistentry>\r\n\r\nI think this change should be put together with \"origin\" parameter, instead of\r\n\"disable_on_error\".\r\n\r\n\r\n0004 patch\r\n==============\r\n1.\r\n+ <para>\r\n+ Now the bidirectional logical replication setup is complete between\r\n+ <literal>node1</literal>, <literal>node2</literal> and\r\n+ <literal>node2</literal>. Any subsequent changes in one node will\r\n+ replicate the changes to the other nodes.\r\n+ </para>\r\n\r\nI think \"node1, node2 and node2\" should be \"node1, node2 and node3\".\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Mon, 6 Jun 2022 08:58:57 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jun 6, 2022 at 2:29 PM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Mon, Jun 6, 2022 1:14 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > The attached v18 patch has the fixes for the same.\n> >\n>\n> Thanks for updating the patch, here are some comments.\n>\n> 0002 patch\n> ==============\n> 1.\n> + <varlistentry>\n> + <term><literal>origin</literal> (<type>string</type>)</term>\n> + <listitem>\n> + <para>\n>\n> It maybe better if the type of \"origin\" parameter is enum, as it cannot be any\n> string and only has two valid values.\n\nCurrently we only support local and any. But this was designed so that\nit can be extended to support origin names. Users can provide a\nparticular origin name to be filtered.\nThe same was also discussed in pg unconference as mentioned in [1]\n\n> 2.\n> @@ -607,6 +626,11 @@ CreateSubscription(ParseState *pstate, CreateSubscriptionStmt *stmt,\n> LOGICALREP_TWOPHASE_STATE_PENDING :\n> LOGICALREP_TWOPHASE_STATE_DISABLED);\n> values[Anum_pg_subscription_subdisableonerr - 1] = BoolGetDatum(opts.disableonerr);\n> + if (opts.origin)\n> + values[Anum_pg_subscription_suborigin - 1] =\n> + CStringGetTextDatum(opts.origin);\n> + else\n> + nulls[Anum_pg_subscription_suborigin - 1] = true;\n> values[Anum_pg_subscription_subconninfo - 1] =\n> CStringGetTextDatum(conninfo);\n> if (opts.slot_name)\n>\n> Document of \"CREATE SUBSCRIPTION\" says, the default value of \"origin\" is \"any\", so why not set\n> suborigin to \"any\" when user doesn't specify this parameter?\n\nModified\n\n>\n> 0003 patch\n> ==============\n> 1.\n> @@ -300,6 +310,11 @@ CREATE SUBSCRIPTION <replaceable class=\"parameter\">subscription_name</replaceabl\n> replication from the publisher. The default is\n> <literal>false</literal>.\n> </para>\n> + <para>\n> + There is some interaction between the <literal>origin</literal>\n> + parameter and <literal>copy_data</literal> parameter. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for details.\n> + </para>\n> </listitem>\n> </varlistentry>\n>\n> I think this change should be put together with \"origin\" parameter, instead of\n> \"disable_on_error\".\n\nModified\n\n>\n> 0004 patch\n> ==============\n> 1.\n> + <para>\n> + Now the bidirectional logical replication setup is complete between\n> + <literal>node1</literal>, <literal>node2</literal> and\n> + <literal>node2</literal>. Any subsequent changes in one node will\n> + replicate the changes to the other nodes.\n> + </para>\n>\n> I think \"node1, node2 and node2\" should be \"node1, node2 and node3\".\n\nModified\n\nThanks for the comments, the attached v19 patch has the changes for the same.\n[1] - https://wiki.postgresql.org/wiki/PgCon_2022_Developer_Unconference#Logical_Replication_Origin_Filtering_and_Consistency\n\nRegards,\nVignesh","msg_date":"Wed, 8 Jun 2022 16:21:52 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Below are some review comments for the patch v18-0003\n\n1. Commit message\n\nThis patch does a couple of things:\nchange 1) Checks and throws an error if the publication tables were also\nsubscribing data in the publisher from other publishers when copy_data\nparameter is specified as 'ON' and origin parameter is specified as\n'local'.\nchange 2) Adds force value for copy_data parameter.\n\nSUGGESTION\nThis patch does a couple of things:\nchange 1) Checks and throws an error if 'copy_data = on' and 'origin =\nlocal' but the publication tables were also subscribing data in the\npublisher from other publishers.\nchange 2) Adds 'force' value for copy_data parameter.\n\n~~~\n\n2. Commit message - about the example\n\nAll my following review comments for the commit message are assuming\nthat the example steps are as they are written in the patch, but\nactually I felt that the example might be more complex than it needed\nto be: e.g\n- You don’t really need the node2 to have data\n- Therefore you don’t need all the added TRUNCATE complications\n\nE.g. I think you only need node1 (with data) and node2 (no data).\n\nThen node1 subscribes node2 with (origin=local, copy_data=off).\nThen node2 subscribes node1 with (origin=local, copy_data=on).\n- Demonstrates exception happens because node1 already had a subscription\n- Demonstrates need for the copy_data=force to override that exception\n\nSo please consider using a simpler example for this commit message\n\n~~~\n\n3. Commit message\n\nThe below help us understand how the first change will be useful:\n\nIf copy_data parameter was used with 'on' in step 5, then an error\nwill be thrown\nto alert the user to prevent inconsistent data being populated:\nCREATE SUBSCRIPTION sub_node2_node1 CONNECTION '<node1 details>'\nPUBLICATION pub_node1 WITH (copy_data = on, origin = local);\nERROR: CREATE/ALTER SUBSCRIPTION with origin and copy_data as true is not\nallowed when the publisher might have replicated data\n\nSUGGESTION\nThe steps below help to demonstrate how the new exception is useful:\n\nThe initial copy phase has no way to know the origin of the row data,\nso if 'copy_data = on' in the step 5 below, then an error will be\nthrown to prevent any potentially non-local data from being copied:\n<blank line>\ne.g\nCREATE SUBSCRIPTION ...\n\n~~~\n\n4. Commit message\n\nThe following will help us understand how the second change will be useful:\nLet's take a simple case where user is trying to setup bidirectional logical\nreplication between node1 and node2 where the two nodes have some pre-existing\ndata like below:\nnode1: Table t1 (c1 int) has data 11, 12, 13, 14\nnode2: Table t1 (c1 int) has data 21, 22, 23, 24\n\nSUGGESTION\nThe following steps help to demonstrate how the 'copy_data = force'\nchange will be useful:\n<blank line>\nLet's take a scenario where the user wants to set up bidirectional\nlogical replication between node1 and node2 where the same table on\neach node has pre-existing data. e.g.\nnode1: Table t1 (c1 int) has data 11, 12, 13, 14\nnode2: Table t1 (c1 int) has data 21, 22, 23, 24\n\n~~~\n\n5. Commit message\n\nstep 4:\nnode2=# CREATE SUBSCRIPTION sub_node2_node1 Connection '<node1 details>'\nnode2-# PUBLICATION pub_node1;\nCREATE SUBSCRIPTION\n\n\"Connection\" => \"CONNECTION\"\n\n~~~\n\n6. Commit message\n\nIf table t1 has a unique key, it will cause a unique key\nviolation and replication won't proceed.\n\nSUGGESTION\nIn case, table t1 has a unique key, it will lead to a unique key\nviolation and replication won't proceed.\n\n~~~\n\n7. Commit message\n\nstep 3: Create a subscription in node1 to subscribe to node2. Use\ncopy_data specified as on so that the existing table data is copied during\ninitial sync:\n\nSUGGESTION\nstep 3: Create a subscription in node1 to subscribe to node2. Use\n'copy_data = on' so that the existing table data is copied during\ninitial sync:\n\n~~~\n\n8. Commit message\n\nstep 4: Adjust the publication publish settings so that truncate is not\npublished to the subscribers and truncate the table data in node2:\n\nSUGGESTION (only added a comma)\nstep 4: Adjust the publication publish settings so that truncate is\nnot published to the subscribers, and truncate the table data in\nnode2:\n\n~~~\n\n9. Commit message\n\nstep 5: Create a subscription in node2 to subscribe to node1. Use copy_data\nspecified as force when creating a subscription to node1 so that the existing\ntable data is copied during initial sync:\n\nSUGGESTION\nstep 5: Create a subscription in node2 to subscribe to node1. Use\n'copy_data = force' when creating a subscription to node1 so that the\nexisting table data is copied during initial sync:\n\n======\n\n10. doc/src/sgml/ref/create_subscription.sgml\n\n@@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n can have non-existent publications.\n </para>\n\n+ <para>\n+ If subscription is created with <literal>origin = local</literal> and\n+ <literal>copy_data = on</literal>, it will check if the publisher tables are\n+ being subscribed to any other publisher and throw an error to prevent\n+ inconsistent data in the subscription. The user can continue with the copy\n+ operation without throwing any error in this case by specifying\n+ <literal>copy_data = force</literal>.\n+ </para>\n\nSUGGESTION (minor rewording)\nIf the subscription is created with <literal>origin = local</literal>\nand <literal>copy_data = on</literal>, it will check if the publisher\ntables are being subscribed to any other publisher and, if so, then\nthrow an error to prevent possible non-local data from being copied.\nThe user can override this check and continue with the copy operation\nby specifying <literal>copy_data = force</literal>.\n\n======\n\n11. src/backend/commands/subscriptioncmds.c - parse_subscription_options\n\n From [1]:\n>> What about also allowing copy_data = 2, and making it equivalent to \"force\"?\n> Vignesh: I felt the existing looks ok, no need to support 2. It might confuse the user.\n\nI don't think it would be confusing, but I also don't feel strongly\nenough to debate it. Anyway, I could not find a similar precedent, so\nyour decision is fine.\n\n~~~\n\n12. src/backend/commands/subscriptioncmds.c - parse_subscription_options\n\n@@ -339,17 +406,16 @@ parse_subscription_options(ParseState *pstate,\nList *stmt_options,\n errmsg(\"%s and %s are mutually exclusive options\",\n \"connect = false\", \"create_slot = true\")));\n\n- if (opts->copy_data &&\n- IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\n+ if (opts->copy_data && IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\n ereport(ERROR,\n\nThis is just a formatting change. Is it needed for this patch? patch.\n\n~~~\n\n13. src/backend/commands/subscriptioncmds.c - AlterSubscription_refresh\n\n@@ -730,7 +798,8 @@ CreateSubscription(ParseState *pstate,\nCreateSubscriptionStmt *stmt,\n * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\n * PUBLICATION to work.\n */\n- if (opts.twophase && !opts.copy_data && tables != NIL)\n+ if (opts.twophase && opts.copy_data == COPY_DATA_OFF &&\n+ tables != NIL)\n twophase_enabled = true;\n\nWhy is this change needed? I thought the original code is OK now since\nCOPY_DATA_OFF = 0\n\n~~~\n\n14. src/backend/commands/subscriptioncmds.c - AlterSubscription\n\n@@ -1265,7 +1337,8 @@ AlterSubscription(ParseState *pstate,\nAlterSubscriptionStmt *stmt,\n *\n * For more details see comments atop worker.c.\n */\n- if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\n+ if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\n+ opts.copy_data)\n\nThis is just a formatting change. Is it needed for this patch?\n\n~~~\n\n15. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n\n+\n+ if (!origin || (strcmp(origin, \"local\") != 0) || copydata != COPY_DATA_ON)\n+ return;\n+\n\nThis condition could be rearranged to put the strcmp last so it is not\ncalled unless absolutely necessary.\n\n~~~\n\n16. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n\n+ appendStringInfoString(&cmd,\n+ \"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\nPS.srrelid as replicated\\n\"\n+ \"FROM pg_publication P,\\n\"\n\nThe line is too long; needs wrapping.\n\n~~~\n\n17. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n\n+ if (!slot_attisnull(slot, 3))\n+ ereport(ERROR,\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n+ errmsg(\"table:%s.%s might have replicated data in the publisher\",\n+ nspname, relname),\n+ errdetail(\"CREATE/ALTER SUBSCRIPTION with origin as local and\ncopy_data as true is not allowed when the publisher might have\nreplicated data.\"),\n+ errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force.\"));\n\nI felt the errmsg may be easier to read using \"=\" instead of \"as\".\nAnyway, it would be more consistent with the errhint. Also, change the\n\"true\" to \"on\" to be consistent with the errhint.\n\nSUGGESTION\nerrdetail(\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data\n= on is not allowed when the publisher might have replicated data.\"),\n\n\n------\n[1] https://www.postgresql.org/message-id/CALDaNm3iLLxP4OV%2ByQHs-c71P6zQ9W8D30DGsve1SQs_1pFsSQ%40mail.gmail.com\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 10 Jun 2022 14:53:27 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Below are some review comments for the patch v18-0004\n\n======\n\n1. Commit message\n\nDocument the steps for the following:\na) Create a two-node bidirectional replication when there is no data in both the\nnodes.\nb) Adding a new node when there is no data in any of the nodes.\nc) Adding a new node when data is present in the existing nodes.\nd) Generic steps to add a new node to the existing set of nodes.\n\nSUGGESTION (minor changes to make the tense consistent)\nDocuments the steps for the following:\na) Creating a two-node bidirectional replication when there is no data\nin both nodes.\nb) Adding a new node when there is no data in any of the nodes.\nc) Adding a new node when data is present in the existing nodes.\nd) Generic steps for adding a new node to an existing set of nodes.\n\n======\n\n2. doc/src/sgml/logical-replication.sgml - blurb\n\n+ <para>\n+ Bidirectional replication is useful in creating a multi-master database\n+ which helps in performing read/write operations from any of the nodes.\n\nSUGGESTION\nBidirectional replication is useful for creating a multi-master\ndatabase environment for replicating read/write operations performed\nby any of the member nodes.\n\n~~~\n\n3. doc/src/sgml/logical-replication.sgml - warning\n\n+ <warning>\n+ <para>\n+ Setting up bidirectional logical replication across nodes\nrequires multiple\n+ steps to be performed on various nodes. Because not all operations are\n+ transactional, the user is advised to take backups.\n+ </para>\n+ </warning>\n\nSUGGESTION (but keep your formatting)\nSetting up bidirectional logical replication requires multiple steps\nto be performed on various nodes. Because...\n\n~~~\n\n4. doc/src/sgml/logical-replication.sgml - Setting bidirectional\nreplication between two nodes\n\n+ <para>\n+ The steps to create a two-node bidirectional replication when there is no\n+ data in both the nodes <literal>node1</literal> and\n+ <literal>node2</literal> are given below:\n+ </para>\n\nSUGGESTION (but keep your formatting)\nThe following steps demonstrate how to create a two-node bidirectional\nreplication when there is no table data present in both nodes node1\nand node2:\n\n~~~\n\n5. doc/src/sgml/logical-replication.sgml - Setting bidirectional\nreplication between two nodes\n\n+ <para>\n+ Lock the required tables of <literal>node1</literal> and\n+ <literal>node2</literal> in <literal>EXCLUSIVE</literal> mode until the\n+ setup is completed.\n+ </para>\n\nInstead of \"the required tables\" shouldn't this just say \"table t1\"?\n\n~~~\n\n6. doc/src/sgml/logical-replication.sgml - Setting bidirectional\nreplication between two nodes\n\n+ <para>\n+ Now the bidirectional logical replication setup is complete between\n+ <literal>node1</literal>, <literal>node2</literal> and\n+ <literal>node2</literal>. Any subsequent changes in one node will\n+ replicate the changes to the other nodes.\n+ </para>\n\nSUGGESTION (for 2nd sentence, and keep your formatting)\nAny incremental changes from node1 will be replicated to node2, and\nany incremental changes from node2 will be replicated to node1.\n\n~~~\n\n7. doc/src/sgml/logical-replication.sgml - Adding a new node when\nthere is no data in any of the nodes\n\n+ <para>\n+ Adding a new node <literal>node3</literal> to the existing\n+ <literal>node1</literal> and <literal>node2</literal> when there is no data\n+ in any of the nodes requires setting up subscription in\n+ <literal>node1</literal> and <literal>node2</literal> to replicate the data\n+ from <literal>node3</literal> and setting up subscription in\n+ <literal>node3</literal> to replicate data from <literal>node1</literal>\n+ and <literal>node2</literal>.\n+ </para>\n\nSUGGESTION (but keep your formatting)\nThe following steps demonstrate adding a new node node3 to the\nexisting node1 and node2 when there is no t1 data in any of the nodes.\nThis requires creating subscriptions in node1 and node2 to replicate\nthe data from node3 and creating subscriptions in node3 to replicate\ndata from node1 and node2.\n\n~~~\n\n8. doc/src/sgml/logical-replication.sgml - Adding a new node when\nthere is no data in any of the nodes\n\n+ <note>\n+ <para>\n+ It is assumed that the bidirectional logical replication between\n+ <literal>node1</literal> and <literal>node2</literal> is already\ncompleted.\n+ </para>\n+ </note>\n\nIMO this note should just be a text note in the previous paragraph\ninstead of an SGML <note>. e.g. \"Note: These steps assume that...\"\n\n~~~\n\n9. doc/src/sgml/logical-replication.sgml - Adding a new node when\nthere is no data in any of the nodes\n\n+ <para>\n+ Lock the required tables of all the nodes <literal>node1</literal>,\n+ <literal>node2</literal> and <literal>node3</literal> in\n+ <literal>EXCLUSIVE</literal> mode until the setup is completed.\n+ </para>\n\nInstead of \"the required tables\" shouldn't this just say \"table t1\"?\n\n~~~\n\n10. doc/src/sgml/logical-replication.sgml - Adding a new node when\nthere is no data in any of the nodes\n\n+ <para>\n+ Now the bidirectional logical replication setup is complete between\n+ <literal>node1</literal>, <literal>node2</literal> and\n+ <literal>node2</literal>. Any subsequent changes in one node will\n+ replicate the changes to the other nodes.\n+ </para>\n\nSUGGESTION (2nd sentence)\nIncremental changes made in any node will be replicated to the other two nodes.\n\n~~~\n\n11. doc/src/sgml/logical-replication.sgml - Adding a new node when\ndata is present in the existing nodes\n\n+ <para>\n+ Adding a new node <literal>node3</literal> which has no data to the\n+ existing <literal>node1</literal> and <literal>node2</literal> when data\n+ is present in existing nodes <literal>node1</literal> and\n+ <literal>node2</literal> needs similar steps. The only change required\n+ here is that <literal>node3</literal> should create a subscription with\n+ <literal>copy_data = force</literal> to one of the existing nodes to\n+ receive the existing data during initial data synchronization.\n+ </para>\n\nSUGGESTION (but keep your formatting)\nThe following steps demonstrate adding a new node node3 which has no\nt1 data to the existing node1 and node2 where t1 data is present. This\nneeds similar steps; the only change required here is that node3\nshould create a subscription with copy_data = force to one of the\nexisting nodes so it can receive the existing t1 data during initial\ndata synchronization.\n\n~~~\n\n12 doc/src/sgml/logical-replication.sgml - Adding a new node when data\nis present in the existing nodes\n\n+ <note>\n+ <para>\n+ It is assumed that the bidirectional logical replication between\n+ <literal>node1</literal> and <literal>node2</literal> is already\ncompleted.\n+ The nodes <literal>node1</literal> and <literal>node2</literal> has some\n+ pre-existing data in table t1 that is synchronized in both the nodes.\n+ </para>\n+ </note>\n\n12a.\nIMO this note should just be text in the previous paragraph instead of\nan SGML <note>. e.g. \"Note: These steps assume that...\"\n\n12b.\nSUGGESTION (minor rewording; keep your formatting)\nNote: These steps assume that the bidirectional logical replication\nbetween node1 and node2 is already completed, and the pre-existing\ndata in table t1 is already synchronized in both those nodes.\n\n~~~\n\n13. doc/src/sgml/logical-replication.sgml - Adding a new node when\ndata is present in the existing nodes\n\n+ <para>\n+ Lock the required tables of <literal>node2</literal> and\n+ <literal>node3</literal> in <literal>EXCLUSIVE</literal> mode until the\n+ setup is completed. No need to lock the tables in <literal>node1</literal>\n+ as any data changes made will be synchronized while creating the\n+ subscription with <literal>copy_data</literal> specified as\n+ <literal>force</literal>.\n+ </para>\n\n13a.\nInstead of \"the required tables\" shouldn't this just say \"table t1\"?\n\n13b.\nSUGGESTION (2nd sentence; keep your formatting)\nThere is no need to lock table t1 in node1 because any data changes\nmade will be synchronized while creating the subscription with\ncopy_data = force.\n\n~~~\n\n14. doc/src/sgml/logical-replication.sgml - Adding a new node when\ndata is present in the existing nodes\n\n+ <para>\n+ Create a subscription in <literal>node3</literal> to subscribe to\n+ <literal>node1</literal>. Use <literal>copy_data</literal> specified as\n+ <literal>force</literal> so that the existing table data is\n+ copied during initial sync:\n\nSUGGESTION (2nd sentence; keep your formatting)\nUse copy_data = force so that the existing table data is copied during\nthe initial sync:\n\n~~~\n\n15. doc/src/sgml/logical-replication.sgml - Adding a new node when\ndata is present in the existing nodes\n\n+ <para>\n+ Create a subscription in <literal>node3</literal> to subscribe to\n+ <literal>node2</literal>. Use <literal>copy_data</literal> specified as\n+ <literal>off</literal> because the initial table data would have been\n+ already copied in the previous step:\n\nSUGGESTION (2nd sentence; keep your formatting)\nUse copy_data = off because the initial table data would have been\nalready copied in the previous step:\n\n~~~\n\n16. doc/src/sgml/logical-replication.sgml - Adding a new node when\ndata is present in the existing nodes\n\n+ <para>\n+ Now the bidirectional logical replication setup is complete between\n+ <literal>node1</literal>, <literal>node2</literal> and\n+ <literal>node2</literal>. Any subsequent changes in one node will\n+ replicate the changes to the other nodes.\n+ </para>\n\n16a.\nShould say node1, node2 and node3\n\n16b.\nSUGGESTION (2nd sentence – same as the previous comment)\nIncremental changes made in any node will be replicated to the other two nodes.\n\n~~~\n\n17. doc/src/sgml/logical-replication.sgml - Adding a new node when\ndata is present in the new node\n\n+ <warning>\n+ <para>\n+ Adding a new node <literal>node3</literal> to the existing\n+ <literal>node1</literal> and <literal>node2</literal> when data is present\n+ in the new node <literal>node3</literal> is not possible.\n+ </para>\n+ </warning>\n\n17a.\nIMO\n- Not really necessary to name the nodes, because this is a generic statement\n- Maybe say \"not supported\" instead of \"not possible\". e.g. there is\nno ERROR check for this case is there?\n\nSUGGESTION\nAdding a new node when data is present in the new node tables is not supported.\n\n17b.\nI am not sure but I felt this advice seemed more like an SGML <note>;\nnote a <warning>\n\n~~~\n\n18. doc/src/sgml/logical-replication.sgml - Generic steps to add a new\nnode to the existing set of nodes\n\n+ <title>Generic steps to add a new node to the existing set of nodes</title>\n\nSUGGESTION\nGeneric steps for adding a new node to an existing set of nodes\n\n~~~\n\n19. doc/src/sgml/logical-replication.sgml - Generic steps to add a new\nnode to the existing set of nodes\n\n+ <para>\n+ 1. Create the required publication on the new node.\n+ </para>\n+ <para>\n+ 2. Lock the required tables of the new node in <literal>EXCLUSIVE</literal>\n+ mode until the setup is complete. This is required to prevent any\n+ modifications from happening in the new node. If data modifications occur\n+ after step-3, there is a chance that the modifications will be published to\n+ the first node and then synchronized back to the new node while creating\n+ subscription in step-5 resulting in inconsistent data.\n+ </para>\n+ <para>\n+ 3. Create subscriptions on existing nodes pointing to publication on\n+ the new node with <literal>origin</literal> parameter specified as\n+ <literal>local</literal> and <literal>copy_data</literal> specified as\n+ <literal>off</literal>.\n+ </para>\n+ <para>\n+ 4. Lock the required tables of the existing nodes except the first node in\n+ <literal>EXCLUSIVE</literal> mode until the setup is complete. This is\n+ required to prevent any modifications from happening. If data modifications\n+ occur, there is a chance that the modifications done in between step-5 and\n+ step-6 will not be synchronized to the new node resulting in inconsistent\n+ data. No need to lock the tables in the first node as any data changes\n+ made will be synchronized while creating the subscription with\n+ <literal>copy_data</literal> specified as <literal>force</literal>.\n+ </para>\n+ <para>\n+ 5. Create a subscription on the new node pointing to publication on the\n+ first node with <literal>origin</literal> parameter specified as\n+ <literal>local</literal> and <literal>copy_data</literal> parameter\n+ specified as <literal>force</literal>.\n+ </para>\n+ <para>\n+ 6. Create subscriptions on the new node pointing to publications on the\n+ remaining nodes with <literal>origin</literal> parameter specified as\n+ <literal>local</literal> and <literal>copy_data</literal> parameter\n+ specifiedas <literal>off</literal>.\n+ </para>\n\nFollowing suggestions make the following changes:\n- Some minor rewording\n- Change the step names\n- Use \"=\" instead of \"specified as\" for the parameters\n- I felt it is more readable if the explanatory notes are separate\nparagraphs from the steps. Maybe if you could indent them or something\nit would be even better.\n- Added a couple more explanatory notes\n\nSUGGESTION (using those above changes; please keep your formatting)\n\nStep-1: Create a publication on the new node.\n\nStep-2: Lock the required tables of the new node in EXCLUSIVE mode\nuntil the setup is complete.\n\n(This lock is necessary to prevent any modifications from happening in\nthe new node because if data modifications occurred after Step-3,\nthere is a chance that the modifications will be published to the\nfirst node and then synchronized back to the new node while creating\nthe subscription in Step-5. This would result in inconsistent data).\n\nStep-3. Create subscriptions on existing nodes to the publication on\nthe new node with origin = local and copy_data = off.\n\n(The copy_data = off is OK here because it is asserted that the\npublished tables of the new node will have no pre-existing data).\n\nStep-4. Lock the required tables of the existing nodes except the\nfirst node in EXCLUSIVE mode until the setup is complete.\n\n(This lock is necessary to prevent any modifications from happening.\nIf data modifications occur, there is a chance that modifications done\nbetween Step-5 and Step-6 will not be synchronized to the new node.\nThis would result in inconsistent data. There is no need to lock table\nt1 in node1 because any data changes made will be synchronized while\ncreating the subscription with copy_data = force).\n\nStep-5. Create a subscription on the new node to the publication on\nthe first node with origin = local and copy_data = force.\n\n(This will copy the same table data from the existing nodes to the new node)\n\nStep-6. Create subscriptions on the new node to publications on the\nremaining nodes with origin = local and copy_data = off.\n\n(The copy_data = off is OK here because the existing node data was\nalready copied to the new node in Step-5)\n\n======\n\n20. doc/src/sgml/ref/create_subscription.sgml\n\n- <literal>copy_data = force</literal>.\n+ <literal>copy_data = force</literal>. Refer to the\n+ <xref linkend=\"logical-replication-bidirectional\"/> on how\n+ <literal>copy_data</literal> and <literal>origin</literal> can be used\n+ in bidirectional replication.\n </para>\n\nSUGGESTION (keep your formatting)\nRefer to <xref linkend=\"logical-replication-bidirectional\"/> for how\n<literal>copy_data</literal> and <literal>origin</literal> can be used\nin bidirectional replication.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 10 Jun 2022 18:39:04 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jun 10, 2022 at 10:23 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Below are some review comments for the patch v18-0003\n>\n> 1. Commit message\n>\n> This patch does a couple of things:\n> change 1) Checks and throws an error if the publication tables were also\n> subscribing data in the publisher from other publishers when copy_data\n> parameter is specified as 'ON' and origin parameter is specified as\n> 'local'.\n> change 2) Adds force value for copy_data parameter.\n>\n> SUGGESTION\n> This patch does a couple of things:\n> change 1) Checks and throws an error if 'copy_data = on' and 'origin =\n> local' but the publication tables were also subscribing data in the\n> publisher from other publishers.\n> change 2) Adds 'force' value for copy_data parameter.\n\nModified\n\n> ~~~\n>\n> 2. Commit message - about the example\n>\n> All my following review comments for the commit message are assuming\n> that the example steps are as they are written in the patch, but\n> actually I felt that the example might be more complex than it needed\n> to be: e.g\n> - You don’t really need the node2 to have data\n> - Therefore you don’t need all the added TRUNCATE complications\n>\n> E.g. I think you only need node1 (with data) and node2 (no data).\n>\n> Then node1 subscribes node2 with (origin=local, copy_data=off).\n> Then node2 subscribes node1 with (origin=local, copy_data=on).\n> - Demonstrates exception happens because node1 already had a subscription\n> - Demonstrates need for the copy_data=force to override that exception\n>\n> So please consider using a simpler example for this commit message\n\nModified\n\n> ~~~\n>\n> 3. Commit message\n>\n> The below help us understand how the first change will be useful:\n>\n> If copy_data parameter was used with 'on' in step 5, then an error\n> will be thrown\n> to alert the user to prevent inconsistent data being populated:\n> CREATE SUBSCRIPTION sub_node2_node1 CONNECTION '<node1 details>'\n> PUBLICATION pub_node1 WITH (copy_data = on, origin = local);\n> ERROR: CREATE/ALTER SUBSCRIPTION with origin and copy_data as true is not\n> allowed when the publisher might have replicated data\n>\n> SUGGESTION\n> The steps below help to demonstrate how the new exception is useful:\n>\n> The initial copy phase has no way to know the origin of the row data,\n> so if 'copy_data = on' in the step 5 below, then an error will be\n> thrown to prevent any potentially non-local data from being copied:\n> <blank line>\n> e.g\n> CREATE SUBSCRIPTION ...\n\nModified\n\n> ~~~\n>\n> 4. Commit message\n>\n> The following will help us understand how the second change will be useful:\n> Let's take a simple case where user is trying to setup bidirectional logical\n> replication between node1 and node2 where the two nodes have some pre-existing\n> data like below:\n> node1: Table t1 (c1 int) has data 11, 12, 13, 14\n> node2: Table t1 (c1 int) has data 21, 22, 23, 24\n>\n> SUGGESTION\n> The following steps help to demonstrate how the 'copy_data = force'\n> change will be useful:\n> <blank line>\n> Let's take a scenario where the user wants to set up bidirectional\n> logical replication between node1 and node2 where the same table on\n> each node has pre-existing data. e.g.\n> node1: Table t1 (c1 int) has data 11, 12, 13, 14\n> node2: Table t1 (c1 int) has data 21, 22, 23, 24\n\nModified\n\n> ~~~\n>\n> 5. Commit message\n>\n> step 4:\n> node2=# CREATE SUBSCRIPTION sub_node2_node1 Connection '<node1 details>'\n> node2-# PUBLICATION pub_node1;\n> CREATE SUBSCRIPTION\n>\n> \"Connection\" => \"CONNECTION\"\n\nModified\n\n> ~~~\n>\n> 6. Commit message\n>\n> If table t1 has a unique key, it will cause a unique key\n> violation and replication won't proceed.\n>\n> SUGGESTION\n> In case, table t1 has a unique key, it will lead to a unique key\n> violation and replication won't proceed.\n\nModified\n\n> ~~~\n>\n> 7. Commit message\n>\n> step 3: Create a subscription in node1 to subscribe to node2. Use\n> copy_data specified as on so that the existing table data is copied during\n> initial sync:\n>\n> SUGGESTION\n> step 3: Create a subscription in node1 to subscribe to node2. Use\n> 'copy_data = on' so that the existing table data is copied during\n> initial sync:\n\nModified\n\n> ~~~\n>\n> 8. Commit message\n>\n> step 4: Adjust the publication publish settings so that truncate is not\n> published to the subscribers and truncate the table data in node2:\n>\n> SUGGESTION (only added a comma)\n> step 4: Adjust the publication publish settings so that truncate is\n> not published to the subscribers, and truncate the table data in\n> node2:\n\nThis content is not required any more, I have removed it.\n\n> ~~~\n>\n> 9. Commit message\n>\n> step 5: Create a subscription in node2 to subscribe to node1. Use copy_data\n> specified as force when creating a subscription to node1 so that the existing\n> table data is copied during initial sync:\n>\n> SUGGESTION\n> step 5: Create a subscription in node2 to subscribe to node1. Use\n> 'copy_data = force' when creating a subscription to node1 so that the\n> existing table data is copied during initial sync:\n\nModified\n\n> ======\n>\n> 10. doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> can have non-existent publications.\n> </para>\n>\n> + <para>\n> + If subscription is created with <literal>origin = local</literal> and\n> + <literal>copy_data = on</literal>, it will check if the publisher tables are\n> + being subscribed to any other publisher and throw an error to prevent\n> + inconsistent data in the subscription. The user can continue with the copy\n> + operation without throwing any error in this case by specifying\n> + <literal>copy_data = force</literal>.\n> + </para>\n>\n> SUGGESTION (minor rewording)\n> If the subscription is created with <literal>origin = local</literal>\n> and <literal>copy_data = on</literal>, it will check if the publisher\n> tables are being subscribed to any other publisher and, if so, then\n> throw an error to prevent possible non-local data from being copied.\n> The user can override this check and continue with the copy operation\n> by specifying <literal>copy_data = force</literal>.\n\nModified\n\n> ======\n>\n> 11. src/backend/commands/subscriptioncmds.c - parse_subscription_options\n>\n> From [1]:\n> >> What about also allowing copy_data = 2, and making it equivalent to \"force\"?\n> > Vignesh: I felt the existing looks ok, no need to support 2. It might confuse the user.\n>\n> I don't think it would be confusing, but I also don't feel strongly\n> enough to debate it. Anyway, I could not find a similar precedent, so\n> your decision is fine.\n\nOk\n\n> ~~~\n>\n> 12. src/backend/commands/subscriptioncmds.c - parse_subscription_options\n>\n> @@ -339,17 +406,16 @@ parse_subscription_options(ParseState *pstate,\n> List *stmt_options,\n> errmsg(\"%s and %s are mutually exclusive options\",\n> \"connect = false\", \"create_slot = true\")));\n>\n> - if (opts->copy_data &&\n> - IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\n> + if (opts->copy_data && IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\n> ereport(ERROR,\n>\n> This is just a formatting change. Is it needed for this patch? patch.\n\nModified\n\n> ~~~\n>\n> 13. src/backend/commands/subscriptioncmds.c - AlterSubscription_refresh\n>\n> @@ -730,7 +798,8 @@ CreateSubscription(ParseState *pstate,\n> CreateSubscriptionStmt *stmt,\n> * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\n> * PUBLICATION to work.\n> */\n> - if (opts.twophase && !opts.copy_data && tables != NIL)\n> + if (opts.twophase && opts.copy_data == COPY_DATA_OFF &&\n> + tables != NIL)\n> twophase_enabled = true;\n>\n> Why is this change needed? I thought the original code is OK now since\n> COPY_DATA_OFF = 0\n\nModified\n\n> ~~~\n>\n> 14. src/backend/commands/subscriptioncmds.c - AlterSubscription\n>\n> @@ -1265,7 +1337,8 @@ AlterSubscription(ParseState *pstate,\n> AlterSubscriptionStmt *stmt,\n> *\n> * For more details see comments atop worker.c.\n> */\n> - if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\n> + if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\n> + opts.copy_data)\n>\n> This is just a formatting change. Is it needed for this patch?\n\nModified\n\n> ~~~\n>\n> 15. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> +\n> + if (!origin || (strcmp(origin, \"local\") != 0) || copydata != COPY_DATA_ON)\n> + return;\n> +\n>\n> This condition could be rearranged to put the strcmp last so it is not\n> called unless absolutely necessary.\n\nModified\n\n> ~~~\n>\n> 16. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> + appendStringInfoString(&cmd,\n> + \"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\n> PS.srrelid as replicated\\n\"\n> + \"FROM pg_publication P,\\n\"\n>\n> The line is too long; needs wrapping.\n\nModified\n\n> ~~~\n>\n> 17. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> + if (!slot_attisnull(slot, 3))\n> + ereport(ERROR,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"table:%s.%s might have replicated data in the publisher\",\n> + nspname, relname),\n> + errdetail(\"CREATE/ALTER SUBSCRIPTION with origin as local and\n> copy_data as true is not allowed when the publisher might have\n> replicated data.\"),\n> + errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force.\"));\n>\n> I felt the errmsg may be easier to read using \"=\" instead of \"as\".\n> Anyway, it would be more consistent with the errhint. Also, change the\n> \"true\" to \"on\" to be consistent with the errhint.\n>\n> SUGGESTION\n> errdetail(\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data\n> = on is not allowed when the publisher might have replicated data.\"),\n\nModified\n\nThanks for the comments, the attached v20 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Tue, 14 Jun 2022 17:34:52 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jun 10, 2022 at 2:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Below are some review comments for the patch v18-0004\n>\n> ======\n>\n> 1. Commit message\n>\n> Document the steps for the following:\n> a) Create a two-node bidirectional replication when there is no data in both the\n> nodes.\n> b) Adding a new node when there is no data in any of the nodes.\n> c) Adding a new node when data is present in the existing nodes.\n> d) Generic steps to add a new node to the existing set of nodes.\n>\n> SUGGESTION (minor changes to make the tense consistent)\n> Documents the steps for the following:\n> a) Creating a two-node bidirectional replication when there is no data\n> in both nodes.\n> b) Adding a new node when there is no data in any of the nodes.\n> c) Adding a new node when data is present in the existing nodes.\n> d) Generic steps for adding a new node to an existing set of nodes.\n\nModified\n\n> ======\n>\n> 2. doc/src/sgml/logical-replication.sgml - blurb\n>\n> + <para>\n> + Bidirectional replication is useful in creating a multi-master database\n> + which helps in performing read/write operations from any of the nodes.\n>\n> SUGGESTION\n> Bidirectional replication is useful for creating a multi-master\n> database environment for replicating read/write operations performed\n> by any of the member nodes.\n\nModified\n\n> ~~~\n>\n> 3. doc/src/sgml/logical-replication.sgml - warning\n>\n> + <warning>\n> + <para>\n> + Setting up bidirectional logical replication across nodes\n> requires multiple\n> + steps to be performed on various nodes. Because not all operations are\n> + transactional, the user is advised to take backups.\n> + </para>\n> + </warning>\n>\n> SUGGESTION (but keep your formatting)\n> Setting up bidirectional logical replication requires multiple steps\n> to be performed on various nodes. Because...\n\nModified\n\n> ~~~\n>\n> 4. doc/src/sgml/logical-replication.sgml - Setting bidirectional\n> replication between two nodes\n>\n> + <para>\n> + The steps to create a two-node bidirectional replication when there is no\n> + data in both the nodes <literal>node1</literal> and\n> + <literal>node2</literal> are given below:\n> + </para>\n>\n> SUGGESTION (but keep your formatting)\n> The following steps demonstrate how to create a two-node bidirectional\n> replication when there is no table data present in both nodes node1\n> and node2:\n\nModified\n\n> ~~~\n>\n> 5. doc/src/sgml/logical-replication.sgml - Setting bidirectional\n> replication between two nodes\n>\n> + <para>\n> + Lock the required tables of <literal>node1</literal> and\n> + <literal>node2</literal> in <literal>EXCLUSIVE</literal> mode until the\n> + setup is completed.\n> + </para>\n>\n> Instead of \"the required tables\" shouldn't this just say \"table t1\"?\n\nModified\n\n> ~~~\n>\n> 6. doc/src/sgml/logical-replication.sgml - Setting bidirectional\n> replication between two nodes\n>\n> + <para>\n> + Now the bidirectional logical replication setup is complete between\n> + <literal>node1</literal>, <literal>node2</literal> and\n> + <literal>node2</literal>. Any subsequent changes in one node will\n> + replicate the changes to the other nodes.\n> + </para>\n>\n> SUGGESTION (for 2nd sentence, and keep your formatting)\n> Any incremental changes from node1 will be replicated to node2, and\n> any incremental changes from node2 will be replicated to node1.\n\nModified\n\n> ~~~\n>\n> 7. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> there is no data in any of the nodes\n>\n> + <para>\n> + Adding a new node <literal>node3</literal> to the existing\n> + <literal>node1</literal> and <literal>node2</literal> when there is no data\n> + in any of the nodes requires setting up subscription in\n> + <literal>node1</literal> and <literal>node2</literal> to replicate the data\n> + from <literal>node3</literal> and setting up subscription in\n> + <literal>node3</literal> to replicate data from <literal>node1</literal>\n> + and <literal>node2</literal>.\n> + </para>\n>\n> SUGGESTION (but keep your formatting)\n> The following steps demonstrate adding a new node node3 to the\n> existing node1 and node2 when there is no t1 data in any of the nodes.\n> This requires creating subscriptions in node1 and node2 to replicate\n> the data from node3 and creating subscriptions in node3 to replicate\n> data from node1 and node2.\n\nModified\n\n> ~~~\n>\n> 8. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> there is no data in any of the nodes\n>\n> + <note>\n> + <para>\n> + It is assumed that the bidirectional logical replication between\n> + <literal>node1</literal> and <literal>node2</literal> is already\n> completed.\n> + </para>\n> + </note>\n>\n> IMO this note should just be a text note in the previous paragraph\n> instead of an SGML <note>. e.g. \"Note: These steps assume that...\"\n\nModified\n\n> ~~~\n>\n> 9. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> there is no data in any of the nodes\n>\n> + <para>\n> + Lock the required tables of all the nodes <literal>node1</literal>,\n> + <literal>node2</literal> and <literal>node3</literal> in\n> + <literal>EXCLUSIVE</literal> mode until the setup is completed.\n> + </para>\n>\n> Instead of \"the required tables\" shouldn't this just say \"table t1\"?\n\nModified\n\n> ~~~\n>\n> 10. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> there is no data in any of the nodes\n>\n> + <para>\n> + Now the bidirectional logical replication setup is complete between\n> + <literal>node1</literal>, <literal>node2</literal> and\n> + <literal>node2</literal>. Any subsequent changes in one node will\n> + replicate the changes to the other nodes.\n> + </para>\n>\n> SUGGESTION (2nd sentence)\n> Incremental changes made in any node will be replicated to the other two nodes.\n\nModified\n\n> ~~~\n>\n> 11. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> data is present in the existing nodes\n>\n> + <para>\n> + Adding a new node <literal>node3</literal> which has no data to the\n> + existing <literal>node1</literal> and <literal>node2</literal> when data\n> + is present in existing nodes <literal>node1</literal> and\n> + <literal>node2</literal> needs similar steps. The only change required\n> + here is that <literal>node3</literal> should create a subscription with\n> + <literal>copy_data = force</literal> to one of the existing nodes to\n> + receive the existing data during initial data synchronization.\n> + </para>\n>\n> SUGGESTION (but keep your formatting)\n> The following steps demonstrate adding a new node node3 which has no\n> t1 data to the existing node1 and node2 where t1 data is present. This\n> needs similar steps; the only change required here is that node3\n> should create a subscription with copy_data = force to one of the\n> existing nodes so it can receive the existing t1 data during initial\n> data synchronization.\n\nModified\n\n> ~~~\n>\n> 12 doc/src/sgml/logical-replication.sgml - Adding a new node when data\n> is present in the existing nodes\n>\n> + <note>\n> + <para>\n> + It is assumed that the bidirectional logical replication between\n> + <literal>node1</literal> and <literal>node2</literal> is already\n> completed.\n> + The nodes <literal>node1</literal> and <literal>node2</literal> has some\n> + pre-existing data in table t1 that is synchronized in both the nodes.\n> + </para>\n> + </note>\n>\n> 12a.\n> IMO this note should just be text in the previous paragraph instead of\n> an SGML <note>. e.g. \"Note: These steps assume that...\"\n\nModified\n\n> 12b.\n> SUGGESTION (minor rewording; keep your formatting)\n> Note: These steps assume that the bidirectional logical replication\n> between node1 and node2 is already completed, and the pre-existing\n> data in table t1 is already synchronized in both those nodes.\n\nModified\n\n> ~~~\n>\n> 13. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> data is present in the existing nodes\n>\n> + <para>\n> + Lock the required tables of <literal>node2</literal> and\n> + <literal>node3</literal> in <literal>EXCLUSIVE</literal> mode until the\n> + setup is completed. No need to lock the tables in <literal>node1</literal>\n> + as any data changes made will be synchronized while creating the\n> + subscription with <literal>copy_data</literal> specified as\n> + <literal>force</literal>.\n> + </para>\n>\n> 13a.\n> Instead of \"the required tables\" shouldn't this just say \"table t1\"?\n\nModified\n\n> 13b.\n> SUGGESTION (2nd sentence; keep your formatting)\n> There is no need to lock table t1 in node1 because any data changes\n> made will be synchronized while creating the subscription with\n> copy_data = force.\n\nModified\n\n> ~~~\n>\n> 14. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> data is present in the existing nodes\n>\n> + <para>\n> + Create a subscription in <literal>node3</literal> to subscribe to\n> + <literal>node1</literal>. Use <literal>copy_data</literal> specified as\n> + <literal>force</literal> so that the existing table data is\n> + copied during initial sync:\n>\n> SUGGESTION (2nd sentence; keep your formatting)\n> Use copy_data = force so that the existing table data is copied during\n> the initial sync:\n\nModified\n\n> ~~~\n>\n> 15. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> data is present in the existing nodes\n>\n> + <para>\n> + Create a subscription in <literal>node3</literal> to subscribe to\n> + <literal>node2</literal>. Use <literal>copy_data</literal> specified as\n> + <literal>off</literal> because the initial table data would have been\n> + already copied in the previous step:\n>\n> SUGGESTION (2nd sentence; keep your formatting)\n> Use copy_data = off because the initial table data would have been\n> already copied in the previous step:\n\nModified\n\n> ~~~\n>\n> 16. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> data is present in the existing nodes\n>\n> + <para>\n> + Now the bidirectional logical replication setup is complete between\n> + <literal>node1</literal>, <literal>node2</literal> and\n> + <literal>node2</literal>. Any subsequent changes in one node will\n> + replicate the changes to the other nodes.\n> + </para>\n>\n> 16a.\n> Should say node1, node2 and node3\n\nModified it in v19\n\n> 16b.\n> SUGGESTION (2nd sentence – same as the previous comment)\n> Incremental changes made in any node will be replicated to the other two nodes.\n\nModified\n\n> ~~~\n>\n> 17. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> data is present in the new node\n>\n> + <warning>\n> + <para>\n> + Adding a new node <literal>node3</literal> to the existing\n> + <literal>node1</literal> and <literal>node2</literal> when data is present\n> + in the new node <literal>node3</literal> is not possible.\n> + </para>\n> + </warning>\n>\n> 17a.\n> IMO\n> - Not really necessary to name the nodes, because this is a generic statement\n> - Maybe say \"not supported\" instead of \"not possible\". e.g. there is\n> no ERROR check for this case is there?\n>\n> SUGGESTION\n> Adding a new node when data is present in the new node tables is not supported.\n\nModified\n\n> 17b.\n> I am not sure but I felt this advice seemed more like an SGML <note>;\n> note a <warning>\n\nModified\n\n> ~~~\n>\n> 18. doc/src/sgml/logical-replication.sgml - Generic steps to add a new\n> node to the existing set of nodes\n>\n> + <title>Generic steps to add a new node to the existing set of nodes</title>\n>\n> SUGGESTION\n> Generic steps for adding a new node to an existing set of nodes\n\nModified\n\n> ~~~\n>\n> 19. doc/src/sgml/logical-replication.sgml - Generic steps to add a new\n> node to the existing set of nodes\n>\n> + <para>\n> + 1. Create the required publication on the new node.\n> + </para>\n> + <para>\n> + 2. Lock the required tables of the new node in <literal>EXCLUSIVE</literal>\n> + mode until the setup is complete. This is required to prevent any\n> + modifications from happening in the new node. If data modifications occur\n> + after step-3, there is a chance that the modifications will be published to\n> + the first node and then synchronized back to the new node while creating\n> + subscription in step-5 resulting in inconsistent data.\n> + </para>\n> + <para>\n> + 3. Create subscriptions on existing nodes pointing to publication on\n> + the new node with <literal>origin</literal> parameter specified as\n> + <literal>local</literal> and <literal>copy_data</literal> specified as\n> + <literal>off</literal>.\n> + </para>\n> + <para>\n> + 4. Lock the required tables of the existing nodes except the first node in\n> + <literal>EXCLUSIVE</literal> mode until the setup is complete. This is\n> + required to prevent any modifications from happening. If data modifications\n> + occur, there is a chance that the modifications done in between step-5 and\n> + step-6 will not be synchronized to the new node resulting in inconsistent\n> + data. No need to lock the tables in the first node as any data changes\n> + made will be synchronized while creating the subscription with\n> + <literal>copy_data</literal> specified as <literal>force</literal>.\n> + </para>\n> + <para>\n> + 5. Create a subscription on the new node pointing to publication on the\n> + first node with <literal>origin</literal> parameter specified as\n> + <literal>local</literal> and <literal>copy_data</literal> parameter\n> + specified as <literal>force</literal>.\n> + </para>\n> + <para>\n> + 6. Create subscriptions on the new node pointing to publications on the\n> + remaining nodes with <literal>origin</literal> parameter specified as\n> + <literal>local</literal> and <literal>copy_data</literal> parameter\n> + specifiedas <literal>off</literal>.\n> + </para>\n>\n> Following suggestions make the following changes:\n> - Some minor rewording\n> - Change the step names\n> - Use \"=\" instead of \"specified as\" for the parameters\n> - I felt it is more readable if the explanatory notes are separate\n> paragraphs from the steps. Maybe if you could indent them or something\n> it would be even better.\n> - Added a couple more explanatory notes\n>\n> SUGGESTION (using those above changes; please keep your formatting)\n>\n> Step-1: Create a publication on the new node.\n>\n> Step-2: Lock the required tables of the new node in EXCLUSIVE mode\n> until the setup is complete.\n>\n> (This lock is necessary to prevent any modifications from happening in\n> the new node because if data modifications occurred after Step-3,\n> there is a chance that the modifications will be published to the\n> first node and then synchronized back to the new node while creating\n> the subscription in Step-5. This would result in inconsistent data).\n>\n> Step-3. Create subscriptions on existing nodes to the publication on\n> the new node with origin = local and copy_data = off.\n>\n> (The copy_data = off is OK here because it is asserted that the\n> published tables of the new node will have no pre-existing data).\n>\n> Step-4. Lock the required tables of the existing nodes except the\n> first node in EXCLUSIVE mode until the setup is complete.\n>\n> (This lock is necessary to prevent any modifications from happening.\n> If data modifications occur, there is a chance that modifications done\n> between Step-5 and Step-6 will not be synchronized to the new node.\n> This would result in inconsistent data. There is no need to lock table\n> t1 in node1 because any data changes made will be synchronized while\n> creating the subscription with copy_data = force).\n>\n> Step-5. Create a subscription on the new node to the publication on\n> the first node with origin = local and copy_data = force.\n>\n> (This will copy the same table data from the existing nodes to the new node)\n>\n> Step-6. Create subscriptions on the new node to publications on the\n> remaining nodes with origin = local and copy_data = off.\n>\n> (The copy_data = off is OK here because the existing node data was\n> already copied to the new node in Step-5)\n\nModified\n\n> ======\n>\n> 20. doc/src/sgml/ref/create_subscription.sgml\n>\n> - <literal>copy_data = force</literal>.\n> + <literal>copy_data = force</literal>. Refer to the\n> + <xref linkend=\"logical-replication-bidirectional\"/> on how\n> + <literal>copy_data</literal> and <literal>origin</literal> can be used\n> + in bidirectional replication.\n> </para>\n>\n> SUGGESTION (keep your formatting)\n> Refer to <xref linkend=\"logical-replication-bidirectional\"/> for how\n> <literal>copy_data</literal> and <literal>origin</literal> can be used\n> in bidirectional replication.\n\nModified\n\nThanks for the comments. The comments are handled as part of the v20\npatch attached at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm0XtQVX3taeLKWE-gPQyppqs34ipXawAPOyO%3Dhe37MQSg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 14 Jun 2022 17:38:50 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are some review comments for patch v20-0002\n\n======\n\n1. General comment\n\nSomething subtle but significant changed since I last reviewed v18*.\nNow the describe.c is changed so that the catalog will never display a\nNULL origin column; it would always be \"any\". But now I am not sure if\nit is a good idea to still allow the NULL in this catalog column while\nat the same time you are pretending it is not there. I felt it might\nbe less confusing, and would simplify the code (e.g. remove quite a\nfew null checks) to have just used a single concept of the default -\ne.g. Just assign the default as \"any\" everywhere. The column would be\ndefined as NOT NULL. Most of the following review comments are related\nto this point.\n\n======\n\n2. doc/src/sgml/catalogs.sgml\n\n+ <para>\n+ Possible origin values are <literal>local</literal>,\n+ <literal>any</literal>, or <literal>NULL</literal> if none is specified.\n+ If <literal>local</literal>, the subscription will request the\n+ publisher to only send changes that originated locally. If\n+ <literal>any</literal> (or <literal>NULL</literal>), the publisher sends\n+ any changes regardless of their origin.\n+ </para></entry>\n\nIs NULL still possible? Perhaps it would be better if it was not and\nthe default \"any\" was always written instead.\n\n======\n\n3. src/backend/catalog/pg_subscription.c\n\n+ if (!isnull)\n+ sub->origin = TextDatumGetCString(datum);\n+ else\n+ sub->origin = NULL;\n+\n\nMaybe better to either disallow NULL in the first place or assign the\n\"any\" here instead of NULL.\n\n======\n\n4. src/backend/commands/subscriptioncmds.c - parse_subscription_options\n\n@@ -137,6 +139,8 @@ parse_subscription_options(ParseState *pstate,\nList *stmt_options,\n opts->twophase = false;\n if (IsSet(supported_opts, SUBOPT_DISABLE_ON_ERR))\n opts->disableonerr = false;\n+ if (IsSet(supported_opts, SUBOPT_ORIGIN))\n+ opts->origin = NULL;\n\nIf opt->origin was assigned to \"any\" then other code would be simplified.\n\n~~~\n\n5. src/backend/commands/subscriptioncmds.c - CreateSubscription\n\n@@ -607,6 +626,11 @@ CreateSubscription(ParseState *pstate,\nCreateSubscriptionStmt *stmt,\n LOGICALREP_TWOPHASE_STATE_PENDING :\n LOGICALREP_TWOPHASE_STATE_DISABLED);\n values[Anum_pg_subscription_subdisableonerr - 1] =\nBoolGetDatum(opts.disableonerr);\n+ if (opts.origin)\n+ values[Anum_pg_subscription_suborigin - 1] =\n+ CStringGetTextDatum(opts.origin);\n+ else\n+ values[Anum_pg_subscription_suborigin - 1] =\nCStringGetTextDatum(LOGICALREP_ORIGIN_ANY);\n\nIf NULL was not possible then this would just be one line:\nvalues[Anum_pg_subscription_suborigin - 1] = CStringGetTextDatum(opts.origin);\n\n======\n\n6. src/backend/replication/logical/worker.c\n\n@@ -276,6 +276,10 @@ static TransactionId stream_xid = InvalidTransactionId;\n static XLogRecPtr skip_xact_finish_lsn = InvalidXLogRecPtr;\n #define is_skipping_changes()\n(unlikely(!XLogRecPtrIsInvalid(skip_xact_finish_lsn)))\n\n+/* Macro for comparing string fields that might be NULL */\n+#define equalstr(a, b) \\\n+ (((a) != NULL && (b) != NULL) ? (strcmp((a), (b)) == 0) : (a) == (b))\n+\n\nIf the NULL was not allowed in the first place then I think this macro\nwould just become redundant.\n\n7. src/backend/replication/logical/worker.c - ApplyWorkerMain\n\n@@ -3741,6 +3746,11 @@ ApplyWorkerMain(Datum main_arg)\n options.proto.logical.streaming = MySubscription->stream;\n options.proto.logical.twophase = false;\n\n+ if (MySubscription->origin)\n+ options.proto.logical.origin = pstrdup(MySubscription->origin);\n+ else\n+ options.proto.logical.origin = NULL;\n+\n\nCan't the if/else be avoided if you always assigned the \"any\" default\nin the first place?\n\n======\n\n8. src/backend/replication/pgoutput/pgoutput.c - parse_output_parameters\n\n@@ -287,11 +289,13 @@ parse_output_parameters(List *options, PGOutputData *data)\n bool messages_option_given = false;\n bool streaming_given = false;\n bool two_phase_option_given = false;\n+ bool origin_option_given = false;\n\n data->binary = false;\n data->streaming = false;\n data->messages = false;\n data->two_phase = false;\n+ data->origin = NULL;\n\nConsider assigning default \"any\" here instead of NULL.\n\n======\n\n9. src/bin/pg_dump/pg_dump.c - getSubscriptions\n\n+ /* FIXME: 150000 should be changed to 160000 later for PG16. */\n+ if (fout->remoteVersion >= 150000)\n+ appendPQExpBufferStr(query, \" s.suborigin\\n\");\n+ else\n+ appendPQExpBufferStr(query, \" NULL AS suborigin\\n\");\n+\n\nMaybe say: 'any' AS suborigin?\n\n~~~\n\n10. src/bin/pg_dump/pg_dump.c - getSubscriptions\n\n@@ -4517,6 +4525,11 @@ getSubscriptions(Archive *fout)\n subinfo[i].subdisableonerr =\n pg_strdup(PQgetvalue(res, i, i_subdisableonerr));\n\n+ if (PQgetisnull(res, i, i_suborigin))\n+ subinfo[i].suborigin = NULL;\n+ else\n+ subinfo[i].suborigin = pg_strdup(PQgetvalue(res, i, i_suborigin));\n+\n\nIf you disallow the NULL in the first place this condition maybe is no\nlonger needed.\n\n~~~\n\n11. src/bin/pg_dump/pg_dump.c - dumpSubscription\n\n@@ -4589,6 +4602,9 @@ dumpSubscription(Archive *fout, const\nSubscriptionInfo *subinfo)\n if (strcmp(subinfo->subdisableonerr, \"t\") == 0)\n appendPQExpBufferStr(query, \", disable_on_error = true\");\n\n+ if (subinfo->suborigin)\n+ appendPQExpBuffer(query, \", origin = %s\", subinfo->suborigin);\n+\n\nIf NULL cannot happen then maybe this test is also redundant.\n\n======\n\n12. src/bin/pg_dump/t/002_pg_dump.pl\n\nAFAICT there is a test for no origin (default), and a test for\nexplicit origin = local, but there is no test case for the explicit\norigin = any.\n\n======\n\n13. src/include/catalog/pg_subscription.h\n\n@@ -31,6 +31,9 @@\n #define LOGICALREP_TWOPHASE_STATE_PENDING 'p'\n #define LOGICALREP_TWOPHASE_STATE_ENABLED 'e'\n\n+#define LOGICALREP_ORIGIN_LOCAL \"local\"\n+#define LOGICALREP_ORIGIN_ANY \"any\"\n+\n\nI thought there should be a comment above these new constants.\n\n~~~\n\n14. src/include/catalog/pg_subscription.h\n\n@@ -87,6 +90,9 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\nBKI_SHARED_RELATION BKI_ROW\n\n /* List of publications subscribed to */\n text subpublications[1] BKI_FORCE_NOT_NULL;\n+\n+ /* Only publish data originating from the specified origin */\n+ text suborigin;\n #endif\n } FormData_pg_subscription;\n\nPerhaps it would be better if this new column was also forced to be NOT NULL.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 15 Jun 2022 16:39:29 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are some review comments for patch v20-0003.\n\n======\n\n1. Commit message\n\nIn case, table t1 has a unique key, it will lead to a unique key\nviolation and replication won't proceed.\n\nSUGGESTION\nIf table t1 has a unique key, this will lead to a unique key\nviolation, and replication won't proceed.\n\n~~~\n\n2. Commit message\n\nThis problem can be solved by using...\n\nSUGGESTION\nThis problem can be avoided by using...\n\n~~~\n\n3. Commit message\n\nstep 3: Create a subscription in node1 to subscribe to node2. Use\n'copy_data = on' so that the existing table data is copied during\ninitial sync:\nnode1=# CREATE SUBSCRIPTION sub_node1_node2 CONNECTION '<node2 details>'\nnode1-# PUBLICATION pub_node2 WITH (copy_data = off, origin = local);\nCREATE SUBSCRIPTION\n\nThis is wrong information. The table on node2 has no data, so talking\nabout 'copy_data = on' is inappropriate here.\n\n======\n\n4. Commit message\n\nIMO it might be better to refer to subscription/publication/table \"on\"\nnodeXXX, instead of saying \"in\" nodeXXX.\n\n4a.\n\"the publication tables were also subscribing data in the publisher\nfrom other publishers.\" -> \"the publication tables were also\nsubscribing from other publishers.\n\n4b.\n\"After the subscription is created in node2\" -> \"After the\nsubscription is created on node2\"\n\n4c.\n\"step 3: Create a subscription in node1 to subscribe to node2.\" ->\n\"step 3: Create a subscription on node1 to subscribe to node2.\"\n\n4d.\n\"step 4: Create a subscription in node2 to subscribe to node1.\" ->\n\"step 4: Create a subscription on node2 to subscribe to node1.\"\n\n======\n\n5. doc/src/sgml/ref/create_subscription.sgml\n\n@@ -383,6 +397,15 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n can have non-existent publications.\n </para>\n\n+ <para>\n+ If the subscription is created with <literal>origin = local</literal> and\n+ <literal>copy_data = on</literal>, it will check if the publisher tables are\n+ being subscribed to any other publisher and, if so, then throw an error to\n+ prevent possible non-local data from being copied. The user can override\n+ this check and continue with the copy operation by specifying\n+ <literal>copy_data = force</literal>.\n+ </para>\n\nPerhaps it is better here to say 'copy_data = true' instead of\n'copy_data = on', simply because the value 'true' was mentioned\nearlier on this page (but this would be the first mention of 'on').\n\n======\n\n6. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n\n+ errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force.\"));\n\nSaying \"off or force\" is not consistent with the other message wording\nin this patch, which used \"/\" for multiple enums.\n(e.g. \"connect = false\", \"copy_data = true/force\").\n\nSo perhaps this errhint should be worded similarly:\n\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\"\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 15 Jun 2022 16:41:21 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are some review comments for patch v20-0004.\n\n======\n\n1. General\n\nI thought that it is better to refer to the\nsubscription/publications/table \"on\" the node, rather than \"in\" the\nnode. Most of the review comments below are related to this point.\n\n======\n\n2. Commit message\n\na) Creating a two-node bidirectional replication when there is no data\nin both nodes.\nb) Adding a new node when there is no data in any of the nodes.\nc) Adding a new node when data is present in the existing nodes.\n\n\"in both nodes\" -> \"on both nodes\"\n\"in any of the nodes\" -> \"on any of the nodes\"\n\"in the existing nodes\" -> \"on the existing nodes\"\n\n======\n\n3. doc/src/sgml/logical-replication.sgml - Setting bidirectional\nreplication between two nodes\n\n3a.\n+ <para>\n+ The following steps demonstrate how to create a two-node bidirectional\n+ replication when there is no table data present in both nodes\n+ <literal>node1</literal> and <literal>node2</literal>:\n+ </para>\n-> \"on both nodes\"\n\n3b.\n+ Create a publication in <literal>node1</literal>:\n-> \"on\"\n\n3c.\n+ Create a publication in <literal>node2</literal>:\n-> \"on\"\n\n3d.\n+ <para>\n+ Lock the table <literal>t1</literal> in <literal>node1</literal> and\n+ <literal>node2</literal> in <literal>EXCLUSIVE</literal> mode until the\n+ setup is completed.\n+ </para>\n-> \"on <literal>node1</literal>\"\n\n3e.\n+ Create a subscription in <literal>node2</literal> to subscribe to\n-> \"on\"\n\n3f.\n+ Create a subscription in <literal>node1</literal> to subscribe to\n+ <literal>node2</literal>:\n-> \"on\"\n\n~~~\n\n4. doc/src/sgml/logical-replication.sgml - Adding a new node when\nthere is no data in any of the nodes\n\n4a.\n+ <title>Adding a new node when there is no data in any of the nodes</title>\nSUGGESTION\nAdding a new node when there is no table data on any of the nodes\n\n4b.\n+ <para>\n+ The following steps demonstrate adding a new node <literal>node3</literal>\n+ to the existing <literal>node1</literal> and <literal>node2</literal> when\n+ there is no <literal>t1</literal> data in any of the nodes. This requires\n+ creating subscriptions in <literal>node1</literal> and\n+ <literal>node2</literal> to replicate the data from\n+ <literal>node3</literal> and creating subscriptions in\n+ <literal>node3</literal> to replicate data from <literal>node1</literal>\n+ and <literal>node2</literal>. Note: These steps assume that the\n+ bidirectional logical replication between <literal>node1</literal> and\n+ <literal>node2</literal> is already completed.\n+ </para>\n\n\"data in any of the nodes\" -> \"data on any of the nodes\"\n\"creating subscriptions in <literal>node1</literal>\" -> \"creating\nsubscriptions on <literal>node1</literal>\"\n\"creating subscriptions in <literal>node3</literal>\" -> \"creating\nsubscriptions on <literal>node3</literal>\"\n\n4c.\n+ Create a publication in <literal>node3</literal>:\n-> \"on\"\n\n4d.\n+ Lock table <literal>t1</literal> in all the nodes\n-> \"on\"\n\n4e.\n+ Create a subscription in <literal>node1</literal> to subscribe to\n+ <literal>node3</literal>:\n-> \"on\"\n\n4f.\n+ Create a subscription in <literal>node2</literal> to subscribe to\n+ <literal>node3</literal>:\n-> \"on\"\n\n4g.\n+ Create a subscription in <literal>node3</literal> to subscribe to\n+ <literal>node1</literal>:\n-> \"on\"\n\n4h.\n+ Create a subscription in <literal>node3</literal> to subscribe to\n+ <literal>node2</literal>:\n\n4i.\n+ <literal>node3</literal>. Incremental changes made in any node will be\n+ replicated to the other two nodes.\n\"in any node\" -> \"on any node\"\n\n~~~\n\n5. doc/src/sgml/logical-replication.sgml - Adding a new node when data\nis present in the existing nodes\n\n5a.\n+ <title>Adding a new node when data is present in the existing nodes</title>\nSUGGESTION\nAdding a new node when table data is present on the existing nodes\n\n5b.\n+ during initial data synchronization. Note: These steps assume that the\n+ bidirectional logical replication between <literal>node1</literal> and\n+ <literal>node2</literal> is already completed, and the pre-existing data\n+ in table <literal>t1</literal> is already synchronized in both those\n+ nodes.\n+ </para>\n\"in both those nodes\" -> \"on both those nodes\"\n\n5c.\n+ Create a publication in <literal>node3</literal>\n-> \"on\"\n\n5d.\n+ Lock table <literal>t1</literal> in <literal>node2</literal> and\n-> \"on\"\n\n5e.\n+ Create a subscription in <literal>node1</literal> to subscribe to\n+ <literal>node3</literal>:\n-> \"on\"\n\n5f.\n+ Create a subscription in <literal>node2</literal> to subscribe to\n+ <literal>node3</literal>:\n-> \"on\"\n\n5g.\n+ Create a subscription in <literal>node3</literal> to subscribe to\n+ <literal>node1</literal>. Use <literal>copy_data = force </literal> so that\n+ the existing table data is copied during initial sync:\n-> \"on\"\n\n\n5h.\n+ Create a subscription in <literal>node3</literal> to subscribe to\n+ <literal>node2</literal>. Use <literal>copy_data = off</literal>\n-> \"on\"\n\n5i.\n+ <literal>node3</literal>. Incremental changes made in any node will be\n+ replicated to the other two nodes.\n\"in any node\" -> \"on any node\"\n\n~~~\n\n6. doc/src/sgml/logical-replication.sgml - Adding a new node when data\nis present in the new node\n\n+ <title>Adding a new node when data is present in the new node</title>\nSUGGESTION\nAdding a new node when table data is present on the new node\n\n~~~\n\n7. doc/src/sgml/logical-replication.sgml - Generic steps for adding a\nnew node to an existing set of nodes\n\n7a.\n+ <para>\n+ Step-2: Lock the required tables of the new node in EXCLUSIVE mode until\n+ the setup is complete. (This lock is necessary to prevent any modifications\n+ from happening in the new node because if data modifications occurred after\n+ Step-3, there is a chance that the modifications will be published to the\n+ first node and then synchronized back to the new node while creating the\n+ subscription in Step-5. This would result in inconsistent data).\n+ </para>\n\"happening in the new node\" -> \"happening on the new node\"\n\n7b.\n+ not be synchronized to the new node. This would result in inconsistent\n+ data. There is no need to lock the required tables in\n+ <literal>node1</literal> because any data changes made will be synchronized\n+ while creating the subscription with <literal>copy_data = force</literal>).\n+ </para>\n\"no need to lock the required tables in\" -> \"no need to lock the\nrequired tables on\"\n\n======\n\n8. doc/src/sgml/ref/create_subscription.sgml\n\n@@ -403,7 +403,10 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n being subscribed to any other publisher and, if so, then throw an error to\n prevent possible non-local data from being copied. The user can override\n this check and continue with the copy operation by specifying\n- <literal>copy_data = force</literal>.\n+ <literal>copy_data = force</literal>. Refer to\n+ <xref linkend=\"logical-replication-bidirectional\"/> for how\n+ <literal>copy_data</literal> and <literal>origin</literal> can be used\n+ in bidirectional replication.\n </para>\n\n\"can be used in bidirectional replication\" -> \"can be used to set up\nbidirectional replication\"\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 15 Jun 2022 16:43:35 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"PSA a test script that demonstrates all the documented steps for\nsetting up n-way bidirectional replication. These steps are the same\nas those documented [1] on the new page \"Bidirectional logical\nreplication\".\n\nThis script works using the current latest v20* patch set. Each of the\nsections of 31.11.1 - 31.11.5 (see below) can be run independently\n(just edit the script and at the bottom uncomment the part you want to\ntest):\n31.11.1. Setting bidirectional replication between two nodes\n31.11.2. Adding a new node when there is no data in any of the nodes\n31.11.3. Adding a new node when data is present in the existing nodes\n31.11.4. Adding a new node when data is present in the new node\n31.11.5. Generic steps for adding a new node to an existing set of nodes\n\n~~\n\nSome sample output is also attached.\n\n------\n[1] https://www.postgresql.org/message-id/attachment/134464/v20-0004-Document-bidirectional-logical-replication-steps.patch\n\nKind Regards,\nPeter Smith.\nFujitsu Australia","msg_date":"Wed, 15 Jun 2022 18:56:24 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some review comments for patch v20-0002\n>\n> ======\n>\n> 1. General comment\n>\n> Something subtle but significant changed since I last reviewed v18*.\n> Now the describe.c is changed so that the catalog will never display a\n> NULL origin column; it would always be \"any\". But now I am not sure if\n> it is a good idea to still allow the NULL in this catalog column while\n> at the same time you are pretending it is not there. I felt it might\n> be less confusing, and would simplify the code (e.g. remove quite a\n> few null checks) to have just used a single concept of the default -\n> e.g. Just assign the default as \"any\" everywhere. The column would be\n> defined as NOT NULL. Most of the following review comments are related\n> to this point.\n\nOk, I was initially feeling having NULL value will help in the upgrade\ncases where we upgrade from a lower version which does not have origin\nto current. But setting the default value handles the upgrade scenario\ntoo.\n\n> ======\n>\n> 2. doc/src/sgml/catalogs.sgml\n>\n> + <para>\n> + Possible origin values are <literal>local</literal>,\n> + <literal>any</literal>, or <literal>NULL</literal> if none is specified.\n> + If <literal>local</literal>, the subscription will request the\n> + publisher to only send changes that originated locally. If\n> + <literal>any</literal> (or <literal>NULL</literal>), the publisher sends\n> + any changes regardless of their origin.\n> + </para></entry>\n>\n> Is NULL still possible? Perhaps it would be better if it was not and\n> the default \"any\" was always written instead.\n\nModified\n\n> ======\n>\n> 3. src/backend/catalog/pg_subscription.c\n>\n> + if (!isnull)\n> + sub->origin = TextDatumGetCString(datum);\n> + else\n> + sub->origin = NULL;\n> +\n>\n> Maybe better to either disallow NULL in the first place or assign the\n> \"any\" here instead of NULL.\n\nModified\n\n> ======\n>\n> 4. src/backend/commands/subscriptioncmds.c - parse_subscription_options\n>\n> @@ -137,6 +139,8 @@ parse_subscription_options(ParseState *pstate,\n> List *stmt_options,\n> opts->twophase = false;\n> if (IsSet(supported_opts, SUBOPT_DISABLE_ON_ERR))\n> opts->disableonerr = false;\n> + if (IsSet(supported_opts, SUBOPT_ORIGIN))\n> + opts->origin = NULL;\n>\n> If opt->origin was assigned to \"any\" then other code would be simplified.\n\nModified\n\n> ~~~\n>\n> 5. src/backend/commands/subscriptioncmds.c - CreateSubscription\n>\n> @@ -607,6 +626,11 @@ CreateSubscription(ParseState *pstate,\n> CreateSubscriptionStmt *stmt,\n> LOGICALREP_TWOPHASE_STATE_PENDING :\n> LOGICALREP_TWOPHASE_STATE_DISABLED);\n> values[Anum_pg_subscription_subdisableonerr - 1] =\n> BoolGetDatum(opts.disableonerr);\n> + if (opts.origin)\n> + values[Anum_pg_subscription_suborigin - 1] =\n> + CStringGetTextDatum(opts.origin);\n> + else\n> + values[Anum_pg_subscription_suborigin - 1] =\n> CStringGetTextDatum(LOGICALREP_ORIGIN_ANY);\n>\n> If NULL was not possible then this would just be one line:\n> values[Anum_pg_subscription_suborigin - 1] = CStringGetTextDatum(opts.origin);\n\nModified\n\n> ======\n>\n> 6. src/backend/replication/logical/worker.c\n>\n> @@ -276,6 +276,10 @@ static TransactionId stream_xid = InvalidTransactionId;\n> static XLogRecPtr skip_xact_finish_lsn = InvalidXLogRecPtr;\n> #define is_skipping_changes()\n> (unlikely(!XLogRecPtrIsInvalid(skip_xact_finish_lsn)))\n>\n> +/* Macro for comparing string fields that might be NULL */\n> +#define equalstr(a, b) \\\n> + (((a) != NULL && (b) != NULL) ? (strcmp((a), (b)) == 0) : (a) == (b))\n> +\n>\n> If the NULL was not allowed in the first place then I think this macro\n> would just become redundant.\n\nRemoved it\n\n> 7. src/backend/replication/logical/worker.c - ApplyWorkerMain\n>\n> @@ -3741,6 +3746,11 @@ ApplyWorkerMain(Datum main_arg)\n> options.proto.logical.streaming = MySubscription->stream;\n> options.proto.logical.twophase = false;\n>\n> + if (MySubscription->origin)\n> + options.proto.logical.origin = pstrdup(MySubscription->origin);\n> + else\n> + options.proto.logical.origin = NULL;\n> +\n>\n> Can't the if/else be avoided if you always assigned the \"any\" default\n> in the first place?\n\nModified\n\n> ======\n>\n> 8. src/backend/replication/pgoutput/pgoutput.c - parse_output_parameters\n>\n> @@ -287,11 +289,13 @@ parse_output_parameters(List *options, PGOutputData *data)\n> bool messages_option_given = false;\n> bool streaming_given = false;\n> bool two_phase_option_given = false;\n> + bool origin_option_given = false;\n>\n> data->binary = false;\n> data->streaming = false;\n> data->messages = false;\n> data->two_phase = false;\n> + data->origin = NULL;\n>\n> Consider assigning default \"any\" here instead of NULL.\n\nThis is no more required because of setting default to any and the\nsame value will be passed as an option\n\n> ======\n>\n> 9. src/bin/pg_dump/pg_dump.c - getSubscriptions\n>\n> + /* FIXME: 150000 should be changed to 160000 later for PG16. */\n> + if (fout->remoteVersion >= 150000)\n> + appendPQExpBufferStr(query, \" s.suborigin\\n\");\n> + else\n> + appendPQExpBufferStr(query, \" NULL AS suborigin\\n\");\n> +\n>\n> Maybe say: 'any' AS suborigin?\n\nModified\n\n> ~~~\n>\n> 10. src/bin/pg_dump/pg_dump.c - getSubscriptions\n>\n> @@ -4517,6 +4525,11 @@ getSubscriptions(Archive *fout)\n> subinfo[i].subdisableonerr =\n> pg_strdup(PQgetvalue(res, i, i_subdisableonerr));\n>\n> + if (PQgetisnull(res, i, i_suborigin))\n> + subinfo[i].suborigin = NULL;\n> + else\n> + subinfo[i].suborigin = pg_strdup(PQgetvalue(res, i, i_suborigin));\n> +\n>\n> If you disallow the NULL in the first place this condition maybe is no\n> longer needed.\n\nModified\n\n> ~~~\n>\n> 11. src/bin/pg_dump/pg_dump.c - dumpSubscription\n>\n> @@ -4589,6 +4602,9 @@ dumpSubscription(Archive *fout, const\n> SubscriptionInfo *subinfo)\n> if (strcmp(subinfo->subdisableonerr, \"t\") == 0)\n> appendPQExpBufferStr(query, \", disable_on_error = true\");\n>\n> + if (subinfo->suborigin)\n> + appendPQExpBuffer(query, \", origin = %s\", subinfo->suborigin);\n> +\n>\n> If NULL cannot happen then maybe this test is also redundant.\n\nModified\n\n> ======\n>\n> 12. src/bin/pg_dump/t/002_pg_dump.pl\n>\n> AFAICT there is a test for no origin (default), and a test for\n> explicit origin = local, but there is no test case for the explicit\n> origin = any.\n\nAdded a test for the same\n\n> ======\n>\n> 13. src/include/catalog/pg_subscription.h\n>\n> @@ -31,6 +31,9 @@\n> #define LOGICALREP_TWOPHASE_STATE_PENDING 'p'\n> #define LOGICALREP_TWOPHASE_STATE_ENABLED 'e'\n>\n> +#define LOGICALREP_ORIGIN_LOCAL \"local\"\n> +#define LOGICALREP_ORIGIN_ANY \"any\"\n> +\n>\n> I thought there should be a comment above these new constants.\n\nAdded comments\n\n> ~~~\n>\n> 14. src/include/catalog/pg_subscription.h\n>\n> @@ -87,6 +90,9 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\n> BKI_SHARED_RELATION BKI_ROW\n>\n> /* List of publications subscribed to */\n> text subpublications[1] BKI_FORCE_NOT_NULL;\n> +\n> + /* Only publish data originating from the specified origin */\n> + text suborigin;\n> #endif\n> } FormData_pg_subscription;\n>\n> Perhaps it would be better if this new column was also forced to be NOT NULL.\n\nI have set a default value so need to set NOT NULL.\n\nThanks for the comments, the attached v21 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Thu, 16 Jun 2022 15:47:50 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Jun 15, 2022 at 12:11 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some review comments for patch v20-0003.\n>\n> ======\n>\n> 1. Commit message\n>\n> In case, table t1 has a unique key, it will lead to a unique key\n> violation and replication won't proceed.\n>\n> SUGGESTION\n> If table t1 has a unique key, this will lead to a unique key\n> violation, and replication won't proceed.\n\nModified\n\n> ~~~\n>\n> 2. Commit message\n>\n> This problem can be solved by using...\n>\n> SUGGESTION\n> This problem can be avoided by using...\n\nModified\n\n> ~~~\n>\n> 3. Commit message\n>\n> step 3: Create a subscription in node1 to subscribe to node2. Use\n> 'copy_data = on' so that the existing table data is copied during\n> initial sync:\n> node1=# CREATE SUBSCRIPTION sub_node1_node2 CONNECTION '<node2 details>'\n> node1-# PUBLICATION pub_node2 WITH (copy_data = off, origin = local);\n> CREATE SUBSCRIPTION\n>\n> This is wrong information. The table on node2 has no data, so talking\n> about 'copy_data = on' is inappropriate here.\n\nModified\n\n> ======\n>\n> 4. Commit message\n>\n> IMO it might be better to refer to subscription/publication/table \"on\"\n> nodeXXX, instead of saying \"in\" nodeXXX.\n>\n> 4a.\n> \"the publication tables were also subscribing data in the publisher\n> from other publishers.\" -> \"the publication tables were also\n> subscribing from other publishers.\n\nModified\n\n> 4b.\n> \"After the subscription is created in node2\" -> \"After the\n> subscription is created on node2\"\n\nModified\n\n> 4c.\n> \"step 3: Create a subscription in node1 to subscribe to node2.\" ->\n> \"step 3: Create a subscription on node1 to subscribe to node2.\"\n\nModified\n\n> 4d.\n> \"step 4: Create a subscription in node2 to subscribe to node1.\" ->\n> \"step 4: Create a subscription on node2 to subscribe to node1.\"\n\nModified\n\n> ======\n>\n> 5. doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -383,6 +397,15 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> can have non-existent publications.\n> </para>\n>\n> + <para>\n> + If the subscription is created with <literal>origin = local</literal> and\n> + <literal>copy_data = on</literal>, it will check if the publisher tables are\n> + being subscribed to any other publisher and, if so, then throw an error to\n> + prevent possible non-local data from being copied. The user can override\n> + this check and continue with the copy operation by specifying\n> + <literal>copy_data = force</literal>.\n> + </para>\n>\n> Perhaps it is better here to say 'copy_data = true' instead of\n> 'copy_data = on', simply because the value 'true' was mentioned\n> earlier on this page (but this would be the first mention of 'on').\n\nModified\n\n> ======\n>\n> 6. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> + errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force.\"));\n>\n> Saying \"off or force\" is not consistent with the other message wording\n> in this patch, which used \"/\" for multiple enums.\n> (e.g. \"connect = false\", \"copy_data = true/force\").\n>\n> So perhaps this errhint should be worded similarly:\n> \"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\"\n\nModified\n\nThanks for the comments, the v21 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm3%2B6cey0rcDft1ZUCjSUtLDM0xmU_Q%2BYhcsBrqe1RH8%3Dw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Thu, 16 Jun 2022 15:50:02 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Jun 15, 2022 at 12:13 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some review comments for patch v20-0004.\n>\n> ======\n>\n> 1. General\n>\n> I thought that it is better to refer to the\n> subscription/publications/table \"on\" the node, rather than \"in\" the\n> node. Most of the review comments below are related to this point.\n\nModified\n\n> ======\n>\n> 2. Commit message\n>\n> a) Creating a two-node bidirectional replication when there is no data\n> in both nodes.\n> b) Adding a new node when there is no data in any of the nodes.\n> c) Adding a new node when data is present in the existing nodes.\n>\n> \"in both nodes\" -> \"on both nodes\"\n> \"in any of the nodes\" -> \"on any of the nodes\"\n> \"in the existing nodes\" -> \"on the existing nodes\"\n\nModified\n\n> ======\n>\n> 3. doc/src/sgml/logical-replication.sgml - Setting bidirectional\n> replication between two nodes\n>\n> 3a.\n> + <para>\n> + The following steps demonstrate how to create a two-node bidirectional\n> + replication when there is no table data present in both nodes\n> + <literal>node1</literal> and <literal>node2</literal>:\n> + </para>\n> -> \"on both nodes\"\n\nModified\n\n> 3b.\n> + Create a publication in <literal>node1</literal>:\n> -> \"on\"\n\nModified\n\n> 3c.\n> + Create a publication in <literal>node2</literal>:\n> -> \"on\"\n\nModified\n\n> 3d.\n> + <para>\n> + Lock the table <literal>t1</literal> in <literal>node1</literal> and\n> + <literal>node2</literal> in <literal>EXCLUSIVE</literal> mode until the\n> + setup is completed.\n> + </para>\n> -> \"on <literal>node1</literal>\"\n\nModified\n\n> 3e.\n> + Create a subscription in <literal>node2</literal> to subscribe to\n> -> \"on\"\n\nModified\n\n> 3f.\n> + Create a subscription in <literal>node1</literal> to subscribe to\n> + <literal>node2</literal>:\n> -> \"on\"\n\nModified\n\n> ~~~\n>\n> 4. doc/src/sgml/logical-replication.sgml - Adding a new node when\n> there is no data in any of the nodes\n>\n> 4a.\n> + <title>Adding a new node when there is no data in any of the nodes</title>\n> SUGGESTION\n> Adding a new node when there is no table data on any of the nodes\n\nModified\n\n> 4b.\n> + <para>\n> + The following steps demonstrate adding a new node <literal>node3</literal>\n> + to the existing <literal>node1</literal> and <literal>node2</literal> when\n> + there is no <literal>t1</literal> data in any of the nodes. This requires\n> + creating subscriptions in <literal>node1</literal> and\n> + <literal>node2</literal> to replicate the data from\n> + <literal>node3</literal> and creating subscriptions in\n> + <literal>node3</literal> to replicate data from <literal>node1</literal>\n> + and <literal>node2</literal>. Note: These steps assume that the\n> + bidirectional logical replication between <literal>node1</literal> and\n> + <literal>node2</literal> is already completed.\n> + </para>\n>\n> \"data in any of the nodes\" -> \"data on any of the nodes\"\n> \"creating subscriptions in <literal>node1</literal>\" -> \"creating\n> subscriptions on <literal>node1</literal>\"\n> \"creating subscriptions in <literal>node3</literal>\" -> \"creating\n> subscriptions on <literal>node3</literal>\"\n\nModified\n\n> 4c.\n> + Create a publication in <literal>node3</literal>:\n> -> \"on\"\n\nModified\n\n> 4d.\n> + Lock table <literal>t1</literal> in all the nodes\n> -> \"on\"\n\nModified\n\n> 4e.\n> + Create a subscription in <literal>node1</literal> to subscribe to\n> + <literal>node3</literal>:\n> -> \"on\"\n\nModified\n\n> 4f.\n> + Create a subscription in <literal>node2</literal> to subscribe to\n> + <literal>node3</literal>:\n> -> \"on\"\n\nModified\n\n> 4g.\n> + Create a subscription in <literal>node3</literal> to subscribe to\n> + <literal>node1</literal>:\n> -> \"on\"\n\nModified\n\n> 4h.\n> + Create a subscription in <literal>node3</literal> to subscribe to\n> + <literal>node2</literal>:\n\nModified\n\n> 4i.\n> + <literal>node3</literal>. Incremental changes made in any node will be\n> + replicated to the other two nodes.\n> \"in any node\" -> \"on any node\"\n\nModified\n\n> ~~~\n>\n> 5. doc/src/sgml/logical-replication.sgml - Adding a new node when data\n> is present in the existing nodes\n>\n> 5a.\n> + <title>Adding a new node when data is present in the existing nodes</title>\n> SUGGESTION\n> Adding a new node when table data is present on the existing nodes\n\nModified\n\n> 5b.\n> + during initial data synchronization. Note: These steps assume that the\n> + bidirectional logical replication between <literal>node1</literal> and\n> + <literal>node2</literal> is already completed, and the pre-existing data\n> + in table <literal>t1</literal> is already synchronized in both those\n> + nodes.\n> + </para>\n> \"in both those nodes\" -> \"on both those nodes\"\n\nModified\n\n> 5c.\n> + Create a publication in <literal>node3</literal>\n> -> \"on\"\n\nModified\n\n> 5d.\n> + Lock table <literal>t1</literal> in <literal>node2</literal> and\n> -> \"on\"\n\nModified\n\n> 5e.\n> + Create a subscription in <literal>node1</literal> to subscribe to\n> + <literal>node3</literal>:\n> -> \"on\"\n\nModified\n\n> 5f.\n> + Create a subscription in <literal>node2</literal> to subscribe to\n> + <literal>node3</literal>:\n> -> \"on\"\n\nModified\n\n> 5g.\n> + Create a subscription in <literal>node3</literal> to subscribe to\n> + <literal>node1</literal>. Use <literal>copy_data = force </literal> so that\n> + the existing table data is copied during initial sync:\n> -> \"on\"\n\nModified\n\n>\n> 5h.\n> + Create a subscription in <literal>node3</literal> to subscribe to\n> + <literal>node2</literal>. Use <literal>copy_data = off</literal>\n> -> \"on\"\n\nModified\n\n> 5i.\n> + <literal>node3</literal>. Incremental changes made in any node will be\n> + replicated to the other two nodes.\n> \"in any node\" -> \"on any node\"\n\nModified\n\n> ~~~\n>\n> 6. doc/src/sgml/logical-replication.sgml - Adding a new node when data\n> is present in the new node\n>\n> + <title>Adding a new node when data is present in the new node</title>\n> SUGGESTION\n> Adding a new node when table data is present on the new node\n\nModified\n\n> ~~~\n>\n> 7. doc/src/sgml/logical-replication.sgml - Generic steps for adding a\n> new node to an existing set of nodes\n>\n> 7a.\n> + <para>\n> + Step-2: Lock the required tables of the new node in EXCLUSIVE mode until\n> + the setup is complete. (This lock is necessary to prevent any modifications\n> + from happening in the new node because if data modifications occurred after\n> + Step-3, there is a chance that the modifications will be published to the\n> + first node and then synchronized back to the new node while creating the\n> + subscription in Step-5. This would result in inconsistent data).\n> + </para>\n> \"happening in the new node\" -> \"happening on the new node\"\n\nModified\n\n> 7b.\n> + not be synchronized to the new node. This would result in inconsistent\n> + data. There is no need to lock the required tables in\n> + <literal>node1</literal> because any data changes made will be synchronized\n> + while creating the subscription with <literal>copy_data = force</literal>).\n> + </para>\n> \"no need to lock the required tables in\" -> \"no need to lock the\n> required tables on\"\n\nModified\n\n> ======\n>\n> 8. doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -403,7 +403,10 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> being subscribed to any other publisher and, if so, then throw an error to\n> prevent possible non-local data from being copied. The user can override\n> this check and continue with the copy operation by specifying\n> - <literal>copy_data = force</literal>.\n> + <literal>copy_data = force</literal>. Refer to\n> + <xref linkend=\"logical-replication-bidirectional\"/> for how\n> + <literal>copy_data</literal> and <literal>origin</literal> can be used\n> + in bidirectional replication.\n> </para>\n>\n> \"can be used in bidirectional replication\" -> \"can be used to set up\n> bidirectional replication\"\n\nModified\n\nThanks for the comments, the v21 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm3%2B6cey0rcDft1ZUCjSUtLDM0xmU_Q%2BYhcsBrqe1RH8%3Dw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Thu, 16 Jun 2022 15:53:18 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jun 16, 2022 6:18 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> Thanks for the comments, the attached v21 patch has the changes for the\r\n> same.\r\n> \r\n\r\nThanks for updating the patch. Here are some comments.\r\n\r\n0002 patch\r\n==============\r\n1.\r\n+ publisher to only send changes that originated locally. Setting\r\n+ <literal>origin</literal> to <literal>any</literal> means that that\r\n+ the publisher sends any changes regardless of their origin. The\r\n+ default is <literal>any</literal>.\r\n\r\nIt seems there's a redundant \"that\" at the end of second line.\r\n\r\n2.\r\n+ <row>\r\n+ <entry role=\"catalog_table_entry\"><para role=\"column_definition\">\r\n+ <structfield>suborigin</structfield> <type>text</type>\r\n+ </para>\r\n+ <para>\r\n+ Possible origin values are <literal>local</literal> or\r\n+ <literal>any</literal>. The default is <literal>any</literal>.\r\n+ If <literal>local</literal>, the subscription will request the publisher\r\n+ to only send changes that originated locally. If <literal>any</literal>\r\n+ the publisher sends any changes regardless of their origin.\r\n+ </para></entry>\r\n+ </row>.\r\n\r\nA comma can be added after \"If any\".\r\n\r\n3.\r\n@@ -4589,6 +4598,8 @@ dumpSubscription(Archive *fout, const SubscriptionInfo *subinfo)\r\n \tif (strcmp(subinfo->subdisableonerr, \"t\") == 0)\r\n \t\tappendPQExpBufferStr(query, \", disable_on_error = true\");\r\n \r\n+\tappendPQExpBuffer(query, \", origin = %s\", subinfo->suborigin);\r\n+\r\n \tif (strcmp(subinfo->subsynccommit, \"off\") != 0)\r\n \t\tappendPQExpBuffer(query, \", synchronous_commit = %s\", fmtId(subinfo->subsynccommit));\r\n\r\nDo we need to append anything if it's the default value (\"any\")? I saw that some\r\nother parameters, they will be appended only if they are not the default value.\r\n\r\n0003 patch\r\n==============\r\n1. \r\nin create_subscription.sgml:\r\n (You cannot combine setting <literal>connect</literal>\r\n to <literal>false</literal> with\r\n setting <literal>create_slot</literal>, <literal>enabled</literal>,\r\n or <literal>copy_data</literal> to <literal>true</literal>.)\r\n\r\nIn this description about \"connect\" parameter in CREATE SUBSCIPTION document,\r\nmaybe it would be better to change \"copy_data to true\" to \"copy_data to\r\ntrue/force\".\r\n\r\n2.\r\n+\tappendStringInfoString(&cmd,\r\n+\t\t\t\t\t\t \"SELECT DISTINCT N.nspname AS schemaname,\\n\"\r\n+\t\t\t\t\t\t \"\t\t\t\tC.relname AS tablename,\\n\"\r\n+\t\t\t\t\t\t \"\t\t\t\tPS.srrelid as replicated\\n\"\r\n+\t\t\t\t\t\t \"FROM pg_publication P,\\n\"\r\n+\t\t\t\t\t\t \"\t LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\r\n+\t\t\t\t\t\t \"\t LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\r\n+\t\t\t\t\t\t \"\t pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\r\n+\t\t\t\t\t\t \"WHERE C.oid = GPT.relid AND P.pubname IN (\");\r\n\r\n\"PS.srrelid as replicated\" can be modified to \"PS.srrelid AS replicated\".\r\n\r\nBesides, I think we can filter out the tables which are not subscribing data in\r\nthis SQL statement, then later processing can be simplified.\r\n\r\nSomething like:\r\nSELECT DISTINCT N.nspname AS schemaname,\r\n C.relname AS tablename\r\nFROM pg_publication P,\r\n LATERAL pg_get_publication_tables(P.pubname) GPT\r\n LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\r\n pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\r\nWHERE C.oid = GPT.relid AND P.pubname IN ('pa') AND PS.srrelid IS NOT NULL;\r\n\r\n0004 patch\r\n==============\r\n1. Generic steps for adding a new node to an existing set of nodes\r\n\r\n+ Step-2: Lock the required tables of the new node in EXCLUSIVE mode until\r\n+ the setup is complete. (This lock is necessary to prevent any modifications\r\n\r\n+ Step-4. Lock the required tables of the existing nodes except the first node\r\n+ in EXCLUSIVE mode until the setup is complete. (This lock is necessary to\r\n\r\nShould \"in EXCLUSIVE mode\" be modified to \"in <literal>EXCLUSIVE</literal>\r\nmode\"?\r\n\r\n2. Generic steps for adding a new node to an existing set of nodes\r\n\r\n+ data. There is no need to lock the required tables on\r\n+ <literal>node1</literal> because any data changes made will be synchronized\r\n+ while creating the subscription with <literal>copy_data = force</literal>).\r\n\r\nI think it would be better to say \"on the first node\" here, instead of \"node1\",\r\nbecause in this section, node1 is not mentioned before.\r\n\r\nRegards,\r\nShi yu\r\n\r\n\r\n","msg_date":"Mon, 20 Jun 2022 03:52:01 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jun 16, 2022 at 3:48 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n\n> Thanks for the comments, the attached v21 patch has the changes for the same.\n\nI have done some basic review of v21 and I have a few comments,\n\n1.\n+/*\n+ * The subscription will request the publisher to only send changes that\n+ * originated locally.\n+ */\n+#define LOGICALREP_ORIGIN_LOCAL \"local\"\n+\n+/*\n+ * The subscription will request the publisher to send any changes regardless\n+ * of their origin\n+ */\n+#define LOGICALREP_ORIGIN_ANY \"any\"\n\nAre we planning to extend this to more options or are we planning to\nsupport the actual origin name here? If not then why isn't it just\nbool? I think the comments and the patch commit message should\nexplain the details behind it if it has been already discussed and\nconcluded.\n\n2.\n+/*\n+ * Check and throw an error if the publisher has subscribed to the same table\n+ * from some other publisher. This check is required only if copydata is ON and\n+ * the origin is local.\n+ */\n\nI think it should also explain why this combination is not allowed and\nif it is already explained in code\nthen this code can add comments to refer to that part of the code.\n\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 20 Jun 2022 14:37:32 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jun 20, 2022 at 2:37 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>\n> On Thu, Jun 16, 2022 at 3:48 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> > >\n>\n> > Thanks for the comments, the attached v21 patch has the changes for the same.\n>\n> I have done some basic review of v21 and I have a few comments,\n>\n> 1.\n> +/*\n> + * The subscription will request the publisher to only send changes that\n> + * originated locally.\n> + */\n> +#define LOGICALREP_ORIGIN_LOCAL \"local\"\n> +\n> +/*\n> + * The subscription will request the publisher to send any changes regardless\n> + * of their origin\n> + */\n> +#define LOGICALREP_ORIGIN_ANY \"any\"\n>\n> Are we planning to extend this to more options or are we planning to\n> support the actual origin name here? If not then why isn't it just\n> bool? I think the comments and the patch commit message should\n> explain the details behind it if it has been already discussed and\n> concluded.\n\nCurrently we only support local and any. But this was designed to\naccept string instead of boolean type, so that it can be extended\nlater to support filtering of origin names specified by the user in\nthe later versions. The same was also discussed in pg unconference as\nmentioned in [1]. I will add it to the commit message and a comment\nfor the same in the next version.\n\n> 2.\n> +/*\n> + * Check and throw an error if the publisher has subscribed to the same table\n> + * from some other publisher. This check is required only if copydata is ON and\n> + * the origin is local.\n> + */\n>\n> I think it should also explain why this combination is not allowed and\n> if it is already explained in code\n> then this code can add comments to refer to that part of the code.\n\nIn the same function, the reason for this is mentioned detailly just\nabove the place where error is thrown. I think that should be enough.\nHave a look and let me know if that is not sufficient:\n+ /*\n+ * Throw an error if the publisher has subscribed to\nthe same table\n+ * from some other publisher. We cannot differentiate\nbetween the\n+ * local and non-local data that is present in the\nHEAP during the\n+ * initial sync. Identification of local data can be\ndone only from\n+ * the WAL by using the origin id. XXX: For\nsimplicity, we don't check\n+ * whether the table has any data or not. If the table\ndoesn't have\n+ * any data then we don't need to distinguish between local and\n+ * non-local data so we can avoid throwing error in that case.\n+ */\n+ if (!slot_attisnull(slot, 3))\n+ ereport(ERROR,\n+\nerrcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n+ errmsg(\"table:%s.%s might have\nreplicated data in the publisher\",\n+ nspname, relname),\n+ errdetail(\"CREATE/ALTER\nSUBSCRIPTION with origin = local and copy_data = on is not allowed\nwhen the publisher might have replicated data.\"),\n+ errhint(\"Use CREATE/ALTER\nSUBSCRIPTION with copy_data = off/force.\"));\n\n\n[1] - https://wiki.postgresql.org/wiki/PgCon_2022_Developer_Unconference#Logical_Replication_Origin_Filtering_and_Consistency\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 20 Jun 2022 15:16:43 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jun 20, 2022 at 3:16 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Jun 20, 2022 at 2:37 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >\n> > On Thu, Jun 16, 2022 at 3:48 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> > > >\n> >\n> > > Thanks for the comments, the attached v21 patch has the changes for the same.\n> >\n> > I have done some basic review of v21 and I have a few comments,\n> >\n> > 1.\n> > +/*\n> > + * The subscription will request the publisher to only send changes that\n> > + * originated locally.\n> > + */\n> > +#define LOGICALREP_ORIGIN_LOCAL \"local\"\n> > +\n> > +/*\n> > + * The subscription will request the publisher to send any changes regardless\n> > + * of their origin\n> > + */\n> > +#define LOGICALREP_ORIGIN_ANY \"any\"\n> >\n> > Are we planning to extend this to more options or are we planning to\n> > support the actual origin name here? If not then why isn't it just\n> > bool? I think the comments and the patch commit message should\n> > explain the details behind it if it has been already discussed and\n> > concluded.\n>\n> Currently we only support local and any. But this was designed to\n> accept string instead of boolean type, so that it can be extended\n> later to support filtering of origin names specified by the user in\n> the later versions. The same was also discussed in pg unconference as\n> mentioned in [1]. I will add it to the commit message and a comment\n> for the same in the next version.\n>\n> > 2.\n> > +/*\n> > + * Check and throw an error if the publisher has subscribed to the same table\n> > + * from some other publisher. This check is required only if copydata is ON and\n> > + * the origin is local.\n> > + */\n> >\n> > I think it should also explain why this combination is not allowed and\n> > if it is already explained in code\n> > then this code can add comments to refer to that part of the code.\n>\n> In the same function, the reason for this is mentioned detailly just\n> above the place where error is thrown. I think that should be enough.\n> Have a look and let me know if that is not sufficient:\n\nI think that should be sufficient, thanks.\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 20 Jun 2022 16:02:13 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jun 20, 2022 at 9:22 AM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Thu, Jun 16, 2022 6:18 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached v21 patch has the changes for the\n> > same.\n> >\n>\n> Thanks for updating the patch. Here are some comments.\n>\n> 0002 patch\n> ==============\n> 1.\n> + publisher to only send changes that originated locally. Setting\n> + <literal>origin</literal> to <literal>any</literal> means that that\n> + the publisher sends any changes regardless of their origin. The\n> + default is <literal>any</literal>.\n>\n> It seems there's a redundant \"that\" at the end of second line.\n\nModified\n\n> 2.\n> + <row>\n> + <entry role=\"catalog_table_entry\"><para role=\"column_definition\">\n> + <structfield>suborigin</structfield> <type>text</type>\n> + </para>\n> + <para>\n> + Possible origin values are <literal>local</literal> or\n> + <literal>any</literal>. The default is <literal>any</literal>.\n> + If <literal>local</literal>, the subscription will request the publisher\n> + to only send changes that originated locally. If <literal>any</literal>\n> + the publisher sends any changes regardless of their origin.\n> + </para></entry>\n> + </row>.\n>\n> A comma can be added after \"If any\".\n\nModified\n\n> 3.\n> @@ -4589,6 +4598,8 @@ dumpSubscription(Archive *fout, const SubscriptionInfo *subinfo)\n> if (strcmp(subinfo->subdisableonerr, \"t\") == 0)\n> appendPQExpBufferStr(query, \", disable_on_error = true\");\n>\n> + appendPQExpBuffer(query, \", origin = %s\", subinfo->suborigin);\n> +\n> if (strcmp(subinfo->subsynccommit, \"off\") != 0)\n> appendPQExpBuffer(query, \", synchronous_commit = %s\", fmtId(subinfo->subsynccommit));\n>\n> Do we need to append anything if it's the default value (\"any\")? I saw that some\n> other parameters, they will be appended only if they are not the default value.\n\nModified\n\n> 0003 patch\n> ==============\n> 1.\n> in create_subscription.sgml:\n> (You cannot combine setting <literal>connect</literal>\n> to <literal>false</literal> with\n> setting <literal>create_slot</literal>, <literal>enabled</literal>,\n> or <literal>copy_data</literal> to <literal>true</literal>.)\n>\n> In this description about \"connect\" parameter in CREATE SUBSCIPTION document,\n> maybe it would be better to change \"copy_data to true\" to \"copy_data to\n> true/force\".\n\nModified\n\n> 2.\n> + appendStringInfoString(&cmd,\n> + \"SELECT DISTINCT N.nspname AS schemaname,\\n\"\n> + \" C.relname AS tablename,\\n\"\n> + \" PS.srrelid as replicated\\n\"\n> + \"FROM pg_publication P,\\n\"\n> + \" LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n> + \" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n> + \" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n> + \"WHERE C.oid = GPT.relid AND P.pubname IN (\");\n>\n> \"PS.srrelid as replicated\" can be modified to \"PS.srrelid AS replicated\".\n>\n> Besides, I think we can filter out the tables which are not subscribing data in\n> this SQL statement, then later processing can be simplified.\n>\n> Something like:\n> SELECT DISTINCT N.nspname AS schemaname,\n> C.relname AS tablename\n> FROM pg_publication P,\n> LATERAL pg_get_publication_tables(P.pubname) GPT\n> LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n> pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\n> WHERE C.oid = GPT.relid AND P.pubname IN ('pa') AND PS.srrelid IS NOT NULL;\n\nModified\n\n> 0004 patch\n> ==============\n> 1. Generic steps for adding a new node to an existing set of nodes\n>\n> + Step-2: Lock the required tables of the new node in EXCLUSIVE mode until\n> + the setup is complete. (This lock is necessary to prevent any modifications\n>\n> + Step-4. Lock the required tables of the existing nodes except the first node\n> + in EXCLUSIVE mode until the setup is complete. (This lock is necessary to\n>\n> Should \"in EXCLUSIVE mode\" be modified to \"in <literal>EXCLUSIVE</literal>\n> mode\"?\n\nModified\n\n> 2. Generic steps for adding a new node to an existing set of nodes\n>\n> + data. There is no need to lock the required tables on\n> + <literal>node1</literal> because any data changes made will be synchronized\n> + while creating the subscription with <literal>copy_data = force</literal>).\n>\n> I think it would be better to say \"on the first node\" here, instead of \"node1\",\n> because in this section, node1 is not mentioned before.\n\nModified\n\nThanks for the comment, the v22 patch attached has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Mon, 20 Jun 2022 17:25:24 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jun 20, 2022 at 3:16 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Jun 20, 2022 at 2:37 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >\n> > On Thu, Jun 16, 2022 at 3:48 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> > > >\n> >\n> > > Thanks for the comments, the attached v21 patch has the changes for the same.\n> >\n> > I have done some basic review of v21 and I have a few comments,\n> >\n> > 1.\n> > +/*\n> > + * The subscription will request the publisher to only send changes that\n> > + * originated locally.\n> > + */\n> > +#define LOGICALREP_ORIGIN_LOCAL \"local\"\n> > +\n> > +/*\n> > + * The subscription will request the publisher to send any changes regardless\n> > + * of their origin\n> > + */\n> > +#define LOGICALREP_ORIGIN_ANY \"any\"\n> >\n> > Are we planning to extend this to more options or are we planning to\n> > support the actual origin name here? If not then why isn't it just\n> > bool? I think the comments and the patch commit message should\n> > explain the details behind it if it has been already discussed and\n> > concluded.\n>\n> Currently we only support local and any. But this was designed to\n> accept string instead of boolean type, so that it can be extended\n> later to support filtering of origin names specified by the user in\n> the later versions. The same was also discussed in pg unconference as\n> mentioned in [1]. I will add it to the commit message and a comment\n> for the same in the next version.\n\nThanks for the comment, the v22 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm1h-9UNi_Jo_K%2BPK34tXBmV7fhj5C_nB8YzGA9rmUwHEA%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 20 Jun 2022 17:27:29 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for the v22* patch set.\n\n========\nv22-0001\n========\n\nNo comments. LGTM\n\n========\nV22-0002\n========\n\n2.1 doc/src/sgml/catalogs.sgml\n\n+ Possible origin values are <literal>local</literal> or\n+ <literal>any</literal>. The default is <literal>any</literal>.\n\nIMO the word \"Possible\" here is giving a sense of vagueness.\n\nSUGGESTION\nThe origin value must be either <literal>local</literal> or\n<literal>any</literal>.\n\n~~~\n\n2.2 src/backend/commands/subscriptioncmds.c\n\n@@ -265,6 +271,29 @@ parse_subscription_options(ParseState *pstate,\nList *stmt_options,\n opts->specified_opts |= SUBOPT_DISABLE_ON_ERR;\n opts->disableonerr = defGetBoolean(defel);\n }\n+ else if (IsSet(supported_opts, SUBOPT_ORIGIN) &&\n+ strcmp(defel->defname, \"origin\") == 0)\n+ {\n+ if (IsSet(opts->specified_opts, SUBOPT_ORIGIN))\n+ errorConflictingDefElem(defel, pstate);\n+\n+ opts->specified_opts |= SUBOPT_ORIGIN;\n+\n+ /*\n+ * Even though \"origin\" parameter allows only \"local\" and \"any\"\n+ * values, the \"origin\" parameter type is implemented as string\n+ * type instead of boolean to extend the \"origin\" parameter to\n+ * support filtering of origin name specified by the user in the\n+ * later versions.\n+ */\n+ opts->origin = defGetString(defel);\n+\n+ if ((strcmp(opts->origin, LOGICALREP_ORIGIN_LOCAL) != 0) &&\n+ (strcmp(opts->origin, LOGICALREP_ORIGIN_ANY) != 0))\n+ ereport(ERROR,\n+ errcode(ERRCODE_SYNTAX_ERROR),\n+ errmsg(\"unrecognized origin value: \\\"%s\\\"\", opts->origin));\n+ }\n\nI was wondering if it might be wise now to do a pfree(opts->origin)\nhere before setting the new option value which overwrites the\nstrdup-ed default \"any\". OTOH maybe it is overkill to worry about the\ntiny leak? I am not sure what is the convention for this.\n\n~~~\n\n2.3 src/backend/replication/pgoutput/pgoutput.c\n\n@@ -1698,12 +1719,16 @@ pgoutput_message(LogicalDecodingContext *ctx,\nReorderBufferTXN *txn,\n }\n\n /*\n- * Currently we always forward.\n+ * Return true if the data source (origin) is remote and user has requested\n+ * only local data, false otherwise.\n */\n static bool\n pgoutput_origin_filter(LogicalDecodingContext *ctx,\n RepOriginId origin_id)\n\n\"user\" -> \"the user\"\n\n~~~\n\n2.4 src/include/catalog/pg_subscription.h\n\n+/*\n+ * The subscription will request the publisher to send any changes regardless\n+ * of their origin\n+ */\n+#define LOGICALREP_ORIGIN_ANY \"any\"\n\nSUGGESTION (remove 'any'; add period)\nThe subscription will request the publisher to send changes regardless\nof their origin.\n\n========\nv22-0003\n========\n\n3.1 Commit message\n\nchange 1) Checks and throws an error if 'copy_data = on' and 'origin =\nlocal' but the publication tables were also subscribing from other publishers.\n\n\"also subscribing from other publishers.\" -> \"also replicating from\nother publishers.\"\n\n~~~\n\n3.2 doc/src/sgml/ref/alter_subscription.sgml\n\n+ <para>\n+ There is some interaction between the <literal>origin</literal>\n+ parameter and <literal>copy_data</literal> parameter.\n\n\"and copy_data parameter.\" -> \"and the copy_data parameter.\"\n\n~~~\n\n3.3 doc/src/sgml/ref/create_subscription.sgml\n\n+ <para>\n+ There is some interaction between the <literal>origin</literal>\n+ parameter and <literal>copy_data</literal> parameter.\n\n\"and copy_data parameter.\" -> \"and the copy_data parameter.\"\n\n~~~\n\n3.4 doc/src/sgml/ref/create_subscription.sgml\n\n+ <para>\n+ There is some interaction between the <literal>origin</literal>\n+ parameter and <literal>copy_data</literal> parameter. Refer to the\n+ <xref linkend=\"sql-createsubscription-notes\" /> for details.\n+ </para>\n\n\"and copy_data parameter.\" -> \"and the copy_data parameter.\"\n\n~~~\n\n3.5 doc/src/sgml/ref/create_subscription.sgml\n\n@@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n can have non-existent publications.\n </para>\n\n+ <para>\n+ If the subscription is created with <literal>origin = local</literal> and\n+ <literal>copy_data = true</literal>, it will check if the\npublisher tables are\n+ being subscribed to any other publisher and, if so, then throw an error to\n+ prevent possible non-local data from being copied. The user can override\n+ this check and continue with the copy operation by specifying\n+ <literal>copy_data = force</literal>.\n+ </para>\n+\n\nFor the part \"it will check if the publisher tables are being\nsubscribed to any other publisher...\", the \"being subscribed to\"\nsounds a bit strange. Maybe it is better to reword this like you\nalready worded some of the code comments.\n\nSUGGESTION\n-> \"... it will check if the publisher has subscribed to the same\ntable from other publishers and ...\"\n\n~~~\n\n3.6 src/backend/commands/subscriptioncmds.c\n\n+ /*\n+ * Throw an error if the publisher has subscribed to the same table\n+ * from some other publisher. We cannot differentiate between the\n+ * local and non-local data that is present in the HEAP during the\n+ * initial sync. Identification of local data can be done only from\n+ * the WAL by using the origin id. XXX: For simplicity, we don't check\n+ * whether the table has any data or not. If the table doesn't have\n+ * any data then we don't need to distinguish between local and\n+ * non-local data so we can avoid throwing error in that case.\n+ */\n\nI think the XXX part should be after a blank like so it is more\nprominent, instead of being buried in the other text.\n\ne.g.\n\n+ /*\n+ * Throw an error if the publisher has subscribed to the same table\n+ * from some other publisher. We cannot differentiate between the\n+ * local and non-local data that is present in the HEAP during the\n+ * initial sync. Identification of local data can be done only from\n+ * the WAL by using the origin id.\n+ *\n+ * XXX: For simplicity, we don't check\n+ * whether the table has any data or not. If the table doesn't have\n+ * any data then we don't need to distinguish between local and\n+ * non-local data so we can avoid throwing error in that case.\n+ */\n\n~~~\n\n3.7 src/test/regress/sql/subscription.sql\n\nElsewhere, there was code in this patch that apparently accepts\n'copy_data' parameter but with no parameter value specified. But this\ncombination has no test case.\n\n========\nv22-0004\n========\n\n4.1 doc/src/sgml/logical-replication.sgml\n\n+ <sect2 id=\"add-node-data-present-on-new-node\">\n+ <title>Adding a new node when data is present on the new node</title>\n\n\"when data is present\" -> \"when table data is present\"\n\n~~~\n\n4.2 doc/src/sgml/logical-replication.sgml\n\n+ <note>\n+ <para>\n+ Adding a new node when data is present on the new node tables is not\n+ supported.\n+ </para>\n+ </note>\n\nI think the note text should match the title text (see #4.1)\n\nSUGGESTION\nAdding a new node when table data is present on the new node is not supported.\n\n~~~\n\n4.3\n\n+ <para>\n+ Step-2: Lock the required tables of the new node in\n+ <literal>EXCLUSIVE</literal> mode untilthe setup is complete. (This lock is\n+ necessary to prevent any modifications from happening on the new node\n+ because if data modifications occurred after Step-3, there is a chance that\n+ the modifications will be published to the first node and then synchronized\n+ back to the new node while creating the subscription in Step-5. This would\n+ result in inconsistent data).\n+ </para>\n+ <para>\n\n4.3.a\ntypo: \"untilthe\" -> \"until the\"\n\n4.3.b\nSUGGESTION (just for the 2nd sentence)\nThis lock is necessary to prevent any modifications from happening on\nthe new node. If data modifications occurred after Step-3, there is a\nchance they could be published to the first node and then synchronized\nback to the new node while creating the subscription in Step-5.\n\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 22 Jun 2022 16:46:15 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jun 20, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> Thanks for the comment, the v22 patch attached has the changes for the\r\n> same.\r\n\r\nThanks for updating the patch, here are some comments on 0003 patch.\r\n\r\n1. 032_origin.pl\r\n+###############################################################################\r\n+# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\r\n+# replication setup when the existing nodes (node_A & node_B) and the new node\r\n+# (node_C) does not have any data.\r\n+###############################################################################\r\n\r\n\"does not have any data\" should be \"do not have any data\" I think.\r\n\r\n2.\r\nThe comment for 032_origin.pl:\r\n\r\n# Test the CREATE SUBSCRIPTION 'origin' parameter.\r\n\r\nAfter applying this patch, this file tests no only 'origin' parameter, but also\r\n\"copy_data\" parameter, so should we modify this comment?\r\n\r\nBesides, should we change the file name in this patch? It looks more like test\r\ncases for bidirectional logical replication.\r\n\r\n3. subscriptioncmds.c\r\n\t/* Set default values for the boolean supported options. */\r\n...\r\n\tif (IsSet(supported_opts, SUBOPT_CREATE_SLOT))\r\n \t\topts->create_slot = true;\r\n \tif (IsSet(supported_opts, SUBOPT_COPY_DATA))\r\n-\t\topts->copy_data = true;\r\n+\t\topts->copy_data = COPY_DATA_ON;\r\n \tif (IsSet(supported_opts, SUBOPT_REFRESH))\r\n \t\topts->refresh = true;\r\n \tif (IsSet(supported_opts, SUBOPT_BINARY))\r\n\r\n\"copy_data\" option is not Boolean now, which is inconsistent with the comment.\r\nSo maybe we can change the comment here? (\"the boolean supported options\" ->\r\n\"the supported options\")\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Thu, 23 Jun 2022 01:35:15 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Jun 22, 2022 at 12:16 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for the v22* patch set.\n>\n> ========\n> v22-0001\n> ========\n>\n> No comments. LGTM\n>\n> ========\n> V22-0002\n> ========\n>\n> 2.1 doc/src/sgml/catalogs.sgml\n>\n> + Possible origin values are <literal>local</literal> or\n> + <literal>any</literal>. The default is <literal>any</literal>.\n>\n> IMO the word \"Possible\" here is giving a sense of vagueness.\n>\n> SUGGESTION\n> The origin value must be either <literal>local</literal> or\n> <literal>any</literal>.\n\nModified\n\n> ~~~\n>\n> 2.2 src/backend/commands/subscriptioncmds.c\n>\n> @@ -265,6 +271,29 @@ parse_subscription_options(ParseState *pstate,\n> List *stmt_options,\n> opts->specified_opts |= SUBOPT_DISABLE_ON_ERR;\n> opts->disableonerr = defGetBoolean(defel);\n> }\n> + else if (IsSet(supported_opts, SUBOPT_ORIGIN) &&\n> + strcmp(defel->defname, \"origin\") == 0)\n> + {\n> + if (IsSet(opts->specified_opts, SUBOPT_ORIGIN))\n> + errorConflictingDefElem(defel, pstate);\n> +\n> + opts->specified_opts |= SUBOPT_ORIGIN;\n> +\n> + /*\n> + * Even though \"origin\" parameter allows only \"local\" and \"any\"\n> + * values, the \"origin\" parameter type is implemented as string\n> + * type instead of boolean to extend the \"origin\" parameter to\n> + * support filtering of origin name specified by the user in the\n> + * later versions.\n> + */\n> + opts->origin = defGetString(defel);\n> +\n> + if ((strcmp(opts->origin, LOGICALREP_ORIGIN_LOCAL) != 0) &&\n> + (strcmp(opts->origin, LOGICALREP_ORIGIN_ANY) != 0))\n> + ereport(ERROR,\n> + errcode(ERRCODE_SYNTAX_ERROR),\n> + errmsg(\"unrecognized origin value: \\\"%s\\\"\", opts->origin));\n> + }\n>\n> I was wondering if it might be wise now to do a pfree(opts->origin)\n> here before setting the new option value which overwrites the\n> strdup-ed default \"any\". OTOH maybe it is overkill to worry about the\n> tiny leak? I am not sure what is the convention for this.\n\nI have freed the memory to avoid the leak, I felt it is better to\nclean the memory as it is a positive flow.\n\n> ~~~\n>\n> 2.3 src/backend/replication/pgoutput/pgoutput.c\n>\n> @@ -1698,12 +1719,16 @@ pgoutput_message(LogicalDecodingContext *ctx,\n> ReorderBufferTXN *txn,\n> }\n>\n> /*\n> - * Currently we always forward.\n> + * Return true if the data source (origin) is remote and user has requested\n> + * only local data, false otherwise.\n> */\n> static bool\n> pgoutput_origin_filter(LogicalDecodingContext *ctx,\n> RepOriginId origin_id)\n>\n> \"user\" -> \"the user\"\n\nModified\n\n> ~~~\n>\n> 2.4 src/include/catalog/pg_subscription.h\n>\n> +/*\n> + * The subscription will request the publisher to send any changes regardless\n> + * of their origin\n> + */\n> +#define LOGICALREP_ORIGIN_ANY \"any\"\n>\n> SUGGESTION (remove 'any'; add period)\n> The subscription will request the publisher to send changes regardless\n> of their origin.\n\nModified\n\n> ========\n> v22-0003\n> ========\n>\n> 3.1 Commit message\n>\n> change 1) Checks and throws an error if 'copy_data = on' and 'origin =\n> local' but the publication tables were also subscribing from other publishers.\n>\n> \"also subscribing from other publishers.\" -> \"also replicating from\n> other publishers.\"\n\nModified\n\n> ~~~\n>\n> 3.2 doc/src/sgml/ref/alter_subscription.sgml\n>\n> + <para>\n> + There is some interaction between the <literal>origin</literal>\n> + parameter and <literal>copy_data</literal> parameter.\n>\n> \"and copy_data parameter.\" -> \"and the copy_data parameter.\"\n\nModified\n\n> ~~~\n>\n> 3.3 doc/src/sgml/ref/create_subscription.sgml\n>\n> + <para>\n> + There is some interaction between the <literal>origin</literal>\n> + parameter and <literal>copy_data</literal> parameter.\n>\n> \"and copy_data parameter.\" -> \"and the copy_data parameter.\"\n\nModified\n\n> ~~~\n>\n> 3.4 doc/src/sgml/ref/create_subscription.sgml\n>\n> + <para>\n> + There is some interaction between the <literal>origin</literal>\n> + parameter and <literal>copy_data</literal> parameter. Refer to the\n> + <xref linkend=\"sql-createsubscription-notes\" /> for details.\n> + </para>\n>\n> \"and copy_data parameter.\" -> \"and the copy_data parameter.\"\n\nModified\n\n> ~~~\n>\n> 3.5 doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> can have non-existent publications.\n> </para>\n>\n> + <para>\n> + If the subscription is created with <literal>origin = local</literal> and\n> + <literal>copy_data = true</literal>, it will check if the\n> publisher tables are\n> + being subscribed to any other publisher and, if so, then throw an error to\n> + prevent possible non-local data from being copied. The user can override\n> + this check and continue with the copy operation by specifying\n> + <literal>copy_data = force</literal>.\n> + </para>\n> +\n>\n> For the part \"it will check if the publisher tables are being\n> subscribed to any other publisher...\", the \"being subscribed to\"\n> sounds a bit strange. Maybe it is better to reword this like you\n> already worded some of the code comments.\n>\n> SUGGESTION\n> -> \"... it will check if the publisher has subscribed to the same\n> table from other publishers and ...\"\n\nModified\n\n> ~~~\n>\n> 3.6 src/backend/commands/subscriptioncmds.c\n>\n> + /*\n> + * Throw an error if the publisher has subscribed to the same table\n> + * from some other publisher. We cannot differentiate between the\n> + * local and non-local data that is present in the HEAP during the\n> + * initial sync. Identification of local data can be done only from\n> + * the WAL by using the origin id. XXX: For simplicity, we don't check\n> + * whether the table has any data or not. If the table doesn't have\n> + * any data then we don't need to distinguish between local and\n> + * non-local data so we can avoid throwing error in that case.\n> + */\n>\n> I think the XXX part should be after a blank like so it is more\n> prominent, instead of being buried in the other text.\n>\n> e.g.\n>\n> + /*\n> + * Throw an error if the publisher has subscribed to the same table\n> + * from some other publisher. We cannot differentiate between the\n> + * local and non-local data that is present in the HEAP during the\n> + * initial sync. Identification of local data can be done only from\n> + * the WAL by using the origin id.\n> + *\n> + * XXX: For simplicity, we don't check\n> + * whether the table has any data or not. If the table doesn't have\n> + * any data then we don't need to distinguish between local and\n> + * non-local data so we can avoid throwing error in that case.\n> + */\n\nModified\n\n> ~~~\n>\n> 3.7 src/test/regress/sql/subscription.sql\n>\n> Elsewhere, there was code in this patch that apparently accepts\n> 'copy_data' parameter but with no parameter value specified. But this\n> combination has no test case.\n\nI have added a test for the same\n\n> ========\n> v22-0004\n> ========\n>\n> 4.1 doc/src/sgml/logical-replication.sgml\n>\n> + <sect2 id=\"add-node-data-present-on-new-node\">\n> + <title>Adding a new node when data is present on the new node</title>\n>\n> \"when data is present\" -> \"when table data is present\"\n\nModified\n\n> ~~~\n>\n> 4.2 doc/src/sgml/logical-replication.sgml\n>\n> + <note>\n> + <para>\n> + Adding a new node when data is present on the new node tables is not\n> + supported.\n> + </para>\n> + </note>\n>\n> I think the note text should match the title text (see #4.1)\n>\n> SUGGESTION\n> Adding a new node when table data is present on the new node is not supported.\n\nModified\n\n> ~~~\n>\n> 4.3\n>\n> + <para>\n> + Step-2: Lock the required tables of the new node in\n> + <literal>EXCLUSIVE</literal> mode untilthe setup is complete. (This lock is\n> + necessary to prevent any modifications from happening on the new node\n> + because if data modifications occurred after Step-3, there is a chance that\n> + the modifications will be published to the first node and then synchronized\n> + back to the new node while creating the subscription in Step-5. This would\n> + result in inconsistent data).\n> + </para>\n> + <para>\n>\n> 4.3.a\n> typo: \"untilthe\" -> \"until the\"\n\nModified\n\n> 4.3.b\n> SUGGESTION (just for the 2nd sentence)\n> This lock is necessary to prevent any modifications from happening on\n> the new node. If data modifications occurred after Step-3, there is a\n> chance they could be published to the first node and then synchronized\n> back to the new node while creating the subscription in Step-5.\n\nModified\n\nThanks for the comments, the attached v23 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Thu, 23 Jun 2022 15:36:45 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jun 23, 2022 at 7:05 AM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Mon, Jun 20, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comment, the v22 patch attached has the changes for the\n> > same.\n>\n> Thanks for updating the patch, here are some comments on 0003 patch.\n>\n> 1. 032_origin.pl\n> +###############################################################################\n> +# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\n> +# replication setup when the existing nodes (node_A & node_B) and the new node\n> +# (node_C) does not have any data.\n> +###############################################################################\n>\n> \"does not have any data\" should be \"do not have any data\" I think.\n\nI felt the existing is ok, I did not find it gramatically wrong. I did\nnot make any changes for this.\n\n> 2.\n> The comment for 032_origin.pl:\n>\n> # Test the CREATE SUBSCRIPTION 'origin' parameter.\n>\n> After applying this patch, this file tests no only 'origin' parameter, but also\n> \"copy_data\" parameter, so should we modify this comment?\n\nModified\n\n> Besides, should we change the file name in this patch? It looks more like test\n> cases for bidirectional logical replication.\n\nI felt let's keep it as origin, as most of the tests are based on\n'origin' parameter, also later features will be able to filter a\nparticular origin, those tests can easily fit in the same file.\n\n> 3. subscriptioncmds.c\n> /* Set default values for the boolean supported options. */\n> ...\n> if (IsSet(supported_opts, SUBOPT_CREATE_SLOT))\n> opts->create_slot = true;\n> if (IsSet(supported_opts, SUBOPT_COPY_DATA))\n> - opts->copy_data = true;\n> + opts->copy_data = COPY_DATA_ON;\n> if (IsSet(supported_opts, SUBOPT_REFRESH))\n> opts->refresh = true;\n> if (IsSet(supported_opts, SUBOPT_BINARY))\n>\n> \"copy_data\" option is not Boolean now, which is inconsistent with the comment.\n> So maybe we can change the comment here? (\"the boolean supported options\" ->\n> \"the supported options\")\n\nModified\n\nThanks for the comments, the v23 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm1-ZrG%3DhaAoiB2yFKYc%2Bckcd1NLaU8QB3SWs32wPsph4w%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Thu, 23 Jun 2022 15:38:50 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for the v23* patch set.\n\n========\nv23-0001\n========\n\nNo comments. LGTM\n\n========\nV23-0002\n========\n\n2.1 src/backend/commands/subscriptioncmds.c\n\n+ opts->origin = defGetString(defel);\n+\n+ if ((strcmp(opts->origin, LOGICALREP_ORIGIN_LOCAL) != 0) &&\n+ (strcmp(opts->origin, LOGICALREP_ORIGIN_ANY) != 0))\n+ ereport(ERROR,\n+ errcode(ERRCODE_SYNTAX_ERROR),\n+ errmsg(\"unrecognized origin value: \\\"%s\\\"\", opts->origin));\n+ }\n\nI thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\n\n~~~\n\n2.2 src/backend/replication/pgoutput/pgoutput.c\n\n+ data->origin = defGetString(defel);\n+ if (strcmp(data->origin, LOGICALREP_ORIGIN_LOCAL) == 0)\n+ publish_local_origin = true;\n+ else if (strcmp(data->origin, LOGICALREP_ORIGIN_ANY) == 0)\n+ publish_local_origin = false;\n+ else\n+ ereport(ERROR,\n+ errcode(ERRCODE_SYNTAX_ERROR),\n+ errmsg(\"unrecognized origin value: \\\"%s\\\"\", data->origin));\n+ }\n\nI thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\n\n========\nv23-0003\n========\n\n3.1 Commit message\n\nAfter the subscription is created on node2, node1 will be synced to\nnode2 and the newly synced data will be sent to node2, this process of\nnode1 sending data to node2 and node2 sending data to node1 will repeat\ninfinitely. If table t1 has a unique key, this will lead to a unique key\nviolation, and replication won't proceed.\n\n~\n\n31a.\n\"node2, this process of\" -> \"node2; this process of\"\nOR\n\"node2, this process of\" -> \"node2. This process of\"\n\n31b.\nAlso, my grammar checker recommends removing the comma after \"violation\"\n\n~~~\n\n3.2 doc/src/sgml/ref/create_subscription.sgml\n\n@@ -115,7 +115,8 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n (You cannot combine setting <literal>connect</literal>\n to <literal>false</literal> with\n setting <literal>create_slot</literal>, <literal>enabled</literal>,\n- or <literal>copy_data</literal> to <literal>true</literal>.)\n+ or <literal>copy_data</literal> to\n+ <literal>true</literal>/<literal>force</literal>.)\n </para>\n\nI am not sure why that last sentence needs to be in parentheses, but\nOTOH it seems to be that way already in PG15.\n\n~~~\n\n3.3 doc/src/sgml/ref/create_subscription.sgml\n\n@@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n can have non-existent publications.\n </para>\n\n+ <para>\n+ If the subscription is created with <literal>origin = local</literal> and\n+ <literal>copy_data = true</literal>, it will check if the publisher has\n+ subscribed to the same table from other publishers and, if so, then throw an\n+ error to prevent possible non-local data from being copied. The user can\n+ override this check and continue with the copy operation by specifying\n+ <literal>copy_data = force</literal>.\n+ </para>\n\n\"and, if so, then throw...\" -> \"and, if so, throw...\"\n\n~~~\n\n3.4 src/backend/commands/subscriptioncmds.c\n\n+ ereport(ERROR,\n+ errcode(ERRCODE_SYNTAX_ERROR),\n+ errmsg(\"%s requires a boolean or \\\"force\\\"\", def->defname));\n+ return COPY_DATA_OFF; /* keep compiler quiet */\n+}\n\nI thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\n\n========\nv23-0004\n========\n\n4.1 Commit message\n\nDocument the steps for the following:\na) Creating a two-node bidirectional replication when there is no data\non both nodes.\nb) Adding a new node when there is no data on any of the nodes.\nc) Adding a new node when data is present on the existing nodes.\nd) Generic steps for adding a new node to an existing set of nodes.\n\n~\n\nThese pgdocs titles have changed slightly. I think this commit message\ntext should use match the current pgdocs titles.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 24 Jun 2022 14:50:19 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jun 24, 2022 at 10:20 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for the v23* patch set.\n>\n> ========\n> v23-0001\n> ========\n>\n> No comments. LGTM\n>\n> ========\n> V23-0002\n> ========\n>\n> 2.1 src/backend/commands/subscriptioncmds.c\n>\n> + opts->origin = defGetString(defel);\n> +\n> + if ((strcmp(opts->origin, LOGICALREP_ORIGIN_LOCAL) != 0) &&\n> + (strcmp(opts->origin, LOGICALREP_ORIGIN_ANY) != 0))\n> + ereport(ERROR,\n> + errcode(ERRCODE_SYNTAX_ERROR),\n> + errmsg(\"unrecognized origin value: \\\"%s\\\"\", opts->origin));\n> + }\n>\n> I thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\n\nModified\n\n> ~~~\n>\n> 2.2 src/backend/replication/pgoutput/pgoutput.c\n>\n> + data->origin = defGetString(defel);\n> + if (strcmp(data->origin, LOGICALREP_ORIGIN_LOCAL) == 0)\n> + publish_local_origin = true;\n> + else if (strcmp(data->origin, LOGICALREP_ORIGIN_ANY) == 0)\n> + publish_local_origin = false;\n> + else\n> + ereport(ERROR,\n> + errcode(ERRCODE_SYNTAX_ERROR),\n> + errmsg(\"unrecognized origin value: \\\"%s\\\"\", data->origin));\n> + }\n>\n> I thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\n\nModified\n\n> ========\n> v23-0003\n> ========\n>\n> 3.1 Commit message\n>\n> After the subscription is created on node2, node1 will be synced to\n> node2 and the newly synced data will be sent to node2, this process of\n> node1 sending data to node2 and node2 sending data to node1 will repeat\n> infinitely. If table t1 has a unique key, this will lead to a unique key\n> violation, and replication won't proceed.\n>\n> ~\n>\n> 31a.\n> \"node2, this process of\" -> \"node2; this process of\"\n> OR\n> \"node2, this process of\" -> \"node2. This process of\"\n\nModified\n\n> 31b.\n> Also, my grammar checker recommends removing the comma after \"violation\"\n\nModified\n\n> ~~~\n>\n> 3.2 doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -115,7 +115,8 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> (You cannot combine setting <literal>connect</literal>\n> to <literal>false</literal> with\n> setting <literal>create_slot</literal>, <literal>enabled</literal>,\n> - or <literal>copy_data</literal> to <literal>true</literal>.)\n> + or <literal>copy_data</literal> to\n> + <literal>true</literal>/<literal>force</literal>.)\n> </para>\n>\n> I am not sure why that last sentence needs to be in parentheses, but\n> OTOH it seems to be that way already in PG15.\n\nI feel since it is like that since PG15, let's keep it in parenthesis\nlike earlier. I have not made any changes for this.\n\n> ~~~\n>\n> 3.3 doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> can have non-existent publications.\n> </para>\n>\n> + <para>\n> + If the subscription is created with <literal>origin = local</literal> and\n> + <literal>copy_data = true</literal>, it will check if the publisher has\n> + subscribed to the same table from other publishers and, if so, then throw an\n> + error to prevent possible non-local data from being copied. The user can\n> + override this check and continue with the copy operation by specifying\n> + <literal>copy_data = force</literal>.\n> + </para>\n>\n> \"and, if so, then throw...\" -> \"and, if so, throw...\"\n\nModified\n\n> ~~~\n>\n> 3.4 src/backend/commands/subscriptioncmds.c\n>\n> + ereport(ERROR,\n> + errcode(ERRCODE_SYNTAX_ERROR),\n> + errmsg(\"%s requires a boolean or \\\"force\\\"\", def->defname));\n> + return COPY_DATA_OFF; /* keep compiler quiet */\n> +}\n>\n> I thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\n\nModified\n\n> ========\n> v23-0004\n> ========\n>\n> 4.1 Commit message\n>\n> Document the steps for the following:\n> a) Creating a two-node bidirectional replication when there is no data\n> on both nodes.\n> b) Adding a new node when there is no data on any of the nodes.\n> c) Adding a new node when data is present on the existing nodes.\n> d) Generic steps for adding a new node to an existing set of nodes.\n>\n> ~\n>\n> These pgdocs titles have changed slightly. I think this commit message\n> text should use match the current pgdocs titles.\n\nModified\n\nThanks for the comments, the attached v24 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Mon, 27 Jun 2022 19:08:44 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for the v24* patch set.\n\nNow, these few comments are all trivial and non-functional. Apart from\nthese, everything looks good to me.\n\n========\nv24-0001\n========\n\nNo comments. LGTM\n\n========\nV24-0002\n========\n\n2.1 doc/src/sgml/ref/create_subscription.sgml\n\n+ <para>\n+ Specifies whether the subscription will request the publisher to only\n+ send changes that originated locally, or to send any changes\n+ regardless of origin. Setting <literal>origin</literal> to\n+ <literal>local</literal> means that the subscription will request the\n+ publisher to only send changes that originated locally. Setting\n+ <literal>origin</literal> to <literal>any</literal> means that the\n+ publisher sends any changes regardless of their origin. The default\n+ is <literal>any</literal>.\n+ </para>\n\n2.1a.\nIMO remove the word \"any\" from \"any changes\". Then the text will match\nwhat is written in catalogs.sgml.\n\"send any changes regardless of origin\" -> \"send changes regardless of\norigin\" (occurs 2x)\n\n2.1b.\nThis same text is cut/paste to the commit message so that can also be updated.\n\n\n~~~\n\n2.2 src/backend/commands/subscriptioncmds.c\n\n+ /*\n+ * Even though \"origin\" parameter allows only \"local\" and \"any\"\n+ * values, the \"origin\" parameter type is implemented as string\n+ * type instead of boolean to extend the \"origin\" parameter to\n+ * support filtering of origin name specified by the user in the\n+ * later versions.\n+ */\n\n2.2a.\nSUGGESTION\nEven though \"origin\" parameter allows only \"local\" and \"any\" values,\nit is implemented as a string type so that the parameter can be\nextended in future versions to support filtering using origin names\nspecified by the user.\n\n2.2b.\nThis same text is cut/paste to the commit message so that can also be updated.\n\n\n========\nv24-0003\n========\n\n3.1 Commit message\n\nThis patch does a couple of things:\nchange 1) Checks and throws an error if 'copy_data = on' and 'origin =\nlocal' but the publication tables were also replicating from other publishers.\nchange 2) Adds 'force' value for copy_data parameter.\n\n~\n\n\"replicating\" -> \"replicated\"\n\"change 1)\" -> \"1)\"\n\"change 2)\" -> \"2)\"\n\n========\nv24-0004\n========\n\nNo comments. LGTM\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 28 Jun 2022 11:58:49 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jun 28, 2022 at 7:29 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for the v24* patch set.\n>\n> Now, these few comments are all trivial and non-functional. Apart from\n> these, everything looks good to me.\n>\n> ========\n> v24-0001\n> ========\n>\n> No comments. LGTM\n>\n> ========\n> V24-0002\n> ========\n>\n> 2.1 doc/src/sgml/ref/create_subscription.sgml\n>\n> + <para>\n> + Specifies whether the subscription will request the publisher to only\n> + send changes that originated locally, or to send any changes\n> + regardless of origin. Setting <literal>origin</literal> to\n> + <literal>local</literal> means that the subscription will request the\n> + publisher to only send changes that originated locally. Setting\n> + <literal>origin</literal> to <literal>any</literal> means that the\n> + publisher sends any changes regardless of their origin. The default\n> + is <literal>any</literal>.\n> + </para>\n>\n> 2.1a.\n> IMO remove the word \"any\" from \"any changes\". Then the text will match\n> what is written in catalogs.sgml.\n> \"send any changes regardless of origin\" -> \"send changes regardless of\n> origin\" (occurs 2x)\n\nModified\n\n> 2.1b.\n> This same text is cut/paste to the commit message so that can also be updated.\n\nModified\n\n>\n> ~~~\n>\n> 2.2 src/backend/commands/subscriptioncmds.c\n>\n> + /*\n> + * Even though \"origin\" parameter allows only \"local\" and \"any\"\n> + * values, the \"origin\" parameter type is implemented as string\n> + * type instead of boolean to extend the \"origin\" parameter to\n> + * support filtering of origin name specified by the user in the\n> + * later versions.\n> + */\n>\n> 2.2a.\n> SUGGESTION\n> Even though \"origin\" parameter allows only \"local\" and \"any\" values,\n> it is implemented as a string type so that the parameter can be\n> extended in future versions to support filtering using origin names\n> specified by the user.\n\nModified\n\n> 2.2b.\n> This same text is cut/paste to the commit message so that can also be updated.\n\nModified\n\n>\n> ========\n> v24-0003\n> ========\n>\n> 3.1 Commit message\n>\n\n> v24-0004\n> ========\n>\n> No comments. LGTM\n>\n> ------\n> Kind Regards,\n> Peter Smith.\n> This patch does a couple of things:\n> change 1) Checks and throws an error if 'copy_data = on' and 'origin =\n> local' but the publication tables were also replicating from other publishers.\n> change 2) Adds 'force' value for copy_data parameter.\n>\n> ~\n>\n> \"replicating\" -> \"replicated\"\n> \"change 1)\" -> \"1)\"\n> \"change 2)\" -> \"2)\"\n\nModified\n\nThanks for the comments, the attached v25 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Tue, 28 Jun 2022 11:47:53 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jun 28, 2022 2:18 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> Thanks for the comments, the attached v25 patch has the changes for the\r\n> same.\r\n> \r\n\r\nThanks for updating the patch. Here are some comments.\r\n\r\n0002 patch:\r\n==============\r\n1.\r\n+# Test the CREATE SUBSCRIPTION 'origin' parameter and its interaction with\r\n+# 'copy_data' parameter.\r\n\r\nIt seems we should move \"and its interaction with 'copy_data' parameter\" to\r\n0003 patch.\r\n\r\n0003 patch\r\n==============\r\n1.\r\nWhen using ALTER SUBSCRIPTION ... REFRESH, subscription will throw an error if\r\nany table is subscribed in publisher, even if the table has been subscribed\r\nbefore refresh (which won't do the initial copy when refreshing). It looks the\r\npreviously subscribed tables don't need this check. Would it be better that we\r\nonly check the tables which need to do the initial copy?\r\n\r\n2.\r\n+\t\t\t\terrmsg(\"table:%s.%s might have replicated data in the publisher\",\r\n+\t\t\t\t\t nspname, relname),\r\n\r\nI think the table name needs to be enclosed in double quotes, which is\r\nconsistent with other messages.\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Thu, 30 Jun 2022 03:47:24 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jun 30, 2022 at 9:17 AM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Tue, Jun 28, 2022 2:18 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached v25 patch has the changes for the\n> > same.\n> >\n>\n> Thanks for updating the patch. Here are some comments.\n>\n> 0002 patch:\n> ==============\n> 1.\n> +# Test the CREATE SUBSCRIPTION 'origin' parameter and its interaction with\n> +# 'copy_data' parameter.\n>\n> It seems we should move \"and its interaction with 'copy_data' parameter\" to\n> 0003 patch.\n\nModified\n\n> 0003 patch\n> ==============\n> 1.\n> When using ALTER SUBSCRIPTION ... REFRESH, subscription will throw an error if\n> any table is subscribed in publisher, even if the table has been subscribed\n> before refresh (which won't do the initial copy when refreshing). It looks the\n> previously subscribed tables don't need this check. Would it be better that we\n> only check the tables which need to do the initial copy?\n\nModified\n\n> 2.\n> + errmsg(\"table:%s.%s might have replicated data in the publisher\",\n> + nspname, relname),\n>\n> I think the table name needs to be enclosed in double quotes, which is\n> consistent with other messages.\n\nModified\n\nThanks for the comments, the attached v26 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Thu, 30 Jun 2022 21:40:14 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jun 30, 2022 at 9:40 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Thu, Jun 30, 2022 at 9:17 AM shiy.fnst@fujitsu.com\n> <shiy.fnst@fujitsu.com> wrote:\n> >\n\nThe first patch that adds a test case for existing functionality looks\ngood to me and I'll push that early next week (by Tuesday) unless\nthere are more comments on it.\n\nFew minor comments on 0002\n========================\n1.\n+ /* FIXME: 150000 should be changed to 160000 later for PG16. */\n+ if (options->proto.logical.origin &&\n+ PQserverVersion(conn->streamConn) >= 150000)\n+ appendStringInfo(&cmd, \", origin '%s'\",\n+ options->proto.logical.origin);\n\n...\n...\n+ /* FIXME: 150000 should be changed to 160000 later for PG16. */\n+ if (fout->remoteVersion >= 150000)\n+ appendPQExpBufferStr(query, \" s.suborigin\\n\");\n+ else\n+ appendPQExpBuffer(query, \" '%s' AS suborigin\\n\", LOGICALREP_ORIGIN_ANY);\n...\n...\n+ /* FIXME: 150000 should be changed to 160000 later for PG16 */\n+ if (pset.sversion >= 150000)\n+ appendPQExpBuffer(&buf,\n+ \", suborigin AS \\\"%s\\\"\\n\",\n+ gettext_noop(\"Origin\"));\n\nAll these should now change to 16.\n\n2.\n/* ALTER SUBSCRIPTION <name> SET ( */\n else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\nTailMatches(\"SET\", \"(\"))\n- COMPLETE_WITH(\"binary\", \"slot_name\", \"streaming\",\n\"synchronous_commit\", \"disable_on_error\");\n+ COMPLETE_WITH(\"binary\", \"origin\", \"slot_name\", \"streaming\",\n\"synchronous_commit\", \"disable_on_error\");\n /* ALTER SUBSCRIPTION <name> SKIP ( */\n else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\nTailMatches(\"SKIP\", \"(\"))\n COMPLETE_WITH(\"lsn\");\n@@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\n /* Complete \"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\" */\n else if (HeadMatches(\"CREATE\", \"SUBSCRIPTION\") && TailMatches(\"WITH\", \"(\"))\n COMPLETE_WITH(\"binary\", \"connect\", \"copy_data\", \"create_slot\",\n- \"enabled\", \"slot_name\", \"streaming\",\n+ \"enabled\", \"origin\", \"slot_name\", \"streaming\",\n \"synchronous_commit\", \"two_phase\", \"disable_on_error\");\n\nWhy do you choose to add a new option in-between other parameters\ninstead of at the end which we normally do? The one possible reason I\ncan think of is that all the parameters at the end are boolean so you\nwant to add this before those but then why before slot_name, and again\nI don't see such a rule being followed for other parameters.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Sat, 2 Jul 2022 12:12:35 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sat, Jul 2, 2022 at 12:12 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Thu, Jun 30, 2022 at 9:40 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Thu, Jun 30, 2022 at 9:17 AM shiy.fnst@fujitsu.com\n> > <shiy.fnst@fujitsu.com> wrote:\n> > >\n>\n> The first patch that adds a test case for existing functionality looks\n> good to me and I'll push that early next week (by Tuesday) unless\n> there are more comments on it.\n>\n> Few minor comments on 0002\n> ========================\n> 1.\n> + /* FIXME: 150000 should be changed to 160000 later for PG16. */\n> + if (options->proto.logical.origin &&\n> + PQserverVersion(conn->streamConn) >= 150000)\n> + appendStringInfo(&cmd, \", origin '%s'\",\n> + options->proto.logical.origin);\n>\n> ...\n> ...\n> + /* FIXME: 150000 should be changed to 160000 later for PG16. */\n> + if (fout->remoteVersion >= 150000)\n> + appendPQExpBufferStr(query, \" s.suborigin\\n\");\n> + else\n> + appendPQExpBuffer(query, \" '%s' AS suborigin\\n\", LOGICALREP_ORIGIN_ANY);\n> ...\n> ...\n> + /* FIXME: 150000 should be changed to 160000 later for PG16 */\n> + if (pset.sversion >= 150000)\n> + appendPQExpBuffer(&buf,\n> + \", suborigin AS \\\"%s\\\"\\n\",\n> + gettext_noop(\"Origin\"));\n>\n> All these should now change to 16.\n\nModified\n\n> 2.\n> /* ALTER SUBSCRIPTION <name> SET ( */\n> else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\n> TailMatches(\"SET\", \"(\"))\n> - COMPLETE_WITH(\"binary\", \"slot_name\", \"streaming\",\n> \"synchronous_commit\", \"disable_on_error\");\n> + COMPLETE_WITH(\"binary\", \"origin\", \"slot_name\", \"streaming\",\n> \"synchronous_commit\", \"disable_on_error\");\n> /* ALTER SUBSCRIPTION <name> SKIP ( */\n> else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\n> TailMatches(\"SKIP\", \"(\"))\n> COMPLETE_WITH(\"lsn\");\n> @@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\n> /* Complete \"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\" */\n> else if (HeadMatches(\"CREATE\", \"SUBSCRIPTION\") && TailMatches(\"WITH\", \"(\"))\n> COMPLETE_WITH(\"binary\", \"connect\", \"copy_data\", \"create_slot\",\n> - \"enabled\", \"slot_name\", \"streaming\",\n> + \"enabled\", \"origin\", \"slot_name\", \"streaming\",\n> \"synchronous_commit\", \"two_phase\", \"disable_on_error\");\n>\n> Why do you choose to add a new option in-between other parameters\n> instead of at the end which we normally do? The one possible reason I\n> can think of is that all the parameters at the end are boolean so you\n> want to add this before those but then why before slot_name, and again\n> I don't see such a rule being followed for other parameters.\n\nI was not sure if it should be maintained in alphabetical order,\nanyway since the last option \"disable_on_error\" is at the end, I have\nchanged it to the end.\n\nThanks for the comments, the attached v27 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Sun, 3 Jul 2022 20:29:37 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 4, 2022 at 12:59 AM vignesh C <vignesh21@gmail.com> wrote:\n...\n> > 2.\n> > /* ALTER SUBSCRIPTION <name> SET ( */\n> > else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\n> > TailMatches(\"SET\", \"(\"))\n> > - COMPLETE_WITH(\"binary\", \"slot_name\", \"streaming\",\n> > \"synchronous_commit\", \"disable_on_error\");\n> > + COMPLETE_WITH(\"binary\", \"origin\", \"slot_name\", \"streaming\",\n> > \"synchronous_commit\", \"disable_on_error\");\n> > /* ALTER SUBSCRIPTION <name> SKIP ( */\n> > else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\n> > TailMatches(\"SKIP\", \"(\"))\n> > COMPLETE_WITH(\"lsn\");\n> > @@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\n> > /* Complete \"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\" */\n> > else if (HeadMatches(\"CREATE\", \"SUBSCRIPTION\") && TailMatches(\"WITH\", \"(\"))\n> > COMPLETE_WITH(\"binary\", \"connect\", \"copy_data\", \"create_slot\",\n> > - \"enabled\", \"slot_name\", \"streaming\",\n> > + \"enabled\", \"origin\", \"slot_name\", \"streaming\",\n> > \"synchronous_commit\", \"two_phase\", \"disable_on_error\");\n> >\n> > Why do you choose to add a new option in-between other parameters\n> > instead of at the end which we normally do? The one possible reason I\n> > can think of is that all the parameters at the end are boolean so you\n> > want to add this before those but then why before slot_name, and again\n> > I don't see such a rule being followed for other parameters.\n>\n> I was not sure if it should be maintained in alphabetical order,\n> anyway since the last option \"disable_on_error\" is at the end, I have\n> changed it to the end.\n>\n\nAlthough it seems it is not a hard rule, mostly the COMPLETE_WITH are\ncoded using alphabetical order. Anyway, I think that was a clear\nintention here too since 13 of 14 parameters were already in\nalphabetical order; it is actually only that \"disable_on_error\"\nparameter that was misplaced; not the new \"origin\" parameter.\n\nAlso, in practice, on <tab> those completions will get output in\nalphabetical order, so IMO it makes more sense for the code to be\nconsistent with the output:\n\ne.g.\ntest_sub=# create subscription sub xxx connection '' publication pub WITH (\nBINARY DISABLE_ON_ERROR STREAMING\nCONNECT ENABLED SYNCHRONOUS_COMMIT\nCOPY_DATA ORIGIN TWO_PHASE\nCREATE_SLOT SLOT_NAME\ntest_sub=#\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 4 Jul 2022 08:29:03 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 4, 2022 at 3:59 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Mon, Jul 4, 2022 at 12:59 AM vignesh C <vignesh21@gmail.com> wrote:\n> ...\n> > > 2.\n> > > /* ALTER SUBSCRIPTION <name> SET ( */\n> > > else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\n> > > TailMatches(\"SET\", \"(\"))\n> > > - COMPLETE_WITH(\"binary\", \"slot_name\", \"streaming\",\n> > > \"synchronous_commit\", \"disable_on_error\");\n> > > + COMPLETE_WITH(\"binary\", \"origin\", \"slot_name\", \"streaming\",\n> > > \"synchronous_commit\", \"disable_on_error\");\n> > > /* ALTER SUBSCRIPTION <name> SKIP ( */\n> > > else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\n> > > TailMatches(\"SKIP\", \"(\"))\n> > > COMPLETE_WITH(\"lsn\");\n> > > @@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\n> > > /* Complete \"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\" */\n> > > else if (HeadMatches(\"CREATE\", \"SUBSCRIPTION\") && TailMatches(\"WITH\", \"(\"))\n> > > COMPLETE_WITH(\"binary\", \"connect\", \"copy_data\", \"create_slot\",\n> > > - \"enabled\", \"slot_name\", \"streaming\",\n> > > + \"enabled\", \"origin\", \"slot_name\", \"streaming\",\n> > > \"synchronous_commit\", \"two_phase\", \"disable_on_error\");\n> > >\n> > > Why do you choose to add a new option in-between other parameters\n> > > instead of at the end which we normally do? The one possible reason I\n> > > can think of is that all the parameters at the end are boolean so you\n> > > want to add this before those but then why before slot_name, and again\n> > > I don't see such a rule being followed for other parameters.\n> >\n> > I was not sure if it should be maintained in alphabetical order,\n> > anyway since the last option \"disable_on_error\" is at the end, I have\n> > changed it to the end.\n> >\n>\n> Although it seems it is not a hard rule, mostly the COMPLETE_WITH are\n> coded using alphabetical order. Anyway, I think that was a clear\n> intention here too since 13 of 14 parameters were already in\n> alphabetical order; it is actually only that \"disable_on_error\"\n> parameter that was misplaced; not the new \"origin\" parameter.\n>\n\nAgreed, but let's not change disable_on_error as part of this patch.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 4 Jul 2022 09:22:58 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sun, Jul 3, 2022 11:00 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> Thanks for the comments, the attached v27 patch has the changes for the\r\n> same.\r\n> \r\n\r\nThanks for updating the patch.\r\n\r\nA comment on 0003 patch:\r\n+\t\t/*\r\n+\t\t * No need to throw an error for the tables that are in ready state,\r\n+\t\t * as the walsender will send the changes from WAL in case of tables\r\n+\t\t * in ready state.\r\n+\t\t */\r\n+\t\tif (isreadytable)\r\n+\t\t\tcontinue;\r\n+\r\n...\r\n+\t\tereport(ERROR,\r\n+\t\t\t\terrcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\r\n+\t\t\t\terrmsg(\"table: \\\"%s.%s\\\" might have replicated data in the publisher\",\r\n+\t\t\t\t\t nspname, relname),\r\n+\t\t\t\terrdetail(\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data = on is not allowed when the publisher might have replicated data.\"),\r\n+\t\t\t\terrhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\"));\r\n+\r\n+\t\tExecClearTuple(slot);\r\n\r\nI think we should call ExecClearTuple() before getting next tuple, so it should\r\nbe called if the table is in ready state. How about modifying it to:\r\n\t\tif (!isreadytable)\r\n\t\t\tereport(ERROR,\r\n\t\t\t\t\terrcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\r\n\t\t\t\t\terrmsg(\"table: \\\"%s.%s\\\" might have replicated data in the publisher\",\r\n\t\t\t\t\t\t nspname, relname),\r\n\t\t\t\t\terrdetail(\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data = on is not allowed when the publisher might have replicated data.\"),\r\n\t\t\t\t\terrhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\"));\r\n\r\n\t\tExecClearTuple(slot);\r\n\r\nRegards,\r\nShi yu\r\n\r\n","msg_date":"Mon, 4 Jul 2022 08:16:45 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 4, 2022 4:17 PM shiy.fnst@fujitsu.com <shiy.fnst@fujitsu.com> wrote:\r\n> \r\n> On Sun, Jul 3, 2022 11:00 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> >\r\n> > Thanks for the comments, the attached v27 patch has the changes for the\r\n> > same.\r\n> >\r\n> \r\n> Thanks for updating the patch.\r\n> \r\n> A comment on 0003 patch:\r\n> \r\n> I think we should call ExecClearTuple() before getting next tuple, so it should\r\n> be called if the table is in ready state. How about modifying it to:\r\n> \r\n\r\nBy the way, I have tested pg_dump/psql changes in this patch with older version\r\n(server: pg10 ~ pg15, pg_dump/psql: pg16), and it worked ok.\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Mon, 4 Jul 2022 08:41:17 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sun, Jul 3, 2022 at 8:29 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n\nReview comments\n===============\n1.\n@@ -530,7 +557,7 @@ CreateSubscription(ParseState *pstate,\nCreateSubscriptionStmt *stmt,\n SUBOPT_SLOT_NAME | SUBOPT_COPY_DATA |\n SUBOPT_SYNCHRONOUS_COMMIT | SUBOPT_BINARY |\n SUBOPT_STREAMING | SUBOPT_TWOPHASE_COMMIT |\n- SUBOPT_DISABLE_ON_ERR);\n+ SUBOPT_DISABLE_ON_ERR | SUBOPT_ORIGIN);\n parse_subscription_options(pstate, stmt->options, supported_opts, &opts);\n\n /*\n@@ -606,6 +633,8 @@ CreateSubscription(ParseState *pstate,\nCreateSubscriptionStmt *stmt,\n LOGICALREP_TWOPHASE_STATE_PENDING :\n LOGICALREP_TWOPHASE_STATE_DISABLED);\n values[Anum_pg_subscription_subdisableonerr - 1] =\nBoolGetDatum(opts.disableonerr);\n+ values[Anum_pg_subscription_suborigin - 1] =\n+ CStringGetTextDatum(opts.origin);\n values[Anum_pg_subscription_subconninfo - 1] =\n CStringGetTextDatum(conninfo);\n if (opts.slot_name)\n...\n...\n\n /* List of publications subscribed to */\n text subpublications[1] BKI_FORCE_NOT_NULL;\n+\n+ /* Only publish data originating from the specified origin */\n+ text suborigin BKI_DEFAULT(LOGICALREP_ORIGIN_ANY);\n #endif\n } FormData_pg_subscription;\n\nThe order of declaration and assignment for 'suborigin' should match\nin above usage.\n\n2. Similarly the changes in GetSubscription() should also match the\ndeclaration of the origin column.\n\n3.\n GRANT SELECT (oid, subdbid, subskiplsn, subname, subowner, subenabled,\n- subbinary, substream, subtwophasestate,\nsubdisableonerr, subslotname,\n- subsynccommit, subpublications)\n+ subbinary, substream, subtwophasestate, subdisableonerr,\n+ suborigin, subslotname, subsynccommit, subpublications)\n ON pg_subscription TO public;\n\nThis should also match the order of columns as in pg_subscription.h\nunless there is a reason for not doing so.\n\n4.\n+ /*\n+ * Even though \"origin\" parameter allows only \"local\" and \"any\"\n+ * values, it is implemented as a string type so that the parameter\n+ * can be extended in future versions to support filtering using\n+ * origin names specified by the user.\n\n/Even though \"origin\" .../Even though the \"origin\" parameter ...\n\n5.\n+\n+# Create tables on node_A\n+$node_A->safe_psql('postgres', \"CREATE TABLE tab_full (a int PRIMARY KEY)\");\n+\n+# Create the same tables on node_B\n+$node_B->safe_psql('postgres', \"CREATE TABLE tab_full (a int PRIMARY KEY)\");\n\nIn both the above comments, you should use table instead of tables as\nthe test creates only one table.\n\n6.\n+$node_A->safe_psql('postgres', \"DELETE FROM tab_full;\");\n\nAfter this, the test didn't ensure that this operation is replicated.\nCan't that lead to unpredictable results for the other tests after\nthis test?\n\n7.\n+# Setup logical replication\n+# node_C (pub) -> node_B (sub)\n+my $node_C_connstr = $node_C->connstr . ' dbname=postgres';\n+$node_C->safe_psql('postgres',\n+ \"CREATE PUBLICATION tap_pub_C FOR TABLE tab_full\");\n+\n+my $appname_B2 = 'tap_sub_B2';\n+$node_B->safe_psql(\n+ 'postgres', \"\n+ CREATE SUBSCRIPTION tap_sub_B2\n+ CONNECTION '$node_C_connstr application_name=$appname_B2'\n+ PUBLICATION tap_pub_C\n+ WITH (origin = local)\");\n+\n+$node_C->wait_for_catchup($appname_B2);\n+\n+$node_C->poll_query_until('postgres', $synced_query)\n+ or die \"Timed out while waiting for subscriber to synchronize data\";\n\nThis test allows the publisher (node_C) to poll for sync but it should\nbe the subscriber (node_B) that needs to poll to allow the initial\nsync to finish.\n\n8. Do you think it makes sense to see if this new option can also be\nsupported by pg_recvlogical? I see that previously we have not\nextended pg_recvlogical for all the newly added options but I feel we\nshould keep pg_recvlogical up to date w.r.t new options. We can do\nthis as a separate patch if we agree?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 4 Jul 2022 15:48:33 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"> From d8f8844f877806527b6f3f45320b6ba55a8e3154 Mon Sep 17 00:00:00 2001\n> From: Vigneshwaran C <vignesh21@gmail.com>\n> Date: Thu, 26 May 2022 19:29:33 +0530\n> Subject: [PATCH v27 1/4] Add a missing test to verify only-local parameter in\n> test_decoding plugin.\n> \n> Add a missing test to verify only-local parameter in test_decoding plugin.\n\nI don't get it. replorigin.sql already has some lines to test\nlocal-only. What is your patch adding that is new? Maybe instead of\nadding some more lines at the end of the script, you should add lines\nwhere this stuff is already being tested. But that assumes that there\nis something new that is being tested; if so what is it?\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\"Investigación es lo que hago cuando no sé lo que estoy haciendo\"\n(Wernher von Braun)\n\n\n","msg_date":"Mon, 4 Jul 2022 16:14:02 +0200","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 4, 2022 at 3:48 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Sun, Jul 3, 2022 at 8:29 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n>\n> Review comments\n> ===============\n> 1.\n> @@ -530,7 +557,7 @@ CreateSubscription(ParseState *pstate,\n> CreateSubscriptionStmt *stmt,\n> SUBOPT_SLOT_NAME | SUBOPT_COPY_DATA |\n> SUBOPT_SYNCHRONOUS_COMMIT | SUBOPT_BINARY |\n> SUBOPT_STREAMING | SUBOPT_TWOPHASE_COMMIT |\n> - SUBOPT_DISABLE_ON_ERR);\n> + SUBOPT_DISABLE_ON_ERR | SUBOPT_ORIGIN);\n> parse_subscription_options(pstate, stmt->options, supported_opts, &opts);\n>\n> /*\n> @@ -606,6 +633,8 @@ CreateSubscription(ParseState *pstate,\n> CreateSubscriptionStmt *stmt,\n> LOGICALREP_TWOPHASE_STATE_PENDING :\n> LOGICALREP_TWOPHASE_STATE_DISABLED);\n> values[Anum_pg_subscription_subdisableonerr - 1] =\n> BoolGetDatum(opts.disableonerr);\n> + values[Anum_pg_subscription_suborigin - 1] =\n> + CStringGetTextDatum(opts.origin);\n> values[Anum_pg_subscription_subconninfo - 1] =\n> CStringGetTextDatum(conninfo);\n> if (opts.slot_name)\n> ...\n> ...\n>\n> /* List of publications subscribed to */\n> text subpublications[1] BKI_FORCE_NOT_NULL;\n> +\n> + /* Only publish data originating from the specified origin */\n> + text suborigin BKI_DEFAULT(LOGICALREP_ORIGIN_ANY);\n> #endif\n> } FormData_pg_subscription;\n>\n> The order of declaration and assignment for 'suborigin' should match\n> in above usage.\n\nModified\n\n> 2. Similarly the changes in GetSubscription() should also match the\n> declaration of the origin column.\n\nModified\n\n> 3.\n> GRANT SELECT (oid, subdbid, subskiplsn, subname, subowner, subenabled,\n> - subbinary, substream, subtwophasestate,\n> subdisableonerr, subslotname,\n> - subsynccommit, subpublications)\n> + subbinary, substream, subtwophasestate, subdisableonerr,\n> + suborigin, subslotname, subsynccommit, subpublications)\n> ON pg_subscription TO public;\n>\n> This should also match the order of columns as in pg_subscription.h\n> unless there is a reason for not doing so.\n\nModified\n\n> 4.\n> + /*\n> + * Even though \"origin\" parameter allows only \"local\" and \"any\"\n> + * values, it is implemented as a string type so that the parameter\n> + * can be extended in future versions to support filtering using\n> + * origin names specified by the user.\n>\n> /Even though \"origin\" .../Even though the \"origin\" parameter ...\n\nModified\n\n> 5.\n> +\n> +# Create tables on node_A\n> +$node_A->safe_psql('postgres', \"CREATE TABLE tab_full (a int PRIMARY KEY)\");\n> +\n> +# Create the same tables on node_B\n> +$node_B->safe_psql('postgres', \"CREATE TABLE tab_full (a int PRIMARY KEY)\");\n>\n> In both the above comments, you should use table instead of tables as\n> the test creates only one table.\n\nModified\n\n> 6.\n> +$node_A->safe_psql('postgres', \"DELETE FROM tab_full;\");\n>\n> After this, the test didn't ensure that this operation is replicated.\n> Can't that lead to unpredictable results for the other tests after\n> this test?\n\nModified to include wait_for_catchup and verified the data at the\nbeginning of the next test\n\n> 7.\n> +# Setup logical replication\n> +# node_C (pub) -> node_B (sub)\n> +my $node_C_connstr = $node_C->connstr . ' dbname=postgres';\n> +$node_C->safe_psql('postgres',\n> + \"CREATE PUBLICATION tap_pub_C FOR TABLE tab_full\");\n> +\n> +my $appname_B2 = 'tap_sub_B2';\n> +$node_B->safe_psql(\n> + 'postgres', \"\n> + CREATE SUBSCRIPTION tap_sub_B2\n> + CONNECTION '$node_C_connstr application_name=$appname_B2'\n> + PUBLICATION tap_pub_C\n> + WITH (origin = local)\");\n> +\n> +$node_C->wait_for_catchup($appname_B2);\n> +\n> +$node_C->poll_query_until('postgres', $synced_query)\n> + or die \"Timed out while waiting for subscriber to synchronize data\";\n>\n> This test allows the publisher (node_C) to poll for sync but it should\n> be the subscriber (node_B) that needs to poll to allow the initial\n> sync to finish.\n\nModified\n\n> 8. Do you think it makes sense to see if this new option can also be\n> supported by pg_recvlogical? I see that previously we have not\n> extended pg_recvlogical for all the newly added options but I feel we\n> should keep pg_recvlogical up to date w.r.t new options. We can do\n> this as a separate patch if we agree?\n\nI will analyze this and post my analysis soon.\n\nThanks for the comments, the v28 patch attached has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Mon, 4 Jul 2022 23:29:27 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 4, 2022 at 1:46 PM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Sun, Jul 3, 2022 11:00 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached v27 patch has the changes for the\n> > same.\n> >\n>\n> Thanks for updating the patch.\n>\n> A comment on 0003 patch:\n> + /*\n> + * No need to throw an error for the tables that are in ready state,\n> + * as the walsender will send the changes from WAL in case of tables\n> + * in ready state.\n> + */\n> + if (isreadytable)\n> + continue;\n> +\n> ...\n> + ereport(ERROR,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"table: \\\"%s.%s\\\" might have replicated data in the publisher\",\n> + nspname, relname),\n> + errdetail(\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data = on is not allowed when the publisher might have replicated data.\"),\n> + errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\"));\n> +\n> + ExecClearTuple(slot);\n>\n> I think we should call ExecClearTuple() before getting next tuple, so it should\n> be called if the table is in ready state. How about modifying it to:\n> if (!isreadytable)\n> ereport(ERROR,\n> errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> errmsg(\"table: \\\"%s.%s\\\" might have replicated data in the publisher\",\n> nspname, relname),\n> errdetail(\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data = on is not allowed when the publisher might have replicated data.\"),\n> errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\"));\n>\n> ExecClearTuple(slot);\n\nThanks for the comment, I have modified this in the v28 patch attached at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm3MNK2hMYroTiHGS9HkSxiA-az1QC1mpa0YwDZ8nGmmZg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 4 Jul 2022 23:30:53 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 4, 2022 at 9:23 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Jul 4, 2022 at 3:59 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > On Mon, Jul 4, 2022 at 12:59 AM vignesh C <vignesh21@gmail.com> wrote:\n> > ...\n> > > > 2.\n> > > > /* ALTER SUBSCRIPTION <name> SET ( */\n> > > > else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\n> > > > TailMatches(\"SET\", \"(\"))\n> > > > - COMPLETE_WITH(\"binary\", \"slot_name\", \"streaming\",\n> > > > \"synchronous_commit\", \"disable_on_error\");\n> > > > + COMPLETE_WITH(\"binary\", \"origin\", \"slot_name\", \"streaming\",\n> > > > \"synchronous_commit\", \"disable_on_error\");\n> > > > /* ALTER SUBSCRIPTION <name> SKIP ( */\n> > > > else if (HeadMatches(\"ALTER\", \"SUBSCRIPTION\", MatchAny) &&\n> > > > TailMatches(\"SKIP\", \"(\"))\n> > > > COMPLETE_WITH(\"lsn\");\n> > > > @@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\n> > > > /* Complete \"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\" */\n> > > > else if (HeadMatches(\"CREATE\", \"SUBSCRIPTION\") && TailMatches(\"WITH\", \"(\"))\n> > > > COMPLETE_WITH(\"binary\", \"connect\", \"copy_data\", \"create_slot\",\n> > > > - \"enabled\", \"slot_name\", \"streaming\",\n> > > > + \"enabled\", \"origin\", \"slot_name\", \"streaming\",\n> > > > \"synchronous_commit\", \"two_phase\", \"disable_on_error\");\n> > > >\n> > > > Why do you choose to add a new option in-between other parameters\n> > > > instead of at the end which we normally do? The one possible reason I\n> > > > can think of is that all the parameters at the end are boolean so you\n> > > > want to add this before those but then why before slot_name, and again\n> > > > I don't see such a rule being followed for other parameters.\n> > >\n> > > I was not sure if it should be maintained in alphabetical order,\n> > > anyway since the last option \"disable_on_error\" is at the end, I have\n> > > changed it to the end.\n> > >\n> >\n> > Although it seems it is not a hard rule, mostly the COMPLETE_WITH are\n> > coded using alphabetical order. Anyway, I think that was a clear\n> > intention here too since 13 of 14 parameters were already in\n> > alphabetical order; it is actually only that \"disable_on_error\"\n> > parameter that was misplaced; not the new \"origin\" parameter.\n> >\n>\n> Agreed, but let's not change disable_on_error as part of this patch.\n\nI have changed this in the v28 patch attached at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm3MNK2hMYroTiHGS9HkSxiA-az1QC1mpa0YwDZ8nGmmZg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 4 Jul 2022 23:31:55 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 4, 2022 at 7:44 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n>\n> > From d8f8844f877806527b6f3f45320b6ba55a8e3154 Mon Sep 17 00:00:00 2001\n> > From: Vigneshwaran C <vignesh21@gmail.com>\n> > Date: Thu, 26 May 2022 19:29:33 +0530\n> > Subject: [PATCH v27 1/4] Add a missing test to verify only-local parameter in\n> > test_decoding plugin.\n> >\n> > Add a missing test to verify only-local parameter in test_decoding plugin.\n>\n> I don't get it. replorigin.sql already has some lines to test\n> local-only. What is your patch adding that is new? Maybe instead of\n> adding some more lines at the end of the script, you should add lines\n> where this stuff is already being tested. But that assumes that there\n> is something new that is being tested; if so what is it?\n\nThe test is to check that remote origin data (i.e. replication origin\nbeing set) will be filtered when only-local parameter is set. I felt\nthat this scenario is not covered, unless I'm missing something. I\nhave moved the test as suggested.\nI have changed this in the v28 patch attached at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm3MNK2hMYroTiHGS9HkSxiA-az1QC1mpa0YwDZ8nGmmZg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 4 Jul 2022 23:33:13 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"I checked again the v26* patch set.\n\nI had no more comments for v26-0001, v26-0002, or v26-0004, but below\nare some comments for v26-0003.\n\n========\nv26-0003\n========\n\n3.1 src/backend/catalog/pg_subscription.c\n\n3.1.a\n+/*\n+ * Get all relations for subscription that are in a ready state.\n+ *\n+ * Returned list is palloc'ed in current memory context.\n+ */\n+List *\n+GetSubscriptionReadyRelations(Oid subid)\n\n\"subscription\" -> \"the subscription\"\n\nAlso, It might be better to say in the function header what kind of\nstructures are in the returned List. E.g. Firstly, I'd assumed it was\nthe same return type as the other function\nGetSubscriptionReadyRelations, but it isn’t.\n\n3.1.b\nI think there might be a case to be made for *combining* those\nSubscriptionRelState and SubscripotionRel structs into a single common\nstruct. Then all those GetSubscriptionNotReadyRelations and\nGetSubscriptionNotReadyRelations (also GetSubscriptionRelations?) can\nbe merged to be just one common function that takes a parameter to say\ndo you want to return the relation List of ALL / READY / NOT_READY?\nWhat do you think?\n\n======\n\n3.2 src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n\n+/*\n+ * Check and throw an error if the publisher has subscribed to the same table\n+ * from some other publisher. This check is required only if copydata is ON and\n+ * the origin is local.\n+ */\n+static void\n+check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\n+ CopyData copydata, char *origin, Oid subid)\n\nThe function comment probably should say something about why the new\nREADY logic was added in v26.\n\n~~~\n\n3.3\n\n3.3.a\n+ /*\n+ * The subid will be valid only for ALTER SUBSCRIPTION ... REFRESH\n+ * PUBLICATION. Get the ready relations for the subscription only in case\n+ * of ALTER SUBSCRIPTION case as there will be no relations in ready state\n+ * while the subscription is created.\n+ */\n+ if (subid != InvalidOid)\n+ subreadyrels = GetSubscriptionReadyRelations(subid);\n\nThe word \"case\" is 2x in the same sentence. I also paraphrased my\nunderstanding of this comment below. Maybe it is simpler?\n\nSUGGESTION\nGet the ready relations for the subscription. The subid will be valid\nonly for ALTER SUBSCRIPTION ... REFRESH because there will be no\nrelations in ready state while the subscription is created.\n\n3.3.b\n+ if (subid != InvalidOid)\n+ subreadyrels = GetSubscriptionReadyRelations(subid);\n\nSUGGESTION\nif (OidIsValid(subid))\n\n======\n\n3.4 src/test/subscription/t/032_localonly.pl\n\nNow all the test cases are re-using the same data (e.g. 11,12,13) and\nyou are deleting the data between the tests. I guess it is OK, but IMO\nthe tests are easier to read when each test part was using unique data\n(e.g. 11,12,13, then 12,22,32, then 13,23,33 etc)\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 5 Jul 2022 10:44:32 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 4, 2022 at 11:33 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Jul 4, 2022 at 7:44 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> >\n> > > From d8f8844f877806527b6f3f45320b6ba55a8e3154 Mon Sep 17 00:00:00 2001\n> > > From: Vigneshwaran C <vignesh21@gmail.com>\n> > > Date: Thu, 26 May 2022 19:29:33 +0530\n> > > Subject: [PATCH v27 1/4] Add a missing test to verify only-local parameter in\n> > > test_decoding plugin.\n> > >\n> > > Add a missing test to verify only-local parameter in test_decoding plugin.\n> >\n> > I don't get it. replorigin.sql already has some lines to test\n> > local-only. What is your patch adding that is new? Maybe instead of\n> > adding some more lines at the end of the script, you should add lines\n> > where this stuff is already being tested. But that assumes that there\n> > is something new that is being tested; if so what is it?\n>\n> The test is to check that remote origin data (i.e. replication origin\n> being set) will be filtered when only-local parameter is set. I felt\n> that this scenario is not covered, unless I'm missing something.\n>\n\nIf we change only-local option to '0' in the below part of the test,\nwe can see a different result:\n-- and magically the replayed xact will be filtered!\nSELECT data FROM pg_logical_slot_get_changes('regression_slot', NULL,\nNULL, 'include-xids', '0', 'skip-empty-xacts', '1', 'only-local',\n'1');\n\nSo, I think only-local functionality is being tested but I feel your\ntest is clear in the sense that it clearly shows that remote data can\nonly be fetched with 'only-local' as '0' when the origin is set. It\nseems to me that in existing tests, we are not testing the combination\nwhere the origin is set and 'only-local' is '0'. So, there is some\nvalue to having the test you are proposing but if Alvaro, you, or\nothers don't think it is worth adding a new test for it then we can\nprobably drop this one.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 5 Jul 2022 13:59:28 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 5, 2022 at 1:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Jul 4, 2022 at 11:33 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Mon, Jul 4, 2022 at 7:44 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\n> > >\n> > > > From d8f8844f877806527b6f3f45320b6ba55a8e3154 Mon Sep 17 00:00:00 2001\n> > > > From: Vigneshwaran C <vignesh21@gmail.com>\n> > > > Date: Thu, 26 May 2022 19:29:33 +0530\n> > > > Subject: [PATCH v27 1/4] Add a missing test to verify only-local parameter in\n> > > > test_decoding plugin.\n> > > >\n> > > > Add a missing test to verify only-local parameter in test_decoding plugin.\n> > >\n> > > I don't get it. replorigin.sql already has some lines to test\n> > > local-only. What is your patch adding that is new? Maybe instead of\n> > > adding some more lines at the end of the script, you should add lines\n> > > where this stuff is already being tested. But that assumes that there\n> > > is something new that is being tested; if so what is it?\n> >\n> > The test is to check that remote origin data (i.e. replication origin\n> > being set) will be filtered when only-local parameter is set. I felt\n> > that this scenario is not covered, unless I'm missing something.\n> >\n>\n> If we change only-local option to '0' in the below part of the test,\n> we can see a different result:\n> -- and magically the replayed xact will be filtered!\n> SELECT data FROM pg_logical_slot_get_changes('regression_slot', NULL,\n> NULL, 'include-xids', '0', 'skip-empty-xacts', '1', 'only-local',\n> '1');\n>\n> So, I think only-local functionality is being tested but I feel your\n> test is clear in the sense that it clearly shows that remote data can\n> only be fetched with 'only-local' as '0' when the origin is set. It\n> seems to me that in existing tests, we are not testing the combination\n> where the origin is set and 'only-local' is '0'. So, there is some\n> value to having the test you are proposing but if Alvaro, you, or\n> others don't think it is worth adding a new test for it then we can\n> probably drop this one.\n\nSince the existing test is already handling the verification of this\nscenario, I felt no need to add the test. Updated v29 patch removes\nthe 0001 patch which had the test case.\n\nRegards,\nVignesh","msg_date":"Tue, 5 Jul 2022 21:33:18 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 5, 2022 at 6:14 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> I checked again the v26* patch set.\n>\n> I had no more comments for v26-0001, v26-0002, or v26-0004, but below\n> are some comments for v26-0003.\n>\n> ========\n> v26-0003\n> ========\n>\n> 3.1 src/backend/catalog/pg_subscription.c\n>\n> 3.1.a\n> +/*\n> + * Get all relations for subscription that are in a ready state.\n> + *\n> + * Returned list is palloc'ed in current memory context.\n> + */\n> +List *\n> +GetSubscriptionReadyRelations(Oid subid)\n>\n> \"subscription\" -> \"the subscription\"\n>\n> Also, It might be better to say in the function header what kind of\n> structures are in the returned List. E.g. Firstly, I'd assumed it was\n> the same return type as the other function\n> GetSubscriptionReadyRelations, but it isn’t.\n\nThis function has been removed and GetSubscriptionRelations is used\nwith slight changes.\n\n> 3.1.b\n> I think there might be a case to be made for *combining* those\n> SubscriptionRelState and SubscripotionRel structs into a single common\n> struct. Then all those GetSubscriptionNotReadyRelations and\n> GetSubscriptionNotReadyRelations (also GetSubscriptionRelations?) can\n> be merged to be just one common function that takes a parameter to say\n> do you want to return the relation List of ALL / READY / NOT_READY?\n> What do you think?\n\nI have used a common function that can get all relations, ready\nrelations and not ready relations instead of the 3 functions.\n\n> ======\n>\n> 3.2 src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> +/*\n> + * Check and throw an error if the publisher has subscribed to the same table\n> + * from some other publisher. This check is required only if copydata is ON and\n> + * the origin is local.\n> + */\n> +static void\n> +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\n> + CopyData copydata, char *origin, Oid subid)\n>\n> The function comment probably should say something about why the new\n> READY logic was added in v26.\n\nAdded a comment for the same.\n\n> ~~~\n>\n> 3.3\n>\n> 3.3.a\n> + /*\n> + * The subid will be valid only for ALTER SUBSCRIPTION ... REFRESH\n> + * PUBLICATION. Get the ready relations for the subscription only in case\n> + * of ALTER SUBSCRIPTION case as there will be no relations in ready state\n> + * while the subscription is created.\n> + */\n> + if (subid != InvalidOid)\n> + subreadyrels = GetSubscriptionReadyRelations(subid);\n>\n> The word \"case\" is 2x in the same sentence. I also paraphrased my\n> understanding of this comment below. Maybe it is simpler?\n>\n> SUGGESTION\n> Get the ready relations for the subscription. The subid will be valid\n> only for ALTER SUBSCRIPTION ... REFRESH because there will be no\n> relations in ready state while the subscription is created.\n\nModified\n\n> 3.3.b\n> + if (subid != InvalidOid)\n> + subreadyrels = GetSubscriptionReadyRelations(subid);\n>\n> SUGGESTION\n> if (OidIsValid(subid))\n\nModified\n\n> ======\n>\n> 3.4 src/test/subscription/t/032_localonly.pl\n>\n> Now all the test cases are re-using the same data (e.g. 11,12,13) and\n> you are deleting the data between the tests. I guess it is OK, but IMO\n> the tests are easier to read when each test part was using unique data\n> (e.g. 11,12,13, then 12,22,32, then 13,23,33 etc)\n\nModified\n\nThanks for the comments, the v29 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm1T5utq60qVx%3DRN60rHcg7wt2psM7PpCQ2fDiB-R8oLGg%40mail.gmail.com\n\nRegards,\n Vignesh\n\n\n","msg_date":"Tue, 5 Jul 2022 21:36:17 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 5, 2022 at 9:33 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> Since the existing test is already handling the verification of this\n> scenario, I felt no need to add the test. Updated v29 patch removes\n> the 0001 patch which had the test case.\n>\n\nI have again looked at the first and it looks good to me. I would like\nto push it after some more review but before that, I would like to\nknow if someone else has any suggestions/objections to this patch, so\nlet me summarize the idea of the first patch. It will allow users to\nskip the replication of remote data. Here remote data is the data that\nthe publisher node has received from some other node. The primary use\ncase is to avoid loops (infinite replication of the same data) among\nnodes as shown in the initial email of this thread.\n\nTo achieve that this patch adds a new SUBSCRIPTION parameter \"origin\".\nIt specifies whether the subscription will request the publisher to\nonly send changes that originated locally or to send changes\nregardless of origin. Setting it to \"local\" means that the\nsubscription will request the publisher to only send changes that\noriginated locally. Setting it to \"any\" means that the publisher sends\nchanges regardless of their origin. The default is \"any\".\n\nUsage:\nCREATE SUBSCRIPTION sub1 CONNECTION 'dbname=postgres port=9999'\nPUBLICATION pub1 WITH (origin = local);\n\nFor now, even though the \"origin\" parameter allows only \"local\" and\n\"any\" values, it is implemented as a string type so that the parameter\ncan be extended in future versions to support filtering using origin\nnames specified by the user.\n\nThis feature allows filtering only the replication data originated\nfrom WAL but for initial sync (initial copy of table data) we don't\nhave such a facility as we can only distinguish the data based on\norigin from WAL. As a separate patch (v29-0002*), we are planning to\nforbid the initial sync if we notice that the publication tables were\nalso replicated from other publishers to avoid duplicate data or\nloops. They will be allowed to copy with the 'force' option in such\ncases.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 6 Jul 2022 17:09:30 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 5, 2022 at 9:33 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> Since the existing test is already handling the verification of this\n> scenario, I felt no need to add the test. Updated v29 patch removes\n> the 0001 patch which had the test case.\n>\n\nI am not able to apply 0001.\npatching file src/bin/psql/tab-complete.c\nHunk #1 FAILED at 1873.\nHunk #2 FAILED at 3152.\n\nFew comments on 0002\n=====================\n1.\n+ <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n\nIs there a need for space before / in above? If not, please remove it\nwith similar extra space from other similar usages.\n\n2.\n+ <para>\n+ There is some interaction between the <literal>origin</literal>\n+ parameter and the <literal>copy_data</literal> parameter. Refer to\n+ the <command>CREATE SUBSCRIPTION</command>\n+ <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n+ details and usage of <literal>force</literal> for\n+ <literal>copy_data</literal> parameter.\n </para>\n\nI think this is bit long. We can try to reduce this by something like:\nRefer <xref linkend=\"sql-createsubscription-notes\"/> for the usage of\n<literal>force</literal> option and its interaction with the\n<literal>origin</literal> parameter.\n\nAlso, adopt the same other places if you agree with the above change.\n\n4.\n@@ -601,16 +549,28 @@ GetSubscriptionNotReadyRelations(Oid subid)\n SysScanDesc scan;\n\n rel = table_open(SubscriptionRelRelationId, AccessShareLock);\n-\n ScanKeyInit(&skey[nkeys++],\n Anum_pg_subscription_rel_srsubid,\n\nSpurious line removal.\n\n5.\n+static void\n+check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\n+ CopyData copydata, char *origin, Oid subid)\n{\n...\n...\n+ /*\n+ * Get the ready relations for the subscription. The subid will be valid\n+ * only for ALTER SUBSCRIPTION ... REFRESH because there will be no\n+ * relations in ready state while the subscription is created.\n+ */\n+ if (OidIsValid(subid))\n+ subreadyrels = GetSubscriptionRelations(subid, READY_STATE);\n\nWhy do we want to consider only READY_STATE relations here? If you see\nits caller AlterSubscription_refresh(), we don't consider copying the\nrelation if it exists on subscribers in any state. If my observation\nis correct then you probably don't need to introduce SubRelStateType.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 8 Jul 2022 16:38:57 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jul 8, 2022 at 4:39 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Jul 5, 2022 at 9:33 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Since the existing test is already handling the verification of this\n> > scenario, I felt no need to add the test. Updated v29 patch removes\n> > the 0001 patch which had the test case.\n> >\n>\n> I am not able to apply 0001.\n> patching file src/bin/psql/tab-complete.c\n> Hunk #1 FAILED at 1873.\n> Hunk #2 FAILED at 3152.\n\nI have rebased the patch on top of HEAD.\n\n> Few comments on 0002\n> =====================\n> 1.\n> + <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n>\n> Is there a need for space before / in above? If not, please remove it\n> with similar extra space from other similar usages.\n\nModified\n\n> 2.\n> + <para>\n> + There is some interaction between the <literal>origin</literal>\n> + parameter and the <literal>copy_data</literal> parameter. Refer to\n> + the <command>CREATE SUBSCRIPTION</command>\n> + <xref linkend=\"sql-createsubscription-notes\" /> for interaction\n> + details and usage of <literal>force</literal> for\n> + <literal>copy_data</literal> parameter.\n> </para>\n>\n> I think this is bit long. We can try to reduce this by something like:\n> Refer <xref linkend=\"sql-createsubscription-notes\"/> for the usage of\n> <literal>force</literal> option and its interaction with the\n> <literal>origin</literal> parameter.\n>\n> Also, adopt the same other places if you agree with the above change.\n\nModified with slight rewording.\n\n> 4.\n> @@ -601,16 +549,28 @@ GetSubscriptionNotReadyRelations(Oid subid)\n> SysScanDesc scan;\n>\n> rel = table_open(SubscriptionRelRelationId, AccessShareLock);\n> -\n> ScanKeyInit(&skey[nkeys++],\n> Anum_pg_subscription_rel_srsubid,\n>\n> Spurious line removal.\n\nI have removed this\n\n> 5.\n> +static void\n> +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\n> + CopyData copydata, char *origin, Oid subid)\n> {\n> ...\n> ...\n> + /*\n> + * Get the ready relations for the subscription. The subid will be valid\n> + * only for ALTER SUBSCRIPTION ... REFRESH because there will be no\n> + * relations in ready state while the subscription is created.\n> + */\n> + if (OidIsValid(subid))\n> + subreadyrels = GetSubscriptionRelations(subid, READY_STATE);\n>\n> Why do we want to consider only READY_STATE relations here? If you see\n> its caller AlterSubscription_refresh(), we don't consider copying the\n> relation if it exists on subscribers in any state. If my observation\n> is correct then you probably don't need to introduce SubRelStateType.\n\nI have changed it similar to the caller function and removed SubRelStateType\n\nThanks for the comments, the v30 patch attached has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Sat, 9 Jul 2022 20:10:56 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for the v30* patches:\n\n========\nv30-0001\n========\n\n1.1 <general>\n\nI was wondering if it is better to implement a new defGetOrigin method\nnow instead of just using the defGetString to process the 'origin',\nsince you may need to do that in future anyway if the 'origin' name is\nplanned to become specifiable by the user. OTOH maybe you prefer to\nchange this code later when the time comes. I am not sure what way is\nbest.\n\n~~~\n\n1.2. src/include/replication/walreceiver.h\n\n@@ -183,6 +183,8 @@ typedef struct\n bool streaming; /* Streaming of large transactions */\n bool twophase; /* Streaming of two-phase transactions at\n * prepare time */\n+ char *origin; /* Only publish data originating from the\n+ * specified origin */\n } logical;\n } proto;\n } WalRcvStreamOptions;\n\nShould the new comments be aligned with the other ones?\n\n========\nv30-0002\n========\n\n2.1 doc/src/sgml/ref/alter_subscription.sgml\n\n+ <para>\n+ Refer <xref linkend=\"sql-createsubscription-notes\"/> for the usage of\n+ <literal>force</literal> for <literal>copy_data</literal> parameter\n+ and its interaction with the <literal>origin</literal> parameter.\n </para>\n\nIMO it's better to say \"Refer to the Notes\" or \"Refer to CREATE\nSUBSCRIPTION Notes\" instead of just \"Refer Notes\"\n\n~~~\n\n2.2 doc/src/sgml/ref/create_subscription.sgml\n\n2.2.a\n+ <para>\n+ Refer <xref linkend=\"sql-createsubscription-notes\"/> for the usage of\n+ <literal>force</literal> for <literal>copy_data</literal> parameter\n+ and its interaction with the <literal>origin</literal> parameter.\n+ </para>\n\nIMO it's better to say \"Refer to the Notes\" (same as other xref on\nthis page) instead of \"Refer Notes\"\n\n2.2.b\n@@ -316,6 +324,11 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n publisher sends changes regardless of their origin. The default is\n <literal>any</literal>.\n </para>\n+ <para>\n+ Refer <xref linkend=\"sql-createsubscription-notes\"/> for the usage of\n+ <literal>force</literal> for <literal>copy_data</literal> parameter\n+ and its interaction with the <literal>origin</literal> parameter.\n+ </para>\n\nDitto\n\n~~~\n\n2.3 src/backend/commands/subscriptioncmds.c - DefGetCopyData\n\n+/*\n+ * Validate the value specified for copy_data parameter.\n+ */\n+static CopyData\n+DefGetCopyData(DefElem *def)\n\n~~~\n\n2.4\n\n+ /*\n+ * If no parameter given, assume \"true\" is meant.\n+ */\n\nPlease modify the comment to match the recent push [1].\n\n~~~\n\n2.5 src/test/subscription/t/032_localonly.pl\n\n2.5.a\n+# Check Alter subscription ... refresh publication when there is a new\n+# table that is subscribing data from a different publication\n+$node_A->safe_psql('postgres', \"CREATE TABLE tab_full1 (a int PRIMARY KEY)\");\n+$node_B->safe_psql('postgres', \"CREATE TABLE tab_full1 (a int PRIMARY KEY)\");\n\nUnfortunately, I think tab_full1 is a terrible table name because in\nmy screen font the 'l' and the '1' look exactly the same so it just\nlooks like a typo. Maybe change it to \"tab_new\" or something?\n\n2.5b\nWhat exactly is the purpose of \"full\" in all these test table names?\nAFAIK \"full\" is just some name baggage inherited from completely\ndifferent tests which were doing full versus partial table\nreplication. I'm not sure it is relevant here.\n\n========\nv30-0002\n========\n\nNo comments.\n\n------\n[1] https://github.com/postgres/postgres/commit/8445f5a21d40b969673ca03918c74b4fbc882bf4\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 11 Jul 2022 13:41:26 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sat, Jul 9, 2022 at 8:11 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n\nThanks, a few more comments on v30_0002*\n1.\n+/*\n+ * Represents whether copy_data parameter is specified with off, on or force.\n\nA comma is required after on.\n\n2.\n qsort(subrel_local_oids, list_length(subrel_states),\n sizeof(Oid), oid_cmp);\n\n+ check_pub_table_subscribed(wrconn, sub->publications, copy_data,\n+ sub->origin, subrel_local_oids,\n+ list_length(subrel_states));\n\nWe can avoid using list_length by using an additional variable in this case.\n\n3.\nerrmsg(\"table: \\\"%s.%s\\\" might have replicated data in the publisher\",\n+ nspname, relname),\n\nWhy ':' is used after the table in the above message? I don't see such\na convention at other places in the code. Also, having might in the\nerror messages makes it less clear, so, can we slightly change the\nmessage as in the attached patch?\n\n4. I have made some additional changes in the comments, kindly check\nthe attached and merge those if you are okay.\n\n5.\n+$node_C->safe_psql(\n+ 'postgres', \"\n+ DELETE FROM tab_full\");\n+$node_B->safe_psql(\n+ 'postgres', \"\n+ DELETE FROM tab_full where a = 13\");\n\nDon't we need to wait for these operations to replicate?\n\n-- \nWith Regards,\nAmit Kapila.","msg_date":"Mon, 11 Jul 2022 17:03:34 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 11, 2022 at 9:11 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for the v30* patches:\n>\n> ========\n> v30-0001\n> ========\n>\n> 1.1 <general>\n>\n> I was wondering if it is better to implement a new defGetOrigin method\n> now instead of just using the defGetString to process the 'origin',\n> since you may need to do that in future anyway if the 'origin' name is\n> planned to become specifiable by the user. OTOH maybe you prefer to\n> change this code later when the time comes. I am not sure what way is\n> best.\n\nI preferred to do that change when the feature is getting extended.\n\n> ~~~\n>\n> 1.2. src/include/replication/walreceiver.h\n>\n> @@ -183,6 +183,8 @@ typedef struct\n> bool streaming; /* Streaming of large transactions */\n> bool twophase; /* Streaming of two-phase transactions at\n> * prepare time */\n> + char *origin; /* Only publish data originating from the\n> + * specified origin */\n> } logical;\n> } proto;\n> } WalRcvStreamOptions;\n>\n> Should the new comments be aligned with the other ones?\n\nI kept it like this as pgindent also is aligning it as the current code.\n\n> ========\n> v30-0002\n> ========\n>\n> 2.1 doc/src/sgml/ref/alter_subscription.sgml\n>\n> + <para>\n> + Refer <xref linkend=\"sql-createsubscription-notes\"/> for the usage of\n> + <literal>force</literal> for <literal>copy_data</literal> parameter\n> + and its interaction with the <literal>origin</literal> parameter.\n> </para>\n>\n> IMO it's better to say \"Refer to the Notes\" or \"Refer to CREATE\n> SUBSCRIPTION Notes\" instead of just \"Refer Notes\"\n\nModified to \"Refer to the Notes\"\n\n> ~~~\n>\n> 2.2 doc/src/sgml/ref/create_subscription.sgml\n>\n> 2.2.a\n> + <para>\n> + Refer <xref linkend=\"sql-createsubscription-notes\"/> for the usage of\n> + <literal>force</literal> for <literal>copy_data</literal> parameter\n> + and its interaction with the <literal>origin</literal> parameter.\n> + </para>\n>\n> IMO it's better to say \"Refer to the Notes\" (same as other xref on\n> this page) instead of \"Refer Notes\"\n\nModified to \"Refer to the Notes\"\n\n> 2.2.b\n> @@ -316,6 +324,11 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> publisher sends changes regardless of their origin. The default is\n> <literal>any</literal>.\n> </para>\n> + <para>\n> + Refer <xref linkend=\"sql-createsubscription-notes\"/> for the usage of\n> + <literal>force</literal> for <literal>copy_data</literal> parameter\n> + and its interaction with the <literal>origin</literal> parameter.\n> + </para>\n>\n> Ditto\n\nModified to \"Refer to the Notes\"\n\n> ~~~\n>\n> 2.3 src/backend/commands/subscriptioncmds.c - DefGetCopyData\n>\n> +/*\n> + * Validate the value specified for copy_data parameter.\n> + */\n> +static CopyData\n> +DefGetCopyData(DefElem *def)\n\nChanged it to defGetCopyData to keep the naming similar to others\n\n> ~~~\n>\n> 2.4\n>\n> + /*\n> + * If no parameter given, assume \"true\" is meant.\n> + */\n>\n> Please modify the comment to match the recent push [1].\n\nModified\n\n> ~~~\n>\n> 2.5 src/test/subscription/t/032_localonly.pl\n>\n> 2.5.a\n> +# Check Alter subscription ... refresh publication when there is a new\n> +# table that is subscribing data from a different publication\n> +$node_A->safe_psql('postgres', \"CREATE TABLE tab_full1 (a int PRIMARY KEY)\");\n> +$node_B->safe_psql('postgres', \"CREATE TABLE tab_full1 (a int PRIMARY KEY)\");\n>\n> Unfortunately, I think tab_full1 is a terrible table name because in\n> my screen font the 'l' and the '1' look exactly the same so it just\n> looks like a typo. Maybe change it to \"tab_new\" or something?\n\nModified to tab_new\n\n> 2.5b\n> What exactly is the purpose of \"full\" in all these test table names?\n> AFAIK \"full\" is just some name baggage inherited from completely\n> different tests which were doing full versus partial table\n> replication. I'm not sure it is relevant here.\n\nModified to tab\n\nThanks for the comments, the v31 patch attached has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Tue, 12 Jul 2022 08:42:55 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 11, 2022 at 5:03 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Sat, Jul 9, 2022 at 8:11 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n>\n> Thanks, a few more comments on v30_0002*\n> 1.\n> +/*\n> + * Represents whether copy_data parameter is specified with off, on or force.\n>\n> A comma is required after on.\n\nModified\n\n> 2.\n> qsort(subrel_local_oids, list_length(subrel_states),\n> sizeof(Oid), oid_cmp);\n>\n> + check_pub_table_subscribed(wrconn, sub->publications, copy_data,\n> + sub->origin, subrel_local_oids,\n> + list_length(subrel_states));\n>\n> We can avoid using list_length by using an additional variable in this case.\n\nModified\n\n> 3.\n> errmsg(\"table: \\\"%s.%s\\\" might have replicated data in the publisher\",\n> + nspname, relname),\n>\n> Why ':' is used after the table in the above message? I don't see such\n> a convention at other places in the code. Also, having might in the\n> error messages makes it less clear, so, can we slightly change the\n> message as in the attached patch?\n\nModified as suggested\n\n> 4. I have made some additional changes in the comments, kindly check\n> the attached and merge those if you are okay.\n\nI have merged the changes\n\n> 5.\n> +$node_C->safe_psql(\n> + 'postgres', \"\n> + DELETE FROM tab_full\");\n> +$node_B->safe_psql(\n> + 'postgres', \"\n> + DELETE FROM tab_full where a = 13\");\n>\n> Don't we need to wait for these operations to replicate?\n\nModified to include wait\n\nThanks for the comments, the v31 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm2-D860yULtcmZAzDbdiof-Dg6Y_YaY4owbO6Rj%3DXEHMw%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 12 Jul 2022 08:47:06 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 12, 2022 at 8:43 AM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Jul 11, 2022 at 9:11 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Here are my review comments for the v30* patches:\n> >\n> > ========\n> > v30-0001\n> > ========\n> >\n> > 1.1 <general>\n> >\n> > I was wondering if it is better to implement a new defGetOrigin method\n> > now instead of just using the defGetString to process the 'origin',\n> > since you may need to do that in future anyway if the 'origin' name is\n> > planned to become specifiable by the user. OTOH maybe you prefer to\n> > change this code later when the time comes. I am not sure what way is\n> > best.\n>\n> I preferred to do that change when the feature is getting extended.\n>\n\n+1.\n\n*\n+$node_C->safe_psql(\n+ 'postgres', \"\n+ DELETE FROM tab\");\n+$node_B->safe_psql(\n+ 'postgres', \"\n+ DELETE FROM tab where a = 32\");\n+\n+$node_A->wait_for_catchup($subname_BA);\n+$node_B->wait_for_catchup($subname_AB);\n\nHere, don't we need to use node_C instead of node_A for waiting as we\nhave performed an operation on node_C?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 12 Jul 2022 09:51:31 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Jul 12, 2022 at 8:43 AM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Mon, Jul 11, 2022 at 9:11 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > >\n> > > Here are my review comments for the v30* patches:\n> > >\n> > > ========\n> > > v30-0001\n> > > ========\n> > >\n> > > 1.1 <general>\n> > >\n> > > I was wondering if it is better to implement a new defGetOrigin method\n> > > now instead of just using the defGetString to process the 'origin',\n> > > since you may need to do that in future anyway if the 'origin' name is\n> > > planned to become specifiable by the user. OTOH maybe you prefer to\n> > > change this code later when the time comes. I am not sure what way is\n> > > best.\n> >\n> > I preferred to do that change when the feature is getting extended.\n> >\n>\n> +1.\n>\n> *\n> +$node_C->safe_psql(\n> + 'postgres', \"\n> + DELETE FROM tab\");\n> +$node_B->safe_psql(\n> + 'postgres', \"\n> + DELETE FROM tab where a = 32\");\n> +\n> +$node_A->wait_for_catchup($subname_BA);\n> +$node_B->wait_for_catchup($subname_AB);\n>\n> Here, don't we need to use node_C instead of node_A for waiting as we\n> have performed an operation on node_C?\n\nNo need to wait for node_C as we have dropped the subscription before\nthe delete operation. I have added the comment to make it clear.\nThe attached v32 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Tue, 12 Jul 2022 14:58:04 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n\nI find one thing confusing about this patch. Basically, this has two\noption 'local' and 'any', so I would assume that all the local server\nchanges should be covered under the 'local' but now if we set some\norigin using 'select pg_replication_origin_session_setup('aa');' then\nchanges from that session will be ignored because it has an origin id.\nI think actually the name is creating confusion, because by local it\nseems like a change which originated locally and the document is also\nspecifying the same.\n\n+ If <literal>local</literal>, the subscription will request the publisher\n+ to only send changes that originated locally. If <literal>any</literal>,\n\nI think if we want to keep the option local then we should look up all\nthe origin in the replication origin catalog and identify whether it\nis a local origin id or remote origin id and based on that filter out\nthe changes.\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 13 Jul 2022 16:49:56 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>\n> On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> I find one thing confusing about this patch. Basically, this has two\n> option 'local' and 'any', so I would assume that all the local server\n> changes should be covered under the 'local' but now if we set some\n> origin using 'select pg_replication_origin_session_setup('aa');' then\n> changes from that session will be ignored because it has an origin id.\n> I think actually the name is creating confusion, because by local it\n> seems like a change which originated locally and the document is also\n> specifying the same.\n>\n> + If <literal>local</literal>, the subscription will request the publisher\n> + to only send changes that originated locally. If <literal>any</literal>,\n>\n> I think if we want to keep the option local then we should look up all\n> the origin in the replication origin catalog and identify whether it\n> is a local origin id or remote origin id and based on that filter out\n> the changes.\n\nOn the other hand if we are interested in receiving the changes which\nare generated without any origin then I think we should change 'local'\nto 'none' and then in future we can provide a new option which can\nsend the changes generated by all the local origin? I think other\nthan this the patch LGTM.\n\n\n--\nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 14 Jul 2022 11:26:32 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>\n> On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >\n> > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > I find one thing confusing about this patch. Basically, this has two\n> > option 'local' and 'any', so I would assume that all the local server\n> > changes should be covered under the 'local' but now if we set some\n> > origin using 'select pg_replication_origin_session_setup('aa');' then\n> > changes from that session will be ignored because it has an origin id.\n> > I think actually the name is creating confusion, because by local it\n> > seems like a change which originated locally and the document is also\n> > specifying the same.\n> >\n> > + If <literal>local</literal>, the subscription will request the publisher\n> > + to only send changes that originated locally. If <literal>any</literal>,\n> >\n> > I think if we want to keep the option local then we should look up all\n> > the origin in the replication origin catalog and identify whether it\n> > is a local origin id or remote origin id and based on that filter out\n> > the changes.\n>\n> On the other hand if we are interested in receiving the changes which\n> are generated without any origin then I think we should change 'local'\n> to 'none' and then in future we can provide a new option which can\n> send the changes generated by all the local origin? I think other\n> than this the patch LGTM.\n\nThanks for the comment. The attached v33 patch has the changes to\nspecify origin as 'none' instead of 'local' which will not publish the\ndata having any origin.\n\nRegards,\nVignesh","msg_date":"Thu, 14 Jul 2022 18:34:02 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, 14 Jul 2022 at 6:34 PM, vignesh C <vignesh21@gmail.com> wrote:\n\n> On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com>\n> wrote:\n> >\n> > On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com>\n> wrote:\n> > >\n> > > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\n> > > >\n> > > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com>\n> wrote:\n> > >\n> > > I find one thing confusing about this patch. Basically, this has two\n> > > option 'local' and 'any', so I would assume that all the local server\n> > > changes should be covered under the 'local' but now if we set some\n> > > origin using 'select pg_replication_origin_session_setup('aa');' then\n> > > changes from that session will be ignored because it has an origin id.\n> > > I think actually the name is creating confusion, because by local it\n> > > seems like a change which originated locally and the document is also\n> > > specifying the same.\n> > >\n> > > + If <literal>local</literal>, the subscription will request the\n> publisher\n> > > + to only send changes that originated locally. If\n> <literal>any</literal>,\n> > >\n> > > I think if we want to keep the option local then we should look up all\n> > > the origin in the replication origin catalog and identify whether it\n> > > is a local origin id or remote origin id and based on that filter out\n> > > the changes.\n> >\n> > On the other hand if we are interested in receiving the changes which\n> > are generated without any origin then I think we should change 'local'\n> > to 'none' and then in future we can provide a new option which can\n> > send the changes generated by all the local origin? I think other\n> > than this the patch LGTM.\n>\n> Thanks for the comment. The attached v33 patch has the changes to\n> specify origin as 'none' instead of 'local' which will not publish the\n> data having any origin.\n\n\nI think the ‘none’ might have problem from expand ability pov? what if in\nfuture we support the actual origin name and than what none mean? no origin\nor origin name none? Should we just give origin name empty name ‘’? Or is\nthere some other issue?\n\n>\n—\nDilip\n\n> --\nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\nOn Thu, 14 Jul 2022 at 6:34 PM, vignesh C <vignesh21@gmail.com> wrote:On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>\n> On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >\n> > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > I find one thing confusing about this patch. Basically, this has two\n> > option 'local' and 'any', so I would assume that all the local server\n> > changes should be covered under the 'local' but now if we set some\n> > origin using 'select pg_replication_origin_session_setup('aa');' then\n> > changes from that session will be ignored because it has an origin id.\n> > I think actually the name is creating confusion, because by local it\n> > seems like a change which originated locally and the document is also\n> > specifying the same.\n> >\n> > + If <literal>local</literal>, the subscription will request the publisher\n> > + to only send changes that originated locally. If <literal>any</literal>,\n> >\n> > I think if we want to keep the option local then we should look up all\n> > the origin in the replication origin catalog and identify whether it\n> > is a local origin id or remote origin id and based on that filter out\n> > the changes.\n>\n> On the other hand if we are interested in receiving the changes which\n> are generated without any origin then I think we should change 'local'\n> to 'none' and then in future we can provide a new option which can\n> send the changes generated by all the local origin? I think other\n> than this the patch LGTM.\n\nThanks for the comment. The attached v33 patch has the changes to\nspecify origin as 'none' instead of 'local' which will not publish the\ndata having any origin.I think the ‘none’ might have problem from expand ability pov? what if in future we support the actual origin name and than what none mean? no origin or origin name none? Should we just give origin name empty name ‘’? Or is there some other issue?—Dilip-- Regards,Dilip KumarEnterpriseDB: http://www.enterprisedb.com","msg_date":"Thu, 14 Jul 2022 18:42:11 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jul 14, 2022 at 6:42 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>\n> On Thu, 14 Jul 2022 at 6:34 PM, vignesh C <vignesh21@gmail.com> wrote:\n>>\n>> On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>> >\n>> > On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>> > >\n>> > > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\n>> > > >\n>> > > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>> > >\n>> > > I find one thing confusing about this patch. Basically, this has two\n>> > > option 'local' and 'any', so I would assume that all the local server\n>> > > changes should be covered under the 'local' but now if we set some\n>> > > origin using 'select pg_replication_origin_session_setup('aa');' then\n>> > > changes from that session will be ignored because it has an origin id.\n>> > > I think actually the name is creating confusion, because by local it\n>> > > seems like a change which originated locally and the document is also\n>> > > specifying the same.\n>> > >\n>> > > + If <literal>local</literal>, the subscription will request the publisher\n>> > > + to only send changes that originated locally. If <literal>any</literal>,\n>> > >\n>> > > I think if we want to keep the option local then we should look up all\n>> > > the origin in the replication origin catalog and identify whether it\n>> > > is a local origin id or remote origin id and based on that filter out\n>> > > the changes.\n>> >\n>> > On the other hand if we are interested in receiving the changes which\n>> > are generated without any origin then I think we should change 'local'\n>> > to 'none' and then in future we can provide a new option which can\n>> > send the changes generated by all the local origin? I think other\n>> > than this the patch LGTM.\n>>\n>> Thanks for the comment. The attached v33 patch has the changes to\n>> specify origin as 'none' instead of 'local' which will not publish the\n>> data having any origin.\n>\n>\n> I think the ‘none’ might have problem from expand ability pov? what if in future we support the actual origin name and than what none mean? no origin or origin name none? Should we just give origin name empty name ‘’? Or is there some other issue?\n\nCurrently there is no restriction in the name we can specify for\norigin, ex any, none, local, etc all are allowed as origin name. How\nabout extending it with another parameter \"origin_name\" when we\nsupport filtering of a particular origin like:\n1) origin = name, origin_name = 'orig1' -- Means that the publisher\nwill filter the changes having origin name as 'orig1' and send the\nother changes.\n2) origin = any -- Means that the publisher sends all changes\nregardless of their origin.\n3) origin = none -- Means that the subscription will request the\npublisher to only send changes that have no origin associated.\nWhen we need to specify filtering of a particular origin name we will\nhave to use both origin and origin_name option, like origin = name,\norigin_name = 'orig1' as in the first example.\n\nI'm not sure if there is a simpler way to do this with only a single\noption as both 'none' and 'any' can be specified as the origin name.\nThoughts?\n\nRegards,\nVignesh\n\n\n","msg_date":"Sat, 16 Jul 2022 09:04:54 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sat, Jul 16, 2022 at 9:05 AM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Thu, Jul 14, 2022 at 6:42 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >\n> > On Thu, 14 Jul 2022 at 6:34 PM, vignesh C <vignesh21@gmail.com> wrote:\n> >>\n> >> On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >> >\n> >> > On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >> > >\n> >> > > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\n> >> > > >\n> >> > > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >> > >\n> >> > > I find one thing confusing about this patch. Basically, this has two\n> >> > > option 'local' and 'any', so I would assume that all the local server\n> >> > > changes should be covered under the 'local' but now if we set some\n> >> > > origin using 'select pg_replication_origin_session_setup('aa');' then\n> >> > > changes from that session will be ignored because it has an origin id.\n> >> > > I think actually the name is creating confusion, because by local it\n> >> > > seems like a change which originated locally and the document is also\n> >> > > specifying the same.\n> >> > >\n> >> > > + If <literal>local</literal>, the subscription will request the publisher\n> >> > > + to only send changes that originated locally. If <literal>any</literal>,\n> >> > >\n> >> > > I think if we want to keep the option local then we should look up all\n> >> > > the origin in the replication origin catalog and identify whether it\n> >> > > is a local origin id or remote origin id and based on that filter out\n> >> > > the changes.\n> >> >\n> >> > On the other hand if we are interested in receiving the changes which\n> >> > are generated without any origin then I think we should change 'local'\n> >> > to 'none' and then in future we can provide a new option which can\n> >> > send the changes generated by all the local origin? I think other\n> >> > than this the patch LGTM.\n> >>\n> >> Thanks for the comment. The attached v33 patch has the changes to\n> >> specify origin as 'none' instead of 'local' which will not publish the\n> >> data having any origin.\n> >\n> >\n> > I think the ‘none’ might have problem from expand ability pov? what if in future we support the actual origin name and than what none mean? no origin or origin name none? Should we just give origin name empty name ‘’? Or is there some other issue?\n>\n> Currently there is no restriction in the name we can specify for\n> origin, ex any, none, local, etc all are allowed as origin name. How\n> about extending it with another parameter \"origin_name\" when we\n> support filtering of a particular origin like:\n> 1) origin = name, origin_name = 'orig1' -- Means that the publisher\n> will filter the changes having origin name as 'orig1' and send the\n> other changes.\n> 2) origin = any -- Means that the publisher sends all changes\n> regardless of their origin.\n> 3) origin = none -- Means that the subscription will request the\n> publisher to only send changes that have no origin associated.\n> When we need to specify filtering of a particular origin name we will\n> have to use both origin and origin_name option, like origin = name,\n> origin_name = 'orig1' as in the first example.\n>\n> I'm not sure if there is a simpler way to do this with only a single\n> option as both 'none' and 'any' can be specified as the origin name.\n> Thoughts?\n\nI think giving two options would be really confusing from the\nusability perspective. I think what we should be doing here is to\nkeep these three names 'none', 'any' and 'local' as reserved names for\nthe origin name so that those are not allowed to be set by the user\nand they have some internal meaning. And I don't think this is going\nto create too much trouble for anyone because those are not really the\nnames someone wants to use for their replication origin. So I think\npg_replication_origin_create() we can also check for the reserved\nnames just for the replication origin.\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Sat, 16 Jul 2022 10:29:11 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sat, Jul 16, 2022 at 10:29 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>\n> I think giving two options would be really confusing from the\n> usability perspective. I think what we should be doing here is to\n> keep these three names 'none', 'any' and 'local' as reserved names for\n> the origin name so that those are not allowed to be set by the user\n> and they have some internal meaning.\n>\n\nThis makes sense to me. I think we can avoid reserving 'local' for now\ntill we agree on its use case and implementation. One similar point\nabout slots is that we treat 'none' slot_name in subscription commands\nas a special value indicating no slot name whereas we do allow\ncreating a slot with the name 'none' with\npg_create_logical_replication_slot(). So, if we want to follow a\nsimilar convention here, we may not need to add any restriction for\norigin names but personally, I think it is better to add such a\nrestriction to avoid confusion and in fact, as a separate patch we\nshould even disallow creating slot name as 'none'.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 18 Jul 2022 10:22:50 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 18, 2022 at 10:23 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Sat, Jul 16, 2022 at 10:29 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >\n> > I think giving two options would be really confusing from the\n> > usability perspective. I think what we should be doing here is to\n> > keep these three names 'none', 'any' and 'local' as reserved names for\n> > the origin name so that those are not allowed to be set by the user\n> > and they have some internal meaning.\n> >\n>\n> This makes sense to me. I think we can avoid reserving 'local' for now\n> till we agree on its use case and implementation. One similar point\n> about slots is that we treat 'none' slot_name in subscription commands\n> as a special value indicating no slot name whereas we do allow\n> creating a slot with the name 'none' with\n> pg_create_logical_replication_slot(). So, if we want to follow a\n> similar convention here, we may not need to add any restriction for\n> origin names but personally, I think it is better to add such a\n> restriction to avoid confusion and in fact, as a separate patch we\n> should even disallow creating slot name as 'none'.\n\nI have made changes to disallow the name \"any\" and \"none\" in\npg_replication_origin_create. The attached v34 patch has the changes\nfor the same. I will post a separate patch to disallow creation of\nslots with names as 'none' separately later.\n\nRegards,\nVignesh","msg_date":"Mon, 18 Jul 2022 16:58:32 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 18, 2022 at 4:58 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Jul 18, 2022 at 10:23 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Sat, Jul 16, 2022 at 10:29 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> > >\n> > > I think giving two options would be really confusing from the\n> > > usability perspective. I think what we should be doing here is to\n> > > keep these three names 'none', 'any' and 'local' as reserved names for\n> > > the origin name so that those are not allowed to be set by the user\n> > > and they have some internal meaning.\n> > >\n> >\n> > This makes sense to me. I think we can avoid reserving 'local' for now\n> > till we agree on its use case and implementation. One similar point\n> > about slots is that we treat 'none' slot_name in subscription commands\n> > as a special value indicating no slot name whereas we do allow\n> > creating a slot with the name 'none' with\n> > pg_create_logical_replication_slot(). So, if we want to follow a\n> > similar convention here, we may not need to add any restriction for\n> > origin names but personally, I think it is better to add such a\n> > restriction to avoid confusion and in fact, as a separate patch we\n> > should even disallow creating slot name as 'none'.\n>\n> I have made changes to disallow the name \"any\" and \"none\" in\n> pg_replication_origin_create. The attached v34 patch has the changes\n> for the same. I will post a separate patch to disallow creation of\n> slots with names as 'none' separately later.\n\nI have updated the patch to handle the origin value case\ninsensitively. The attached patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Mon, 18 Jul 2022 21:46:37 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 18, 2022 at 9:46 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> I have updated the patch to handle the origin value case\n> insensitively. The attached patch has the changes for the same.\n>\n\nThanks, the patch looks mostly good to me. I have made a few changes\nin 0001 patch which are as follows: (a) make a comparison of origin\nnames in maybe_reread_subscription similar to slot names as in future\nwe may support origin names other than 'any' and 'none', (b) made\ncomment changes at few places and minor change in one of the error\nmessage, (c) ran pgindent and slightly changed the commit message.\n\nI am planning to push this day after tomorrow unless there are any\ncomments/suggestions.\n\n-- \nWith Regards,\nAmit Kapila.","msg_date":"Tue, 19 Jul 2022 17:04:37 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 19, 2022 at 11:34 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Jul 18, 2022 at 9:46 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > I have updated the patch to handle the origin value case\n> > insensitively. The attached patch has the changes for the same.\n> >\n>\n> Thanks, the patch looks mostly good to me. I have made a few changes\n> in 0001 patch which are as follows: (a) make a comparison of origin\n> names in maybe_reread_subscription similar to slot names as in future\n> we may support origin names other than 'any' and 'none', (b) made\n> comment changes at few places and minor change in one of the error\n> message, (c) ran pgindent and slightly changed the commit message.\n>\n> I am planning to push this day after tomorrow unless there are any\n> comments/suggestions.\n\nFYI, the function name in the comment is not same as the function name here:\n\n+/*\n+ * IsReservedName\n+ * True iff name is either \"none\" or \"any\".\n+ */\n+static bool\n+IsReservedOriginName(const char *name)\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 20 Jul 2022 17:08:36 +1200","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Jul 20, 2022 at 10:38 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Tue, Jul 19, 2022 at 11:34 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Mon, Jul 18, 2022 at 9:46 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > I have updated the patch to handle the origin value case\n> > > insensitively. The attached patch has the changes for the same.\n> > >\n> >\n> > Thanks, the patch looks mostly good to me. I have made a few changes\n> > in 0001 patch which are as follows: (a) make a comparison of origin\n> > names in maybe_reread_subscription similar to slot names as in future\n> > we may support origin names other than 'any' and 'none', (b) made\n> > comment changes at few places and minor change in one of the error\n> > message, (c) ran pgindent and slightly changed the commit message.\n> >\n> > I am planning to push this day after tomorrow unless there are any\n> > comments/suggestions.\n>\n> FYI, the function name in the comment is not same as the function name here:\n>\n> +/*\n> + * IsReservedName\n> + * True iff name is either \"none\" or \"any\".\n> + */\n> +static bool\n> +IsReservedOriginName(const char *name)\n\nModified. Apart from this I have run pgperltidy on the perl file and\nrenamed 032_origin.pl to 030_origin.pl as currently there is\n029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\nThanks for the comment, the attached patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Wed, 20 Jul 2022 14:33:16 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Jul 20, 2022 at 2:33 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> Modified. Apart from this I have run pgperltidy on the perl file and\n> renamed 032_origin.pl to 030_origin.pl as currently there is\n> 029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\n> Thanks for the comment, the attached patch has the changes for the same.\n>\n\nPushed. Kindly rebase the remaining patches.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 21 Jul 2022 14:06:31 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jul 21, 2022 at 2:06 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Jul 20, 2022 at 2:33 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Modified. Apart from this I have run pgperltidy on the perl file and\n> > renamed 032_origin.pl to 030_origin.pl as currently there is\n> > 029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\n> > Thanks for the comment, the attached patch has the changes for the same.\n> >\n>\n> Pushed. Kindly rebase the remaining patches.\n\nThanks for pushing the patch.\nThe attached v37 version contains the rebased patch for the remaining patches.\n\nRegards,\nVignesh","msg_date":"Thu, 21 Jul 2022 16:04:22 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi,\r\n\r\nOn 7/21/22 6:34 AM, vignesh C wrote:\r\n> On Thu, Jul 21, 2022 at 2:06 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n>>\r\n>> On Wed, Jul 20, 2022 at 2:33 PM vignesh C <vignesh21@gmail.com> wrote:\r\n>>>\r\n>>> Modified. Apart from this I have run pgperltidy on the perl file and\r\n>>> renamed 032_origin.pl to 030_origin.pl as currently there is\r\n>>> 029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\r\n>>> Thanks for the comment, the attached patch has the changes for the same.\r\n>>>\r\n>>\r\n>> Pushed. Kindly rebase the remaining patches.\r\n> \r\n> Thanks for pushing the patch.\r\n> The attached v37 version contains the rebased patch for the remaining patches.\r\n\r\nThanks for the work on this feature -- this is definitely very helpful \r\ntowards supporting more types of use cases with logical replication!\r\n\r\nI've read through the proposed documentation and did some light testing \r\nof the patch. I have two general comments about the docs as they \r\ncurrently read:\r\n\r\n1. I'm concerned by calling this \"Bidirectional replication\" in the docs \r\nthat we are overstating the current capabilities. I think this is \r\naccentuated int he opening paragraph:\r\n\r\n==snip==\r\n Bidirectional replication is useful for creating a multi-master database\r\n environment for replicating read/write operations performed by any of the\r\n member nodes.\r\n==snip==\r\n\r\nFor one, we're not replicating reads, we're replicating writes. Amongst \r\nthe writes, at this point we're only replicating DML. A reader could \r\nthink that deploying can work for a full bidirectional solution.\r\n\r\n(Even if we're aspirationally calling this section \"Bidirectional \r\nreplication\", that does make it sound like we're limited to two nodes, \r\nwhen we can support more than two).\r\n\r\nPerhaps \"Logical replication between writers\" or \"Logical replication \r\nbetween primaries\" or \"Replicating changes between primaries\", or \r\nsomething better.\r\n\r\n2. There is no mention of conflicts in the documentation, e.g. \r\nreferencing the \"Conflicts\" section of the documentation. It's very easy \r\nto create a conflicting transaction that causes a subscriber to be \r\nunable to continue to apply transactions:\r\n\r\n -- DB 1\r\n CREATE TABLE abc (id int);\r\n CREATE PUBLICATION node1 FOR ALL TABLES ;\r\n\r\n -- DB2\r\n CREATE TABLE abc (id int);\r\n CREATE PUBLICATION node2 FOR ALL TABLES ;\r\n CREATE SUBSCRIPTION node2_node1\r\n CONNECTION 'dbname=logi port=5433'\r\n PUBLICATION node1\r\n WITH (copy_data = off, origin = none);\r\n\r\n -- DB1\r\n CREATE SUBSCRIPTION node1_node2\r\n CONNECTION 'dbname=logi port=5434'\r\n PUBLICATION node2\r\n WITH (copy_data = off, origin = none);\r\n INSERT INTO abc VALUES (1);\r\n\r\n -- DB2\r\n INSERT INTO abc VALUES (2);\r\n\r\n -- DB1\r\n ALTER TABLE abc ADD PRIMARY KEY id;\r\n INSERT INTO abc VALUES (3);\r\n\r\n -- DB2\r\n INSERT INTO abc VALUES (3);\r\n\r\n -- DB1 cannot apply the transactions\r\n\r\nAt a minimum, I think we should reference the documentation we have in \r\nthe logical replication section on conflicts. We may also want to advise \r\nthat a user is responsible for designing their schemas in a way to \r\nminimize the risk of conflicts.\r\n\r\nThanks,\r\n\r\nJonathan","msg_date":"Thu, 21 Jul 2022 16:08:54 -0400","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> Thanks for the work on this feature -- this is definitely very helpful\n> towards supporting more types of use cases with logical replication!\n>\n> I've read through the proposed documentation and did some light testing\n> of the patch. I have two general comments about the docs as they\n> currently read:\n>\n> 1. I'm concerned by calling this \"Bidirectional replication\" in the docs\n> that we are overstating the current capabilities. I think this is\n> accentuated int he opening paragraph:\n>\n> ==snip==\n> Bidirectional replication is useful for creating a multi-master database\n> environment for replicating read/write operations performed by any of the\n> member nodes.\n> ==snip==\n>\n> For one, we're not replicating reads, we're replicating writes. Amongst\n> the writes, at this point we're only replicating DML. A reader could\n> think that deploying can work for a full bidirectional solution.\n>\n> (Even if we're aspirationally calling this section \"Bidirectional\n> replication\", that does make it sound like we're limited to two nodes,\n> when we can support more than two).\n>\n\nRight, I think the system can support N-Way replication.\n\n> Perhaps \"Logical replication between writers\" or \"Logical replication\n> between primaries\" or \"Replicating changes between primaries\", or\n> something better.\n>\n\nAmong the above \"Replicating changes between primaries\" sounds good to\nme or simply \"Replication between primaries\". As this is a sub-section\non the Logical Replication page, I feel it is okay to not use Logical\nin the title.\n\n> 2. There is no mention of conflicts in the documentation, e.g.\n> referencing the \"Conflicts\" section of the documentation. It's very easy\n> to create a conflicting transaction that causes a subscriber to be\n> unable to continue to apply transactions:\n>\n> -- DB 1\n> CREATE TABLE abc (id int);\n> CREATE PUBLICATION node1 FOR ALL TABLES ;\n>\n> -- DB2\n> CREATE TABLE abc (id int);\n> CREATE PUBLICATION node2 FOR ALL TABLES ;\n> CREATE SUBSCRIPTION node2_node1\n> CONNECTION 'dbname=logi port=5433'\n> PUBLICATION node1\n> WITH (copy_data = off, origin = none);\n>\n> -- DB1\n> CREATE SUBSCRIPTION node1_node2\n> CONNECTION 'dbname=logi port=5434'\n> PUBLICATION node2\n> WITH (copy_data = off, origin = none);\n> INSERT INTO abc VALUES (1);\n>\n> -- DB2\n> INSERT INTO abc VALUES (2);\n>\n> -- DB1\n> ALTER TABLE abc ADD PRIMARY KEY id;\n> INSERT INTO abc VALUES (3);\n>\n> -- DB2\n> INSERT INTO abc VALUES (3);\n>\n> -- DB1 cannot apply the transactions\n>\n> At a minimum, I think we should reference the documentation we have in\n> the logical replication section on conflicts. We may also want to advise\n> that a user is responsible for designing their schemas in a way to\n> minimize the risk of conflicts.\n>\n\nThis sounds reasonable to me.\n\nOne more point about docs, it appears to be added as the last\nsub-section on the Logical Replication page. Is there a reason for\ndoing so? I feel this should be third sub-section after describing\nPublication and Subscription.\n\nBTW, do you have any opinion on the idea of the first remaining patch\nwhere we accomplish two things: a) Checks and throws an error if\n'copy_data = on' and 'origin = none' but the publication tables were\nalso replicated from other publishers. b) Adds 'force' value for\ncopy_data parameter to allow copying in such a case. The primary\nreason for this patch is to avoid loops or duplicate data in the\ninitial phase. We can't skip copying based on origin as we can do\nwhile replicating changes from WAL. So, we detect that the publisher\nalready has data from some other node and doesn't allow replication\nunless the user uses the 'force' option for copy_data.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 22 Jul 2022 10:17:01 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> Hi,\n>\n> On 7/21/22 6:34 AM, vignesh C wrote:\n> > On Thu, Jul 21, 2022 at 2:06 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >>\n> >> On Wed, Jul 20, 2022 at 2:33 PM vignesh C <vignesh21@gmail.com> wrote:\n> >>>\n> >>> Modified. Apart from this I have run pgperltidy on the perl file and\n> >>> renamed 032_origin.pl to 030_origin.pl as currently there is\n> >>> 029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\n> >>> Thanks for the comment, the attached patch has the changes for the same.\n> >>>\n> >>\n> >> Pushed. Kindly rebase the remaining patches.\n> >\n> > Thanks for pushing the patch.\n> > The attached v37 version contains the rebased patch for the remaining patches.\n>\n> Thanks for the work on this feature -- this is definitely very helpful\n> towards supporting more types of use cases with logical replication!\n>\n> I've read through the proposed documentation and did some light testing\n> of the patch. I have two general comments about the docs as they\n> currently read:\n>\n> 1. I'm concerned by calling this \"Bidirectional replication\" in the docs\n> that we are overstating the current capabilities. I think this is\n> accentuated int he opening paragraph:\n>\n> ==snip==\n> Bidirectional replication is useful for creating a multi-master database\n> environment for replicating read/write operations performed by any of the\n> member nodes.\n> ==snip==\n>\n> For one, we're not replicating reads, we're replicating writes. Amongst\n> the writes, at this point we're only replicating DML. A reader could\n> think that deploying can work for a full bidirectional solution.\n\nI have changed read/write operations to write operations. I have also\nadded a note saying \"The logical replication restrictions applies to\nthe replication between primaries also.\", to clarify that non DML\noperations and other restrictions apply in this case too.\n\n> (Even if we're aspirationally calling this section \"Bidirectional\n> replication\", that does make it sound like we're limited to two nodes,\n> when we can support more than two).\n>\n> Perhaps \"Logical replication between writers\" or \"Logical replication\n> between primaries\" or \"Replicating changes between primaries\", or\n> something better.\n\nI have changed it to \"Replication between primaries\".\n\n> 2. There is no mention of conflicts in the documentation, e.g.\n> referencing the \"Conflicts\" section of the documentation. It's very easy\n> to create a conflicting transaction that causes a subscriber to be\n> unable to continue to apply transactions:\n>\n> -- DB 1\n> CREATE TABLE abc (id int);\n> CREATE PUBLICATION node1 FOR ALL TABLES ;\n>\n> -- DB2\n> CREATE TABLE abc (id int);\n> CREATE PUBLICATION node2 FOR ALL TABLES ;\n> CREATE SUBSCRIPTION node2_node1\n> CONNECTION 'dbname=logi port=5433'\n> PUBLICATION node1\n> WITH (copy_data = off, origin = none);\n>\n> -- DB1\n> CREATE SUBSCRIPTION node1_node2\n> CONNECTION 'dbname=logi port=5434'\n> PUBLICATION node2\n> WITH (copy_data = off, origin = none);\n> INSERT INTO abc VALUES (1);\n>\n> -- DB2\n> INSERT INTO abc VALUES (2);\n>\n> -- DB1\n> ALTER TABLE abc ADD PRIMARY KEY id;\n> INSERT INTO abc VALUES (3);\n>\n> -- DB2\n> INSERT INTO abc VALUES (3);\n>\n> -- DB1 cannot apply the transactions\n>\n> At a minimum, I think we should reference the documentation we have in\n> the logical replication section on conflicts. We may also want to advise\n> that a user is responsible for designing their schemas in a way to\n> minimize the risk of conflicts.\n\nAdded a note for the same and referred it to the conflicts section.\n\nThanks for the comments, the attached v38 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Sat, 23 Jul 2022 22:57:56 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jul 22, 2022 at 10:17 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n> >\n> > Thanks for the work on this feature -- this is definitely very helpful\n> > towards supporting more types of use cases with logical replication!\n> >\n> > I've read through the proposed documentation and did some light testing\n> > of the patch. I have two general comments about the docs as they\n> > currently read:\n> >\n> > 1. I'm concerned by calling this \"Bidirectional replication\" in the docs\n> > that we are overstating the current capabilities. I think this is\n> > accentuated int he opening paragraph:\n> >\n> > ==snip==\n> > Bidirectional replication is useful for creating a multi-master database\n> > environment for replicating read/write operations performed by any of the\n> > member nodes.\n> > ==snip==\n> >\n> > For one, we're not replicating reads, we're replicating writes. Amongst\n> > the writes, at this point we're only replicating DML. A reader could\n> > think that deploying can work for a full bidirectional solution.\n> >\n> > (Even if we're aspirationally calling this section \"Bidirectional\n> > replication\", that does make it sound like we're limited to two nodes,\n> > when we can support more than two).\n> >\n>\n> Right, I think the system can support N-Way replication.\n>\n> > Perhaps \"Logical replication between writers\" or \"Logical replication\n> > between primaries\" or \"Replicating changes between primaries\", or\n> > something better.\n> >\n>\n> Among the above \"Replicating changes between primaries\" sounds good to\n> me or simply \"Replication between primaries\". As this is a sub-section\n> on the Logical Replication page, I feel it is okay to not use Logical\n> in the title.\n\nI have changed it to \"Replication between primaries\".\n\n> > 2. There is no mention of conflicts in the documentation, e.g.\n> > referencing the \"Conflicts\" section of the documentation. It's very easy\n> > to create a conflicting transaction that causes a subscriber to be\n> > unable to continue to apply transactions:\n> >\n> > -- DB 1\n> > CREATE TABLE abc (id int);\n> > CREATE PUBLICATION node1 FOR ALL TABLES ;\n> >\n> > -- DB2\n> > CREATE TABLE abc (id int);\n> > CREATE PUBLICATION node2 FOR ALL TABLES ;\n> > CREATE SUBSCRIPTION node2_node1\n> > CONNECTION 'dbname=logi port=5433'\n> > PUBLICATION node1\n> > WITH (copy_data = off, origin = none);\n> >\n> > -- DB1\n> > CREATE SUBSCRIPTION node1_node2\n> > CONNECTION 'dbname=logi port=5434'\n> > PUBLICATION node2\n> > WITH (copy_data = off, origin = none);\n> > INSERT INTO abc VALUES (1);\n> >\n> > -- DB2\n> > INSERT INTO abc VALUES (2);\n> >\n> > -- DB1\n> > ALTER TABLE abc ADD PRIMARY KEY id;\n> > INSERT INTO abc VALUES (3);\n> >\n> > -- DB2\n> > INSERT INTO abc VALUES (3);\n> >\n> > -- DB1 cannot apply the transactions\n> >\n> > At a minimum, I think we should reference the documentation we have in\n> > the logical replication section on conflicts. We may also want to advise\n> > that a user is responsible for designing their schemas in a way to\n> > minimize the risk of conflicts.\n> >\n>\n> This sounds reasonable to me.\n>\n> One more point about docs, it appears to be added as the last\n> sub-section on the Logical Replication page. Is there a reason for\n> doing so? I feel this should be third sub-section after describing\n> Publication and Subscription.\n\nI had initially kept it at the end since we demonstrated the various\nsteps to create the replication between the primaries. Since it gives\nan introduction about the \"Replication between primaries\" and then\nstates the steps, it looks ok to move it as suggested. I have modified\nthis in the v38 patch attached at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm01x0sLz8YzfCSjxcMFxM4NDQxcFzZa%2B4eesUmD40DdTg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Sat, 23 Jul 2022 23:02:31 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On 7/22/22 12:47 AM, Amit Kapila wrote:\r\n> On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\r\n\r\n>> 1. I'm concerned by calling this \"Bidirectional replication\" in the docs\r\n>> that we are overstating the current capabilities. I think this is\r\n>> accentuated int he opening paragraph:\r\n>>\r\n>> ==snip==\r\n>> Bidirectional replication is useful for creating a multi-master database\r\n>> environment for replicating read/write operations performed by any of the\r\n>> member nodes.\r\n>> ==snip==\r\n>>\r\n>> For one, we're not replicating reads, we're replicating writes. Amongst\r\n>> the writes, at this point we're only replicating DML. A reader could\r\n>> think that deploying can work for a full bidirectional solution.\r\n>>\r\n>> (Even if we're aspirationally calling this section \"Bidirectional\r\n>> replication\", that does make it sound like we're limited to two nodes,\r\n>> when we can support more than two).\r\n>>\r\n> \r\n> Right, I think the system can support N-Way replication.\r\n\r\nI did some more testing of the feature, i.e. doing 3-node and 4-node \r\ntests. While logical replication today can handle replicating between \r\nmultiple nodes (N-way), the \"origin = none\" does require setting up \r\nsubscribers between each of the nodes.\r\n\r\nFor example, if I have 4 nodes A, B, C, D and I want to replicate the \r\nsame table between all of them, I need to set up subscribers between all \r\nof them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node \r\ncan replicate between each other in a way that's convenient (vs. having \r\nto do something funky with partitions) so this is still a big step forward.\r\n\r\nThis is a long way of saying that I do think it's fair to say we support \r\n\"N-way\" replication so long as you are set up in a mesh (vs. a ring, \r\ne.g. A=>B=>C=>D=>A).\r\n\r\n> Among the above \"Replicating changes between primaries\" sounds good to\r\n> me or simply \"Replication between primaries\". As this is a sub-section\r\n> on the Logical Replication page, I feel it is okay to not use Logical\r\n> in the title.\r\n\r\nAgreed, I think that's fine.\r\n\r\n>> At a minimum, I think we should reference the documentation we have in\r\n>> the logical replication section on conflicts. We may also want to advise\r\n>> that a user is responsible for designing their schemas in a way to\r\n>> minimize the risk of conflicts.\r\n>>\r\n> \r\n> This sounds reasonable to me.\r\n> \r\n> One more point about docs, it appears to be added as the last\r\n> sub-section on the Logical Replication page. Is there a reason for\r\n> doing so? I feel this should be third sub-section after describing\r\n> Publication and Subscription.\r\n\r\nWhen I first reviewed, I had not built the docs. Did so on this pass.\r\n\r\nI agree with the positioning argument, i.e. it should go after \r\n\"Subscription\" in the table of contents -- but it makes me question a \r\ncouple of things:\r\n\r\n1. The general ordering of the docs\r\n2. How we describe that section (more on that in a sec)\r\n3. If \"row filters\" should be part of \"subscription\" instead of its own \r\nsection.\r\n\r\nIf you look at the current order, \"Quick setup\" is the last section; one \r\nwould think the \"quick\" portion goes first :) Given a lot of this is for \r\nthe current docs, I may start a separate discussion on -docs for this part.\r\n\r\nFor the time being, I agree it should be moved to the section after \r\n\"Subscription\".\r\n\r\nI think what this section describes is \"Configuring Replication Between \r\nNodes\" as it covers a few different scenarios.\r\n\r\nI do think we need to iterate on these docs -- the examples with the \r\ncommands are generally OK and easy to follow, but a few things I noticed:\r\n\r\n1. The general description of the section needs work. We may want to \r\nrefine the description of the use cases, and in the warning, link to \r\ninstructions on how to take backups.\r\n2. We put the \"case not supported\" in the middle, not at the end.\r\n3. The \"generic steps for adding a new node...\" section uses a \r\nconvention for steps that is not found in the docs. We also don't \r\nprovide an example for this section, and this is the most complicated \r\nscenario to set up.\r\n\r\nI may be able to propose some suggestions in a few days.\r\n\r\n> BTW, do you have any opinion on the idea of the first remaining patch\r\n> where we accomplish two things: a) Checks and throws an error if\r\n> 'copy_data = on' and 'origin = none' but the publication tables were\r\n> also replicated from other publishers. b) Adds 'force' value for\r\n> copy_data parameter to allow copying in such a case. The primary\r\n> reason for this patch is to avoid loops or duplicate data in the\r\n> initial phase. We can't skip copying based on origin as we can do\r\n> while replicating changes from WAL. So, we detect that the publisher\r\n> already has data from some other node and doesn't allow replication\r\n> unless the user uses the 'force' option for copy_data.\r\n\r\nIn general, I agree with the patch; but I'm not sure why we are calling \r\n\"copy_data = force\" in this case and how it varies from \"on\". I \r\nunderstand the goal is to prevent the infinite loop, but is there some \r\ntechnical restriction why we can't set \"origin = none, copy_data = on\" \r\nand have this work (and apologies if I missed that upthread)?\r\n\r\nThe other concern I'll note is that we're changing a boolean parameter \r\nto an enum and I want to be sensitive to folks who are already using \r\n\"copy_data\" to be sure we don't break them.\r\n\r\nJonathan","msg_date":"Sun, 24 Jul 2022 12:51:48 -0400","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> On 7/22/22 12:47 AM, Amit Kapila wrote:\n> > On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> >> 1. I'm concerned by calling this \"Bidirectional replication\" in the docs\n> >> that we are overstating the current capabilities. I think this is\n> >> accentuated int he opening paragraph:\n> >>\n> >> ==snip==\n> >> Bidirectional replication is useful for creating a multi-master database\n> >> environment for replicating read/write operations performed by any of the\n> >> member nodes.\n> >> ==snip==\n> >>\n> >> For one, we're not replicating reads, we're replicating writes. Amongst\n> >> the writes, at this point we're only replicating DML. A reader could\n> >> think that deploying can work for a full bidirectional solution.\n> >>\n> >> (Even if we're aspirationally calling this section \"Bidirectional\n> >> replication\", that does make it sound like we're limited to two nodes,\n> >> when we can support more than two).\n> >>\n> >\n> > Right, I think the system can support N-Way replication.\n>\n> I did some more testing of the feature, i.e. doing 3-node and 4-node\n> tests. While logical replication today can handle replicating between\n> multiple nodes (N-way), the \"origin = none\" does require setting up\n> subscribers between each of the nodes.\n>\n> For example, if I have 4 nodes A, B, C, D and I want to replicate the\n> same table between all of them, I need to set up subscribers between all\n> of them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node\n> can replicate between each other in a way that's convenient (vs. having\n> to do something funky with partitions) so this is still a big step forward.\n>\n> This is a long way of saying that I do think it's fair to say we support\n> \"N-way\" replication so long as you are set up in a mesh (vs. a ring,\n> e.g. A=>B=>C=>D=>A).\n>\n\nSorry, but I don't get your point of mesh vs. ring? I think with some\ncare users can set up replication even in a ring topology.\n\n> > Among the above \"Replicating changes between primaries\" sounds good to\n> > me or simply \"Replication between primaries\". As this is a sub-section\n> > on the Logical Replication page, I feel it is okay to not use Logical\n> > in the title.\n>\n> Agreed, I think that's fine.\n>\n> >> At a minimum, I think we should reference the documentation we have in\n> >> the logical replication section on conflicts. We may also want to advise\n> >> that a user is responsible for designing their schemas in a way to\n> >> minimize the risk of conflicts.\n> >>\n> >\n> > This sounds reasonable to me.\n> >\n> > One more point about docs, it appears to be added as the last\n> > sub-section on the Logical Replication page. Is there a reason for\n> > doing so? I feel this should be third sub-section after describing\n> > Publication and Subscription.\n>\n> When I first reviewed, I had not built the docs. Did so on this pass.\n>\n> I agree with the positioning argument, i.e. it should go after\n> \"Subscription\" in the table of contents -- but it makes me question a\n> couple of things:\n>\n> 1. The general ordering of the docs\n> 2. How we describe that section (more on that in a sec)\n> 3. If \"row filters\" should be part of \"subscription\" instead of its own\n> section.\n>\n\nI don't think it is a good idea to keep \"row filters\" as a part of\nsubscription because we define those at publisher but there are\ncertain things like initial sync or combining of row filters that are\nrelated to subscriptions. So, probably having it in a separate\nsub-section seems okay to me. I have also thought about keeping it as\na part of Publication or Subscription but left it due to the reasons\nmentioned.\n\n> If you look at the current order, \"Quick setup\" is the last section; one\n> would think the \"quick\" portion goes first :) Given a lot of this is for\n> the current docs, I may start a separate discussion on -docs for this part.\n>\n> For the time being, I agree it should be moved to the section after\n> \"Subscription\".\n>\n\nOkay, thanks!\n\n> I think what this section describes is \"Configuring Replication Between\n> Nodes\" as it covers a few different scenarios.\n>\n> I do think we need to iterate on these docs -- the examples with the\n> commands are generally OK and easy to follow, but a few things I noticed:\n>\n> 1. The general description of the section needs work. We may want to\n> refine the description of the use cases, and in the warning, link to\n> instructions on how to take backups.\n> 2. We put the \"case not supported\" in the middle, not at the end.\n> 3. The \"generic steps for adding a new node...\" section uses a\n> convention for steps that is not found in the docs. We also don't\n> provide an example for this section, and this is the most complicated\n> scenario to set up.\n>\n\nI agree with all these points, especially your point related to a\ncomplicated setup. This won't be easy for users without a very clear\ndescription and examples. However, after this work, it will at least\nbe possible for users to set up an N-Way replication with some\nrestrictions and care. BTW, while working on this we have noticed\nthat it is normally a lot of work for users to set up N-way\nreplication among 3, 4, or more nodes and that is why there is another\nproposal to make that set up easier by providing a simpler APIs which\nwill internally do similar to all the manual steps the 0002 patch is\ntrying to describe. See [1] (please don't be confused with the thread\ntitle but the real intent is to allow users to provide an easier way\nto set up an N-Way replication by having all current restrictions of\nLogical Replication like it doesn't support DDL replication)\n\n> I may be able to propose some suggestions in a few days.\n>\n\nOkay, thanks! The plan is to get the 0001 patch of Vignesh. Then work\non docs to describe how users can set up replication among primary\nnodes. After that, if we get consensus on providing simpler APIs for\nsetting up replication among primary nodes as is being discussed in\nthe thread [1], then work on it.\n\n> > BTW, do you have any opinion on the idea of the first remaining patch\n> > where we accomplish two things: a) Checks and throws an error if\n> > 'copy_data = on' and 'origin = none' but the publication tables were\n> > also replicated from other publishers. b) Adds 'force' value for\n> > copy_data parameter to allow copying in such a case. The primary\n> > reason for this patch is to avoid loops or duplicate data in the\n> > initial phase. We can't skip copying based on origin as we can do\n> > while replicating changes from WAL. So, we detect that the publisher\n> > already has data from some other node and doesn't allow replication\n> > unless the user uses the 'force' option for copy_data.\n>\n> In general, I agree with the patch; but I'm not sure why we are calling\n> \"copy_data = force\" in this case and how it varies from \"on\". I\n> understand the goal is to prevent the infinite loop, but is there some\n> technical restriction why we can't set \"origin = none, copy_data = on\"\n> and have this work (and apologies if I missed that upthread)?\n>\n\nThe technical restriction is that we want to throw an error when users\nset \"origin = none, copy_data = on\" and we find that the publication\ntables were also replicated from other publishers. Now, we can give\nWARNING to users in this case but users won't have any way to avoid\nduplicate data which can lead to constraint violations. So, we decided\nto throw ERROR and allow users to perform it with a \"force\" option.\n\n> The other concern I'll note is that we're changing a boolean parameter\n> to an enum and I want to be sensitive to folks who are already using\n> \"copy_data\" to be sure we don't break them.\n>\n\nOkay, but AFAIU, the patch still allows users to specify on/off, so\nwon't that be sufficient?\n\n[1] - https://www.postgresql.org/message-id/CAHut%2BPuwRAoWY9pz%3DEubps3ooQCOBFiYPU9Yi%3DVB-U%2ByORU7OA%40mail.gmail.com\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 25 Jul 2022 09:50:58 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Firstly, I have some (case-sensitivity) questions about the previous\npatch which was already pushed [1].\n\nQ1. create_subscription docs\n\nI did not understand why the docs refer to slot_name = NONE, yet the\nnewly added option says origin = none/any. I think that saying origin\n= NONE/ANY would be more consistent with the existing usage of NONE in\nthis documentation.\n\n~~~\n\nQ2. parse_subscription_options\n\nSimilarly, in the code (parse_subscription_options), I did not\nunderstand why the checks for special name values are implemented\ndifferently:\n\nThe new 'origin' code is using pg_strcmpcase to check special values\n(none/any), and the old 'slot_name' code uses case-sensitive strcmp to\ncheck the special value (none).\n\nFWIW, here I thought the new origin code is the correct one.\n\n======\n\nNow, here are some review comments for the patch v38-0001:\n\n1. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n\n@@ -1781,6 +1858,121 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\n table_close(rel, RowExclusiveLock);\n }\n\n+/*\n+ * Check and throw an error if the publisher has subscribed to the same table\n+ * from some other publisher. This check is required only if copydata is ON and\n+ * the origin is local for CREATE SUBSCRIPTION and\n+ * ALTER SUBSCRIPTION ... REFRESH statements to avoid replicating remote data\n+ * from the publisher.\n+ *\n+ * This check need not be performed on the tables that are already added as\n+ * incremental sync for such tables will happen through WAL and the origin of\n+ * the data can be identified from the WAL records.\n+ *\n+ * subrel_local_oids contains the list of relation oids that are already\n+ * present on the subscriber.\n+ */\n+static void\n+check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\n+ CopyData copydata, char *origin,\n+ Oid *subrel_local_oids, int subrel_count)\n\n1a.\n\"copydata is ON\" --> \"copy_data = on\" (because the comment is talking\nabout the CREATE/ALTER statements, so it seemed a bit confusing to\nrefer to the copydata function param instead of the copy_data\nsubscription parameter)\n\n1b.\n\"the origin is local\" ?? But, \"local\" was the old special name value.\nNow it is \"none\", so I think this part needs minor rewording.\n\n~~~\n\n2.\n\n+ if (copydata != COPY_DATA_ON || !origin ||\n+ (pg_strcasecmp(origin, \"none\") != 0))\n+ return;\n\nShould this be using the constant LOGICALREP_ORIGIN_NONE?\n\n~~~\n\n3.\n\n+ /*\n+ * Throw an error if the publisher has subscribed to the same table\n+ * from some other publisher. We cannot differentiate between the\n+ * origin and non-origin data that is present in the HEAP during the\n+ * initial sync. Identification of non-origin data can be done only\n+ * from the WAL by using the origin id.\n+ *\n+ * XXX: For simplicity, we don't check whether the table has any data\n+ * or not. If the table doesn't have any data then we don't need to\n+ * distinguish between local and non-local data so we can avoid\n+ * throwing an error in that case.\n+ */\n\n3a.\nWhen the special origin value changed from \"local\" to \"none\" this\ncomment's first part seems to have got a bit lost in translation.\n\nSUGGESTION:\nThrow an error if the publisher has subscribed to the same table from\nsome other publisher. We cannot know the origin of data during the\ninitial sync. Data origins can be found only from the WAL by looking\nat the origin id.\n\n3b.\nI think referring to \"local and non-local\" data in the XXX part of\nthis comment also needs some minor rewording now that \"local\" is not a\nspecial origin name anymore.\n\n------\n[1] https://github.com/postgres/postgres/commit/366283961ac0ed6d89014444c6090f3fd02fce0a\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 25 Jul 2022 17:28:40 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> On 7/22/22 12:47 AM, Amit Kapila wrote:\n> > On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> >> 1. I'm concerned by calling this \"Bidirectional replication\" in the docs\n> >> that we are overstating the current capabilities. I think this is\n> >> accentuated int he opening paragraph:\n> >>\n> >> ==snip==\n> >> Bidirectional replication is useful for creating a multi-master database\n> >> environment for replicating read/write operations performed by any of the\n> >> member nodes.\n> >> ==snip==\n> >>\n> >> For one, we're not replicating reads, we're replicating writes. Amongst\n> >> the writes, at this point we're only replicating DML. A reader could\n> >> think that deploying can work for a full bidirectional solution.\n> >>\n> >> (Even if we're aspirationally calling this section \"Bidirectional\n> >> replication\", that does make it sound like we're limited to two nodes,\n> >> when we can support more than two).\n> >>\n> >\n> > Right, I think the system can support N-Way replication.\n>\n> I did some more testing of the feature, i.e. doing 3-node and 4-node\n> tests. While logical replication today can handle replicating between\n> multiple nodes (N-way), the \"origin = none\" does require setting up\n> subscribers between each of the nodes.\n>\n> For example, if I have 4 nodes A, B, C, D and I want to replicate the\n> same table between all of them, I need to set up subscribers between all\n> of them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node\n> can replicate between each other in a way that's convenient (vs. having\n> to do something funky with partitions) so this is still a big step forward.\n>\n> This is a long way of saying that I do think it's fair to say we support\n> \"N-way\" replication so long as you are set up in a mesh (vs. a ring,\n> e.g. A=>B=>C=>D=>A).\n>\n> > Among the above \"Replicating changes between primaries\" sounds good to\n> > me or simply \"Replication between primaries\". As this is a sub-section\n> > on the Logical Replication page, I feel it is okay to not use Logical\n> > in the title.\n>\n> Agreed, I think that's fine.\n>\n> >> At a minimum, I think we should reference the documentation we have in\n> >> the logical replication section on conflicts. We may also want to advise\n> >> that a user is responsible for designing their schemas in a way to\n> >> minimize the risk of conflicts.\n> >>\n> >\n> > This sounds reasonable to me.\n> >\n> > One more point about docs, it appears to be added as the last\n> > sub-section on the Logical Replication page. Is there a reason for\n> > doing so? I feel this should be third sub-section after describing\n> > Publication and Subscription.\n>\n> When I first reviewed, I had not built the docs. Did so on this pass.\n>\n> I agree with the positioning argument, i.e. it should go after\n> \"Subscription\" in the table of contents -- but it makes me question a\n> couple of things:\n>\n> 1. The general ordering of the docs\n> 2. How we describe that section (more on that in a sec)\n> 3. If \"row filters\" should be part of \"subscription\" instead of its own\n> section.\n>\n> If you look at the current order, \"Quick setup\" is the last section; one\n> would think the \"quick\" portion goes first :) Given a lot of this is for\n> the current docs, I may start a separate discussion on -docs for this part.\n>\n> For the time being, I agree it should be moved to the section after\n> \"Subscription\".\n>\n> I think what this section describes is \"Configuring Replication Between\n> Nodes\" as it covers a few different scenarios.\n>\n> I do think we need to iterate on these docs -- the examples with the\n> commands are generally OK and easy to follow, but a few things I noticed:\n>\n> 1. The general description of the section needs work. We may want to\n> refine the description of the use cases, and in the warning, link to\n> instructions on how to take backups.\n> 2. We put the \"case not supported\" in the middle, not at the end.\n> 3. The \"generic steps for adding a new node...\" section uses a\n> convention for steps that is not found in the docs. We also don't\n> provide an example for this section, and this is the most complicated\n> scenario to set up.\n>\n> I may be able to propose some suggestions in a few days.\n>\n> > BTW, do you have any opinion on the idea of the first remaining patch\n> > where we accomplish two things: a) Checks and throws an error if\n> > 'copy_data = on' and 'origin = none' but the publication tables were\n> > also replicated from other publishers. b) Adds 'force' value for\n> > copy_data parameter to allow copying in such a case. The primary\n> > reason for this patch is to avoid loops or duplicate data in the\n> > initial phase. We can't skip copying based on origin as we can do\n> > while replicating changes from WAL. So, we detect that the publisher\n> > already has data from some other node and doesn't allow replication\n> > unless the user uses the 'force' option for copy_data.\n>\n> In general, I agree with the patch; but I'm not sure why we are calling\n> \"copy_data = force\" in this case and how it varies from \"on\". I\n> understand the goal is to prevent the infinite loop, but is there some\n> technical restriction why we can't set \"origin = none, copy_data = on\"\n> and have this work (and apologies if I missed that upthread)?\n\nLet's take a simple case to understand why copy_data = force is\nrequired to replicate between two primaries for table t1 which has\ndata as given below:\nNode-1:\nTable t1 (c1 int) has data\n1, 2, 3, 4\n\nNode-2:\nTable t1 (c1 int) has data\n5, 6, 7, 8\n\nstep1 - Node-1\n#Publication for t1\nCreate Publication pub1_2 For Table t1;\n\nstep2 - Node-2\n#Publication for t1,\nCreate Publication pub2_1 For Table t1;\n\nstep3 - Node-1:\nCreate Subscription sub1 Connection '<node-2 details>' publication\npub2_1 with (origin = none);\n\nAfter this the data will be something like this:\nNode-1:\n1, 2, 3, 4, 5, 6, 7, 8\n\nNode-2:\n5, 6, 7, 8\n\nstep4 - Node-2:\nCreate Subscription sub2 Connection '<node-1 details>' Publication\npub1_2 with (origin = none, copy_data=on);\nIf we had allowed the create subscription to be successful with\ncopy_data = on. After this the data will be something like this:\nNode-1:\n1, 2, 3, 4, 5, 6, 7, 8\n\nNode-2:\n1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\n\nSo, you can see that data on Node-2 (5, 6, 7, 8) is duplicated. In\ncase, table t1 has a unique key, it will lead to a unique key\nviolation and replication won't proceed.\n\nTo avoid this we will throw an error:\nERROR: could not replicate table \"public.t1\"\nDETAIL: CREATE/ALTER SUBSCRIPTION with origin = none and copy_data =\non is not allowed when the publisher has subscribed same table.\nHINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\n\nUsers can then overcome this problem by using the following steps:\nstep1 to step3 is the same as above.\n\nstep4 - Node-2\n# Disallow truncates to be published and then truncate the table\nAlter Publication pub2_1 Set (publish = 'insert, update, delete');\nTruncate t1;\n\nAfter this the data will be like this:\nNode-1:\n1, 2, 3, 4, 5, 6, 7, 8\n\nNode-2: no data\n\nstep5 - Node-2\nCreate Subscription sub2 Connection '<node-1 details>' Publication\npub1_2 with (origin = none, copy_data = force);\n\nAfter this the data will be in sync:\nNode-1:\n1, 2, 3, 4, 5, 6, 7, 8\n\nNode-2:\n1, 2, 3, 4, 5, 6, 7, 8\n\nstep6 - Node-1\n# Now include truncates to be published\nAlter Publication pub1_2 Set (publish = 'insert, update, delete, truncate');\n\nNow the replication setup between two primaries node1 and node2 is\ncomplete. Any incremental changes from node1 will be replicated to\nnode2, and any incremental changes from node2 will be replicated to\nnode1.\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 25 Jul 2022 14:24:00 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Firstly, I have some (case-sensitivity) questions about the previous\n> patch which was already pushed [1].\n>\n> Q1. create_subscription docs\n>\n> I did not understand why the docs refer to slot_name = NONE, yet the\n> newly added option says origin = none/any. I think that saying origin\n> = NONE/ANY would be more consistent with the existing usage of NONE in\n> this documentation.\n>\n> ~~~\n>\n> Q2. parse_subscription_options\n>\n> Similarly, in the code (parse_subscription_options), I did not\n> understand why the checks for special name values are implemented\n> differently:\n>\n> The new 'origin' code is using pg_strcmpcase to check special values\n> (none/any), and the old 'slot_name' code uses case-sensitive strcmp to\n> check the special value (none).\n>\n\nWe have a restriction for slot_names for lower case letters (aka\n\"Replication slot names may only contain lower case letters, numbers,\nand the underscore character.\") whereas there is no such restriction\nin origin name, that's why the check is different. So, if you try with\nslot_name = 'NONE', you will get the error.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 25 Jul 2022 14:24:43 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Firstly, I have some (case-sensitivity) questions about the previous\n> patch which was already pushed [1].\n>\n> Q1. create_subscription docs\n>\n> I did not understand why the docs refer to slot_name = NONE, yet the\n> newly added option says origin = none/any. I think that saying origin\n> = NONE/ANY would be more consistent with the existing usage of NONE in\n> this documentation.\n\nBoth NONE and none are ok in the case of origin, if you want I can\nchange it to NONE/ANY in case of origin to keep it consistent.\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 25 Jul 2022 15:03:25 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 25, 2022 at 2:24 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n> >\n> > On 7/22/22 12:47 AM, Amit Kapila wrote:\n> > > On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n> >\n> > >> 1. I'm concerned by calling this \"Bidirectional replication\" in the docs\n> > >> that we are overstating the current capabilities. I think this is\n> > >> accentuated int he opening paragraph:\n> > >>\n> > >> ==snip==\n> > >> Bidirectional replication is useful for creating a multi-master database\n> > >> environment for replicating read/write operations performed by any of the\n> > >> member nodes.\n> > >> ==snip==\n> > >>\n> > >> For one, we're not replicating reads, we're replicating writes. Amongst\n> > >> the writes, at this point we're only replicating DML. A reader could\n> > >> think that deploying can work for a full bidirectional solution.\n> > >>\n> > >> (Even if we're aspirationally calling this section \"Bidirectional\n> > >> replication\", that does make it sound like we're limited to two nodes,\n> > >> when we can support more than two).\n> > >>\n> > >\n> > > Right, I think the system can support N-Way replication.\n> >\n> > I did some more testing of the feature, i.e. doing 3-node and 4-node\n> > tests. While logical replication today can handle replicating between\n> > multiple nodes (N-way), the \"origin = none\" does require setting up\n> > subscribers between each of the nodes.\n> >\n> > For example, if I have 4 nodes A, B, C, D and I want to replicate the\n> > same table between all of them, I need to set up subscribers between all\n> > of them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node\n> > can replicate between each other in a way that's convenient (vs. having\n> > to do something funky with partitions) so this is still a big step forward.\n> >\n> > This is a long way of saying that I do think it's fair to say we support\n> > \"N-way\" replication so long as you are set up in a mesh (vs. a ring,\n> > e.g. A=>B=>C=>D=>A).\n> >\n> > > Among the above \"Replicating changes between primaries\" sounds good to\n> > > me or simply \"Replication between primaries\". As this is a sub-section\n> > > on the Logical Replication page, I feel it is okay to not use Logical\n> > > in the title.\n> >\n> > Agreed, I think that's fine.\n> >\n> > >> At a minimum, I think we should reference the documentation we have in\n> > >> the logical replication section on conflicts. We may also want to advise\n> > >> that a user is responsible for designing their schemas in a way to\n> > >> minimize the risk of conflicts.\n> > >>\n> > >\n> > > This sounds reasonable to me.\n> > >\n> > > One more point about docs, it appears to be added as the last\n> > > sub-section on the Logical Replication page. Is there a reason for\n> > > doing so? I feel this should be third sub-section after describing\n> > > Publication and Subscription.\n> >\n> > When I first reviewed, I had not built the docs. Did so on this pass.\n> >\n> > I agree with the positioning argument, i.e. it should go after\n> > \"Subscription\" in the table of contents -- but it makes me question a\n> > couple of things:\n> >\n> > 1. The general ordering of the docs\n> > 2. How we describe that section (more on that in a sec)\n> > 3. If \"row filters\" should be part of \"subscription\" instead of its own\n> > section.\n> >\n> > If you look at the current order, \"Quick setup\" is the last section; one\n> > would think the \"quick\" portion goes first :) Given a lot of this is for\n> > the current docs, I may start a separate discussion on -docs for this part.\n> >\n> > For the time being, I agree it should be moved to the section after\n> > \"Subscription\".\n> >\n> > I think what this section describes is \"Configuring Replication Between\n> > Nodes\" as it covers a few different scenarios.\n> >\n> > I do think we need to iterate on these docs -- the examples with the\n> > commands are generally OK and easy to follow, but a few things I noticed:\n> >\n> > 1. The general description of the section needs work. We may want to\n> > refine the description of the use cases, and in the warning, link to\n> > instructions on how to take backups.\n> > 2. We put the \"case not supported\" in the middle, not at the end.\n> > 3. The \"generic steps for adding a new node...\" section uses a\n> > convention for steps that is not found in the docs. We also don't\n> > provide an example for this section, and this is the most complicated\n> > scenario to set up.\n> >\n> > I may be able to propose some suggestions in a few days.\n> >\n> > > BTW, do you have any opinion on the idea of the first remaining patch\n> > > where we accomplish two things: a) Checks and throws an error if\n> > > 'copy_data = on' and 'origin = none' but the publication tables were\n> > > also replicated from other publishers. b) Adds 'force' value for\n> > > copy_data parameter to allow copying in such a case. The primary\n> > > reason for this patch is to avoid loops or duplicate data in the\n> > > initial phase. We can't skip copying based on origin as we can do\n> > > while replicating changes from WAL. So, we detect that the publisher\n> > > already has data from some other node and doesn't allow replication\n> > > unless the user uses the 'force' option for copy_data.\n> >\n> > In general, I agree with the patch; but I'm not sure why we are calling\n> > \"copy_data = force\" in this case and how it varies from \"on\". I\n> > understand the goal is to prevent the infinite loop, but is there some\n> > technical restriction why we can't set \"origin = none, copy_data = on\"\n> > and have this work (and apologies if I missed that upthread)?\n>\n> Let's take a simple case to understand why copy_data = force is\n> required to replicate between two primaries for table t1 which has\n> data as given below:\n> Node-1:\n> Table t1 (c1 int) has data\n> 1, 2, 3, 4\n>\n> Node-2:\n> Table t1 (c1 int) has data\n> 5, 6, 7, 8\n>\n> step1 - Node-1\n> #Publication for t1\n> Create Publication pub1_2 For Table t1;\n>\n> step2 - Node-2\n> #Publication for t1,\n> Create Publication pub2_1 For Table t1;\n>\n> step3 - Node-1:\n> Create Subscription sub1 Connection '<node-2 details>' publication\n> pub2_1 with (origin = none);\n>\n> After this the data will be something like this:\n> Node-1:\n> 1, 2, 3, 4, 5, 6, 7, 8\n>\n> Node-2:\n> 5, 6, 7, 8\n>\n> step4 - Node-2:\n> Create Subscription sub2 Connection '<node-1 details>' Publication\n> pub1_2 with (origin = none, copy_data=on);\n> If we had allowed the create subscription to be successful with\n> copy_data = on. After this the data will be something like this:\n> Node-1:\n> 1, 2, 3, 4, 5, 6, 7, 8\n>\n> Node-2:\n> 1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\n>\n> So, you can see that data on Node-2 (5, 6, 7, 8) is duplicated. In\n> case, table t1 has a unique key, it will lead to a unique key\n> violation and replication won't proceed.\n>\n> To avoid this we will throw an error:\n> ERROR: could not replicate table \"public.t1\"\n> DETAIL: CREATE/ALTER SUBSCRIPTION with origin = none and copy_data =\n> on is not allowed when the publisher has subscribed same table.\n> HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\n>\n> Users can then overcome this problem by using the following steps:\n> step1 to step3 is the same as above.\n>\n> step4 - Node-2\n> # Disallow truncates to be published and then truncate the table\n> Alter Publication pub2_1 Set (publish = 'insert, update, delete');\n> Truncate t1;\n>\n> After this the data will be like this:\n> Node-1:\n> 1, 2, 3, 4, 5, 6, 7, 8\n>\n> Node-2: no data\n>\n> step5 - Node-2\n> Create Subscription sub2 Connection '<node-1 details>' Publication\n> pub1_2 with (origin = none, copy_data = force);\n>\n> After this the data will be in sync:\n> Node-1:\n> 1, 2, 3, 4, 5, 6, 7, 8\n>\n> Node-2:\n> 1, 2, 3, 4, 5, 6, 7, 8\n>\n> step6 - Node-1\n> # Now include truncates to be published\n> Alter Publication pub1_2 Set (publish = 'insert, update, delete, truncate');\n>\n> Now the replication setup between two primaries node1 and node2 is\n> complete. Any incremental changes from node1 will be replicated to\n> node2, and any incremental changes from node2 will be replicated to\n> node1.\n\nIn the above steps, sorry that I mentioned Node-1 instead of Node-2 in\nthe last step step6.\nThe below:\n step6 - Node-1\n # Now include truncates to be published\n Alter Publication pub1_2 Set (publish = 'insert, update, delete, truncate');\n\nshould be:\nstep6 - Node-2\n# Now include truncates to be published\nAlter Publication pub2_1 Set (publish = 'insert, update, delete, truncate');\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 25 Jul 2022 20:14:28 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 25, 2022 at 6:54 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n...\n> >\n> > Q2. parse_subscription_options\n> >\n> > Similarly, in the code (parse_subscription_options), I did not\n> > understand why the checks for special name values are implemented\n> > differently:\n> >\n> > The new 'origin' code is using pg_strcmpcase to check special values\n> > (none/any), and the old 'slot_name' code uses case-sensitive strcmp to\n> > check the special value (none).\n> >\n>\n> We have a restriction for slot_names for lower case letters (aka\n> \"Replication slot names may only contain lower case letters, numbers,\n> and the underscore character.\") whereas there is no such restriction\n> in origin name, that's why the check is different. So, if you try with\n> slot_name = 'NONE', you will get the error.\n>\n\n2022-07-26 09:06:06.380 AEST [3630] STATEMENT: create subscription\nmysub connection 'host=localhost' publication mypub with\n(slot_name='None', enabled=false, create_slot=false);\nERROR: replication slot name \"None\" contains invalid character\nHINT: Replication slot names may only contain lower case letters,\nnumbers, and the underscore character.\n\nYou are right. Thanks for the explanation.\n\n(Aside: Probably that error message wording ought to say \"contains\ninvalid characters\" instead of \"contains invalid character\")\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 26 Jul 2022 09:13:17 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 25, 2022 at 7:33 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Firstly, I have some (case-sensitivity) questions about the previous\n> > patch which was already pushed [1].\n> >\n> > Q1. create_subscription docs\n> >\n> > I did not understand why the docs refer to slot_name = NONE, yet the\n> > newly added option says origin = none/any. I think that saying origin\n> > = NONE/ANY would be more consistent with the existing usage of NONE in\n> > this documentation.\n>\n> Both NONE and none are ok in the case of origin, if you want I can\n> change it to NONE/ANY in case of origin to keep it consistent.\n>\n\nI preferred the special origin values should be documented as NONE/ANY\nfor better consistency, but let's see what others think about it.\n\nThere will also be associated minor changes needed for a few\nerror/hint messages.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 26 Jul 2022 09:34:11 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On 7/25/22 4:54 AM, vignesh C wrote:\r\n> On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\r\n>>\r\n>> On 7/22/22 12:47 AM, Amit Kapila wrote:\r\n>>> On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\r\n\r\n>>> BTW, do you have any opinion on the idea of the first remaining patch\r\n>>> where we accomplish two things: a) Checks and throws an error if\r\n>>> 'copy_data = on' and 'origin = none' but the publication tables were\r\n>>> also replicated from other publishers. b) Adds 'force' value for\r\n>>> copy_data parameter to allow copying in such a case. The primary\r\n>>> reason for this patch is to avoid loops or duplicate data in the\r\n>>> initial phase. We can't skip copying based on origin as we can do\r\n>>> while replicating changes from WAL. So, we detect that the publisher\r\n>>> already has data from some other node and doesn't allow replication\r\n>>> unless the user uses the 'force' option for copy_data.\r\n>>\r\n>> In general, I agree with the patch; but I'm not sure why we are calling\r\n>> \"copy_data = force\" in this case and how it varies from \"on\". I\r\n>> understand the goal is to prevent the infinite loop, but is there some\r\n>> technical restriction why we can't set \"origin = none, copy_data = on\"\r\n>> and have this work (and apologies if I missed that upthread)?\r\n> \r\n> Let's take a simple case to understand why copy_data = force is\r\n> required to replicate between two primaries for table t1 which has\r\n> data as given below:\r\n\r\n> step4 - Node-2:\r\n> Create Subscription sub2 Connection '<node-1 details>' Publication\r\n> pub1_2 with (origin = none, copy_data=on);\r\n> If we had allowed the create subscription to be successful with\r\n> copy_data = on. After this the data will be something like this:\r\n> Node-1:\r\n> 1, 2, 3, 4, 5, 6, 7, 8\r\n> \r\n> Node-2:\r\n> 1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\r\n> \r\n> So, you can see that data on Node-2 (5, 6, 7, 8) is duplicated. In\r\n> case, table t1 has a unique key, it will lead to a unique key\r\n> violation and replication won't proceed.\r\n> \r\n> To avoid this we will throw an error:\r\n> ERROR: could not replicate table \"public.t1\"\r\n> DETAIL: CREATE/ALTER SUBSCRIPTION with origin = none and copy_data =\r\n> on is not allowed when the publisher has subscribed same table.\r\n> HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\r\n\r\nThanks for the example. I agree that it is fairly simple to reproduce.\r\n\r\nI understand that \"copy_data = force\" is meant to protect a user from \r\nhurting themself. I'm not convinced that this is the best way to do so.\r\n\r\nFor example today I can subscribe to multiple publications that write to \r\nthe same table. If I have a primary key on that table, and two of the \r\nsubscriptions try to write an identical ID, we conflict. We don't have \r\nany special flags or modes to guard against that from happening, though \r\nwe do have documentation on conflicts and managing them.\r\n\r\nAFAICT the same issue with \"copy_data\" also exists in the above scenario \r\ntoo, even without the \"origin\" attribute. However, I think this case is \r\nmore noticeable for \"origin=none\" because we currently default \r\n\"copy_data\" to \"true\" and in this case data can be copied in two \r\ndirections.\r\n\r\nThat said, this introduces a new restriction for this particular \r\nscenario that doesn't exist on other scenarios. Instead, I would \r\nadvocate we document how to correctly set up the two-way replication \r\nscenario (which we have a draft!), document the warnings around the \r\nconflicts, perhaps include Vignesh's instructions on how to remediate a \r\nconflict on initial sync, and consider throwing a WARNING as you suggested.\r\n\r\nThoughts?\r\n\r\nThanks,\r\n\r\nJonathan","msg_date":"Mon, 25 Jul 2022 21:43:37 -0400","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are some review comments for the patch v38-0002:\n\n======\n\n<General> - terminology\n\nThere seemed to be an inconsistent alternation of the terms\n\"primaries\" and \"nodes\"... For example \"Setting replication between\ntwo primaries\" versus \"Adding a new node...\" (instead of \"Adding a new\nprimary...\"?). I have included suggested minor rewording changes in\nthe review comments below, but please check in case I miss something.\nBecause I suggested changes to some titles maybe you will also want to\nchange the section ids too.\n\n~~~\n\n1. Commit message\n\nThe documentation was recently modified to remove the term\n\"bidirectional replication\" and replace it all with \"replication\nbetween primaries\", so this commit message (and also the patch name\nitself) should be similarly modified.\n\n~~~\n\n2.\n+ <para>\n+ Replication between primaries is useful for creating a multi-master\n+ database environment for replicating write operations performed by any of\n+ the member nodes. The steps to create replication between primaries in\n+ various scenarios are given below. Note: User is responsible for designing\n+ their schemas in a way to minimize the risk of conflicts. See\n+ <xref linkend=\"logical-replication-conflicts\"/> for the details of logical\n+ replication conflicts. The logical replication restrictions applies to\n+ the replication between primaries also. See\n+ <xref linkend=\"logical-replication-restrictions\"/> for the details of\n+ logical replication restrictions.\n+ </para>\n\n2a.\n\"User\" -> \"The user\"\n\n2b.\n\"The logical replication restrictions applies to...\" --> \"The logical\nreplication restrictions apply to...\"\n\n2c.\nThese are important notes. Instead of just being part of the text\nblurb, perhaps these should be rendered as SGML <note> (or put them\nboth in a single <note> if you want)\n\n~~~\n\n3. Setting replication between two primaries\n\n+ <title>Setting replication between two primaries</title>\n+ <para>\n+ The following steps demonstrate how to setup replication between two\n+ primaries when there is no table data present on both nodes\n+ <literal>node1</literal> and <literal>node2</literal>:\n+ </para>\n\nSUGGESTED\nThe following steps demonstrate how to set up replication between two\nprimaries (node1 and node2) when there is no table data present on\nboth nodes:\n\n~~~\n\n4.\n+ <para>\n+ Now the replication setup between two primaries <literal>node1</literal>\n+ and <literal>node2</literal> is complete. Any incremental changes from\n+ <literal>node1</literal> will be replicated to <literal>node2</literal>,\n+ and any incremental changes from <literal>node2</literal> will be\n+ replicated to <literal>node1</literal>.\n+ </para>\n\n\"between two primaries\" -> \"between primaries\"\n\n~~~\n\n5. Adding a new node when there is no table data on any of the nodes\n\nSUGGESTION (title)\nAdding a new primary when there is no table data on any of the primaries\n\n~~~\n\n6.\n+ <para>\n+ The following steps demonstrate adding a new node <literal>node3</literal>\n+ to the existing <literal>node1</literal> and <literal>node2</literal> when\n+ there is no <literal>t1</literal> data on any of the nodes. This requires\n\nSUGGESTION\nThe following steps demonstrate adding a new primary (node3) to the\nexisting primaries (node1 and node2) when there is no t1 data on any\nof the nodes.\n\n~~~\n\n7. Adding a new node when table data is present on the existing nodes\n\nSUGGESTION (title)\nAdding a new primary when table data is present on the existing primaries\n\n~~~\n\n8.\n+ <para>\n+ The following steps demonstrate adding a new node <literal>node3</literal>\n+ which has no <literal>t1</literal> data to the existing\n+ <literal>node1</literal> and <literal>node2</literal> where\n+ <literal>t1</literal> data is present. This needs similar steps; the only\n\nSUGGESTION\nThe following steps demonstrate adding a new primary (node3) that has\nno t1 data to the existing primaries (node1 and node2) where t1 data\nis present.\n\n~~~\n\n9. Adding a new node when table data is present on the new node\n\nSUGGESTION (title)\nAdding a new primary that has existing table data\n\n~~~\n\n10.\n+ <note>\n+ <para>\n+ Adding a new node when table data is present on the new node is not\n+ supported.\n+ </para>\n+ </note>\n\nSUGGESTION\nAdding a new primary that has existing table data is not supported.\n\n~~~\n\n11. Generic steps for adding a new node to an existing set of primaries\n\nSUGGESTION (title)\nGeneric steps for adding a new primary to an existing set of primaries\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 26 Jul 2022 11:45:52 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 11:43 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n...\n> That said, this introduces a new restriction for this particular\n> scenario that doesn't exist on other scenarios. Instead, I would\n> advocate we document how to correctly set up the two-way replication\n> scenario (which we have a draft!), document the warnings around the\n> conflicts, perhaps include Vignesh's instructions on how to remediate a\n> conflict on initial sync, and consider throwing a WARNING as you suggested.\n>\n> Thoughts?\n\nPerhaps a WARNING can be useful if the SUBSCRIPTION was created with\nenabled=false because then the user still has a chance to reconsider,\nbut otherwise, I don't see what good a warning does if the potentially\nharmful initial copy is going to proceed anyway; isn't that like\nputting a warning sign at the bottom of a cliff?\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 26 Jul 2022 12:18:05 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 7:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> On 7/25/22 4:54 AM, vignesh C wrote:\n> >\n> > Let's take a simple case to understand why copy_data = force is\n> > required to replicate between two primaries for table t1 which has\n> > data as given below:\n>\n> > step4 - Node-2:\n> > Create Subscription sub2 Connection '<node-1 details>' Publication\n> > pub1_2 with (origin = none, copy_data=on);\n> > If we had allowed the create subscription to be successful with\n> > copy_data = on. After this the data will be something like this:\n> > Node-1:\n> > 1, 2, 3, 4, 5, 6, 7, 8\n> >\n> > Node-2:\n> > 1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\n> >\n> > So, you can see that data on Node-2 (5, 6, 7, 8) is duplicated. In\n> > case, table t1 has a unique key, it will lead to a unique key\n> > violation and replication won't proceed.\n> >\n> > To avoid this we will throw an error:\n> > ERROR: could not replicate table \"public.t1\"\n> > DETAIL: CREATE/ALTER SUBSCRIPTION with origin = none and copy_data =\n> > on is not allowed when the publisher has subscribed same table.\n> > HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\n>\n> Thanks for the example. I agree that it is fairly simple to reproduce.\n>\n> I understand that \"copy_data = force\" is meant to protect a user from\n> hurting themself. I'm not convinced that this is the best way to do so.\n>\n> For example today I can subscribe to multiple publications that write to\n> the same table. If I have a primary key on that table, and two of the\n> subscriptions try to write an identical ID, we conflict. We don't have\n> any special flags or modes to guard against that from happening, though\n> we do have documentation on conflicts and managing them.\n>\n> AFAICT the same issue with \"copy_data\" also exists in the above scenario\n> too, even without the \"origin\" attribute.\n>\n\nThat's true but there is no parameter like origin = NONE which\nindicates that constraint violations or duplicate data problems won't\noccur due to replication. In the current case, I think the situation\nis different because a user has specifically asked not to replicate\nany remote data by specifying origin = NONE, which should be dealt\ndifferently. Note that current users or their setup won't see any\ndifference/change unless they specify the new parameter origin as\nNONE.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 26 Jul 2022 09:07:04 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 7:48 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Tue, Jul 26, 2022 at 11:43 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n> >\n> ...\n> > That said, this introduces a new restriction for this particular\n> > scenario that doesn't exist on other scenarios. Instead, I would\n> > advocate we document how to correctly set up the two-way replication\n> > scenario (which we have a draft!), document the warnings around the\n> > conflicts, perhaps include Vignesh's instructions on how to remediate a\n> > conflict on initial sync, and consider throwing a WARNING as you suggested.\n> >\n> > Thoughts?\n>\n> Perhaps a WARNING can be useful if the SUBSCRIPTION was created with\n> enabled=false because then the user still has a chance to reconsider,\n>\n\nAgreed. I think but in that case, when the user enables it, we need to\nensure that we won't allow replicating (during initial sync) remote\ndata. If this is really required/preferred, it can be done as a\nseparate enhancement.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 26 Jul 2022 09:12:23 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 5:04 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Mon, Jul 25, 2022 at 7:33 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> > >\n> > > Firstly, I have some (case-sensitivity) questions about the previous\n> > > patch which was already pushed [1].\n> > >\n> > > Q1. create_subscription docs\n> > >\n> > > I did not understand why the docs refer to slot_name = NONE, yet the\n> > > newly added option says origin = none/any. I think that saying origin\n> > > = NONE/ANY would be more consistent with the existing usage of NONE in\n> > > this documentation.\n> >\n> > Both NONE and none are ok in the case of origin, if you want I can\n> > change it to NONE/ANY in case of origin to keep it consistent.\n> >\n>\n> I preferred the special origin values should be documented as NONE/ANY\n> for better consistency, but let's see what others think about it.\n>\n> There will also be associated minor changes needed for a few\n> error/hint messages.\n>\n\nI am not really sure how much we gain by maintaining consistency with\nslot_name because if due to this we have to change the error messages\nas well then it can create an inconsistency with reserved origin\nnames. Consider message: DETAIL: Origin names \"any\", \"none\", and\nnames starting with \"pg_\" are reserved. Now, if we change this to\n\"ANY\", \"NONE\" in the above message, it will look a bit odd as \"pg_\"\nstarts with lower case letters.\n\n--\nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 26 Jul 2022 09:39:01 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 2:09 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Jul 26, 2022 at 5:04 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > On Mon, Jul 25, 2022 at 7:33 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> > > >\n> > > > Firstly, I have some (case-sensitivity) questions about the previous\n> > > > patch which was already pushed [1].\n> > > >\n> > > > Q1. create_subscription docs\n> > > >\n> > > > I did not understand why the docs refer to slot_name = NONE, yet the\n> > > > newly added option says origin = none/any. I think that saying origin\n> > > > = NONE/ANY would be more consistent with the existing usage of NONE in\n> > > > this documentation.\n> > >\n> > > Both NONE and none are ok in the case of origin, if you want I can\n> > > change it to NONE/ANY in case of origin to keep it consistent.\n> > >\n> >\n> > I preferred the special origin values should be documented as NONE/ANY\n> > for better consistency, but let's see what others think about it.\n> >\n> > There will also be associated minor changes needed for a few\n> > error/hint messages.\n> >\n>\n> I am not really sure how much we gain by maintaining consistency with\n> slot_name because if due to this we have to change the error messages\n> as well then it can create an inconsistency with reserved origin\n> names. Consider message: DETAIL: Origin names \"any\", \"none\", and\n> names starting with \"pg_\" are reserved. Now, if we change this to\n> \"ANY\", \"NONE\" in the above message, it will look a bit odd as \"pg_\"\n> starts with lower case letters.\n>\n\nSure, the message looks a bit odd with the quotes like you wrote\nabove, but I would not suggest to change it that way - I was thinking\nmore like below (which is similar to the style the slot_name messages\nuse)\n\nCURRENT\nDETAIL: Origin names \"any\", \"none\", and names starting with \"pg_\" are reserved.\n\nSUGGESTED\nDETAIL: Origin names ANY, NONE, and names starting with \"pg_\" are reserved.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 26 Jul 2022 14:53:05 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sun, Jul 24, 2022 1:28 AM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> Added a note for the same and referred it to the conflicts section.\r\n> \r\n> Thanks for the comments, the attached v38 patch has the changes for the\r\n> same.\r\n> \r\n\r\nThanks for updating the patch. A comment on the test in 0001 patch.\r\n\r\n+# Alter subscription ... refresh publication should fail when a new table is\r\n+# subscribing data from a different publication should fail\r\n+($result, $stdout, $stderr) = $node_A->psql(\r\n+\t'postgres', \"\r\n+ ALTER SUBSCRIPTION tap_sub_A2 REFRESH PUBLICATION\");\r\n+like(\r\n+\t$stderr,\r\n+\tqr/ERROR: ( [A-Z0-9]+:)? could not replicate table \"public.tab_new\"/,\r\n+\t\"Create subscription with origin and copy_data having replicated table in publisher\"\r\n+);\r\n\r\nThe comment says \"should fail\" twice, the latter one can be removed.\r\n\r\nBesides, \"Create subscription with origin and copy_data\" should be changed to\r\n\"Alter subscription with origin and copy_data\" I think.\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Tue, 26 Jul 2022 08:05:05 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sun, Jul 24, 2022 1:28 AM vignesh C <vignesh21@gmail.com> wrote:\r\n> Added a note for the same and referred it to the conflicts section.\r\n> \r\n> Thanks for the comments, the attached v38 patch has the changes for the same.\r\n\r\nThanks for your patches.\r\n\r\nTwo slight comments on the below message in the 0001 patch:\r\n\r\nThe error message in the function check_pub_table_subscribed().\r\n+\t\tereport(ERROR,\r\n+\t\t\t\terrcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\r\n+\t\t\t\terrmsg(\"could not replicate table \\\"%s.%s\\\"\",\r\n+\t\t\t\t\t nspname, relname),\r\n+\t\t\t\terrdetail(\"CREATE/ALTER SUBSCRIPTION with origin = none and copy_data = on is not allowed when the publisher has subscribed same table.\"),\r\n+\t\t\t\terrhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\"));\r\n\r\n1.\r\nI think it might be better to use true/false here than on/off.\r\nJust for consistency with another error message\r\n(in function parse_subscription_options) and the description of this parameter\r\nin the PG document.\r\n\r\nIf you agree with this, please also kindly consider the attached\r\n\"slight_modification.diff\" file.\r\n(This is a slight modification for the second patch, just replace \"off\" with\r\n\"false\" in PG document.)\r\n\r\n2.\r\nHow about replacing \"origin = none\" and \"copy_data = on\" in the message with\r\n\"%s\"? I think this might be better for translation. Just like the following\r\nerror message in the function parse_subscription_options:\r\n```\r\n\t\tif (opts->copy_data &&\r\n\t\t\tIsSet(opts->specified_opts, SUBOPT_COPY_DATA))\r\n\t\t\tereport(ERROR,\r\n\t\t\t\t\t(errcode(ERRCODE_SYNTAX_ERROR),\r\n\t\t\t\t\t errmsg(\"%s and %s are mutually exclusive options\",\r\n\t\t\t\t\t\t\t\"connect = false\", \"copy_data = true/force\")));\r\n```\r\n\r\nRegards,\r\nWang wei","msg_date":"Tue, 26 Jul 2022 08:42:09 +0000","msg_from":"\"wangw.fnst@fujitsu.com\" <wangw.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> On 7/22/22 12:47 AM, Amit Kapila wrote:\n> > On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> >> 1. I'm concerned by calling this \"Bidirectional replication\" in the docs\n> >> that we are overstating the current capabilities. I think this is\n> >> accentuated int he opening paragraph:\n> >>\n> >> ==snip==\n> >> Bidirectional replication is useful for creating a multi-master database\n> >> environment for replicating read/write operations performed by any of the\n> >> member nodes.\n> >> ==snip==\n> >>\n> >> For one, we're not replicating reads, we're replicating writes. Amongst\n> >> the writes, at this point we're only replicating DML. A reader could\n> >> think that deploying can work for a full bidirectional solution.\n> >>\n> >> (Even if we're aspirationally calling this section \"Bidirectional\n> >> replication\", that does make it sound like we're limited to two nodes,\n> >> when we can support more than two).\n> >>\n> >\n> > Right, I think the system can support N-Way replication.\n>\n> I did some more testing of the feature, i.e. doing 3-node and 4-node\n> tests. While logical replication today can handle replicating between\n> multiple nodes (N-way), the \"origin = none\" does require setting up\n> subscribers between each of the nodes.\n>\n> For example, if I have 4 nodes A, B, C, D and I want to replicate the\n> same table between all of them, I need to set up subscribers between all\n> of them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node\n> can replicate between each other in a way that's convenient (vs. having\n> to do something funky with partitions) so this is still a big step forward.\n>\n> This is a long way of saying that I do think it's fair to say we support\n> \"N-way\" replication so long as you are set up in a mesh (vs. a ring,\n> e.g. A=>B=>C=>D=>A).\n>\n> > Among the above \"Replicating changes between primaries\" sounds good to\n> > me or simply \"Replication between primaries\". As this is a sub-section\n> > on the Logical Replication page, I feel it is okay to not use Logical\n> > in the title.\n>\n> Agreed, I think that's fine.\n>\n> >> At a minimum, I think we should reference the documentation we have in\n> >> the logical replication section on conflicts. We may also want to advise\n> >> that a user is responsible for designing their schemas in a way to\n> >> minimize the risk of conflicts.\n> >>\n> >\n> > This sounds reasonable to me.\n> >\n> > One more point about docs, it appears to be added as the last\n> > sub-section on the Logical Replication page. Is there a reason for\n> > doing so? I feel this should be third sub-section after describing\n> > Publication and Subscription.\n>\n> When I first reviewed, I had not built the docs. Did so on this pass.\n>\n> I agree with the positioning argument, i.e. it should go after\n> \"Subscription\" in the table of contents -- but it makes me question a\n> couple of things:\n>\n> 1. The general ordering of the docs\n> 2. How we describe that section (more on that in a sec)\n> 3. If \"row filters\" should be part of \"subscription\" instead of its own\n> section.\n>\n> If you look at the current order, \"Quick setup\" is the last section; one\n> would think the \"quick\" portion goes first :) Given a lot of this is for\n> the current docs, I may start a separate discussion on -docs for this part.\n>\n> For the time being, I agree it should be moved to the section after\n> \"Subscription\".\n>\n> I think what this section describes is \"Configuring Replication Between\n> Nodes\" as it covers a few different scenarios.\n>\n> I do think we need to iterate on these docs -- the examples with the\n> commands are generally OK and easy to follow, but a few things I noticed:\n>\n> 1. The general description of the section needs work. We may want to\n> refine the description of the use cases, and in the warning, link to\n> instructions on how to take backups.\n\nModified\n\n> 2. We put the \"case not supported\" in the middle, not at the end.\n\nModified\n\n> 3. The \"generic steps for adding a new node...\" section uses a\n> convention for steps that is not found in the docs. We also don't\n> provide an example for this section, and this is the most complicated\n> scenario to set up.\n\nModified\n\nThanks a lot for the suggestions, I have made the changes for the\nsame in the v39 patch attached.\n\nRegards,\nVignesh","msg_date":"Tue, 26 Jul 2022 23:27:40 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 7:16 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some review comments for the patch v38-0002:\n>\n> ======\n>\n> <General> - terminology\n>\n> There seemed to be an inconsistent alternation of the terms\n> \"primaries\" and \"nodes\"... For example \"Setting replication between\n> two primaries\" versus \"Adding a new node...\" (instead of \"Adding a new\n> primary...\"?). I have included suggested minor rewording changes in\n> the review comments below, but please check in case I miss something.\n> Because I suggested changes to some titles maybe you will also want to\n> change the section ids too.\n>\n> ~~~\n>\n> 1. Commit message\n>\n> The documentation was recently modified to remove the term\n> \"bidirectional replication\" and replace it all with \"replication\n> between primaries\", so this commit message (and also the patch name\n> itself) should be similarly modified.\n\nModified\n\n> ~~~\n>\n> 2.\n> + <para>\n> + Replication between primaries is useful for creating a multi-master\n> + database environment for replicating write operations performed by any of\n> + the member nodes. The steps to create replication between primaries in\n> + various scenarios are given below. Note: User is responsible for designing\n> + their schemas in a way to minimize the risk of conflicts. See\n> + <xref linkend=\"logical-replication-conflicts\"/> for the details of logical\n> + replication conflicts. The logical replication restrictions applies to\n> + the replication between primaries also. See\n> + <xref linkend=\"logical-replication-restrictions\"/> for the details of\n> + logical replication restrictions.\n> + </para>\n>\n> 2a.\n> \"User\" -> \"The user\"\n\nModified\n\n> 2b.\n> \"The logical replication restrictions applies to...\" --> \"The logical\n> replication restrictions apply to...\"\n\nModified\n\n> 2c.\n> These are important notes. Instead of just being part of the text\n> blurb, perhaps these should be rendered as SGML <note> (or put them\n> both in a single <note> if you want)\n\nModified\n\n> ~~~\n>\n> 3. Setting replication between two primaries\n>\n> + <title>Setting replication between two primaries</title>\n> + <para>\n> + The following steps demonstrate how to setup replication between two\n> + primaries when there is no table data present on both nodes\n> + <literal>node1</literal> and <literal>node2</literal>:\n> + </para>\n>\n> SUGGESTED\n> The following steps demonstrate how to set up replication between two\n> primaries (node1 and node2) when there is no table data present on\n> both nodes:.\n\nModified\n\n> ~~~\n>\n> 4.\n> + <para>\n> + Now the replication setup between two primaries <literal>node1</literal>\n> + and <literal>node2</literal> is complete. Any incremental changes from\n> + <literal>node1</literal> will be replicated to <literal>node2</literal>,\n> + and any incremental changes from <literal>node2</literal> will be\n> + replicated to <literal>node1</literal>.\n> + </para>\n>\n> \"between two primaries\" -> \"between primaries\"\n\nModified\n\n> ~~~\n>\n> 5. Adding a new node when there is no table data on any of the nodes\n>\n> SUGGESTION (title)\n> Adding a new primary when there is no table data on any of the primaries\n\nModified\n\n> ~~~\n>\n> 6.\n> + <para>\n> + The following steps demonstrate adding a new node <literal>node3</literal>\n> + to the existing <literal>node1</literal> and <literal>node2</literal> when\n> + there is no <literal>t1</literal> data on any of the nodes. This requires\n>\n> SUGGESTION\n> The following steps demonstrate adding a new primary (node3) to the\n> existing primaries (node1 and node2) when there is no t1 data on any\n> of the nodes.\n\nModified\n\n> ~~~\n>\n> 7. Adding a new node when table data is present on the existing nodes\n>\n> SUGGESTION (title)\n> Adding a new primary when table data is present on the existing primaries\n\nModified\n\n> ~~~\n>\n> 8.\n> + <para>\n> + The following steps demonstrate adding a new node <literal>node3</literal>\n> + which has no <literal>t1</literal> data to the existing\n> + <literal>node1</literal> and <literal>node2</literal> where\n> + <literal>t1</literal> data is present. This needs similar steps; the only\n>\n> SUGGESTION\n> The following steps demonstrate adding a new primary (node3) that has\n> no t1 data to the existing primaries (node1 and node2) where t1 data\n> is present.\n\nModified\n\n> ~~~\n>\n> 9. Adding a new node when table data is present on the new node\n>\n> SUGGESTION (title)\n> Adding a new primary that has existing table data\n\nModified\n\n> ~~~\n>\n> 10.\n> + <note>\n> + <para>\n> + Adding a new node when table data is present on the new node is not\n> + supported.\n> + </para>\n> + </note>\n>\n> SUGGESTION\n> Adding a new primary that has existing table data is not supported.\n\nModified\n\n> ~~~\n>\n> 11. Generic steps for adding a new node to an existing set of primaries\n>\n> SUGGESTION (title)\n> Generic steps for adding a new primary to an existing set of primaries\n\nModified\n\nThanks for the comments, the v39 patch shared at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm2POATc_jwQ-8MBJgGCVZGdUNhnTv8zkBuGzLaY03dM%3DA%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 26 Jul 2022 23:34:50 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 1:35 PM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Sun, Jul 24, 2022 1:28 AM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Added a note for the same and referred it to the conflicts section.\n> >\n> > Thanks for the comments, the attached v38 patch has the changes for the\n> > same.\n> >\n>\n> Thanks for updating the patch. A comment on the test in 0001 patch.\n>\n> +# Alter subscription ... refresh publication should fail when a new table is\n> +# subscribing data from a different publication should fail\n> +($result, $stdout, $stderr) = $node_A->psql(\n> + 'postgres', \"\n> + ALTER SUBSCRIPTION tap_sub_A2 REFRESH PUBLICATION\");\n> +like(\n> + $stderr,\n> + qr/ERROR: ( [A-Z0-9]+:)? could not replicate table \"public.tab_new\"/,\n> + \"Create subscription with origin and copy_data having replicated table in publisher\"\n> +);\n>\n> The comment says \"should fail\" twice, the latter one can be removed.\n\nModified\n\n> Besides, \"Create subscription with origin and copy_data\" should be changed to\n> \"Alter subscription with origin and copy_data\" I think.\n\nModified to \"Refresh publication\"\n\nThanks for the comments, the v39 patch shared at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm2POATc_jwQ-8MBJgGCVZGdUNhnTv8zkBuGzLaY03dM%3DA%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 26 Jul 2022 23:37:10 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 2:12 PM wangw.fnst@fujitsu.com\n<wangw.fnst@fujitsu.com> wrote:\n>\n> On Sun, Jul 24, 2022 1:28 AM vignesh C <vignesh21@gmail.com> wrote:\n> > Added a note for the same and referred it to the conflicts section.\n> >\n> > Thanks for the comments, the attached v38 patch has the changes for the same.\n>\n> Thanks for your patches.\n>\n> Two slight comments on the below message in the 0001 patch:\n>\n> The error message in the function check_pub_table_subscribed().\n> + ereport(ERROR,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"could not replicate table \\\"%s.%s\\\"\",\n> + nspname, relname),\n> + errdetail(\"CREATE/ALTER SUBSCRIPTION with origin = none and copy_data = on is not allowed when the publisher has subscribed same table.\"),\n> + errhint(\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\"));\n>\n> 1.\n> I think it might be better to use true/false here than on/off.\n> Just for consistency with another error message\n> (in function parse_subscription_options) and the description of this parameter\n> in the PG document.\n\nModified\n\n> If you agree with this, please also kindly consider the attached\n> \"slight_modification.diff\" file.\n> (This is a slight modification for the second patch, just replace \"off\" with\n> \"false\" in PG document.)\n\nI have updated the documentation similarly\n\n> 2.\n> How about replacing \"origin = none\" and \"copy_data = on\" in the message with\n> \"%s\"? I think this might be better for translation. Just like the following\n> error message in the function parse_subscription_options:\n> ```\n> if (opts->copy_data &&\n> IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\n> ereport(ERROR,\n> (errcode(ERRCODE_SYNTAX_ERROR),\n> errmsg(\"%s and %s are mutually exclusive options\",\n> \"connect = false\", \"copy_data = true/force\")));\n> ```\n\nModified\n\nThanks for the comments, the v39 patch shared at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm2POATc_jwQ-8MBJgGCVZGdUNhnTv8zkBuGzLaY03dM%3DA%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 26 Jul 2022 23:38:04 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Firstly, I have some (case-sensitivity) questions about the previous\n> patch which was already pushed [1].\n>\n> Q1. create_subscription docs\n>\n> I did not understand why the docs refer to slot_name = NONE, yet the\n> newly added option says origin = none/any. I think that saying origin\n> = NONE/ANY would be more consistent with the existing usage of NONE in\n> this documentation.\n>\n> ~~~\n>\n> Q2. parse_subscription_options\n>\n> Similarly, in the code (parse_subscription_options), I did not\n> understand why the checks for special name values are implemented\n> differently:\n>\n> The new 'origin' code is using pg_strcmpcase to check special values\n> (none/any), and the old 'slot_name' code uses case-sensitive strcmp to\n> check the special value (none).\n>\n> FWIW, here I thought the new origin code is the correct one.\n>\n> ======\n>\n> Now, here are some review comments for the patch v38-0001:\n>\n> 1. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> @@ -1781,6 +1858,121 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\n> table_close(rel, RowExclusiveLock);\n> }\n>\n> +/*\n> + * Check and throw an error if the publisher has subscribed to the same table\n> + * from some other publisher. This check is required only if copydata is ON and\n> + * the origin is local for CREATE SUBSCRIPTION and\n> + * ALTER SUBSCRIPTION ... REFRESH statements to avoid replicating remote data\n> + * from the publisher.\n> + *\n> + * This check need not be performed on the tables that are already added as\n> + * incremental sync for such tables will happen through WAL and the origin of\n> + * the data can be identified from the WAL records.\n> + *\n> + * subrel_local_oids contains the list of relation oids that are already\n> + * present on the subscriber.\n> + */\n> +static void\n> +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\n> + CopyData copydata, char *origin,\n> + Oid *subrel_local_oids, int subrel_count)\n>\n> 1a.\n> \"copydata is ON\" --> \"copy_data = on\" (because the comment is talking\n> about the CREATE/ALTER statements, so it seemed a bit confusing to\n> refer to the copydata function param instead of the copy_data\n> subscription parameter)\n\nModified\n\n> 1b.\n> \"the origin is local\" ?? But, \"local\" was the old special name value.\n> Now it is \"none\", so I think this part needs minor rewording.\n\nModified\n\n>\n> ~~~\n>\n> 2.\n>\n> + if (copydata != COPY_DATA_ON || !origin ||\n> + (pg_strcasecmp(origin, \"none\") != 0))\n> + return;\n>\n> Should this be using the constant LOGICALREP_ORIGIN_NONE?\n\nModified\n\n> ~~~\n>\n> 3.\n>\n> + /*\n> + * Throw an error if the publisher has subscribed to the same table\n> + * from some other publisher. We cannot differentiate between the\n> + * origin and non-origin data that is present in the HEAP during the\n> + * initial sync. Identification of non-origin data can be done only\n> + * from the WAL by using the origin id.\n> + *\n> + * XXX: For simplicity, we don't check whether the table has any data\n> + * or not. If the table doesn't have any data then we don't need to\n> + * distinguish between local and non-local data so we can avoid\n> + * throwing an error in that case.\n> + */\n>\n> 3a.\n> When the special origin value changed from \"local\" to \"none\" this\n> comment's first part seems to have got a bit lost in translation.\n\nModified\n\n> SUGGESTION:\n> Throw an error if the publisher has subscribed to the same table from\n> some other publisher. We cannot know the origin of data during the\n> initial sync. Data origins can be found only from the WAL by looking\n> at the origin id.\n\nModified\n\n> 3b.\n> I think referring to \"local and non-local\" data in the XXX part of\n> this comment also needs some minor rewording now that \"local\" is not a\n> special origin name anymore.\n\nModified\n\nThanks for the comments, the v39 patch shared at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm2POATc_jwQ-8MBJgGCVZGdUNhnTv8zkBuGzLaY03dM%3DA%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 26 Jul 2022 23:43:53 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 10:23 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Tue, Jul 26, 2022 at 2:09 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> >\n> > I am not really sure how much we gain by maintaining consistency with\n> > slot_name because if due to this we have to change the error messages\n> > as well then it can create an inconsistency with reserved origin\n> > names. Consider message: DETAIL: Origin names \"any\", \"none\", and\n> > names starting with \"pg_\" are reserved. Now, if we change this to\n> > \"ANY\", \"NONE\" in the above message, it will look a bit odd as \"pg_\"\n> > starts with lower case letters.\n> >\n>\n> Sure, the message looks a bit odd with the quotes like you wrote\n> above, but I would not suggest to change it that way - I was thinking\n> more like below (which is similar to the style the slot_name messages\n> use)\n>\n> CURRENT\n> DETAIL: Origin names \"any\", \"none\", and names starting with \"pg_\" are reserved.\n>\n> SUGGESTED\n> DETAIL: Origin names ANY, NONE, and names starting with \"pg_\" are reserved.\n>\n\nI see your point but not sure if that is an improvement over the\ncurrent one, so, let's wait and see if we get some other votes in\nfavor of your suggestion.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 27 Jul 2022 08:23:19 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi Vignesh.\n\nFYI the v39* patch fails to apply [1]. Can you please rebase it?\n\n\n[1]\n=== Applying patches on top of PostgreSQL commit ID\n5f858dd3bebd1f3845aef2bff7f4345bfb7b74b3 ===\n=== applying patch\n./v39-0001-Check-and-throw-an-error-if-publication-tables-w.patch\npatching file doc/src/sgml/ref/alter_subscription.sgml\npatching file doc/src/sgml/ref/create_subscription.sgml\npatching file src/backend/commands/subscriptioncmds.c\nHunk #10 FAILED at 886.\n1 out of 14 hunks FAILED -- saving rejects to file\nsrc/backend/commands/subscriptioncmds.c.rej\npatching file src/test/regress/expected/subscription.out\npatching file src/test/regress/sql/subscription.sql\npatching file src/test/subscription/t/030_origin.pl\npatching file src/tools/pgindent/typedefs.list\n\n------\n[1] http://cfbot.cputube.org/patch_38_3610.log\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 28 Jul 2022 15:58:35 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jul 28, 2022 at 11:28 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Hi Vignesh.\n>\n> FYI the v39* patch fails to apply [1]. Can you please rebase it?\n>\n>\n> [1]\n> === Applying patches on top of PostgreSQL commit ID\n> 5f858dd3bebd1f3845aef2bff7f4345bfb7b74b3 ===\n> === applying patch\n> ./v39-0001-Check-and-throw-an-error-if-publication-tables-w.patch\n> patching file doc/src/sgml/ref/alter_subscription.sgml\n> patching file doc/src/sgml/ref/create_subscription.sgml\n> patching file src/backend/commands/subscriptioncmds.c\n> Hunk #10 FAILED at 886.\n> 1 out of 14 hunks FAILED -- saving rejects to file\n> src/backend/commands/subscriptioncmds.c.rej\n> patching file src/test/regress/expected/subscription.out\n> patching file src/test/regress/sql/subscription.sql\n> patching file src/test/subscription/t/030_origin.pl\n> patching file src/tools/pgindent/typedefs.list\n\nThanks for reporting this, I will post an updated version for this soon.\n\nRegards,\nVignesh\n\n\n","msg_date":"Thu, 28 Jul 2022 16:15:42 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Jul 28, 2022 at 11:28 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Hi Vignesh.\n>\n> FYI the v39* patch fails to apply [1]. Can you please rebase it?\n>\n>\n> [1]\n> === Applying patches on top of PostgreSQL commit ID\n> 5f858dd3bebd1f3845aef2bff7f4345bfb7b74b3 ===\n> === applying patch\n> ./v39-0001-Check-and-throw-an-error-if-publication-tables-w.patch\n> patching file doc/src/sgml/ref/alter_subscription.sgml\n> patching file doc/src/sgml/ref/create_subscription.sgml\n> patching file src/backend/commands/subscriptioncmds.c\n> Hunk #10 FAILED at 886.\n> 1 out of 14 hunks FAILED -- saving rejects to file\n> src/backend/commands/subscriptioncmds.c.rej\n> patching file src/test/regress/expected/subscription.out\n> patching file src/test/regress/sql/subscription.sql\n> patching file src/test/subscription/t/030_origin.pl\n> patching file src/tools/pgindent/typedefs.list\n>\n> ------\n\nPlease find the v40 patch attached which is rebased on top of head.\n\nRegards,\nVignesh","msg_date":"Thu, 28 Jul 2022 19:33:31 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are some comments for the patch v40-0001:\n\n======\n\n1. Commit message\n\nIt might be better to always use 'copy_data = true' in favour of\n'copy_data = on' just for consistency with all the docs and the error\nmessages.\n\n======\n\n2. doc/src/sgml/ref/create_subscription.sgml\n\n@@ -386,6 +401,15 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n can have non-existent publications.\n </para>\n\n+ <para>\n+ If the subscription is created with <literal>origin = NONE</literal> and\n+ <literal>copy_data = true</literal>, it will check if the publisher has\n+ subscribed to the same table from other publishers and, if so, throw an\n+ error to prevent possible non-local data from being copied. The user can\n+ override this check and continue with the copy operation by specifying\n+ <literal>copy_data = force</literal>.\n+ </para>\n\n2a.\nIt is interesting that you changed the note to say origin = NONE.\nPersonally, I prefer it written as you did, but I think maybe this\nchange does not belong in this patch. The suggestion for changing from\n\"none\" to NONE is being discussed elsewhere in this thread and\nprobably all such changes should be done together (if at all) as a\nseparate patch. Until then I think this patch 0001 should just stay\nconsistent with whatever is already pushed on HEAD.\n\n2b.\n\"possible no-local data\". Maybe the terminology \"local/non-local\" is a\nhangover from back when the subscription parameter was called local\ninstead of origin. I'm not sure if you want to change this or not, and\nanyway I didn't have any better suggestions – so this comment is just\nto bring it to your attention.\n\n======\n\n3. src/backend/commands/subscriptioncmds.c - DefGetCopyData\n\n+ /*\n+ * The set of strings accepted here should match up with the\n+ * grammar's opt_boolean_or_string production.\n+ */\n+ if (pg_strcasecmp(sval, \"false\") == 0 ||\n+ pg_strcasecmp(sval, \"off\") == 0)\n+ return COPY_DATA_OFF;\n+ if (pg_strcasecmp(sval, \"true\") == 0 ||\n+ pg_strcasecmp(sval, \"on\") == 0)\n+ return COPY_DATA_ON;\n+ if (pg_strcasecmp(sval, \"force\") == 0)\n+ return COPY_DATA_FORCE;\n\nI understand the intention of the comment, but it is not strictly\ncorrect to say \"should match up\" because \"force\" is a new value.\nPerhaps the comment should be as suggested below.\n\nSUGGESTION\nThe set of strings accepted here must include all those accepted by\nthe grammar's opt_boolean_or_string production.\n\n~~~\n\n4. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n\n@@ -1781,6 +1858,122 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\n table_close(rel, RowExclusiveLock);\n }\n\n+/*\n+ * Check and throw an error if the publisher has subscribed to the same table\n+ * from some other publisher. This check is required only if \"copy_data = on\"\n+ * and \"origin = NONE\" for CREATE SUBSCRIPTION and\n+ * ALTER SUBSCRIPTION ... REFRESH statements to avoid the publisher from\n+ * replicating data that has an origin.\n+ *\n+ * This check need not be performed on the tables that are already added as\n+ * incremental sync for such tables will happen through WAL and the origin of\n+ * the data can be identified from the WAL records.\n+ *\n+ * subrel_local_oids contains the list of relation oids that are already\n+ * present on the subscriber.\n+ */\n+static void\n+check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\n+ CopyData copydata, char *origin,\n+ Oid *subrel_local_oids, int subrel_count)\n\n4a.\n\"copy_data = on\" -> \"copy_data = true\" (for consistency with the docs\nand the error messages)\n\n4b.\nThe same NONE/none review comment from #2a applies here too. Probably\nit should be written as none for now unless/until *everything* changes\nto NONE.\n\n4c.\n\"to avoid the publisher from replicating data that has an origin.\" ->\n\"to avoid replicating data that has an origin.\"\n\n4d.\n+ * This check need not be performed on the tables that are already added as\n+ * incremental sync for such tables will happen through WAL and the origin of\n+ * the data can be identified from the WAL records.\n\nSUGGESTION (maybe?)\nThis check need not be performed on the tables that are already added\nbecause incremental sync for those tables will happen through WAL and\nthe origin of the data can be identified from the WAL records.\n\n======\n\n5. src/test/subscription/t/030_origin.pl\n\n+ \"Refresh publication when the publisher has subscribed for the new table\"\n\nSUGGESTION (Just to mention origin = none somehow. Maybe you can\nreword it better than this)\nRefresh publication when the publisher has subscribed for the new\ntable, but the subscriber-side wants origin=none\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 29 Jul 2022 13:01:30 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jul 29, 2022 at 8:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some comments for the patch v40-0001:\n>\n> ======\n>\n> 1. Commit message\n>\n> It might be better to always use 'copy_data = true' in favour of\n> 'copy_data = on' just for consistency with all the docs and the error\n> messages.\n>\n> ======\n\nModified\n\n> 2. doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -386,6 +401,15 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> can have non-existent publications.\n> </para>\n>\n> + <para>\n> + If the subscription is created with <literal>origin = NONE</literal> and\n> + <literal>copy_data = true</literal>, it will check if the publisher has\n> + subscribed to the same table from other publishers and, if so, throw an\n> + error to prevent possible non-local data from being copied. The user can\n> + override this check and continue with the copy operation by specifying\n> + <literal>copy_data = force</literal>.\n> + </para>\n>\n> 2a.\n> It is interesting that you changed the note to say origin = NONE.\n> Personally, I prefer it written as you did, but I think maybe this\n> change does not belong in this patch. The suggestion for changing from\n> \"none\" to NONE is being discussed elsewhere in this thread and\n> probably all such changes should be done together (if at all) as a\n> separate patch. Until then I think this patch 0001 should just stay\n> consistent with whatever is already pushed on HEAD.\n\nModified\n\n> 2b.\n> \"possible no-local data\". Maybe the terminology \"local/non-local\" is a\n> hangover from back when the subscription parameter was called local\n> instead of origin. I'm not sure if you want to change this or not, and\n> anyway I didn't have any better suggestions – so this comment is just\n> to bring it to your attention.\n>\n> ======\n\nModified\n\n> 3. src/backend/commands/subscriptioncmds.c - DefGetCopyData\n>\n> + /*\n> + * The set of strings accepted here should match up with the\n> + * grammar's opt_boolean_or_string production.\n> + */\n> + if (pg_strcasecmp(sval, \"false\") == 0 ||\n> + pg_strcasecmp(sval, \"off\") == 0)\n> + return COPY_DATA_OFF;\n> + if (pg_strcasecmp(sval, \"true\") == 0 ||\n> + pg_strcasecmp(sval, \"on\") == 0)\n> + return COPY_DATA_ON;\n> + if (pg_strcasecmp(sval, \"force\") == 0)\n> + return COPY_DATA_FORCE;\n>\n> I understand the intention of the comment, but it is not strictly\n> correct to say \"should match up\" because \"force\" is a new value.\n> Perhaps the comment should be as suggested below.\n>\n> SUGGESTION\n> The set of strings accepted here must include all those accepted by\n> the grammar's opt_boolean_or_string production.\n>\n> ~~\n\nModified\n\n>\n> 4. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> @@ -1781,6 +1858,122 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\n> table_close(rel, RowExclusiveLock);\n> }\n>\n> +/*\n> + * Check and throw an error if the publisher has subscribed to the same table\n> + * from some other publisher. This check is required only if \"copy_data = on\"\n> + * and \"origin = NONE\" for CREATE SUBSCRIPTION and\n> + * ALTER SUBSCRIPTION ... REFRESH statements to avoid the publisher from\n> + * replicating data that has an origin.\n> + *\n> + * This check need not be performed on the tables that are already added as\n> + * incremental sync for such tables will happen through WAL and the origin of\n> + * the data can be identified from the WAL records.\n> + *\n> + * subrel_local_oids contains the list of relation oids that are already\n> + * present on the subscriber.\n> + */\n> +static void\n> +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\n> + CopyData copydata, char *origin,\n> + Oid *subrel_local_oids, int subrel_count)\n>\n> 4a.\n> \"copy_data = on\" -> \"copy_data = true\" (for consistency with the docs\n> and the error messages)\n\nModified\n\n> 4b.\n> The same NONE/none review comment from #2a applies here too. Probably\n> it should be written as none for now unless/until *everything* changes\n> to NONE.\n\nModified\n\n> 4c.\n> \"to avoid the publisher from replicating data that has an origin.\" ->\n> \"to avoid replicating data that has an origin.\"\n\nModified\n\n> 4d.\n> + * This check need not be performed on the tables that are already added as\n> + * incremental sync for such tables will happen through WAL and the origin of\n> + * the data can be identified from the WAL records.\n>\n> SUGGESTION (maybe?)\n> This check need not be performed on the tables that are already added\n> because incremental sync for those tables will happen through WAL and\n> the origin of the data can be identified from the WAL records.\n>\n> ======\n\nModified\n\n>\n> 5. src/test/subscription/t/030_origin.pl\n>\n> + \"Refresh publication when the publisher has subscribed for the new table\"\n>\n> SUGGESTION (Just to mention origin = none somehow. Maybe you can\n> reword it better than this)\n> Refresh publication when the publisher has subscribed for the new\n> table, but the subscriber-side wants origin=none\n\nModified\n\nThanks for the comments, the attached v41 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Fri, 29 Jul 2022 10:51:40 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> \r\n> Thanks for the comments, the attached v41 patch has the changes for the\r\n> same.\r\n> \r\n\r\nThanks for updating the patch.\r\n\r\nI wonder in the case that the publisher uses PG15 (or before), subscriber uses\r\nPG16, should we have this check (check if publication tables were also\r\nsubscribing from other publishers)? In this case, even if origin=none is\r\nspecified, it doesn't work because the publisher doesn't filter the origin. So\r\nmaybe we don't need the check for initial sync. Thoughts?\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Mon, 1 Aug 2022 05:27:30 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Aug 1, 2022 at 3:27 PM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> >\n> > Thanks for the comments, the attached v41 patch has the changes for the\n> > same.\n> >\n>\n> Thanks for updating the patch.\n>\n> I wonder in the case that the publisher uses PG15 (or before), subscriber uses\n> PG16, should we have this check (check if publication tables were also\n> subscribing from other publishers)? In this case, even if origin=none is\n> specified, it doesn't work because the publisher doesn't filter the origin. So\n> maybe we don't need the check for initial sync. Thoughts?\n>\n\nIIUC for the scenario you've described (subscription origin=none and\npublisher < PG16) the subscriber can end up getting extra data they\ndid not want, right?\n\nSo instead of just \"don't need the check\", maybe this combination\nshould throw ERROR, or at least a log a WARNING?\n\n------\nKInd Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 1 Aug 2022 18:02:21 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Aug 1, 2022 at 1:32 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Mon, Aug 1, 2022 at 3:27 PM shiy.fnst@fujitsu.com\n> <shiy.fnst@fujitsu.com> wrote:\n> >\n> > On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > >\n> > > Thanks for the comments, the attached v41 patch has the changes for the\n> > > same.\n> > >\n> >\n> > Thanks for updating the patch.\n> >\n> > I wonder in the case that the publisher uses PG15 (or before), subscriber uses\n> > PG16, should we have this check (check if publication tables were also\n> > subscribing from other publishers)? In this case, even if origin=none is\n> > specified, it doesn't work because the publisher doesn't filter the origin. So\n> > maybe we don't need the check for initial sync. Thoughts?\n> >\n>\n> IIUC for the scenario you've described (subscription origin=none and\n> publisher < PG16) the subscriber can end up getting extra data they\n> did not want, right?\n>\n\nYes, because publishers won't have 'filtering based on origin' functionality.\n\n> So instead of just \"don't need the check\", maybe this combination\n> should throw ERROR, or at least a log a WARNING?\n>\n\nI am not sure if doing anything (ERROR or WARNING) would make sense\nbecause anyway later during replication there won't be any filtering.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 1 Aug 2022 14:21:56 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Aug 1, 2022 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Aug 1, 2022 at 1:32 PM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > On Mon, Aug 1, 2022 at 3:27 PM shiy.fnst@fujitsu.com\n> > <shiy.fnst@fujitsu.com> wrote:\n> > >\n> > > On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\n> > > >\n> > > >\n> > > > Thanks for the comments, the attached v41 patch has the changes for the\n> > > > same.\n> > > >\n> > >\n> > > Thanks for updating the patch.\n> > >\n> > > I wonder in the case that the publisher uses PG15 (or before), subscriber uses\n> > > PG16, should we have this check (check if publication tables were also\n> > > subscribing from other publishers)? In this case, even if origin=none is\n> > > specified, it doesn't work because the publisher doesn't filter the origin. So\n> > > maybe we don't need the check for initial sync. Thoughts?\n> > >\n> >\n> > IIUC for the scenario you've described (subscription origin=none and\n> > publisher < PG16) the subscriber can end up getting extra data they\n> > did not want, right?\n> >\n>\n> Yes, because publishers won't have 'filtering based on origin' functionality.\n>\n> > So instead of just \"don't need the check\", maybe this combination\n> > should throw ERROR, or at least a log a WARNING?\n> >\n>\n> I am not sure if doing anything (ERROR or WARNING) would make sense\n> because anyway later during replication there won't be any filtering.\n>\n\nI was suggesting stopping that replication from happening at all. If\nthe user specifically asked for 'origin=none' but the publisher could\nnot filter that (because < PG16) then I imagined some logic that would\njust disable the subscription up-front. Isn't it preferable for the\nsubscriber to get no data at all then to get data the user\nspecifically said they did NOT want to get?\n\ne.g. pseudo-code for the worker code something like below:\n\nif (origin != ANY and publisher.server_version < PG16)\n{\n set subscription.option.disable_on_error = true;\n throw ERROR (\"publisher does not support origin=none - disabling\nthe subscription\");\n}\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia.\n\n\n","msg_date":"Tue, 2 Aug 2022 12:01:15 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> On Fri, Jul 29, 2022 at 8:31 AM Peter Smith <smithpb2250@gmail.com>\r\n> wrote:\r\n> >\r\n> \r\n> Thanks for the comments, the attached v41 patch has the changes for the\r\n> same.\r\n> \r\n\r\nThanks for updating the patch.\r\n\r\nA comment for 0002 patch.\r\n\r\nIn the example in section 31.11.4 (Generic steps for adding a new primary to an\r\nexisting set of primaries), I think there should be a Step-6, corresponding to\r\nthe steps mentioned before. And part of Step-5 should actually be part of\r\nStep-6.\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Tue, 2 Aug 2022 09:31:20 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Jul 26, 2022 at 7:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n> >\n> > Thanks for the example. I agree that it is fairly simple to reproduce.\n> >\n> > I understand that \"copy_data = force\" is meant to protect a user from\n> > hurting themself. I'm not convinced that this is the best way to do so.\n> >\n> > For example today I can subscribe to multiple publications that write to\n> > the same table. If I have a primary key on that table, and two of the\n> > subscriptions try to write an identical ID, we conflict. We don't have\n> > any special flags or modes to guard against that from happening, though\n> > we do have documentation on conflicts and managing them.\n> >\n> > AFAICT the same issue with \"copy_data\" also exists in the above scenario\n> > too, even without the \"origin\" attribute.\n> >\n>\n> That's true but there is no parameter like origin = NONE which\n> indicates that constraint violations or duplicate data problems won't\n> occur due to replication. In the current case, I think the situation\n> is different because a user has specifically asked not to replicate\n> any remote data by specifying origin = NONE, which should be dealt\n> differently. Note that current users or their setup won't see any\n> difference/change unless they specify the new parameter origin as\n> NONE.\n>\n\nLet me try to summarize the discussion so that it is easier for others\nto follow. The work in this thread is to avoid loops, and duplicate\ndata in logical replication when the operations happened on the same\ntable in multiple nodes. It has been explained in email [1] with an\nexample of how a logical replication setup can lead to duplicate or\ninconsistent data.\n\nThe idea to solve this problem is that we don't replicate data that is\nnot generated locally which we can normally identify based on origin\nof data in WAL. The commit 366283961a achieves that for replication\nbut still the problem can happen during initial sync which is\nperformed internally via copy. We can't differentiate the data in heap\nbased on origin. So, we decided to prohibit the subscription\noperations that can perform initial sync (ex. Create Subscription,\nAlter Subscription ... Refresh) by detecting that the publisher has\nsubscribed to the same table from some other publisher.\n\nTo prohibit the subscription operations, the currently proposed patch\nthrows an error. Then, it also provides a new copy_data option\n'force' under which the user will still be able to perform the\noperation. This could be useful when the user intentionally wants to\nreplicate the initial data even if it contains data from multiple\nnodes (for example, when in a multi-node setup, one decides to get the\ninitial data from just one node and then allow replication of data to\nproceed from each of respective nodes).\n\nThe other alternative discussed was to just give a warning for\nsubscription operations and probably document the steps for users to\navoid it. But the problem with that is once the user sees this\nwarning, it won't be able to do anything except recreate the setup, so\nwhy not give an error in the first place?\n\nThoughts?\n\n[1] - https://www.postgresql.org/message-id/CALDaNm1eJr6qXT9esVPzgc5Qvy4uMhV4kCCTSmxARKjf%2BMwcnw%40mail.gmail.com\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 2 Aug 2022 17:29:36 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Aug 2, 2022 at 8:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Tue, Jul 26, 2022 at 7:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n> > >\n> > > Thanks for the example. I agree that it is fairly simple to reproduce.\n> > >\n> > > I understand that \"copy_data = force\" is meant to protect a user from\n> > > hurting themself. I'm not convinced that this is the best way to do so.\n> > >\n> > > For example today I can subscribe to multiple publications that write to\n> > > the same table. If I have a primary key on that table, and two of the\n> > > subscriptions try to write an identical ID, we conflict. We don't have\n> > > any special flags or modes to guard against that from happening, though\n> > > we do have documentation on conflicts and managing them.\n> > >\n> > > AFAICT the same issue with \"copy_data\" also exists in the above scenario\n> > > too, even without the \"origin\" attribute.\n> > >\n> >\n> > That's true but there is no parameter like origin = NONE which\n> > indicates that constraint violations or duplicate data problems won't\n> > occur due to replication. In the current case, I think the situation\n> > is different because a user has specifically asked not to replicate\n> > any remote data by specifying origin = NONE, which should be dealt\n> > differently. Note that current users or their setup won't see any\n> > difference/change unless they specify the new parameter origin as\n> > NONE.\n> >\n>\n> Let me try to summarize the discussion so that it is easier for others\n> to follow. The work in this thread is to avoid loops, and duplicate\n> data in logical replication when the operations happened on the same\n> table in multiple nodes. It has been explained in email [1] with an\n> example of how a logical replication setup can lead to duplicate or\n> inconsistent data.\n>\n> The idea to solve this problem is that we don't replicate data that is\n> not generated locally which we can normally identify based on origin\n> of data in WAL. The commit 366283961a achieves that for replication\n> but still the problem can happen during initial sync which is\n> performed internally via copy. We can't differentiate the data in heap\n> based on origin. So, we decided to prohibit the subscription\n> operations that can perform initial sync (ex. Create Subscription,\n> Alter Subscription ... Refresh) by detecting that the publisher has\n> subscribed to the same table from some other publisher.\n>\n> To prohibit the subscription operations, the currently proposed patch\n> throws an error. Then, it also provides a new copy_data option\n> 'force' under which the user will still be able to perform the\n> operation. This could be useful when the user intentionally wants to\n> replicate the initial data even if it contains data from multiple\n> nodes (for example, when in a multi-node setup, one decides to get the\n> initial data from just one node and then allow replication of data to\n> proceed from each of respective nodes).\n>\n> The other alternative discussed was to just give a warning for\n> subscription operations and probably document the steps for users to\n> avoid it. But the problem with that is once the user sees this\n> warning, it won't be able to do anything except recreate the setup, so\n> why not give an error in the first place?\n>\n> Thoughts?\n\nThank you for the summary!\n\nI understand that this feature could help some cases, but I'm really\nnot sure adding new value 'force' for them is worthwhile, and\nconcerned it could reduce the usability.\n\nIIUC this feature would work only when origin = 'none' and copy_data =\n'on', and the purpose is to prevent the data from being\nduplicated/conflicted by the initial table sync. But there are cases\nwhere duplication/conflict doesn't happen even if origin = 'none' and\ncopy_data = 'on'. For instance, the table on the publisher might be\nempty. Also, even with origin = 'any', copy_data = 'on', there is a\npossibility of data duplication/conflict. Why do we need to address\nonly the case where origin = 'none'? I think that using origin =\n'none' doesn't necessarily mean using bi-directional (or N-way)\nreplication. Even when using uni-directional logical replication with\ntwo nodes, they may use origin = 'none'. Therefore, it seems to me\nthat this feature works only for a narrow situation and has false\npositives.\n\nSince it has been the user's responsibility not to try to make the\ndata inconsistent by the initial table sync, I think that it might be\nsufficient if we note the risk in the documentation.\n\nRegards,\n\n--\nMasahiko Sawada\nEDB: https://www.enterprisedb.com/\n\n\n","msg_date":"Thu, 4 Aug 2022 22:00:53 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Aug 4, 2022 at 6:31 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n> On Tue, Aug 2, 2022 at 8:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > Let me try to summarize the discussion so that it is easier for others\n> > to follow. The work in this thread is to avoid loops, and duplicate\n> > data in logical replication when the operations happened on the same\n> > table in multiple nodes. It has been explained in email [1] with an\n> > example of how a logical replication setup can lead to duplicate or\n> > inconsistent data.\n> >\n> > The idea to solve this problem is that we don't replicate data that is\n> > not generated locally which we can normally identify based on origin\n> > of data in WAL. The commit 366283961a achieves that for replication\n> > but still the problem can happen during initial sync which is\n> > performed internally via copy. We can't differentiate the data in heap\n> > based on origin. So, we decided to prohibit the subscription\n> > operations that can perform initial sync (ex. Create Subscription,\n> > Alter Subscription ... Refresh) by detecting that the publisher has\n> > subscribed to the same table from some other publisher.\n> >\n> > To prohibit the subscription operations, the currently proposed patch\n> > throws an error. Then, it also provides a new copy_data option\n> > 'force' under which the user will still be able to perform the\n> > operation. This could be useful when the user intentionally wants to\n> > replicate the initial data even if it contains data from multiple\n> > nodes (for example, when in a multi-node setup, one decides to get the\n> > initial data from just one node and then allow replication of data to\n> > proceed from each of respective nodes).\n> >\n> > The other alternative discussed was to just give a warning for\n> > subscription operations and probably document the steps for users to\n> > avoid it. But the problem with that is once the user sees this\n> > warning, it won't be able to do anything except recreate the setup, so\n> > why not give an error in the first place?\n> >\n> > Thoughts?\n>\n> Thank you for the summary!\n>\n> I understand that this feature could help some cases, but I'm really\n> not sure adding new value 'force' for them is worthwhile, and\n> concerned it could reduce the usability.\n>\n> IIUC this feature would work only when origin = 'none' and copy_data =\n> 'on', and the purpose is to prevent the data from being\n> duplicated/conflicted by the initial table sync. But there are cases\n> where duplication/conflict doesn't happen even if origin = 'none' and\n> copy_data = 'on'. For instance, the table on the publisher might be\n> empty.\n>\n\nRight, but if we want we can check the tables on publishers to ensure\nthat. Now, another case could be where the corresponding subscription\nwas disabled on publisher during create subscription but got enabled\njust before copy, we can even catch that situation as we are doing for\ncolumn lists in fetch_remote_table_info(). Are there other cases of\nfalse positives you can think of? I see your point that we can\ndocument that users should be careful with certain configurations to\navoid data inconsistency but not able completely convince myself about\nthe same. I think the main thing to decide here is how much we want to\nask users to be careful by referring them to docs. Now, if you are not\nconvinced with giving an ERROR here then we can probably show a\nWARNING (that we might copy data for multiple origins during initial\nsync in spite of the user having specified origin as NONE)?\n\n>\n Also, even with origin = 'any', copy_data = 'on', there is a\n> possibility of data duplication/conflict. Why do we need to address\n> only the case where origin = 'none'?\n>\n\nBecause the user has specifically asked not to replicate any remote\ndata by specifying origin = NONE, which should be dealt with\ndifferently whereas 'any' doesn't have such a requirement. Now,\ntomorrow, if we want to support replication based on specific origin\nnames say origin = 'node-1' then also we won't be able to identify the\ndata during initial sync but I think 'none' is a special case where\ngiving some intimation to user won't be a bad idea especially because\nwe can identify the same.\n\n> I think that using origin =\n> 'none' doesn't necessarily mean using bi-directional (or N-way)\n> replication. Even when using uni-directional logical replication with\n> two nodes, they may use origin = 'none'.\n>\n\nIt is possible but still, I think it is a must for bi-directional (or\nN-way) replication, otherwise, there is a risk of loops.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 5 Aug 2022 09:40:02 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Jul 29, 2022 at 10:51 AM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Fri, Jul 29, 2022 at 8:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Here are some comments for the patch v40-0001:\n> >\n> > ======\n> >\n> > 1. Commit message\n> >\n> > It might be better to always use 'copy_data = true' in favour of\n> > 'copy_data = on' just for consistency with all the docs and the error\n> > messages.\n> >\n> > ======\n>\n> Modified\n>\n> > 2. doc/src/sgml/ref/create_subscription.sgml\n> >\n> > @@ -386,6 +401,15 @@ CREATE SUBSCRIPTION <replaceable\n> > class=\"parameter\">subscription_name</replaceabl\n> > can have non-existent publications.\n> > </para>\n> >\n> > + <para>\n> > + If the subscription is created with <literal>origin = NONE</literal> and\n> > + <literal>copy_data = true</literal>, it will check if the publisher has\n> > + subscribed to the same table from other publishers and, if so, throw an\n> > + error to prevent possible non-local data from being copied. The user can\n> > + override this check and continue with the copy operation by specifying\n> > + <literal>copy_data = force</literal>.\n> > + </para>\n> >\n> > 2a.\n> > It is interesting that you changed the note to say origin = NONE.\n> > Personally, I prefer it written as you did, but I think maybe this\n> > change does not belong in this patch. The suggestion for changing from\n> > \"none\" to NONE is being discussed elsewhere in this thread and\n> > probably all such changes should be done together (if at all) as a\n> > separate patch. Until then I think this patch 0001 should just stay\n> > consistent with whatever is already pushed on HEAD.\n>\n> Modified\n>\n> > 2b.\n> > \"possible no-local data\". Maybe the terminology \"local/non-local\" is a\n> > hangover from back when the subscription parameter was called local\n> > instead of origin. I'm not sure if you want to change this or not, and\n> > anyway I didn't have any better suggestions – so this comment is just\n> > to bring it to your attention.\n> >\n> > ======\n>\n> Modified\n>\n> > 3. src/backend/commands/subscriptioncmds.c - DefGetCopyData\n> >\n> > + /*\n> > + * The set of strings accepted here should match up with the\n> > + * grammar's opt_boolean_or_string production.\n> > + */\n> > + if (pg_strcasecmp(sval, \"false\") == 0 ||\n> > + pg_strcasecmp(sval, \"off\") == 0)\n> > + return COPY_DATA_OFF;\n> > + if (pg_strcasecmp(sval, \"true\") == 0 ||\n> > + pg_strcasecmp(sval, \"on\") == 0)\n> > + return COPY_DATA_ON;\n> > + if (pg_strcasecmp(sval, \"force\") == 0)\n> > + return COPY_DATA_FORCE;\n> >\n> > I understand the intention of the comment, but it is not strictly\n> > correct to say \"should match up\" because \"force\" is a new value.\n> > Perhaps the comment should be as suggested below.\n> >\n> > SUGGESTION\n> > The set of strings accepted here must include all those accepted by\n> > the grammar's opt_boolean_or_string production.\n> >\n> > ~~\n>\n> Modified\n>\n> >\n> > 4. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n> >\n> > @@ -1781,6 +1858,122 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\n> > table_close(rel, RowExclusiveLock);\n> > }\n> >\n> > +/*\n> > + * Check and throw an error if the publisher has subscribed to the same table\n> > + * from some other publisher. This check is required only if \"copy_data = on\"\n> > + * and \"origin = NONE\" for CREATE SUBSCRIPTION and\n> > + * ALTER SUBSCRIPTION ... REFRESH statements to avoid the publisher from\n> > + * replicating data that has an origin.\n> > + *\n> > + * This check need not be performed on the tables that are already added as\n> > + * incremental sync for such tables will happen through WAL and the origin of\n> > + * the data can be identified from the WAL records.\n> > + *\n> > + * subrel_local_oids contains the list of relation oids that are already\n> > + * present on the subscriber.\n> > + */\n> > +static void\n> > +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\n> > + CopyData copydata, char *origin,\n> > + Oid *subrel_local_oids, int subrel_count)\n> >\n> > 4a.\n> > \"copy_data = on\" -> \"copy_data = true\" (for consistency with the docs\n> > and the error messages)\n>\n> Modified\n>\n> > 4b.\n> > The same NONE/none review comment from #2a applies here too. Probably\n> > it should be written as none for now unless/until *everything* changes\n> > to NONE.\n>\n> Modified\n>\n> > 4c.\n> > \"to avoid the publisher from replicating data that has an origin.\" ->\n> > \"to avoid replicating data that has an origin.\"\n>\n> Modified\n>\n> > 4d.\n> > + * This check need not be performed on the tables that are already added as\n> > + * incremental sync for such tables will happen through WAL and the origin of\n> > + * the data can be identified from the WAL records.\n> >\n> > SUGGESTION (maybe?)\n> > This check need not be performed on the tables that are already added\n> > because incremental sync for those tables will happen through WAL and\n> > the origin of the data can be identified from the WAL records.\n> >\n> > ======\n>\n> Modified\n>\n> >\n> > 5. src/test/subscription/t/030_origin.pl\n> >\n> > + \"Refresh publication when the publisher has subscribed for the new table\"\n> >\n> > SUGGESTION (Just to mention origin = none somehow. Maybe you can\n> > reword it better than this)\n> > Refresh publication when the publisher has subscribed for the new\n> > table, but the subscriber-side wants origin=none\n>\n> Modified\n>\n> Thanks for the comments, the attached v41 patch has the changes for the same.\n\nThe patch does not apply on Head because of the commit\n\"0c20dd33db1607d6a85ffce24238c1e55e384b49\", I have attached a rebased\nv41 patch on top of the HEAD. I have not yet done the changes to\nchange the error to warning in the first patch, I will change it in\nthe next version once it is concluded.\n\nRegards,\nVignesh","msg_date":"Mon, 8 Aug 2022 10:21:24 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tuesday, August 2, 2022 8:00 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> >\r\n> > On Tue, Jul 26, 2022 at 7:13 AM Jonathan S. Katz <jkatz@postgresql.org>\r\n> wrote:\r\n> > >\r\n> > > Thanks for the example. I agree that it is fairly simple to reproduce.\r\n> > >\r\n> > > I understand that \"copy_data = force\" is meant to protect a user\r\n> > > from hurting themself. I'm not convinced that this is the best way to do so.\r\n> > >\r\n> > > For example today I can subscribe to multiple publications that\r\n> > > write to the same table. If I have a primary key on that table, and\r\n> > > two of the subscriptions try to write an identical ID, we conflict.\r\n> > > We don't have any special flags or modes to guard against that from\r\n> > > happening, though we do have documentation on conflicts and managing\r\n> them.\r\n> > >\r\n> > > AFAICT the same issue with \"copy_data\" also exists in the above\r\n> > > scenario too, even without the \"origin\" attribute.\r\n> > >\r\n> >\r\n> > That's true but there is no parameter like origin = NONE which\r\n> > indicates that constraint violations or duplicate data problems won't\r\n> > occur due to replication. In the current case, I think the situation\r\n> > is different because a user has specifically asked not to replicate\r\n> > any remote data by specifying origin = NONE, which should be dealt\r\n> > differently. Note that current users or their setup won't see any\r\n> > difference/change unless they specify the new parameter origin as\r\n> > NONE.\r\n> >\r\n> \r\n> Let me try to summarize the discussion so that it is easier for others to follow.\r\n> The work in this thread is to avoid loops, and duplicate data in logical\r\n> replication when the operations happened on the same table in multiple nodes.\r\n> It has been explained in email [1] with an example of how a logical replication\r\n> setup can lead to duplicate or inconsistent data.\r\n> \r\n> The idea to solve this problem is that we don't replicate data that is not\r\n> generated locally which we can normally identify based on origin of data in\r\n> WAL. The commit 366283961a achieves that for replication but still the\r\n> problem can happen during initial sync which is performed internally via copy.\r\n> We can't differentiate the data in heap based on origin. So, we decided to\r\n> prohibit the subscription operations that can perform initial sync (ex. Create\r\n> Subscription, Alter Subscription ... Refresh) by detecting that the publisher has\r\n> subscribed to the same table from some other publisher.\r\n> \r\n> To prohibit the subscription operations, the currently proposed patch throws\r\n> an error. Then, it also provides a new copy_data option 'force' under which\r\n> the user will still be able to perform the operation. This could be useful when\r\n> the user intentionally wants to replicate the initial data even if it contains data\r\n> from multiple nodes (for example, when in a multi-node setup, one decides to\r\n> get the initial data from just one node and then allow replication of data to\r\n> proceed from each of respective nodes).\r\n> \r\n> The other alternative discussed was to just give a warning for subscription\r\n> operations and probably document the steps for users to avoid it. But the\r\n> problem with that is once the user sees this warning, it won't be able to do\r\n> anything except recreate the setup, so why not give an error in the first place?\r\n> \r\n> Thoughts?\r\n\r\nThanks for the summary.\r\n\r\nI think it's fine to make the user use the copy_data option more carefully to\r\nprevent duplicate copies by reporting an ERROR.\r\n\r\nBut I also have similar concern with Sawada-san as it's possible for user to\r\nreceive an ERROR in some unexpected cases.\r\n\r\nFor example I want to build bi-directional setup between two nodes:\r\n\r\nNode A: TABLE test (has actual data)\r\nNode B: TABLE test (empty)\r\n\r\nStep 1:\r\nCREATE PUBLICATION on both Node A and B.\r\n\r\nStep 2:\r\nCREATE SUBSCRIPTION on Node A with (copy_data = on)\r\n-- this is fine as there is no data on Node B\r\n\r\nStep 3:\r\nCREATE SUBSCRIPTION on Node B with (copy_data = on)\r\n-- this should be fine as user needs to copy data from Node A to Node B,\r\n-- but we still report an error for this case.\r\n\r\nIt looks a bit strict to report an ERROR in this case and it seems not easy to\r\navoid this. So, personally, I think it might be better to document the correct\r\nsteps to build the bi-directional replication and probably also docuemnt the\r\nsteps to recover if user accidently did duplicate initial copy if not\r\ndocumented yet.\r\n\r\nIn addition, we could also LOG some additional information about the ORIGIN and\r\ninitial copy which might help user to analyze if needed.\r\n\r\n\r\n----- Some other thoughts about the duplicate initial copy problem.\r\n\r\nActually, I feel the better way to address the possible duplicate copy problem\r\nis to provide a command like \"bi_setup\" which can help user build the\r\nbi-directional replication in all nodes and can handle the initial copy\r\nautomtically. But that might be too far.\r\n\r\nAnother naive idea I once thought is that maybe we can add a publication option\r\nlike: data_source_in_bi_group. If data_source_in_bi_group is true, we silently\r\nlet it succeed when we subscribe this publication\r\nwith(origin='NONE',copy_data=on). If data_source_in_bi_group is false, we\r\nreport an ERROR when we subscribe this publication\r\nwith(origin='NONE',copy_data=on). The slight difference in this approach is\r\nthat we don't do any additional check for the publisher and subscriber, we\r\ntrust the user and give the responsibility to choose the node with actual data\r\nto user. Having said that this approach seems not a good approach as it doesn't\r\nsolve the actual problem and can only handle the case when ORIGIN='NONE' and\r\ncannot handle the future case when we support ORIGIN='NODE1'.\r\n\r\nSo, I personally feel that adding document and LOG might be sufficient at this\r\nstage.\r\n\r\nBest regards,\r\nHou zj\r\n","msg_date":"Wed, 17 Aug 2022 03:18:08 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> On Tuesday, August 2, 2022 8:00 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> Thanks for the summary.\n>\n> I think it's fine to make the user use the copy_data option more carefully to\n> prevent duplicate copies by reporting an ERROR.\n>\n> But I also have similar concern with Sawada-san as it's possible for user to\n> receive an ERROR in some unexpected cases.\n>\n> For example I want to build bi-directional setup between two nodes:\n>\n> Node A: TABLE test (has actual data)\n> Node B: TABLE test (empty)\n>\n> Step 1:\n> CREATE PUBLICATION on both Node A and B.\n>\n> Step 2:\n> CREATE SUBSCRIPTION on Node A with (copy_data = on)\n> -- this is fine as there is no data on Node B\n>\n> Step 3:\n> CREATE SUBSCRIPTION on Node B with (copy_data = on)\n> -- this should be fine as user needs to copy data from Node A to Node B,\n> -- but we still report an error for this case.\n>\n> It looks a bit strict to report an ERROR in this case and it seems not easy to\n> avoid this. So, personally, I think it might be better to document the correct\n> steps to build the bi-directional replication and probably also docuemnt the\n> steps to recover if user accidently did duplicate initial copy if not\n> documented yet.\n>\n> In addition, we could also LOG some additional information about the ORIGIN and\n> initial copy which might help user to analyze if needed.\n>\n\nBut why LOG instead of WARNING? I feel in this case there is a chance\nof inconsistent data so a WARNING like \"publication \"pub1\" could have\ndata from multiple origins\" can be given when the user has specified\noptions: \"copy_data = on, origin = NONE\" while creating a\nsubscription. We give a WARNING during subscription creation when the\ncorresponding publication doesn't exist, eg.\n\npostgres=# create subscription sub1 connection 'dbname = postgres'\npublication pub1;\nWARNING: publication \"pub1\" does not exist in the publisher\n\nThen, we can explain in docs how users can avoid data inconsistencies\nwhile setting up replication.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 17 Aug 2022 12:03:43 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 17, 2022 at 4:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n> >\n> > On Tuesday, August 2, 2022 8:00 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > Thanks for the summary.\n> >\n> > I think it's fine to make the user use the copy_data option more carefully to\n> > prevent duplicate copies by reporting an ERROR.\n> >\n> > But I also have similar concern with Sawada-san as it's possible for user to\n> > receive an ERROR in some unexpected cases.\n> >\n> > For example I want to build bi-directional setup between two nodes:\n> >\n> > Node A: TABLE test (has actual data)\n> > Node B: TABLE test (empty)\n> >\n> > Step 1:\n> > CREATE PUBLICATION on both Node A and B.\n> >\n> > Step 2:\n> > CREATE SUBSCRIPTION on Node A with (copy_data = on)\n> > -- this is fine as there is no data on Node B\n> >\n> > Step 3:\n> > CREATE SUBSCRIPTION on Node B with (copy_data = on)\n> > -- this should be fine as user needs to copy data from Node A to Node B,\n> > -- but we still report an error for this case.\n> >\n> > It looks a bit strict to report an ERROR in this case and it seems not easy to\n> > avoid this. So, personally, I think it might be better to document the correct\n> > steps to build the bi-directional replication and probably also docuemnt the\n> > steps to recover if user accidently did duplicate initial copy if not\n> > documented yet.\n> >\n> > In addition, we could also LOG some additional information about the ORIGIN and\n> > initial copy which might help user to analyze if needed.\n> >\n>\n> But why LOG instead of WARNING? I feel in this case there is a chance\n> of inconsistent data so a WARNING like \"publication \"pub1\" could have\n> data from multiple origins\" can be given when the user has specified\n> options: \"copy_data = on, origin = NONE\" while creating a\n> subscription. We give a WARNING during subscription creation when the\n> corresponding publication doesn't exist, eg.\n>\n> postgres=# create subscription sub1 connection 'dbname = postgres'\n> publication pub1;\n> WARNING: publication \"pub1\" does not exist in the publisher\n>\n> Then, we can explain in docs how users can avoid data inconsistencies\n> while setting up replication.\n>\n\nI was wondering if this copy/origin case really should be a NOTICE.\n\nSee table [1]. It says WARNING is meant for \"warnings of likey\nproblems\". But this is not exactly a \"likely\" problem - IIUC we really\ndon't know if there is even any problem at all .... we only know there\nis the *potential* for a problem, but the user has to then judge it\nfor themselves, Perhaps WARNING may be a bit overkill for this\nsituation - it might be unnecessarily scary to give false warnings.\n\nOTOH, NOTICE [1] says it is for \"information that might be helpful to\nusers\" which seems more like what is needed here.\n\n------\n[1] https://www.postgresql.org/docs/current/runtime-config-logging.html#RUNTIME-CONFIG-SEVERITY-LEVELS\n\nKind Regards,\nPeter Smith.\nFujitsu Australia.\n\n\n","msg_date":"Wed, 17 Aug 2022 17:03:50 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 17, 2022 at 12:34 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Wed, Aug 17, 2022 at 4:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\n> > <houzj.fnst@fujitsu.com> wrote:\n> > >\n> > > On Tuesday, August 2, 2022 8:00 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > > > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > >\n> > > Thanks for the summary.\n> > >\n> > > I think it's fine to make the user use the copy_data option more carefully to\n> > > prevent duplicate copies by reporting an ERROR.\n> > >\n> > > But I also have similar concern with Sawada-san as it's possible for user to\n> > > receive an ERROR in some unexpected cases.\n> > >\n> > > For example I want to build bi-directional setup between two nodes:\n> > >\n> > > Node A: TABLE test (has actual data)\n> > > Node B: TABLE test (empty)\n> > >\n> > > Step 1:\n> > > CREATE PUBLICATION on both Node A and B.\n> > >\n> > > Step 2:\n> > > CREATE SUBSCRIPTION on Node A with (copy_data = on)\n> > > -- this is fine as there is no data on Node B\n> > >\n> > > Step 3:\n> > > CREATE SUBSCRIPTION on Node B with (copy_data = on)\n> > > -- this should be fine as user needs to copy data from Node A to Node B,\n> > > -- but we still report an error for this case.\n> > >\n> > > It looks a bit strict to report an ERROR in this case and it seems not easy to\n> > > avoid this. So, personally, I think it might be better to document the correct\n> > > steps to build the bi-directional replication and probably also docuemnt the\n> > > steps to recover if user accidently did duplicate initial copy if not\n> > > documented yet.\n> > >\n> > > In addition, we could also LOG some additional information about the ORIGIN and\n> > > initial copy which might help user to analyze if needed.\n> > >\n> >\n> > But why LOG instead of WARNING? I feel in this case there is a chance\n> > of inconsistent data so a WARNING like \"publication \"pub1\" could have\n> > data from multiple origins\" can be given when the user has specified\n> > options: \"copy_data = on, origin = NONE\" while creating a\n> > subscription. We give a WARNING during subscription creation when the\n> > corresponding publication doesn't exist, eg.\n> >\n> > postgres=# create subscription sub1 connection 'dbname = postgres'\n> > publication pub1;\n> > WARNING: publication \"pub1\" does not exist in the publisher\n> >\n> > Then, we can explain in docs how users can avoid data inconsistencies\n> > while setting up replication.\n> >\n>\n> I was wondering if this copy/origin case really should be a NOTICE.\n>\n\nWe usually give NOTICE for some sort of additional implicit\ninformation, e.g., when we create a slot during CREATE SUBSCRIPTION\ncommand: \"NOTICE: created replication slot \"sub1\" on publisher\". IMO,\nthis is likely to be a problem of data inconsistency so I think here\nwe can choose between WARNING and LOG. I prefer WARNING but okay with\nLOG as well if others feel so. I think we can change this later as\nwell if required. We do have an option to not do anything and just\ndocument it but I feel it is better to give user some indication of\nproblem here because not everyone reads each update of documentation.\n\nJonathan, Sawada-San, Hou-San, and others, what do you think is the\nbest way to move forward here?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 18 Aug 2022 08:43:08 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thursday, August 18, 2022 11:13 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n> \r\n> On Wed, Aug 17, 2022 at 12:34 PM Peter Smith <smithpb2250@gmail.com>\r\n> wrote:\r\n> >\r\n> > On Wed, Aug 17, 2022 at 4:33 PM Amit Kapila <amit.kapila16@gmail.com>\r\n> wrote:\r\n> > >\r\n> > > On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\r\n> > > <houzj.fnst@fujitsu.com> wrote:\r\n> > > >\r\n> > > > On Tuesday, August 2, 2022 8:00 PM Amit Kapila\r\n> <amit.kapila16@gmail.com> wrote:\r\n> > > > > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila\r\n> <amit.kapila16@gmail.com> wrote:\r\n> > > >\r\n> > > > Thanks for the summary.\r\n> > > >\r\n> > > > I think it's fine to make the user use the copy_data option more\r\n> > > > carefully to prevent duplicate copies by reporting an ERROR.\r\n> > > >\r\n> > > > But I also have similar concern with Sawada-san as it's possible\r\n> > > > for user to receive an ERROR in some unexpected cases.\r\n> > > >\r\n> > > > For example I want to build bi-directional setup between two nodes:\r\n> > > >\r\n> > > > Node A: TABLE test (has actual data) Node B: TABLE test (empty)\r\n> > > >\r\n> > > > Step 1:\r\n> > > > CREATE PUBLICATION on both Node A and B.\r\n> > > >\r\n> > > > Step 2:\r\n> > > > CREATE SUBSCRIPTION on Node A with (copy_data = on)\r\n> > > > -- this is fine as there is no data on Node B\r\n> > > >\r\n> > > > Step 3:\r\n> > > > CREATE SUBSCRIPTION on Node B with (copy_data = on)\r\n> > > > -- this should be fine as user needs to copy data from Node A to\r\n> > > > Node B,\r\n> > > > -- but we still report an error for this case.\r\n> > > >\r\n> > > > It looks a bit strict to report an ERROR in this case and it seems\r\n> > > > not easy to avoid this. So, personally, I think it might be better\r\n> > > > to document the correct steps to build the bi-directional\r\n> > > > replication and probably also docuemnt the steps to recover if\r\n> > > > user accidently did duplicate initial copy if not documented yet.\r\n> > > >\r\n> > > > In addition, we could also LOG some additional information about\r\n> > > > the ORIGIN and initial copy which might help user to analyze if needed.\r\n> > > >\r\n> > >\r\n> > > But why LOG instead of WARNING? I feel in this case there is a\r\n> > > chance of inconsistent data so a WARNING like \"publication \"pub1\"\r\n> > > could have data from multiple origins\" can be given when the user\r\n> > > has specified\r\n> > > options: \"copy_data = on, origin = NONE\" while creating a\r\n> > > subscription. We give a WARNING during subscription creation when\r\n> > > the corresponding publication doesn't exist, eg.\r\n> > >\r\n> > > postgres=# create subscription sub1 connection 'dbname = postgres'\r\n> > > publication pub1;\r\n> > > WARNING: publication \"pub1\" does not exist in the publisher\r\n> > >\r\n> > > Then, we can explain in docs how users can avoid data\r\n> > > inconsistencies while setting up replication.\r\n> > >\r\n> >\r\n> > I was wondering if this copy/origin case really should be a NOTICE.\r\n> >\r\n> \r\n> We usually give NOTICE for some sort of additional implicit information, e.g.,\r\n> when we create a slot during CREATE SUBSCRIPTION\r\n> command: \"NOTICE: created replication slot \"sub1\" on publisher\". IMO, this is\r\n> likely to be a problem of data inconsistency so I think here we can choose\r\n> between WARNING and LOG. I prefer WARNING but okay with LOG as well if\r\n> others feel so. I think we can change this later as well if required. We do have an\r\n> option to not do anything and just document it but I feel it is better to give user\r\n> some indication of problem here because not everyone reads each update of\r\n> documentation.\r\n> \r\n> Jonathan, Sawada-San, Hou-San, and others, what do you think is the best way\r\n> to move forward here?\r\n\r\nI think it's fine to throw a WARNING in this case given that there is a\r\nchance of inconsistent data.\r\n\r\nBest regards,\r\nHou zj\r\n\r\n","msg_date":"Mon, 22 Aug 2022 03:49:15 +0000","msg_from":"\"houzj.fnst@fujitsu.com\" <houzj.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Aug 22, 2022 at 9:19 AM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> On Thursday, August 18, 2022 11:13 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Wed, Aug 17, 2022 at 12:34 PM Peter Smith <smithpb2250@gmail.com>\n> > wrote:\n> > >\n> > > On Wed, Aug 17, 2022 at 4:33 PM Amit Kapila <amit.kapila16@gmail.com>\n> > wrote:\n> > > >\n> > > > On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\n> > > > <houzj.fnst@fujitsu.com> wrote:\n> > > > >\n> > > > > On Tuesday, August 2, 2022 8:00 PM Amit Kapila\n> > <amit.kapila16@gmail.com> wrote:\n> > > > > > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila\n> > <amit.kapila16@gmail.com> wrote:\n> > > > >\n> > > > > Thanks for the summary.\n> > > > >\n> > > > > I think it's fine to make the user use the copy_data option more\n> > > > > carefully to prevent duplicate copies by reporting an ERROR.\n> > > > >\n> > > > > But I also have similar concern with Sawada-san as it's possible\n> > > > > for user to receive an ERROR in some unexpected cases.\n> > > > >\n> > > > > For example I want to build bi-directional setup between two nodes:\n> > > > >\n> > > > > Node A: TABLE test (has actual data) Node B: TABLE test (empty)\n> > > > >\n> > > > > Step 1:\n> > > > > CREATE PUBLICATION on both Node A and B.\n> > > > >\n> > > > > Step 2:\n> > > > > CREATE SUBSCRIPTION on Node A with (copy_data = on)\n> > > > > -- this is fine as there is no data on Node B\n> > > > >\n> > > > > Step 3:\n> > > > > CREATE SUBSCRIPTION on Node B with (copy_data = on)\n> > > > > -- this should be fine as user needs to copy data from Node A to\n> > > > > Node B,\n> > > > > -- but we still report an error for this case.\n> > > > >\n> > > > > It looks a bit strict to report an ERROR in this case and it seems\n> > > > > not easy to avoid this. So, personally, I think it might be better\n> > > > > to document the correct steps to build the bi-directional\n> > > > > replication and probably also docuemnt the steps to recover if\n> > > > > user accidently did duplicate initial copy if not documented yet.\n> > > > >\n> > > > > In addition, we could also LOG some additional information about\n> > > > > the ORIGIN and initial copy which might help user to analyze if needed.\n> > > > >\n> > > >\n> > > > But why LOG instead of WARNING? I feel in this case there is a\n> > > > chance of inconsistent data so a WARNING like \"publication \"pub1\"\n> > > > could have data from multiple origins\" can be given when the user\n> > > > has specified\n> > > > options: \"copy_data = on, origin = NONE\" while creating a\n> > > > subscription. We give a WARNING during subscription creation when\n> > > > the corresponding publication doesn't exist, eg.\n> > > >\n> > > > postgres=# create subscription sub1 connection 'dbname = postgres'\n> > > > publication pub1;\n> > > > WARNING: publication \"pub1\" does not exist in the publisher\n> > > >\n> > > > Then, we can explain in docs how users can avoid data\n> > > > inconsistencies while setting up replication.\n> > > >\n> > >\n> > > I was wondering if this copy/origin case really should be a NOTICE.\n> > >\n> >\n> > We usually give NOTICE for some sort of additional implicit information, e.g.,\n> > when we create a slot during CREATE SUBSCRIPTION\n> > command: \"NOTICE: created replication slot \"sub1\" on publisher\". IMO, this is\n> > likely to be a problem of data inconsistency so I think here we can choose\n> > between WARNING and LOG. I prefer WARNING but okay with LOG as well if\n> > others feel so. I think we can change this later as well if required. We do have an\n> > option to not do anything and just document it but I feel it is better to give user\n> > some indication of problem here because not everyone reads each update of\n> > documentation.\n> >\n> > Jonathan, Sawada-San, Hou-San, and others, what do you think is the best way\n> > to move forward here?\n>\n> I think it's fine to throw a WARNING in this case given that there is a\n> chance of inconsistent data.\n\nSince there was no objections to change it to throw a warning, I have\nmade the changes for the same.\n\nI have made one change for the documentation patch.\nThe current steps states that:\n1. Create a publication on primary3.\n2. Lock table t1 on primary2 and primary3 in EXCLUSIVE mode until the\nsetup is completed. There is no need to lock table t1 on primary1\nbecause any data changes made will be synchronized while creating the\nsubscription with copy_data = true.\n3. Create a subscription on primary1 to subscribe to primary3.\n4. Create a subscription on primary2 to subscribe to primary3.\n5. Create a subscription on primary3 to subscribe to primary1. Use\ncopy_data = true so that the existing table data is copied during\ninitial sync.\n6. Create a subscription on primary3 to subscribe to primary2. Use\ncopy_data = false because the initial table data would have been\nalready copied in the previous step.\n7. Now the replication setup between primaries primary1, primary2 and\nprimary3 is complete. Incremental changes made on any primary will be\nreplicated to the other two primaries.\n\nWe found a problem with the proposed steps. Here the problem is that\nwe lock table t1 on primary2/primary3 in step2, then we create a\nsubscription on primary3 in step5. While we create subscription in\nstep5, tablesync worker will try to lock the table in primary3 during\nstep5 before copying, as we have already taken a lock, this process\nwill wait.\nEven if the lock is released in this step just before creating\nsubscription, there is a chance that some other transaction can\nacquire the lock and change the data which will result in inconsistent\ndata. Here the locking solution cannot handle adding a new primary in\na consistent way.\n\nI have changed the lock step to mention that user need to ensure no\ndata changes happen until the setup is completed.\nThe attached v42 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Wed, 24 Aug 2022 19:27:14 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Aug 22, 2022 at 9:19 AM houzj.fnst@fujitsu.com\n<houzj.fnst@fujitsu.com> wrote:\n>\n> > Jonathan, Sawada-San, Hou-San, and others, what do you think is the best way\n> > to move forward here?\n>\n> I think it's fine to throw a WARNING in this case given that there is a\n> chance of inconsistent data.\n\nIMHO, since the user has specifically asked for origin=NONE but we do\nnot have any way to detect the origin during initial sync so I think\nthis could be documented and we can also issue the WARNING. So that\nusers notice that part and carefully set up the replication. OTOH, I\ndo not think that giving an error is very inconvenient because we are\nalready providing a new option \"origin=NONE\" so along with that lets\nforce the user to choose between copy_data=off or copy_data=force and\nwith that, there is no scope for mistakes.\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 26 Aug 2022 09:52:27 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Aug 26, 2022 at 9:52 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n>\n> On Mon, Aug 22, 2022 at 9:19 AM houzj.fnst@fujitsu.com\n> <houzj.fnst@fujitsu.com> wrote:\n> >\n> > > Jonathan, Sawada-San, Hou-San, and others, what do you think is the best way\n> > > to move forward here?\n> >\n> > I think it's fine to throw a WARNING in this case given that there is a\n> > chance of inconsistent data.\n>\n> IMHO, since the user has specifically asked for origin=NONE but we do\n> not have any way to detect the origin during initial sync so I think\n> this could be documented and we can also issue the WARNING. So that\n> users notice that part and carefully set up the replication. OTOH, I\n> do not think that giving an error is very inconvenient because we are\n> already providing a new option \"origin=NONE\" so along with that lets\n> force the user to choose between copy_data=off or copy_data=force and\n> with that, there is no scope for mistakes.\n\nSince Jonathan also had suggested to throw a warning as in [1] and\nHou-san also had suggested to throw a warning as in [2], I have made\nchanges to throw a warning and also documented the following contents\nfor the same:\n+ <para>\n+ If the subscription is created with <literal>origin = none</literal> and\n+ <literal>copy_data = true</literal>, it will check if the publisher has\n+ subscribed to the same table from other publishers and, if so, throw a\n+ warning to notify user to check the publisher tables. The user can ensure\n+ that publisher tables do not have data which has an origin associated before\n+ continuing with any other operations to prevent inconsistent data being\n+ replicated.\n+ </para>\n\nThe changes for the same are available in the v42 patch at [3].\n[1] - https://www.postgresql.org/message-id/afb653ba-e2b1-33a3-a54c-849f4466e1b4%40postgresql.org\n[2] - https://www.postgresql.org/message-id/OS0PR01MB5716C383623ADAD64CE4841194719%40OS0PR01MB5716.jpnprd01.prod.outlook.com\n[3] - https://www.postgresql.org/message-id/CALDaNm2oLWsSYtOFLdOkbrpC94%3DJZzKMCYDJoiZaqAX_Hn%2BU9Q%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 29 Aug 2022 08:24:06 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Aug 29, 2022 at 8:24 AM vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Fri, Aug 26, 2022 at 9:52 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\n> >\n> > IMHO, since the user has specifically asked for origin=NONE but we do\n> > not have any way to detect the origin during initial sync so I think\n> > this could be documented and we can also issue the WARNING. So that\n> > users notice that part and carefully set up the replication. OTOH, I\n> > do not think that giving an error is very inconvenient because we are\n> > already providing a new option \"origin=NONE\" so along with that lets\n> > force the user to choose between copy_data=off or copy_data=force and\n> > with that, there is no scope for mistakes.\n>\n\nInitially, I also thought that giving an error should be okay in this\ncase but later multiple people voted against that idea primarily\nbecause we won't be able to detect whether the initial sync data\ncontains data from multiple origins. Consider that even though the\npublisher is subscribed to some other publisher but it may not have\ngotten any data from it by the time we try to subscribe from it. Also,\none might have removed the subscription that led to data from multiple\norigins on the publisher, in this case, may be one can say that we can\nconsider that the data is now owned by the publisher but I am not\nsure. So, considering this, I think giving a WARNING and documenting\nhow to set up replication correctly sounds like a reasonable way\nforward.\n\n> Since Jonathan also had suggested to throw a warning as in [1] and\n> Hou-san also had suggested to throw a warning as in [2],\n>\n\nAgreed that multiple seems to be in favor of that approach. We can\nalways promote it to ERROR later if others and or users felt so.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 29 Aug 2022 08:40:29 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are some review comments for patch v42-0001:\n\n======\n\n1. Commit message\n\nA later review comment below suggests some changes to the WARNING\nmessage so if those changes are made then the example in this commit\nmessage also needs to be modified.\n\n======\n\n2. doc/src/sgml/ref/alter_subscription.sgml - copy_data\n\nRefer to the Notes for the usage of true for copy_data parameter and\nits interaction with the origin parameter.\n\nI think saying \"usage of true\" sounded a bit strange.\n\nSUGGESTION\nRefer to the Notes about how copy_data = true can interact with the\norigin parameter.\n\n======\n\n3. doc/src/sgml/ref/create_subscription.sgml - copy data\n\nRefer to the Notes for the usage of true for copy_data parameter and\nits interaction with the origin parameter.\n\nSUGGESTION\n(same as #2)\n\n~~\n\n4. doc/src/sgml/ref/create_subscription.sgml - origin\n\nRefer to the Notes for the usage of true for copy_data parameter and\nits interaction with the origin parameter.\n\nSUGGESTION\n(same as #2)\n\n~~~\n\n5. doc/src/sgml/ref/create_subscription.sgml - Notes\n\n+ <para>\n+ If the subscription is created with <literal>origin = none</literal> and\n+ <literal>copy_data = true</literal>, it will check if the publisher has\n+ subscribed to the same table from other publishers and, if so, throw a\n+ warning to notify user to check the publisher tables. The user can ensure\n\n\"throw a warning to notify user\" -> \"log a warning to notify the user\"\n\n~~\n\n6.\n\n+ warning to notify user to check the publisher tables. The user can ensure\n+ that publisher tables do not have data which has an origin associated before\n+ continuing with any other operations to prevent inconsistent data being\n+ replicated.\n+ </para>\n\n6a.\n\nI'm not so sure about this. IMO the warning is not about the\nreplication – really it is about the COPY which has already happened\nanyway. So the user can't really prevent anything from going wrong;\ninstead, they have to take some action to clean up if anything did go\nwrong.\n\nSUGGESTION\nBefore continuing with other operations the user should check that\npublisher tables did not have data with different origins, otherwise\ninconsistent data may have been copied.\n\n~\n\n6b.\n\nI am also wondering what can the user do now. Assuming there was bad\nCOPY then the subscriber table has already got unwanted stuff copied\ninto it. Is there any advice we can give to help users fix this mess?\n\n======\n\n7. src/backend/commands/subscriptioncmds.c - AlterSubscription_refresh\n\n+ /* Check whether we can allow copy of newly added relations. */\n+ check_pub_table_subscribed(wrconn, sub->publications, copy_data,\n+ sub->origin, subrel_local_oids,\n+ subrel_count);\n\n\"whether we can allow\" seems not quite the right wording here anymore,\nbecause now there is no ERROR stopping this - so if there was unwanted\ndata the COPY will proceed to copy it anyhow...\n\n~~~\n\n8. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n\n@@ -1781,6 +1794,121 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\n table_close(rel, RowExclusiveLock);\n }\n\n+/*\n+ * Check and throw a warning if the publisher has subscribed to the same table\n+ * from some other publisher. This check is required only if \"copy_data = true\"\n+ * and \"origin = none\" for CREATE SUBSCRIPTION and\n+ * ALTER SUBSCRIPTION ... REFRESH statements to notify user that data having\n+ * origin might have been copied.\n\n\"throw a warning\" -> \"log a warning\"\n\n\"to notify user\" -> \"to notify the user\" ?\n\n~~~\n\n9.\n\n+ if (copydata == false || !origin ||\n+ (pg_strcasecmp(origin, LOGICALREP_ORIGIN_NONE) != 0))\n+ return;\n\nSUGGESTION\nif (!copydata || !origin ||\n(pg_strcasecmp(origin, LOGICALREP_ORIGIN_NONE) != 0))\n\n~~~\n\n10. (Question)\n\nI can't tell just by reading the code if FOR ALL TABLES case is\nhandled – e.g. will this recognise the case were the publisher might\nhave table data from other origins because it has a subscription on\nsome other node that was publishing \"FOR ALL TABLES\"?\n\n~~~\n\n11.\n\n+ ereport(WARNING,\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n+ errmsg(\"publisher has subscribed table \\\"%s.%s\\\" from some other publisher\",\n+ nspname, relname),\n+ errdetail(\"Publisher might have subscribed one or more tables from\nsome other publisher.\"),\n+ errhint(\"Verify that these publisher tables do not have data that\nhas an origin associated before proceeding to avoid inconsistency.\"));\n+ break;\n\n11a.\nI'm not sure having that \"break\" code is correct logic. Won't that\nmean the user will only get a warning for the first potential problem\nencountered but then other potential problems tables will have no\nwarnings at all so the user may not be made aware of them? Perhaps you\nneed to first gather the list of all the suspicious tables before\nlogging a single warning which shows that list? Or perhaps you need to\nlog multiple warnings – one warning per suspicious table?\n\n~\n\n11b.\nThe WARNING message seems a bit inside-out because the errmsg is\ngiving more details than the errdetail.\n\nSUGGESTIONS (or something similar)\nerrmsg - \"subscription XXX requested origin=NONE but may have copied\ndata that had a different origin.\"\nerrdetail – Publisher YYY has subscribed table \\\"%s.%s\\\" from some\nother publisher\"\n\n~\n\n11c.\nThe errhint sentence is unusual.\n\nBEFORE\n\"Verify that these publisher tables do not have data that has an\norigin associated before proceeding to avoid inconsistency.\"\n\nSUGGESTION (or something similar)\nBefore proceeding, verify that initial data copied from the publisher\ntables did not come from other origins.\n\n======\n\n12. src/test/subscription/t/030_origin.pl\n\n+# have remotely originated data from node_A. We throw a warning, in this case,\n+# to draw attention to there being possible remote data.\n\n\"throw a warning\" -> \"log a warning\" (this occurs 2x)\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 29 Aug 2022 16:29:27 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are some review comments for patch v42-0002:\n\n======\n\n1. doc/src/sgml/logical-replication.sgml\n\n<literal>copy_data = true </literal>\n\nThere are a couple of these tags where there is a trailing space\nbefore the </literal>. I guess it is doing no harm, but it is doing no\ngood either, so IMO better to get rid of the space.\n\n~~\n\n2. doc/src/sgml/logical-replication.sgml - 31.3.1. Setting replication\nbetween two primaries\n\nUser need to ensure that no data changes happen on table t1 on\nprimary1 and primary2 until the setup is completed.\n\nSUGGESTION\nThe user must ensure that no data changes happen on table t1 of\nprimary1 and primary2 until the setup is completed.\n\n~~~\n\n3. doc/src/sgml/logical-replication.sgml - 31.3.2. Adding a new\nprimary when there is no table data on any of the primaries\n\nUser need to ensure that no data changes happen on table t1 on all the\nprimaries primary1, primary2 and primary3 until the setup is\ncompleted.\n\nSUGGESTION\nThe user must ensure that no data changes happen on table t1 of all\nthe primaries primary1, primary2 and primary3 until the setup is\ncompleted.\n\n~~~\n\n4. doc/src/sgml/logical-replication.sgml - 31.3.3. Adding a new\nprimary when table data is present on the existing primaries\n\nUser need to ensure that no operations should be performed on table t1\non primary2 and primary3 until the setup is completed.\n\nSUGGESTION\nThe user must ensure that no operations are performed on table t1 of\nprimary2 and primary3 until the setup is completed.\n\nHere also you changed the wording - \"no data changes happen\" versus\n\"no operations should be performed\". Was that a deliberate difference?\nMaybe they should all be consistent wording? If they are\ninterchangeable IMO the \"no operations are performed...\" wording was\nthe better of the two.\n\n~~~\n\n5. doc/src/sgml/logical-replication.sgml - 31.3.4. Generic steps for\nadding a new primary to an existing set of primaries\n\nStep-2: User need to ensure that no changes happen on the required\ntables of the new primary until the setup is complete.\n\nSUGGESTION\nStep-2: The user must ensure that no changes happen on the required\ntables of the new primary until the setup is complete.\n\n~~~\n\n6.\n\nStep-4. User need to ensure that no changes happen on the required\ntables of the existing primaries except the first primary until the\nsetup is complete.\n\nSUGGESTION\nStep-4. The user must ensure that no changes happen on the required\ntables of the existing primaries except the first primary until the\nsetup is complete.\n\n~~~\n\n7. (the example)\n\n7a.\nStep-2. User need to ensure that no changes happen on table t1 on\nprimary4 until the setup is completed.\n\nSUGGESTION\nStep-2. The user must ensure that no changes happen on table t1 of\nprimary4 until the setup is completed.\n\n~\n\n7b.\nStep-4. User need to ensure that no changes happen on table t1 on\nprimary2 and primary3 until the setup is completed.\n\nSUGGESTION\nStep-4. The user must ensure that no changes happen on table t1 of\nprimary2 and primary3 until the setup is completed.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 29 Aug 2022 16:31:07 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 24, 2022 at 7:27 PM vignesh C <vignesh21@gmail.com> wrote:\n>\n> Since there was no objections to change it to throw a warning, I have\n> made the changes for the same.\n>\n\nReview comments for v42-0001*\n==========================\n1. Can we improve the query in check_pub_table_subscribed() so that it\ndoesn't fetch any table that is already part of the subscription on\nthe subscriber? This may be better than what the patch is doing which\nis to first fetch such information and then skip it. If forming this\nquery turns out to be too complex then we can retain your method as\nwell but I feel it is worth trying to optimize the query used in the\npatch.\n\n2. I thought that it may be better to fetch all the tables that have\nthe possibility to have data from more than one origin but maybe the\nlist will be too long especially for FOR ALL TABLE types of cases. Is\nthis the reason you have decided to give a WARNING as soon as you see\nany such table? If so, probably adding a comment for it can be good.\n\n3.\n+ $node_publisher->wait_for_catchup($sub_name);\n+\n+ # also wait for initial table sync to finish\n+ $node_subscriber->poll_query_until('postgres', $synced_query)\n+ or die \"Timed out while waiting for subscriber to synchronize data\";\n....\n....\n....\n+$node_B->safe_psql(\n+ 'postgres', \"\n+ ALTER SUBSCRIPTION $subname_BA REFRESH PUBLICATION\");\n+\n+$node_B->poll_query_until('postgres', $synced_query)\n+ or die \"Timed out while waiting for subscriber to synchronize data\";\n\nYou can replace this and any similar code in the patch with a method\nused in commit 0c20dd33db.\n\n4.\n+###############################################################################\n+# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\n+# replication setup when the existing nodes (node_A & node_B) has pre-existing\n+# data and the new node (node_C) does not have any data.\n+###############################################################################\n+$result = $node_A->safe_psql('postgres', \"SELECT * FROM tab ORDER BY 1;\");\n+is($result, qq(), 'Check existing data');\n+\n+$result = $node_B->safe_psql('postgres', \"SELECT * FROM tab ORDER BY 1;\");\n+is($result, qq(), 'Check existing data');\n+\n+$result = $node_C->safe_psql('postgres', \"SELECT * FROM tab ORDER BY 1;\");\n+is($result, qq(), 'Check existing data');\n\nThe comments say that for this test, two of the nodes have some\npre-existing data but I don't see the same in the test results. The\ntest following this one will also have a similar effect. BTW, I am not\nsure all the new three node tests added by this patch are required\nbecause I don't see if they have additional code coverage or test\nanything which is not tested without those. This has doubled the\namount of test timing for this test file which is okay if these tests\nmake any useful contribution but otherwise, it may be better to remove\nthese. Am, I missing something?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 29 Aug 2022 16:34:51 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Aug 29, 2022 at 11:59 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some review comments for patch v42-0001:\n>\n> ~~\n>\n> 6.\n>\n> + warning to notify user to check the publisher tables. The user can ensure\n> + that publisher tables do not have data which has an origin associated before\n> + continuing with any other operations to prevent inconsistent data being\n> + replicated.\n> + </para>\n>\n> 6a.\n>\n> I'm not so sure about this. IMO the warning is not about the\n> replication – really it is about the COPY which has already happened\n> anyway. So the user can't really prevent anything from going wrong;\n> instead, they have to take some action to clean up if anything did go\n> wrong.\n>\n> SUGGESTION\n> Before continuing with other operations the user should check that\n> publisher tables did not have data with different origins, otherwise\n> inconsistent data may have been copied.\n>\n> ~\n\nI would like to avoid the use of 'may' here. So, let's change it to\nsomething like: \"Before continuing with other operations the user\nshould check that publisher tables did not have data with different\norigins to prevent data inconsistency issues on the subscriber.\"\n>\n> 6b.\n>\n> I am also wondering what can the user do now. Assuming there was bad\n> COPY then the subscriber table has already got unwanted stuff copied\n> into it. Is there any advice we can give to help users fix this mess?\n>\n\nI don't think the user can do much in that case. She will probably\nneed to re-create the replication.\n\n>\n> 11.\n>\n> + ereport(WARNING,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"publisher has subscribed table \\\"%s.%s\\\" from some other publisher\",\n> + nspname, relname),\n> + errdetail(\"Publisher might have subscribed one or more tables from\n> some other publisher.\"),\n> + errhint(\"Verify that these publisher tables do not have data that\n> has an origin associated before proceeding to avoid inconsistency.\"));\n> + break;\n>\n> 11a.\n> I'm not sure having that \"break\" code is correct logic. Won't that\n> mean the user will only get a warning for the first potential problem\n> encountered but then other potential problems tables will have no\n> warnings at all so the user may not be made aware of them? Perhaps you\n> need to first gather the list of all the suspicious tables before\n> logging a single warning which shows that list? Or perhaps you need to\n> log multiple warnings – one warning per suspicious table?\n>\n> ~\n>\n> 11b.\n> The WARNING message seems a bit inside-out because the errmsg is\n> giving more details than the errdetail.\n>\n> SUGGESTIONS (or something similar)\n> errmsg - \"subscription XXX requested origin=NONE but may have copied\n> data that had a different origin.\"\n> errdetail – Publisher YYY has subscribed table \\\"%s.%s\\\" from some\n> other publisher\"\n>\n\nYour suggestion is better but we can't say 'copied' because the copy\nmay start afterwards by tablesync worker. Also, using 'may' is not\nadvisable in error messages. How about : \"subscription XXX requested\norigin=NONE but might copy data that had a different origin.\"?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 29 Aug 2022 17:21:44 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, 29 Aug 2022 at 11:59, Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some review comments for patch v42-0001:\n>\n> ======\n>\n> 1. Commit message\n>\n> A later review comment below suggests some changes to the WARNING\n> message so if those changes are made then the example in this commit\n> message also needs to be modified.\n\nModified\n\n> ======\n>\n> 2. doc/src/sgml/ref/alter_subscription.sgml - copy_data\n>\n> Refer to the Notes for the usage of true for copy_data parameter and\n> its interaction with the origin parameter.\n>\n> I think saying \"usage of true\" sounded a bit strange.\n>\n> SUGGESTION\n> Refer to the Notes about how copy_data = true can interact with the\n> origin parameter.\n\nModified\n\n> ======\n>\n> 3. doc/src/sgml/ref/create_subscription.sgml - copy data\n>\n> Refer to the Notes for the usage of true for copy_data parameter and\n> its interaction with the origin parameter.\n>\n> SUGGESTION\n> (same as #2)\n\nModified\n\n> ~~\n>\n> 4. doc/src/sgml/ref/create_subscription.sgml - origin\n>\n> Refer to the Notes for the usage of true for copy_data parameter and\n> its interaction with the origin parameter.\n>\n> SUGGESTION\n> (same as #2)\n\nModified\n\n> ~~~\n>\n> 5. doc/src/sgml/ref/create_subscription.sgml - Notes\n>\n> + <para>\n> + If the subscription is created with <literal>origin = none</literal> and\n> + <literal>copy_data = true</literal>, it will check if the publisher has\n> + subscribed to the same table from other publishers and, if so, throw a\n> + warning to notify user to check the publisher tables. The user can ensure\n>\n> \"throw a warning to notify user\" -> \"log a warning to notify the user\"\n\nModified\n\n> ~~\n>\n> 6.\n>\n> + warning to notify user to check the publisher tables. The user can ensure\n> + that publisher tables do not have data which has an origin associated before\n> + continuing with any other operations to prevent inconsistent data being\n> + replicated.\n> + </para>\n>\n> 6a.\n>\n> I'm not so sure about this. IMO the warning is not about the\n> replication – really it is about the COPY which has already happened\n> anyway. So the user can't really prevent anything from going wrong;\n> instead, they have to take some action to clean up if anything did go\n> wrong.\n>\n> SUGGESTION\n> Before continuing with other operations the user should check that\n> publisher tables did not have data with different origins, otherwise\n> inconsistent data may have been copied.\n\nModified based on the suggestion provided by Amit.\n\n> ~\n>\n> 6b.\n>\n> I am also wondering what can the user do now. Assuming there was bad\n> COPY then the subscriber table has already got unwanted stuff copied\n> into it. Is there any advice we can give to help users fix this mess?\n\nThere is nothing much a user can do in this case. Only option would be\nto take a backup before the operation and restore it and then recreate\nthe replication setup. I was not sure if we should document these\nsteps as similar inconsistent data could get created without the\norigin option . Thoughts?\n\n> ======\n>\n> 7. src/backend/commands/subscriptioncmds.c - AlterSubscription_refresh\n>\n> + /* Check whether we can allow copy of newly added relations. */\n> + check_pub_table_subscribed(wrconn, sub->publications, copy_data,\n> + sub->origin, subrel_local_oids,\n> + subrel_count);\n>\n> \"whether we can allow\" seems not quite the right wording here anymore,\n> because now there is no ERROR stopping this - so if there was unwanted\n> data the COPY will proceed to copy it anyhow...\n\nRemoved the comment as the function details the necessary things.\n\n> ~~~\n>\n> 8. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> @@ -1781,6 +1794,121 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\n> table_close(rel, RowExclusiveLock);\n> }\n>\n> +/*\n> + * Check and throw a warning if the publisher has subscribed to the same table\n> + * from some other publisher. This check is required only if \"copy_data = true\"\n> + * and \"origin = none\" for CREATE SUBSCRIPTION and\n> + * ALTER SUBSCRIPTION ... REFRESH statements to notify user that data having\n> + * origin might have been copied.\n>\n> \"throw a warning\" -> \"log a warning\"\n>\n> \"to notify user\" -> \"to notify the user\" ?\n\nModified\n\n> ~~~\n>\n> 9.\n>\n> + if (copydata == false || !origin ||\n> + (pg_strcasecmp(origin, LOGICALREP_ORIGIN_NONE) != 0))\n> + return;\n>\n> SUGGESTION\n> if (!copydata || !origin ||\n> (pg_strcasecmp(origin, LOGICALREP_ORIGIN_NONE) != 0))\n\nModified\n\n> ~~~\n>\n> 10. (Question)\n>\n> I can't tell just by reading the code if FOR ALL TABLES case is\n> handled – e.g. will this recognise the case were the publisher might\n> have table data from other origins because it has a subscription on\n> some other node that was publishing \"FOR ALL TABLES\"?\n\nYes it handles it.\n\n> ~~~\n>\n> 11.\n>\n> + ereport(WARNING,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"publisher has subscribed table \\\"%s.%s\\\" from some other publisher\",\n> + nspname, relname),\n> + errdetail(\"Publisher might have subscribed one or more tables from\n> some other publisher.\"),\n> + errhint(\"Verify that these publisher tables do not have data that\n> has an origin associated before proceeding to avoid inconsistency.\"));\n> + break;\n>\n> 11a.\n> I'm not sure having that \"break\" code is correct logic. Won't that\n> mean the user will only get a warning for the first potential problem\n> encountered but then other potential problems tables will have no\n> warnings at all so the user may not be made aware of them? Perhaps you\n> need to first gather the list of all the suspicious tables before\n> logging a single warning which shows that list? Or perhaps you need to\n> log multiple warnings – one warning per suspicious table?\n\nModified to get the list of all the tables, I have limited it to\ndisplay 100 tables followed by ... to indicate there might be more\ntables. I did not want to loga warning message with many 1000's of\ntables which will not help the readability of the message.\n\n> ~\n>\n> 11b.\n> The WARNING message seems a bit inside-out because the errmsg is\n> giving more details than the errdetail.\n>\n> SUGGESTIONS (or something similar)\n> errmsg - \"subscription XXX requested origin=NONE but may have copied\n> data that had a different origin.\"\n> errdetail – Publisher YYY has subscribed table \\\"%s.%s\\\" from some\n> other publisher\"\n\nWe don't have the tables based on each publisher, it is for all the\npublishers. I have changed the errdetail message slightly.\n\n> ~\n>\n> 11c.\n> The errhint sentence is unusual.\n>\n> BEFORE\n> \"Verify that these publisher tables do not have data that has an\n> origin associated before proceeding to avoid inconsistency.\"\n>\n> SUGGESTION (or something similar)\n> Before proceeding, verify that initial data copied from the publisher\n> tables did not come from other origins.\n\nModified\n\n> ======\n>\n> 12. src/test/subscription/t/030_origin.pl\n>\n> +# have remotely originated data from node_A. We throw a warning, in this case,\n> +# to draw attention to there being possible remote data.\n>\n> \"throw a warning\" -> \"log a warning\" (this occurs 2x)\n\nModified\n\nThe attached v43 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Tue, 30 Aug 2022 22:35:30 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, 29 Aug 2022 at 12:01, Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some review comments for patch v42-0002:\n>\n> ======\n>\n> 1. doc/src/sgml/logical-replication.sgml\n>\n> <literal>copy_data = true </literal>\n>\n> There are a couple of these tags where there is a trailing space\n> before the </literal>. I guess it is doing no harm, but it is doing no\n> good either, so IMO better to get rid of the space.\n\nModified\n\n> ~~\n>\n> 2. doc/src/sgml/logical-replication.sgml - 31.3.1. Setting replication\n> between two primaries\n>\n> User need to ensure that no data changes happen on table t1 on\n> primary1 and primary2 until the setup is completed.\n>\n> SUGGESTION\n> The user must ensure that no data changes happen on table t1 of\n> primary1 and primary2 until the setup is completed.\n\nModified\n\n> ~~~\n>\n> 3. doc/src/sgml/logical-replication.sgml - 31.3.2. Adding a new\n> primary when there is no table data on any of the primaries\n>\n> User need to ensure that no data changes happen on table t1 on all the\n> primaries primary1, primary2 and primary3 until the setup is\n> completed.\n>\n> SUGGESTION\n> The user must ensure that no data changes happen on table t1 of all\n> the primaries primary1, primary2 and primary3 until the setup is\n> completed.\n\nModified\n\n> ~~~\n>\n> 4. doc/src/sgml/logical-replication.sgml - 31.3.3. Adding a new\n> primary when table data is present on the existing primaries\n>\n> User need to ensure that no operations should be performed on table t1\n> on primary2 and primary3 until the setup is completed.\n>\n> SUGGESTION\n> The user must ensure that no operations are performed on table t1 of\n> primary2 and primary3 until the setup is completed.\n>\n> Here also you changed the wording - \"no data changes happen\" versus\n> \"no operations should be performed\". Was that a deliberate difference?\n> Maybe they should all be consistent wording? If they are\n> interchangeable IMO the \"no operations are performed...\" wording was\n> the better of the two.\n\nModified, it was not a deliberate change. I have changed it to \"no\noperations are performed\" in other places too.\n\n> ~~~\n>\n> 5. doc/src/sgml/logical-replication.sgml - 31.3.4. Generic steps for\n> adding a new primary to an existing set of primaries\n>\n> Step-2: User need to ensure that no changes happen on the required\n> tables of the new primary until the setup is complete.\n>\n> SUGGESTION\n> Step-2: The user must ensure that no changes happen on the required\n> tables of the new primary until the setup is complete.\n\nModified\n\n> ~~~\n>\n> 6.\n>\n> Step-4. User need to ensure that no changes happen on the required\n> tables of the existing primaries except the first primary until the\n> setup is complete.\n>\n> SUGGESTION\n> Step-4. The user must ensure that no changes happen on the required\n> tables of the existing primaries except the first primary until the\n> setup is complete.\n\nModified\n\n> ~~~\n>\n> 7. (the example)\n>\n> 7a.\n> Step-2. User need to ensure that no changes happen on table t1 on\n> primary4 until the setup is completed.\n>\n> SUGGESTION\n> Step-2. The user must ensure that no changes happen on table t1 of\n> primary4 until the setup is completed.\n\nModified\n\n> ~\n>\n> 7b.\n> Step-4. User need to ensure that no changes happen on table t1 on\n> primary2 and primary3 until the setup is completed.\n>\n> SUGGESTION\n> Step-4. The user must ensure that no changes happen on table t1 of\n> primary2 and primary3 until the setup is completed.\n\nModified\n\nThanks for the comments, the changes for the same are available in the\nv43 patch attached at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm1V%2BAvzPsUcq%3DmNYG%2BJfAyovTWBe4vWKTknOgH_ko1E0Q%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 30 Aug 2022 22:41:02 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for v43-0001.\n\n======\n\n1. Commit message\n\nThe initial copy phase has no way to know the origin of the row data,\nso if 'copy_data = true' in the step 4 below, log a warning to notify user\nthat potentially non-local data might have been copied.\n\n1a.\n\"in the step 4\" -> \"in step 4\"\n\n~\n\n1b.\n\"notify user\" -> \"notify the user\"\n\n======\n\n2. doc/src/sgml/ref/create_subscription.sgml\n\n+ <para>\n+ If the subscription is created with <literal>origin = none</literal> and\n+ <literal>copy_data = true</literal>, it will check if the publisher has\n+ subscribed to the same table from other publishers and, if so, log a\n+ warning to notify the user to check the publisher tables. Before continuing\n+ with other operations the user should check that publisher tables did not\n+ have data with different origins to prevent data inconsistency issues on the\n+ subscriber.\n+ </para>\n\nI am not sure about that wording saying \"to prevent data inconsistency\nissues\" because I thought maybe is already too late because any\nunwanted origin data is already copied during the initial sync. I\nthink the user can do it try to clean up any problems caused before\nthings become even worse (??)\n\n~~~\n\nYou asked for my thoughts (see [1] 6b):\n\n> There is nothing much a user can do in this case. Only option would be\n> to take a backup before the operation and restore it and then recreate\n> the replication setup. I was not sure if we should document these\n> steps as similar inconsistent data could get created without the\n> origin option . Thoughts?\n\nI think those cases are a bit different:\n\n- For a normal scenario (i.e. origin=ANY) then inconsistent data just\nrefers to problems like maybe PK violations so I think you just get\nwhat you get and have to deal with the errors...\n\n- But when the user said origin=NONE they are specifically saying they\ndo NOT want any non-local data, yet they still might end up getting\ndata contrary to their wishes. So I think maybe there ought to be\nsomething documented to warn about this potentially happening. My\nfirst impression is that all this seems like a kind of terrible\nproblem because if it happens then cleanup could be difficult - if\nthat subscriber in turn was also a publisher then this bad data might\nhave propagated all over the place already! Anyway, I guess all this\ncould be mentioned in some (new ?) section of the\nlogical-replication.sgml file (i.e. patch 0002).\n\nIDEA: I was wondering if the documentation should recommend (or maybe\nwe can even enforce this) that when using origin=NONE the user should\nalways create the subscription with enabled=FALSE. That way if there\nare some warnings about potential bad origin data then there it might\nbe possible to take some action or change their mind *before* any\nunwanted data pollutes everything.\n\n======\n\n3. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n\n+ if (count == 0)\n+ tablenames = makeStringInfo();\n+ else\n+ appendStringInfoString(tablenames, \", \");\n+\n+ appendStringInfo(tablenames, \"%s.%s\", nspname, relname);\n+ count++;\n\nI think the quotes are not correct - each separate table name should be quoted.\n\n~~~\n\n4.\n\n+ /*\n+ * There might be many tables present in this case, we will display a\n+ * maximum of 100 tables followed by \"...\" to indicate there might be\n+ * more tables.\n+ */\n+ if (count == 100)\n+ {\n+ appendStringInfoString(tablenames, \", ...\");\n+ break;\n+ }\n\n4a.\nIMO this code should be part of the previous if/else\n\ne.g.\nif (count == 0) ... makeStringInfo()\nelse if (count > 100) ... append ellipsis \"...\" and break out of the loop\nelse ... append table name to the message\n\nDoing it in this way means \"...\" definitely means there are more than\n100 bad tables.\n\n~\n\n4b.\nChanging like above (#4a) simplifies the comment because it removes\ndoubts like \"might be…\" since you already know you found the 101st\ntable.\n\nSUGGESTION (comment)\nThe message displays a maximum of 100 tables having potentially\nnon-local data. Any more than that will be indicated by \"...\".\n\n~\n\n4c.\nIt still seems a bit disconcerting to me that there can be cases where\nbad data was possibly copied but there is no record of what tables\nhave been polluted. Perhaps when the maximum bad tables are exceeded\nthen there can be some additional message to tell users what SQL they\ncan use to discover what all the other bad tables were.\n\n~~~\n\n5.\n\n+ * XXX: For simplicity, we don't check whether the table has any data\n+ * or not. If the table doesn't have any data then we don't need to\n+ * distinguish between data having origin and data not having origin so\n+ * we can avoid throwing a warning in that case.\n\n\"throwing a warning\" -> \"logging a warning\"\n\n~~~\n\n6.\n\n+ errdetail(\"Subscribed table(s) \\\"%s\\\" has been subscribed from some\nother publisher.\",\n+ tablenames->data),\n\nI don't think the table name list should be quoted in the errdetail,\nbecause each of the individual table names should have been quoted.\nSee previous review comment #3.\n\n======\n\n7. src/test/subscription/t/030_origin.pl\n\n+###############################################################################\n+# Specify origin as 'none' which indicates that the publisher should only\n+# replicate the changes that are generated locally from node_B, but in\n+# this case since the node_B is also subscribing data from node_A, node_B can\n+# have remotely originated data from node_A. We log a warning, in this case,\n+# to draw attention to there being possible remote data.\n+###############################################################################\n\nThis comment wording makes it seem like this is using different\nsubscription parameters from the other tests that preceded it, but in\nfact the prior tests also been using origin=none. (??)\n\nMaybe it just needs a trivial rewording.\n\nSUGGESTION\nSpecifying origin=NONE indicates that...\n\n~~~\n\n8.\n\n+like(\n+ $stderr,\n+ qr/WARNING: ( [A-Z0-9]+:)? subscription tap_sub_a_b_2 requested\norigin=NONE but might copy data that had a different origin/,\n+ \"Create subscription with origin = none and copy_data when the\npublisher has subscribed same table\"\n+);\n\nIs it possible for these checks of the log file to get the errdetail\nmessage too? Then the test will be more readable for the expected\nresult and you will also have the bonus of testing the table-name list\nin the warning more thoroughly.\n\n------\n[1] https://www.postgresql.org/message-id/CALDaNm1V%2BAvzPsUcq%3DmNYG%2BJfAyovTWBe4vWKTknOgH_ko1E0Q%40mail.gmail.com\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 31 Aug 2022 16:05:01 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are some review comments for patch v43-0002:\n\n======\n\n1. doc/src/sgml/ref/create_subscription.sgml\n\n@@ -403,7 +403,9 @@ CREATE SUBSCRIPTION <replaceable\nclass=\"parameter\">subscription_name</replaceabl\n warning to notify the user to check the publisher tables. Before continuing\n with other operations the user should check that publisher tables did not\n have data with different origins to prevent data inconsistency issues on the\n- subscriber.\n+ subscriber. Refer to <xref linkend=\"replication-between-primaries\"/> for\n+ how <literal>copy_data</literal> and <literal>origin</literal> can be used\n+ to set up replication between primaries.\n </para>\n\nRegarding my earlier v43-0001 review (see [1] comment #2) perhaps\nanother pg docs section should be added in the\nlogical-replication.sgml (e.g. \"Specifying origins during CREATE\nSUBSCRIPTION\"), so then this Notes text also should have more added to\nit.\n\nSUGGESTION\nRefer to <XXX_REF> for details about potential initialization\ninconsistency warnings using origin=NONE.\nRefer to <YYY_REF> for how copy_data and origin can be used to set up\nreplication between primaries.\n\n------\n[1] https://www.postgresql.org/message-id/CAHut%2BPvonTd423-cWqoxh0w8Bd_Po3OToqqyxuR1iMNmxSLr_Q%40mail.gmail.com\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 31 Aug 2022 16:15:29 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 31, 2022 at 11:35 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for v43-0001.\n>\n> ======\n>\n> 1. Commit message\n>\n> The initial copy phase has no way to know the origin of the row data,\n> so if 'copy_data = true' in the step 4 below, log a warning to notify user\n> that potentially non-local data might have been copied.\n>\n> 1a.\n> \"in the step 4\" -> \"in step 4\"\n>\n> ~\n>\n> 1b.\n> \"notify user\" -> \"notify the user\"\n>\n> ======\n>\n> 2. doc/src/sgml/ref/create_subscription.sgml\n>\n> + <para>\n> + If the subscription is created with <literal>origin = none</literal> and\n> + <literal>copy_data = true</literal>, it will check if the publisher has\n> + subscribed to the same table from other publishers and, if so, log a\n> + warning to notify the user to check the publisher tables. Before continuing\n> + with other operations the user should check that publisher tables did not\n> + have data with different origins to prevent data inconsistency issues on the\n> + subscriber.\n> + </para>\n>\n> I am not sure about that wording saying \"to prevent data inconsistency\n> issues\" because I thought maybe is already too late because any\n> unwanted origin data is already copied during the initial sync. I\n> think the user can do it try to clean up any problems caused before\n> things become even worse (??)\n>\n\nOh no, it is not too late in all cases. The problem can only occur if\nthe user will try to subscribe from all the different publications\nwith copy_data = true. We are anyway trying to clarify in the second\npatch the right way to accomplish this. So, I am not sure what better\nwe can do here. The only bulletproof way is to provide some APIs where\nusers don't need to bother about all these cases, basically something\nsimilar to what Kuroda-San is working on in the thread [1].\n\n> ~~~\n>\n> You asked for my thoughts (see [1] 6b):\n>\n> > There is nothing much a user can do in this case. Only option would be\n> > to take a backup before the operation and restore it and then recreate\n> > the replication setup. I was not sure if we should document these\n> > steps as similar inconsistent data could get created without the\n> > origin option . Thoughts?\n>\n> I think those cases are a bit different:\n>\n> - For a normal scenario (i.e. origin=ANY) then inconsistent data just\n> refers to problems like maybe PK violations so I think you just get\n> what you get and have to deal with the errors...\n>\n> - But when the user said origin=NONE they are specifically saying they\n> do NOT want any non-local data, yet they still might end up getting\n> data contrary to their wishes. So I think maybe there ought to be\n> something documented to warn about this potentially happening. My\n> first impression is that all this seems like a kind of terrible\n> problem because if it happens then cleanup could be difficult - if\n> that subscriber in turn was also a publisher then this bad data might\n> have propagated all over the place already! Anyway, I guess all this\n> could be mentioned in some (new ?) section of the\n> logical-replication.sgml file (i.e. patch 0002).\n>\n> IDEA: I was wondering if the documentation should recommend (or maybe\n> we can even enforce this) that when using origin=NONE the user should\n> always create the subscription with enabled=FALSE. That way if there\n> are some warnings about potential bad origin data then there it might\n> be possible to take some action or change their mind *before* any\n> unwanted data pollutes everything.\n>\n\nIf we want to give this suggestion, then we need to also say that this\nis only valid when copy_data is true. The other drawback is if users\nalways need to do this to use origin=NONE then this can be a burden to\nthe user. I am not against having such a NOTE but the tone should not\nbe to enforce users to do this.\n\n[1] - https://www.postgresql.org/message-id/CAHut%2BPuwRAoWY9pz%3DEubps3ooQCOBFiYPU9Yi%3DVB-U%2ByORU7OA%40mail.gmail.com\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 31 Aug 2022 16:06:40 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 31, 2022 at 8:36 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Aug 31, 2022 at 11:35 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Here are my review comments for v43-0001.\n> >\n...\n> > ======\n> >\n> > 2. doc/src/sgml/ref/create_subscription.sgml\n> >\n> > + <para>\n> > + If the subscription is created with <literal>origin = none</literal> and\n> > + <literal>copy_data = true</literal>, it will check if the publisher has\n> > + subscribed to the same table from other publishers and, if so, log a\n> > + warning to notify the user to check the publisher tables. Before continuing\n> > + with other operations the user should check that publisher tables did not\n> > + have data with different origins to prevent data inconsistency issues on the\n> > + subscriber.\n> > + </para>\n> >\n> > I am not sure about that wording saying \"to prevent data inconsistency\n> > issues\" because I thought maybe is already too late because any\n> > unwanted origin data is already copied during the initial sync. I\n> > think the user can do it try to clean up any problems caused before\n> > things become even worse (??)\n> >\n>\n> Oh no, it is not too late in all cases. The problem can only occur if\n> the user will try to subscribe from all the different publications\n> with copy_data = true. We are anyway trying to clarify in the second\n> patch the right way to accomplish this. So, I am not sure what better\n> we can do here. The only bulletproof way is to provide some APIs where\n> users don't need to bother about all these cases, basically something\n> similar to what Kuroda-San is working on in the thread [1].\n>\n\nThe point of my review comment was only about the wording of the note\n- specifically, you cannot \"prevent\" something (e,g, data\ninconsistency) if it has already happened.\n\nMaybe modify the wording like below?\n\nBEFORE\nBefore continuing with other operations the user should check that\npublisher tables did not have data with different origins to prevent\ndata inconsistency issues on the subscriber.\n\nAFTER\nIf a publisher table with data from different origins was found to\nhave been copied to the subscriber, then some corrective action may be\nnecessary before continuing with other operations.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Thu, 1 Sep 2022 10:19:19 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 31, 2022 1:06 AM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> The attached v43 patch has the changes for the same.\r\n> \r\n\r\nThanks for updating the patch.\r\n\r\nHere is a comment on the 0001 patch.\r\n\r\n+\t\tif (!isnewtable)\r\n+\t\t{\r\n+\t\t\tpfree(nspname);\r\n+\t\t\tpfree(relname);\r\n+\t\t\tcontinue;\r\n+\t\t}\r\n\r\nIf it is a new table, in which case it would log a warning, should we also call\r\npfree()?\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Thu, 1 Sep 2022 02:31:37 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 31, 2022 at 11:35 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for v43-0001.\n>\n> ======\n>\n> 1. Commit message\n>\n> The initial copy phase has no way to know the origin of the row data,\n> so if 'copy_data = true' in the step 4 below, log a warning to notify user\n> that potentially non-local data might have been copied.\n>\n> 1a.\n> \"in the step 4\" -> \"in step 4\"\n>\n> ~\n>\n> 1b.\n> \"notify user\" -> \"notify the user\"\n>\n> ======\n>\n> 2. doc/src/sgml/ref/create_subscription.sgml\n>\n> + <para>\n> + If the subscription is created with <literal>origin = none</literal> and\n> + <literal>copy_data = true</literal>, it will check if the publisher has\n> + subscribed to the same table from other publishers and, if so, log a\n> + warning to notify the user to check the publisher tables. Before continuing\n> + with other operations the user should check that publisher tables did not\n> + have data with different origins to prevent data inconsistency issues on the\n> + subscriber.\n> + </para>\n>\n> I am not sure about that wording saying \"to prevent data inconsistency\n> issues\" because I thought maybe is already too late because any\n> unwanted origin data is already copied during the initial sync. I\n> think the user can do it try to clean up any problems caused before\n> things become even worse (??)\n>\n> ~~~\n>\n> You asked for my thoughts (see [1] 6b):\n>\n> > There is nothing much a user can do in this case. Only option would be\n> > to take a backup before the operation and restore it and then recreate\n> > the replication setup. I was not sure if we should document these\n> > steps as similar inconsistent data could get created without the\n> > origin option . Thoughts?\n>\n> I think those cases are a bit different:\n>\n> - For a normal scenario (i.e. origin=ANY) then inconsistent data just\n> refers to problems like maybe PK violations so I think you just get\n> what you get and have to deal with the errors...\n>\n> - But when the user said origin=NONE they are specifically saying they\n> do NOT want any non-local data, yet they still might end up getting\n> data contrary to their wishes. So I think maybe there ought to be\n> something documented to warn about this potentially happening. My\n> first impression is that all this seems like a kind of terrible\n> problem because if it happens then cleanup could be difficult - if\n> that subscriber in turn was also a publisher then this bad data might\n> have propagated all over the place already! Anyway, I guess all this\n> could be mentioned in some (new ?) section of the\n> logical-replication.sgml file (i.e. patch 0002).\n>\n> IDEA: I was wondering if the documentation should recommend (or maybe\n> we can even enforce this) that when using origin=NONE the user should\n> always create the subscription with enabled=FALSE. That way if there\n> are some warnings about potential bad origin data then there it might\n> be possible to take some action or change their mind *before* any\n> unwanted data pollutes everything.\n>\n> ======\n>\n> 3. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> + if (count == 0)\n> + tablenames = makeStringInfo();\n> + else\n> + appendStringInfoString(tablenames, \", \");\n> +\n> + appendStringInfo(tablenames, \"%s.%s\", nspname, relname);\n> + count++;\n>\n> I think the quotes are not correct - each separate table name should be quoted.\n>\n> ~~~\n>\n> 4.\n>\n> + /*\n> + * There might be many tables present in this case, we will display a\n> + * maximum of 100 tables followed by \"...\" to indicate there might be\n> + * more tables.\n> + */\n> + if (count == 100)\n> + {\n> + appendStringInfoString(tablenames, \", ...\");\n> + break;\n> + }\n>\n> 4a.\n> IMO this code should be part of the previous if/else\n>\n> e.g.\n> if (count == 0) ... makeStringInfo()\n> else if (count > 100) ... append ellipsis \"...\" and break out of the loop\n> else ... append table name to the message\n>\n> Doing it in this way means \"...\" definitely means there are more than\n> 100 bad tables.\n>\n> ~\n>\n> 4b.\n> Changing like above (#4a) simplifies the comment because it removes\n> doubts like \"might be…\" since you already know you found the 101st\n> table.\n>\n> SUGGESTION (comment)\n> The message displays a maximum of 100 tables having potentially\n> non-local data. Any more than that will be indicated by \"...\".\n>\n> ~\n>\n\nAre we using this style of an error message anywhere else in the code?\nIf not, then I think we should avoid it here as well as it appears a\nbit adhoc to me in the sense that why restrict at 100 tables. Can we\ninstead think of displaying the publications list in the message? So\nerrdetail would be something like: Subscribed publications \\\"%s\\\" has\nbeen subscribed from some other publisher. Then we can probably give a\nSQL query for a user to find tables that may data from multiple\norigins especially if that is not complicated. Also, then we can\nchange the function name to check_publications_origin().\n\nFew other minor comments on v43-0001*\n1.\n+ ereport(WARNING,\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n+ errmsg(\"subscription %s requested origin=NONE but might copy data\nthat had a different origin.\",\n+ subname),\n\nIn error message, we don't use full stop at the end.\n\n2.\n+ ereport(WARNING,\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n+ errmsg(\"subscription %s requested origin=NONE but might copy data\nthat had a different origin.\",\n+ subname),\n+ errdetail(\"Subscribed table(s) \\\"%s\\\" has been subscribed from some\nother publisher.\",\n+ tablenames->data),\n\nIt is better to use errdetail_plural here as there could be one or\nmore objects in this message?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 1 Sep 2022 09:19:20 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, 31 Aug 2022 at 11:35, Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for v43-0001.\n>\n> ======\n>\n> 1. Commit message\n>\n> The initial copy phase has no way to know the origin of the row data,\n> so if 'copy_data = true' in the step 4 below, log a warning to notify user\n> that potentially non-local data might have been copied.\n>\n> 1a.\n> \"in the step 4\" -> \"in step 4\"\n\nModified\n\n> ~\n>\n> 1b.\n> \"notify user\" -> \"notify the user\"\n\nModified\n\n> ======\n>\n> 2. doc/src/sgml/ref/create_subscription.sgml\n>\n> + <para>\n> + If the subscription is created with <literal>origin = none</literal> and\n> + <literal>copy_data = true</literal>, it will check if the publisher has\n> + subscribed to the same table from other publishers and, if so, log a\n> + warning to notify the user to check the publisher tables. Before continuing\n> + with other operations the user should check that publisher tables did not\n> + have data with different origins to prevent data inconsistency issues on the\n> + subscriber.\n> + </para>\n>\n> I am not sure about that wording saying \"to prevent data inconsistency\n> issues\" because I thought maybe is already too late because any\n> unwanted origin data is already copied during the initial sync. I\n> think the user can do it try to clean up any problems caused before\n> things become even worse (??)\n>\n> ~~~\n>\n> You asked for my thoughts (see [1] 6b):\n>\n> > There is nothing much a user can do in this case. Only option would be\n> > to take a backup before the operation and restore it and then recreate\n> > the replication setup. I was not sure if we should document these\n> > steps as similar inconsistent data could get created without the\n> > origin option . Thoughts?\n>\n> I think those cases are a bit different:\n>\n> - For a normal scenario (i.e. origin=ANY) then inconsistent data just\n> refers to problems like maybe PK violations so I think you just get\n> what you get and have to deal with the errors...\n>\n> - But when the user said origin=NONE they are specifically saying they\n> do NOT want any non-local data, yet they still might end up getting\n> data contrary to their wishes. So I think maybe there ought to be\n> something documented to warn about this potentially happening. My\n> first impression is that all this seems like a kind of terrible\n> problem because if it happens then cleanup could be difficult - if\n> that subscriber in turn was also a publisher then this bad data might\n> have propagated all over the place already! Anyway, I guess all this\n> could be mentioned in some (new ?) section of the\n> logical-replication.sgml file (i.e. patch 0002).\n>\n> IDEA: I was wondering if the documentation should recommend (or maybe\n> we can even enforce this) that when using origin=NONE the user should\n> always create the subscription with enabled=FALSE. That way if there\n> are some warnings about potential bad origin data then there it might\n> be possible to take some action or change their mind *before* any\n> unwanted data pollutes everything.\n\nUpdated document based on your suggestion.\n\n> ======\n>\n> 3. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n>\n> + if (count == 0)\n> + tablenames = makeStringInfo();\n> + else\n> + appendStringInfoString(tablenames, \", \");\n> +\n> + appendStringInfo(tablenames, \"%s.%s\", nspname, relname);\n> + count++;\n>\n> I think the quotes are not correct - each separate table name should be quoted.\n\nWe will not be displaying tables anymore, the comment does not apply anymore\n\n> ~~~\n>\n> 4.\n>\n> + /*\n> + * There might be many tables present in this case, we will display a\n> + * maximum of 100 tables followed by \"...\" to indicate there might be\n> + * more tables.\n> + */\n> + if (count == 100)\n> + {\n> + appendStringInfoString(tablenames, \", ...\");\n> + break;\n> + }\n>\n> 4a.\n> IMO this code should be part of the previous if/else\n>\n> e.g.\n> if (count == 0) ... makeStringInfo()\n> else if (count > 100) ... append ellipsis \"...\" and break out of the loop\n> else ... append table name to the message\n>\n> Doing it in this way means \"...\" definitely means there are more than\n> 100 bad tables.\n\nThis code is removed, the comment is no more applicable.\n\n> ~\n>\n> 4b.\n> Changing like above (#4a) simplifies the comment because it removes\n> doubts like \"might be…\" since you already know you found the 101st\n> table.\n>\n> SUGGESTION (comment)\n> The message displays a maximum of 100 tables having potentially\n> non-local data. Any more than that will be indicated by \"...\".\n\nThis code is removed, the comment is no more applicable.\n\n> ~\n>\n> 4c.\n> It still seems a bit disconcerting to me that there can be cases where\n> bad data was possibly copied but there is no record of what tables\n> have been polluted. Perhaps when the maximum bad tables are exceeded\n> then there can be some additional message to tell users what SQL they\n> can use to discover what all the other bad tables were.\n\nWe will not be displaying the tablename anymore, we will be displaying\nthe publication names which has been subscribed from other\npublications.\n\n> ~~~\n>\n> 5.\n>\n> + * XXX: For simplicity, we don't check whether the table has any data\n> + * or not. If the table doesn't have any data then we don't need to\n> + * distinguish between data having origin and data not having origin so\n> + * we can avoid throwing a warning in that case.\n>\n> \"throwing a warning\" -> \"logging a warning\"\n\nModified\n\n> ~~~\n>\n> 6.\n>\n> + errdetail(\"Subscribed table(s) \\\"%s\\\" has been subscribed from some\n> other publisher.\",\n> + tablenames->data),\n>\n> I don't think the table name list should be quoted in the errdetail,\n> because each of the individual table names should have been quoted.\n> See previous review comment #3.\n\nModified\n\n> ======\n>\n> 7. src/test/subscription/t/030_origin.pl\n>\n> +###############################################################################\n> +# Specify origin as 'none' which indicates that the publisher should only\n> +# replicate the changes that are generated locally from node_B, but in\n> +# this case since the node_B is also subscribing data from node_A, node_B can\n> +# have remotely originated data from node_A. We log a warning, in this case,\n> +# to draw attention to there being possible remote data.\n> +###############################################################################\n>\n> This comment wording makes it seem like this is using different\n> subscription parameters from the other tests that preceded it, but in\n> fact the prior tests also been using origin=none. (??)\n>\n> Maybe it just needs a trivial rewording.\n>\n> SUGGESTION\n> Specifying origin=NONE indicates that...\n\nModified\n\n> ~~~\n>\n> 8.\n>\n> +like(\n> + $stderr,\n> + qr/WARNING: ( [A-Z0-9]+:)? subscription tap_sub_a_b_2 requested\n> origin=NONE but might copy data that had a different origin/,\n> + \"Create subscription with origin = none and copy_data when the\n> publisher has subscribed same table\"\n> +);\n>\n> Is it possible for these checks of the log file to get the errdetail\n> message too? Then the test will be more readable for the expected\n> result and you will also have the bonus of testing the table-name list\n> in the warning more thoroughly.\n\nI did not see any tap test doing this, I did not want to make some\nchange which would fail in buildfarm machines. I have not done any\nchange for this.\n\nThanks for the comments, the attached v44 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Fri, 2 Sep 2022 00:09:11 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, 29 Aug 2022 at 16:35, Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Aug 24, 2022 at 7:27 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Since there was no objections to change it to throw a warning, I have\n> > made the changes for the same.\n> >\n>\n> Review comments for v42-0001*\n> ==========================\n> 1. Can we improve the query in check_pub_table_subscribed() so that it\n> doesn't fetch any table that is already part of the subscription on\n> the subscriber? This may be better than what the patch is doing which\n> is to first fetch such information and then skip it. If forming this\n> query turns out to be too complex then we can retain your method as\n> well but I feel it is worth trying to optimize the query used in the\n> patch.\n\nI'm analysing this, I will post a reply for this soon.\n\n> 2. I thought that it may be better to fetch all the tables that have\n> the possibility to have data from more than one origin but maybe the\n> list will be too long especially for FOR ALL TABLE types of cases. Is\n> this the reason you have decided to give a WARNING as soon as you see\n> any such table? If so, probably adding a comment for it can be good.\n\nYes that was the reason, now I have changed it to display the\npublications based on your recent comment whose count will be\nsignificantly very much lesser compared to the number of tables.\n\n> 3.\n> + $node_publisher->wait_for_catchup($sub_name);\n> +\n> + # also wait for initial table sync to finish\n> + $node_subscriber->poll_query_until('postgres', $synced_query)\n> + or die \"Timed out while waiting for subscriber to synchronize data\";\n> ....\n> ....\n> ....\n> +$node_B->safe_psql(\n> + 'postgres', \"\n> + ALTER SUBSCRIPTION $subname_BA REFRESH PUBLICATION\");\n> +\n> +$node_B->poll_query_until('postgres', $synced_query)\n> + or die \"Timed out while waiting for subscriber to synchronize data\";\n>\n> You can replace this and any similar code in the patch with a method\n> used in commit 0c20dd33db.\n\nModified\n\n> 4.\n> +###############################################################################\n> +# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\n> +# replication setup when the existing nodes (node_A & node_B) has pre-existing\n> +# data and the new node (node_C) does not have any data.\n> +###############################################################################\n> +$result = $node_A->safe_psql('postgres', \"SELECT * FROM tab ORDER BY 1;\");\n> +is($result, qq(), 'Check existing data');\n> +\n> +$result = $node_B->safe_psql('postgres', \"SELECT * FROM tab ORDER BY 1;\");\n> +is($result, qq(), 'Check existing data');\n> +\n> +$result = $node_C->safe_psql('postgres', \"SELECT * FROM tab ORDER BY 1;\");\n> +is($result, qq(), 'Check existing data');\n>\n> The comments say that for this test, two of the nodes have some\n> pre-existing data but I don't see the same in the test results. The\n> test following this one will also have a similar effect. BTW, I am not\n> sure all the new three node tests added by this patch are required\n> because I don't see if they have additional code coverage or test\n> anything which is not tested without those. This has doubled the\n> amount of test timing for this test file which is okay if these tests\n> make any useful contribution but otherwise, it may be better to remove\n> these. Am, I missing something?\n\nEarlier we felt with the origin patch we could support joining of\nnodes to existing n-wary replication setup in various scenarios, I had\nadded the tests for the same scenarios. But as the patch evolved, I\ncould understand that this patch cannot handle the add node scenarios.\nI have removed these tests.\n\nThanks for the comments, the v44 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm0NRJ1O1cYcZD%3Df7NgynozFprb7zpJSayFN5rcaS44G6Q%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 2 Sep 2022 00:14:09 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, 31 Aug 2022 at 11:45, Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are some review comments for patch v43-0002:\n>\n> ======\n>\n> 1. doc/src/sgml/ref/create_subscription.sgml\n>\n> @@ -403,7 +403,9 @@ CREATE SUBSCRIPTION <replaceable\n> class=\"parameter\">subscription_name</replaceabl\n> warning to notify the user to check the publisher tables. Before continuing\n> with other operations the user should check that publisher tables did not\n> have data with different origins to prevent data inconsistency issues on the\n> - subscriber.\n> + subscriber. Refer to <xref linkend=\"replication-between-primaries\"/> for\n> + how <literal>copy_data</literal> and <literal>origin</literal> can be used\n> + to set up replication between primaries.\n> </para>\n>\n> Regarding my earlier v43-0001 review (see [1] comment #2) perhaps\n> another pg docs section should be added in the\n> logical-replication.sgml (e.g. \"Specifying origins during CREATE\n> SUBSCRIPTION\"), so then this Notes text also should have more added to\n> it.\n>\n> SUGGESTION\n> Refer to <XXX_REF> for details about potential initialization\n> inconsistency warnings using origin=NONE.\n> Refer to <YYY_REF> for how copy_data and origin can be used to set up\n> replication between primaries.\n\nI have moved all these contents to a separate section in the\nlogical-replication page. I have referred to this link from the\ndocumentation of origin and copy_data parameter. I have also referred\nto \"setting up replication between primaries\" in the newly added\nsection. Since this new section is referred to from other places, I\nfelt we need not provide a link from create_subscription notes. The\nchanges for the same are available at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm0NRJ1O1cYcZD%3Df7NgynozFprb7zpJSayFN5rcaS44G6Q%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 2 Sep 2022 00:18:47 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, 1 Sept 2022 at 08:01, shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Wed, Aug 31, 2022 1:06 AM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > The attached v43 patch has the changes for the same.\n> >\n>\n> Thanks for updating the patch.\n>\n> Here is a comment on the 0001 patch.\n>\n> + if (!isnewtable)\n> + {\n> + pfree(nspname);\n> + pfree(relname);\n> + continue;\n> + }\n>\n> If it is a new table, in which case it would log a warning, should we also call\n> pfree()?\n\nYes it should be added, I have fixed the same in the v44 patch attached at [1].\n[1] - https://www.postgresql.org/message-id/CALDaNm0NRJ1O1cYcZD%3Df7NgynozFprb7zpJSayFN5rcaS44G6Q%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 2 Sep 2022 00:19:54 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, 1 Sept 2022 at 09:19, Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Aug 31, 2022 at 11:35 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Here are my review comments for v43-0001.\n> >\n> > ======\n> >\n> > 1. Commit message\n> >\n> > The initial copy phase has no way to know the origin of the row data,\n> > so if 'copy_data = true' in the step 4 below, log a warning to notify user\n> > that potentially non-local data might have been copied.\n> >\n> > 1a.\n> > \"in the step 4\" -> \"in step 4\"\n> >\n> > ~\n> >\n> > 1b.\n> > \"notify user\" -> \"notify the user\"\n> >\n> > ======\n> >\n> > 2. doc/src/sgml/ref/create_subscription.sgml\n> >\n> > + <para>\n> > + If the subscription is created with <literal>origin = none</literal> and\n> > + <literal>copy_data = true</literal>, it will check if the publisher has\n> > + subscribed to the same table from other publishers and, if so, log a\n> > + warning to notify the user to check the publisher tables. Before continuing\n> > + with other operations the user should check that publisher tables did not\n> > + have data with different origins to prevent data inconsistency issues on the\n> > + subscriber.\n> > + </para>\n> >\n> > I am not sure about that wording saying \"to prevent data inconsistency\n> > issues\" because I thought maybe is already too late because any\n> > unwanted origin data is already copied during the initial sync. I\n> > think the user can do it try to clean up any problems caused before\n> > things become even worse (??)\n> >\n> > ~~~\n> >\n> > You asked for my thoughts (see [1] 6b):\n> >\n> > > There is nothing much a user can do in this case. Only option would be\n> > > to take a backup before the operation and restore it and then recreate\n> > > the replication setup. I was not sure if we should document these\n> > > steps as similar inconsistent data could get created without the\n> > > origin option . Thoughts?\n> >\n> > I think those cases are a bit different:\n> >\n> > - For a normal scenario (i.e. origin=ANY) then inconsistent data just\n> > refers to problems like maybe PK violations so I think you just get\n> > what you get and have to deal with the errors...\n> >\n> > - But when the user said origin=NONE they are specifically saying they\n> > do NOT want any non-local data, yet they still might end up getting\n> > data contrary to their wishes. So I think maybe there ought to be\n> > something documented to warn about this potentially happening. My\n> > first impression is that all this seems like a kind of terrible\n> > problem because if it happens then cleanup could be difficult - if\n> > that subscriber in turn was also a publisher then this bad data might\n> > have propagated all over the place already! Anyway, I guess all this\n> > could be mentioned in some (new ?) section of the\n> > logical-replication.sgml file (i.e. patch 0002).\n> >\n> > IDEA: I was wondering if the documentation should recommend (or maybe\n> > we can even enforce this) that when using origin=NONE the user should\n> > always create the subscription with enabled=FALSE. That way if there\n> > are some warnings about potential bad origin data then there it might\n> > be possible to take some action or change their mind *before* any\n> > unwanted data pollutes everything.\n> >\n> > ======\n> >\n> > 3. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\n> >\n> > + if (count == 0)\n> > + tablenames = makeStringInfo();\n> > + else\n> > + appendStringInfoString(tablenames, \", \");\n> > +\n> > + appendStringInfo(tablenames, \"%s.%s\", nspname, relname);\n> > + count++;\n> >\n> > I think the quotes are not correct - each separate table name should be quoted.\n> >\n> > ~~~\n> >\n> > 4.\n> >\n> > + /*\n> > + * There might be many tables present in this case, we will display a\n> > + * maximum of 100 tables followed by \"...\" to indicate there might be\n> > + * more tables.\n> > + */\n> > + if (count == 100)\n> > + {\n> > + appendStringInfoString(tablenames, \", ...\");\n> > + break;\n> > + }\n> >\n> > 4a.\n> > IMO this code should be part of the previous if/else\n> >\n> > e.g.\n> > if (count == 0) ... makeStringInfo()\n> > else if (count > 100) ... append ellipsis \"...\" and break out of the loop\n> > else ... append table name to the message\n> >\n> > Doing it in this way means \"...\" definitely means there are more than\n> > 100 bad tables.\n> >\n> > ~\n> >\n> > 4b.\n> > Changing like above (#4a) simplifies the comment because it removes\n> > doubts like \"might be…\" since you already know you found the 101st\n> > table.\n> >\n> > SUGGESTION (comment)\n> > The message displays a maximum of 100 tables having potentially\n> > non-local data. Any more than that will be indicated by \"...\".\n> >\n> > ~\n> >\n>\n> Are we using this style of an error message anywhere else in the code?\n> If not, then I think we should avoid it here as well as it appears a\n> bit adhoc to me in the sense that why restrict at 100 tables. Can we\n> instead think of displaying the publications list in the message? So\n> errdetail would be something like: Subscribed publications \\\"%s\\\" has\n> been subscribed from some other publisher. Then we can probably give a\n> SQL query for a user to find tables that may data from multiple\n> origins especially if that is not complicated. Also, then we can\n> change the function name to check_publications_origin().\n\nModified as suggested.\n\n> Few other minor comments on v43-0001*\n> 1.\n> + ereport(WARNING,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"subscription %s requested origin=NONE but might copy data\n> that had a different origin.\",\n> + subname),\n>\n> In error message, we don't use full stop at the end.\n\nModified\n\n> 2.\n> + ereport(WARNING,\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\n> + errmsg(\"subscription %s requested origin=NONE but might copy data\n> that had a different origin.\",\n> + subname),\n> + errdetail(\"Subscribed table(s) \\\"%s\\\" has been subscribed from some\n> other publisher.\",\n> + tablenames->data),\n>\n> It is better to use errdetail_plural here as there could be one or\n> more objects in this message?\n\nModified the new errdetail message to errdetail_plural.\n\nThanks for the comments, the v44 version attached at [1] has the\nchanges for the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm0NRJ1O1cYcZD%3Df7NgynozFprb7zpJSayFN5rcaS44G6Q%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 2 Sep 2022 00:23:16 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for the latest patch v44-0001.\n\n======\n\n1. doc/src/sgml/logical-replication.sgml\n\n+ <sect1 id=\"specifying-origins-for-subscription\">\n+ <title>Specifying origins for subscription</title>\n\nI thought the title is OK, but maybe can be better. OTOH, I am not\nsure if my suggestions below are improvements or not. Anyway, even if\nthe title says the same, the convention is to use uppercase words.\n\nSomething like:\n\"Specifying Origins for Subscriptions\"\n\"Specifying Data Origins for Subscriptions\"\n\"Specifying Data Origins in CREATE SUBSCRIPTION\"\n\"Subscription Data Origins\"\n\n~~~\n\n2.\n\nCurrently, this new section in this patch is only discussing the\n*problems* users might encounter for using copy_data,origin=NONE, but\nI think a section with a generic title about \"subscription origins\"\nshould be able to stand alone. For example, it should give some brief\nmention of what is the meaning/purpose of origin=ANY and origin=NONE.\nAnd it should xref back to CREATE SUBSCRIPTION page.\n\nIMO all this content currently in the patch maybe belongs in a\nsub-section of this new section (e.g. call it something like\n\"Preventing Data Inconsistencies for copy_data,origin=NONE\")\n\n~~~\n\n3.\n\n+ To find which tables might potentially include non-local origins (due to\n+ other subscriptions created on the publisher) try this SQL query by\n+ specifying the publications in IN condition:\n+<programlisting>\n+SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\n+FROM pg_publication P,\n+ LATERAL pg_get_publication_tables(P.pubname) GPT\n+ LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n+ pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\n+WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\n+ P.pubname IN (...);\n+</programlisting>\n+ </para>\n\n3a.\nI think the \"To find...\" sentence should be a new paragraph.\n\n~\n\n3b.\nInstead of saying \"specifying the publications in IN condition\" I\nthink it might be simpler if those instructions are part of the SQL\n\nSUGGESTION\n+<programlisting>\n+# substitute <pub-names> below with your publication name(s) to be queried\n+SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\n+FROM pg_publication P,\n+ LATERAL pg_get_publication_tables(P.pubname) GPT\n+ LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n+ pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\n+WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\n+ P.pubname IN (<pub-names>);\n+</programlisting>\n\n~\n\n3c.\n</programlisting>\n </para>\n\nI recall there was a commit or hackers post sometime earlier this year\nthat reported it is better for these particular closing tags to be\ndone together like </programlisting></para> because otherwise there is\nsome case where the whitespace might not render correctly.\nUnfortunately, I can't seem to find the evidence of that post/commit,\nbut I am sure I saw something about this because I have been using\nthat practice ever since I saw it.\n\n======\n\n4. doc/src/sgml/ref/alter_subscription.sgml\n\n+ <para>\n+ Refer to the <xref\nlinkend=\"specifying-origins-for-subscription\"/> about how\n+ <literal>copy_data = true</literal> can interact with the\n+ <literal>origin</literal> parameter.\n+ </para>\n\nPreviously when this xref pointed to the \"Note\" section the sentence\nlooked OK (it said, \"Refer to the Note about how...\") but now the word\nis not \"Note\" anymore, (e.g. \"Refer to the Section 31.3 about how\ncopy_data = true can interact with the origin parameter. \"), so I\nthink this should be tweaked...\n\n\"Refer to the XXX about how...\" -> \"See XXX for details of how...\"\n\n======\n\n5. doc/src/sgml/ref/create_subscription.sgml\n\nSave as comment #4 (2 times)\n\n======\n\n6. src/backend/commands/subscriptioncmds.c\n\n+ if (bsearch(&relid, subrel_local_oids,\n+ subrel_count, sizeof(Oid), oid_cmp))\n+ isnewtable = false;\n\nSUGGESTION (search PG source - there are examples like this)\nnewtable = bsearch(&relid, subrel_local_oids, subrel_count,\nsizeof(Oid), oid_cmp);\n\n~~~\n\n7.\n\n+ if (list_length(publist))\n+ {\n\nI think the proper way to test for non-empty List is\n\nif (publist != NIL) { ... }\n\nor simply\n\nif (publist) { ... }\n\n~~~\n\n8.\n\n+ errmsg(\"subscription \\\"%s\\\" requested origin=NONE but might copy\ndata that had a different origin\",\n\nDo you think this should say \"requested copy_data with origin=NONE\",\ninstead of just \"requested origin=NONE\"?\n\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 2 Sep 2022 18:21:20 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, Sep 2, 2022 at 6:21 PM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for the latest patch v44-0001.\n>\n...\n>\n> 6. src/backend/commands/subscriptioncmds.c\n>\n> + if (bsearch(&relid, subrel_local_oids,\n> + subrel_count, sizeof(Oid), oid_cmp))\n> + isnewtable = false;\n>\n> SUGGESTION (search PG source - there are examples like this)\n> newtable = bsearch(&relid, subrel_local_oids, subrel_count,\n> sizeof(Oid), oid_cmp);\n>\n\nOops. Of course, I meant !bsearch\n\nSUGGESTION\nnewtable = !bsearch(&relid, subrel_local_oids, subrel_count,\nsizeof(Oid), oid_cmp);\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Fri, 2 Sep 2022 18:26:08 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, 29 Aug 2022 at 16:35, Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Aug 24, 2022 at 7:27 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Since there was no objections to change it to throw a warning, I have\n> > made the changes for the same.\n> >\n>\n> Review comments for v42-0001*\n> ==========================\n> 1. Can we improve the query in check_pub_table_subscribed() so that it\n> doesn't fetch any table that is already part of the subscription on\n> the subscriber? This may be better than what the patch is doing which\n> is to first fetch such information and then skip it. If forming this\n> query turns out to be too complex then we can retain your method as\n> well but I feel it is worth trying to optimize the query used in the\n> patch.\n\nI have modified the query to skip the tables as part of the query. The\nattached v45 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Fri, 2 Sep 2022 15:59:01 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, 2 Sept 2022 at 13:51, Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for the latest patch v44-0001.\n>\n> ======\n>\n> 1. doc/src/sgml/logical-replication.sgml\n>\n> + <sect1 id=\"specifying-origins-for-subscription\">\n> + <title>Specifying origins for subscription</title>\n>\n> I thought the title is OK, but maybe can be better. OTOH, I am not\n> sure if my suggestions below are improvements or not. Anyway, even if\n> the title says the same, the convention is to use uppercase words.\n>\n> Something like:\n> \"Specifying Origins for Subscriptions\"\n> \"Specifying Data Origins for Subscriptions\"\n> \"Specifying Data Origins in CREATE SUBSCRIPTION\"\n> \"Subscription Data Origins\"\n\nModified to \"Preventing Data Inconsistencies for copy_data,\norigin=NONE\" to avoid confusions\n\n> ~~~\n>\n> 2.\n>\n> Currently, this new section in this patch is only discussing the\n> *problems* users might encounter for using copy_data,origin=NONE, but\n> I think a section with a generic title about \"subscription origins\"\n> should be able to stand alone. For example, it should give some brief\n> mention of what is the meaning/purpose of origin=ANY and origin=NONE.\n> And it should xref back to CREATE SUBSCRIPTION page.\n>\n> IMO all this content currently in the patch maybe belongs in a\n> sub-section of this new section (e.g. call it something like\n> \"Preventing Data Inconsistencies for copy_data,origin=NONE\")\n\nI wanted to just document the inconsistency issue when copy_data and\norigin is used. I felt the origin information was already present in\ncreate subscription page. I have not documented the origin again here.\nI have renamed the section to \"Preventing Data Inconsistencies for\ncopy_data, origin=NONE\" to avoid any confusions.\n\n> ~~~\n>\n> 3.\n>\n> + To find which tables might potentially include non-local origins (due to\n> + other subscriptions created on the publisher) try this SQL query by\n> + specifying the publications in IN condition:\n> +<programlisting>\n> +SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\n> +FROM pg_publication P,\n> + LATERAL pg_get_publication_tables(P.pubname) GPT\n> + LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n> + pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\n> +WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\n> + P.pubname IN (...);\n> +</programlisting>\n> + </para>\n>\n> 3a.\n> I think the \"To find...\" sentence should be a new paragraph.\n\nmodified\n\n> ~\n>\n> 3b.\n> Instead of saying \"specifying the publications in IN condition\" I\n> think it might be simpler if those instructions are part of the SQL\n>\n> SUGGESTION\n> +<programlisting>\n> +# substitute <pub-names> below with your publication name(s) to be queried\n> +SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\n> +FROM pg_publication P,\n> + LATERAL pg_get_publication_tables(P.pubname) GPT\n> + LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n> + pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\n> +WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\n> + P.pubname IN (<pub-names>);\n> +</programlisting>\n\n modified, I could not use <pub-name> since it was considering it as a\ntag, instead I have changed it to pub-names\n\n> ~\n>\n> 3c.\n> </programlisting>\n> </para>\n>\n> I recall there was a commit or hackers post sometime earlier this year\n> that reported it is better for these particular closing tags to be\n> done together like </programlisting></para> because otherwise there is\n> some case where the whitespace might not render correctly.\n> Unfortunately, I can't seem to find the evidence of that post/commit,\n> but I am sure I saw something about this because I have been using\n> that practice ever since I saw it.\n\nModified\n\n> ======\n>\n> 4. doc/src/sgml/ref/alter_subscription.sgml\n>\n> + <para>\n> + Refer to the <xref\n> linkend=\"specifying-origins-for-subscription\"/> about how\n> + <literal>copy_data = true</literal> can interact with the\n> + <literal>origin</literal> parameter.\n> + </para>\n>\n> Previously when this xref pointed to the \"Note\" section the sentence\n> looked OK (it said, \"Refer to the Note about how...\") but now the word\n> is not \"Note\" anymore, (e.g. \"Refer to the Section 31.3 about how\n> copy_data = true can interact with the origin parameter. \"), so I\n> think this should be tweaked...\n>\n> \"Refer to the XXX about how...\" -> \"See XXX for details of how...\"\n\nModified\n\n> ======\n>\n> 5. doc/src/sgml/ref/create_subscription.sgml\n>\n> Save as comment #4 (2 times)\n\nModified\n\n> ======\n>\n> 6. src/backend/commands/subscriptioncmds.c\n>\n> + if (bsearch(&relid, subrel_local_oids,\n> + subrel_count, sizeof(Oid), oid_cmp))\n> + isnewtable = false;\n>\n> SUGGESTION (search PG source - there are examples like this)\n> newtable = bsearch(&relid, subrel_local_oids, subrel_count,\n> sizeof(Oid), oid_cmp);\n\nThis code has been removed as part of another comment, the comment no\nmore applies.\n\n> ~~~\n>\n> 7.\n>\n> + if (list_length(publist))\n> + {\n>\n> I think the proper way to test for non-empty List is\n>\n> if (publist != NIL) { ... }\n>\n> or simply\n>\n> if (publist) { ... }\n\nModified\n\n> ~~~\n>\n> 8.\n>\n> + errmsg(\"subscription \\\"%s\\\" requested origin=NONE but might copy\n> data that had a different origin\",\n>\n> Do you think this should say \"requested copy_data with origin=NONE\",\n> instead of just \"requested origin=NONE\"?\n\nModified\n\nThanks for the comments, the v45 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm3fwW4mHGfpZfVLaHe_tSavOSPqntD5XPwO%2B_jwScrj_g%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Fri, 2 Sep 2022 16:03:44 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Here are my review comments for v45-0001:\n\n======\n\n1. doc/src/sgml/logical-replication.sgml\n\n <para>\n To find which tables might potentially include non-local origins (due to\n other subscriptions created on the publisher) try this SQL query:\n<programlisting>\nSELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\nFROM pg_publication P,\n LATERAL pg_get_publication_tables(P.pubname) GPT\n LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\nWHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\n P.pubname IN (pub-names);\n</programlisting></para>\n\n1a.\nTo use \"<pub-names>\" with the <> then simply put meta characters in the SGML.\ne.g.\n<pub-names>\n\n~\n\n1b.\nThe patch forgot to add the SQL \"#\" instruction as suggested by my\nprevious comment (see [1] #3b)\n\n~~~\n\n2.\n\n <sect1 id=\"preventing-inconsistencies-for-copy_data-origin\">\n <title>Preventing Data Inconsistencies for copy_data, origin=NONE</title>\n\nThe title is OK, but I think this should all be a <sect2> sub-section\nof \"31.2 Subscription\"\n\n======\n\n3. src/backend/commands/subscriptioncmds.c - check_publications_origin\n\n+ initStringInfo(&cmd);\n+ appendStringInfoString(&cmd,\n+ \"SELECT DISTINCT P.pubname AS pubname\\n\"\n+ \"FROM pg_publication P,\\n\"\n+ \" LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n+ \" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n+ \" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n+ \"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\");\n+ get_publications_str(publications, &cmd, true);\n+ appendStringInfoChar(&cmd, ')');\n+ get_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\n\n(see from get_skip_tables_str)\n\n+ appendStringInfoString(dest, \" AND C.oid NOT IN (SELECT C.oid FROM\npg_class C JOIN pg_namespace N on N.oid = C.relnamespace where \");\n\n\nIMO the way you are using get_skip_tables_str should be modified. I\nwill show by way of example below.\n- \"where\" -> \"WHERE\"\n- put the SELECT at the caller instead of inside the function\n- handle the \")\" at the caller\n\nAll this will also make the body of the 'get_skip_tables_str' function\nmuch simpler (see the next review comments)\n\nSUGGESTION\nif (subrel_count > 0)\n{\n/* TODO - put some explanatory comment here about skipping the tables */\nappendStringInfo(&cmd,\n\" AND C.oid NOT IN (\\n\"\n\"SELECT C.oid FROM pg_class C\\n\"\n\"JOIN pg_namespace N on N.oid = C.relnamespace WHERE \");\nget_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\nappendStringInf(&cmd, “)”);\n}\n\n~~~\n\n4. src/backend/commands/subscriptioncmds.c - get_skip_tables_str\n\n+/*\n+ * Add the table names that should be skipped.\n+ */\n\nThis comment does not have enough detail to know really what the\nfunction is for. Perhaps you only need to say that this is a helper\nfunction for 'check_publications_origin' and then where it is called\nyou can give a comment (e.g. see my review comment #3)\n\n~~\n\n5. get_skip_tables_str (body)\n\n5a. Variable 'count' is not a very good name; IMO just say 'i' for\nindex, and it can be declared C99 style.\n\n~\n\n5b. Variable 'first' is not necessary - it's same as (i == 0)\n\n~\n\n5c. The whole \"SELECT\" part and the \")\" parts are more simply done at\nthe caller (see the review comment #3)\n\n~\n\n5d.\n\n+ appendStringInfo(dest, \"(C.relname = '%s' and N.nspname = '%s')\",\n+ tablename, schemaname);\n\nIt makes no difference but I thought it would feel more natural if the\nSQL would compare the schema name *before* the table name.\n\n~\n\n5e.\n\"and\" -> \"AND\"\n\n~\n\nDoing all 5a,b,c,d, and e means overall having a much simpler function body:\n\nSUGGESTION\n+ for (int i = 0; i < subrel_count; i++)\n+ {\n+ Oid relid = subrel_local_oids[i];\n+ char *schemaname = get_namespace_name(get_rel_namespace(relid));\n+ char *tablename = get_rel_name(relid);\n+\n+ if (i > 0)\n+ appendStringInfoString(dest, \" OR \");\n+\n+ appendStringInfo(dest, \"(N.nspname = '%s' AND C.relname = '%s')\",\n+ schemaname, tablename);\n+ }\n\n------\n[1] https://www.postgresql.org/message-id/CAHut%2BPsku25%2BVjz7HiohWxc2WU07O_ZV4voFG%2BU7WzwKhUzpGQ%40mail.gmail.com\n\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Mon, 5 Sep 2022 14:17:05 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, Sep 5, 2022 at 9:47 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for v45-0001:\n>\n> ======\n>\n> 1. doc/src/sgml/logical-replication.sgml\n>\n> <para>\n> To find which tables might potentially include non-local origins (due to\n> other subscriptions created on the publisher) try this SQL query:\n> <programlisting>\n> SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\n> FROM pg_publication P,\n> LATERAL pg_get_publication_tables(P.pubname) GPT\n> LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n> pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\n> WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\n> P.pubname IN (pub-names);\n> </programlisting></para>\n>\n> 1a.\n> To use \"<pub-names>\" with the <> then simply put meta characters in the SGML.\n> e.g.\n> <pub-names>\n>\n> ~\n>\n> 1b.\n> The patch forgot to add the SQL \"#\" instruction as suggested by my\n> previous comment (see [1] #3b)\n>\n> ~~~\n>\n> 2.\n>\n> <sect1 id=\"preventing-inconsistencies-for-copy_data-origin\">\n> <title>Preventing Data Inconsistencies for copy_data, origin=NONE</title>\n>\n> The title is OK, but I think this should all be a <sect2> sub-section\n> of \"31.2 Subscription\"\n>\n> ======\n>\n\nIt is recommended to create the subscription\n+ using <literal>enabled=false</literal>, so that if the origin WARNING occurs\n+ no copy has happened yet. Otherwise some corrective steps might be needed to\n+ remove any unwanted data that got copied.\n\nI am not completely sure of this part of the docs as this can add\nadditional steps for users while working on subscriptions even when\nthe same is not required. I suggest for now we can remove this part.\nLater based on some feedback on this feature, we can extend the docs\nif required.\n\nAlso, instead of having it as a separate section, let's keep this as\npart of create_subscription.sgml\n\n*\n+ errhint(\"Verify that initial data copied from the publisher tables\ndid not come from other origins. Some corrective action may be\nnecessary.\"));\n\nThe second sentence in this message doesn't seem to be required.\n\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 5 Sep 2022 15:10:41 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, 5 Sept 2022 at 09:47, Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> Here are my review comments for v45-0001:\n>\n> ======\n>\n> 1. doc/src/sgml/logical-replication.sgml\n>\n> <para>\n> To find which tables might potentially include non-local origins (due to\n> other subscriptions created on the publisher) try this SQL query:\n> <programlisting>\n> SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\n> FROM pg_publication P,\n> LATERAL pg_get_publication_tables(P.pubname) GPT\n> LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n> pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\n> WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\n> P.pubname IN (pub-names);\n> </programlisting></para>\n>\n> 1a.\n> To use \"<pub-names>\" with the <> then simply put meta characters in the SGML.\n> e.g.\n> <pub-names>\n\nModified\n\n> ~\n>\n> 1b.\n> The patch forgot to add the SQL \"#\" instruction as suggested by my\n> previous comment (see [1] #3b)\n\nModified\n\n> ~~~\n>\n> 2.\n>\n> <sect1 id=\"preventing-inconsistencies-for-copy_data-origin\">\n> <title>Preventing Data Inconsistencies for copy_data, origin=NONE</title>\n>\n> The title is OK, but I think this should all be a <sect2> sub-section\n> of \"31.2 Subscription\"\n\nI have moved it to create subscription notes based on a recent comment\nfrom Amit.\n\n> ======\n>\n> 3. src/backend/commands/subscriptioncmds.c - check_publications_origin\n>\n> + initStringInfo(&cmd);\n> + appendStringInfoString(&cmd,\n> + \"SELECT DISTINCT P.pubname AS pubname\\n\"\n> + \"FROM pg_publication P,\\n\"\n> + \" LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n> + \" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n> + \" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n> + \"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\");\n> + get_publications_str(publications, &cmd, true);\n> + appendStringInfoChar(&cmd, ')');\n> + get_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\n>\n> (see from get_skip_tables_str)\n>\n> + appendStringInfoString(dest, \" AND C.oid NOT IN (SELECT C.oid FROM\n> pg_class C JOIN pg_namespace N on N.oid = C.relnamespace where \");\n>\n>\n> IMO the way you are using get_skip_tables_str should be modified. I\n> will show by way of example below.\n> - \"where\" -> \"WHERE\"\n> - put the SELECT at the caller instead of inside the function\n> - handle the \")\" at the caller\n>\n> All this will also make the body of the 'get_skip_tables_str' function\n> much simpler (see the next review comments)\n>\n> SUGGESTION\n> if (subrel_count > 0)\n> {\n> /* TODO - put some explanatory comment here about skipping the tables */\n> appendStringInfo(&cmd,\n> \" AND C.oid NOT IN (\\n\"\n> \"SELECT C.oid FROM pg_class C\\n\"\n> \"JOIN pg_namespace N on N.oid = C.relnamespace WHERE \");\n> get_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\n> appendStringInf(&cmd, “)”);\n> }\n\nModified\n\n> ~~~\n>\n> 4. src/backend/commands/subscriptioncmds.c - get_skip_tables_str\n>\n> +/*\n> + * Add the table names that should be skipped.\n> + */\n>\n> This comment does not have enough detail to know really what the\n> function is for. Perhaps you only need to say that this is a helper\n> function for 'check_publications_origin' and then where it is called\n> you can give a comment (e.g. see my review comment #3)\n\nModified\n\n> ~~\n>\n> 5. get_skip_tables_str (body)\n>\n> 5a. Variable 'count' is not a very good name; IMO just say 'i' for\n> index, and it can be declared C99 style.\n\nModified\n\n> ~\n>\n> 5b. Variable 'first' is not necessary - it's same as (i == 0)\n\nModified\n\n> ~\n>\n> 5c. The whole \"SELECT\" part and the \")\" parts are more simply done at\n> the caller (see the review comment #3)\n\nModified\n\n> ~\n>\n> 5d.\n>\n> + appendStringInfo(dest, \"(C.relname = '%s' and N.nspname = '%s')\",\n> + tablename, schemaname);\n>\n> It makes no difference but I thought it would feel more natural if the\n> SQL would compare the schema name *before* the table name.\n\nModified\n\n> ~\n>\n> 5e.\n> \"and\" -> \"AND\"\n\nModified\n\n> ~\n>\n> Doing all 5a,b,c,d, and e means overall having a much simpler function body:\n>\n> SUGGESTION\n> + for (int i = 0; i < subrel_count; i++)\n> + {\n> + Oid relid = subrel_local_oids[i];\n> + char *schemaname = get_namespace_name(get_rel_namespace(relid));\n> + char *tablename = get_rel_name(relid);\n> +\n> + if (i > 0)\n> + appendStringInfoString(dest, \" OR \");\n> +\n> + appendStringInfo(dest, \"(N.nspname = '%s' AND C.relname = '%s')\",\n> + schemaname, tablename);\n> + }\n\nModified\n\nThanks for the comments, the attached patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Tue, 6 Sep 2022 08:43:38 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Mon, 5 Sept 2022 at 15:10, Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Sep 5, 2022 at 9:47 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > Here are my review comments for v45-0001:\n> >\n> > ======\n> >\n> > 1. doc/src/sgml/logical-replication.sgml\n> >\n> > <para>\n> > To find which tables might potentially include non-local origins (due to\n> > other subscriptions created on the publisher) try this SQL query:\n> > <programlisting>\n> > SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\n> > FROM pg_publication P,\n> > LATERAL pg_get_publication_tables(P.pubname) GPT\n> > LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n> > pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\n> > WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\n> > P.pubname IN (pub-names);\n> > </programlisting></para>\n> >\n> > 1a.\n> > To use \"<pub-names>\" with the <> then simply put meta characters in the SGML.\n> > e.g.\n> > <pub-names>\n> >\n> > ~\n> >\n> > 1b.\n> > The patch forgot to add the SQL \"#\" instruction as suggested by my\n> > previous comment (see [1] #3b)\n> >\n> > ~~~\n> >\n> > 2.\n> >\n> > <sect1 id=\"preventing-inconsistencies-for-copy_data-origin\">\n> > <title>Preventing Data Inconsistencies for copy_data, origin=NONE</title>\n> >\n> > The title is OK, but I think this should all be a <sect2> sub-section\n> > of \"31.2 Subscription\"\n> >\n> > ======\n> >\n>\n> It is recommended to create the subscription\n> + using <literal>enabled=false</literal>, so that if the origin WARNING occurs\n> + no copy has happened yet. Otherwise some corrective steps might be needed to\n> + remove any unwanted data that got copied.\n>\n> I am not completely sure of this part of the docs as this can add\n> additional steps for users while working on subscriptions even when\n> the same is not required. I suggest for now we can remove this part.\n> Later based on some feedback on this feature, we can extend the docs\n> if required.\n\nModified\n\n> Also, instead of having it as a separate section, let's keep this as\n> part of create_subscription.sgml\n\nModified\n\n> *\n> + errhint(\"Verify that initial data copied from the publisher tables\n> did not come from other origins. Some corrective action may be\n> necessary.\"));\n>\n> The second sentence in this message doesn't seem to be required.\n\nModified\n\nThanks for the comments, the v46 patch at [1] has the changes for the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm3TTGdCCkeDsN8hqtF_2z-8%2B%3D3tc9Gh5xOKAQ_BRMCkMA%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Tue, 6 Sep 2022 08:45:45 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Sep 6, 2022 11:14 AM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> Thanks for the comments, the attached patch has the changes for the same.\r\n> \r\n\r\nThanks for updating the patch. Here are some comments.\r\n\r\n1.\r\n+\tif (subrel_count)\r\n+\t{\r\n+\t\t/*\r\n+\t\t * In case of ALTER SUBSCRIPTION ... REFRESH, subrel_local_oids\r\n+\t\t * contains the list of relation oids that are already present on the\r\n+\t\t * subscriber. This check should be skipped for these tables.\r\n+\t\t */\r\n+\t\tappendStringInfo(&cmd, \"AND C.oid NOT IN (\\n\"\r\n+\t\t\t\t\t\t \"SELECT C.oid \\n\"\r\n+\t\t\t\t\t\t \"FROM pg_class C\\n\"\r\n+\t\t\t\t\t\t \" JOIN pg_namespace N ON N.oid = C.relnamespace\\n\"\r\n+\t\t\t\t\t\t \"WHERE \");\r\n+\t\tget_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\r\n+\t\tappendStringInfo(&cmd, \")\");\r\n+\t}\r\n\r\nI think we can skip the tables without using a subquery. See the SQL below.\r\nWould it be better?\r\n\r\nSELECT DISTINCT P.pubname AS pubname\r\nFROM pg_publication P,\r\n LATERAL pg_get_publication_tables(P.pubname) GPT\r\n LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\r\n pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\r\nWHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN ('p1', 'p3')\r\nAND NOT (N.nspname = 'public' AND C.relname = 'tbl')\r\nAND NOT (N.nspname = 'public' AND C.relname = 't1');\r\n\r\n\r\n2.\r\n+ When using a subscription parameter combination of\r\n+ <literal>copy_data=true</literal> and <literal>origin=NONE</literal>, the\r\n\r\nCould we use \"copy_data = true\" and \"origin = NONE\" (add two spaces around the\r\nequals sign)? I think it looks clearer that way. And it is consistent with other\r\nplaces in this file.\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Tue, 6 Sep 2022 04:00:47 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tues, 6 Sept 2022 at 11:14, vignesh C <vignesh21@gmail.com> wrote:\r\n> Thanks for the comments, the attached patch has the changes for the same.\r\n\r\nThanks for updating the patch.\r\n\r\nHere is one comment for 0001 patch.\r\n1. The query in function check_publications_origin.\r\n+\tappendStringInfoString(&cmd,\r\n+\t\t\t\t\t\t \"SELECT DISTINCT P.pubname AS pubname\\n\"\r\n+\t\t\t\t\t\t \"FROM pg_publication P,\\n\"\r\n+\t\t\t\t\t\t \"\t LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\r\n+\t\t\t\t\t\t \"\t LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\r\n+\t\t\t\t\t\t \"\t pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\r\n+\t\t\t\t\t\t \"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\");\r\n\r\nSince I found that we only use \"PS.srrelid\" in the WHERE statement by\r\nspecifying \"PS.srrelid IS NOT NULL\", could we just use \"[INNER] JOIN\" to join\r\nthe table pg_subscription_rel?\r\n\r\nRegards,\r\nWang wei\r\n","msg_date":"Tue, 6 Sep 2022 04:01:04 +0000","msg_from":"\"wangw.fnst@fujitsu.com\" <wangw.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"Thanks for addressing my previous review comments. I have no more\ncomments for v46*.\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Tue, 6 Sep 2022 15:11:16 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Sep 6, 2022 at 9:31 AM wangw.fnst@fujitsu.com\n<wangw.fnst@fujitsu.com> wrote:\n>\n> On Tues, 6 Sept 2022 at 11:14, vignesh C <vignesh21@gmail.com> wrote:\n> > Thanks for the comments, the attached patch has the changes for the same.\n>\n> Thanks for updating the patch.\n>\n> Here is one comment for 0001 patch.\n> 1. The query in function check_publications_origin.\n> + appendStringInfoString(&cmd,\n> + \"SELECT DISTINCT P.pubname AS pubname\\n\"\n> + \"FROM pg_publication P,\\n\"\n> + \" LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n> + \" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n> + \" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n> + \"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\");\n>\n> Since I found that we only use \"PS.srrelid\" in the WHERE statement by\n> specifying \"PS.srrelid IS NOT NULL\", could we just use \"[INNER] JOIN\" to join\n> the table pg_subscription_rel?\n>\n\nI also think we can use INNER JOIN here but maybe there is a reason\nwhy that is not used in the first place. If we change it in the code\nthen also let's change the same in the docs section as well.\n\nFew minor points:\n===============\n1.\nThis scenario is detected and a WARNING is logged to the\n+ user, but the warning is only an indication of a potential problem; it is\n+ the user responsibility to make the necessary checks to ensure the copied\n+ data origins are really as wanted or not.\n+ </para>\n\n/user/user's\n\n2. How about a commit message like:\nRaise a warning if there is a possibility of data from multiple origins.\n\nThis commit raises a warning message for a combination of options\n('copy_data = true' and 'origin = none') during CREATE/ALTER subscription\noperations if the publication tables were also replicated from other\npublishers.\n\nDuring replication, we can skip the data from other origins as we have that\ninformation in WAL but that is not possible during initial sync so we raise\na warning if there is such a possibility.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 6 Sep 2022 15:25:10 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, 6 Sept 2022 at 09:31, shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Tue, Sep 6, 2022 11:14 AM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached patch has the changes for the same.\n> >\n>\n> Thanks for updating the patch. Here are some comments.\n>\n> 1.\n> + if (subrel_count)\n> + {\n> + /*\n> + * In case of ALTER SUBSCRIPTION ... REFRESH, subrel_local_oids\n> + * contains the list of relation oids that are already present on the\n> + * subscriber. This check should be skipped for these tables.\n> + */\n> + appendStringInfo(&cmd, \"AND C.oid NOT IN (\\n\"\n> + \"SELECT C.oid \\n\"\n> + \"FROM pg_class C\\n\"\n> + \" JOIN pg_namespace N ON N.oid = C.relnamespace\\n\"\n> + \"WHERE \");\n> + get_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\n> + appendStringInfo(&cmd, \")\");\n> + }\n>\n> I think we can skip the tables without using a subquery. See the SQL below.\n> Would it be better?\n>\n> SELECT DISTINCT P.pubname AS pubname\n> FROM pg_publication P,\n> LATERAL pg_get_publication_tables(P.pubname) GPT\n> LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\n> pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\n> WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN ('p1', 'p3')\n> AND NOT (N.nspname = 'public' AND C.relname = 'tbl')\n> AND NOT (N.nspname = 'public' AND C.relname = 't1');\n\nModified\n\n> 2.\n> + When using a subscription parameter combination of\n> + <literal>copy_data=true</literal> and <literal>origin=NONE</literal>, the\n>\n> Could we use \"copy_data = true\" and \"origin = NONE\" (add two spaces around the\n> equals sign)? I think it looks clearer that way. And it is consistent with other\n> places in this file.\n\nModified\n\nThanks for the comments, the attached v47 patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Wed, 7 Sep 2022 09:52:53 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, 6 Sept 2022 at 09:31, wangw.fnst@fujitsu.com\n<wangw.fnst@fujitsu.com> wrote:\n>\n> On Tues, 6 Sept 2022 at 11:14, vignesh C <vignesh21@gmail.com> wrote:\n> > Thanks for the comments, the attached patch has the changes for the same.\n>\n> Thanks for updating the patch.\n>\n> Here is one comment for 0001 patch.\n> 1. The query in function check_publications_origin.\n> + appendStringInfoString(&cmd,\n> + \"SELECT DISTINCT P.pubname AS pubname\\n\"\n> + \"FROM pg_publication P,\\n\"\n> + \" LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n> + \" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n> + \" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n> + \"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\");\n>\n> Since I found that we only use \"PS.srrelid\" in the WHERE statement by\n> specifying \"PS.srrelid IS NOT NULL\", could we just use \"[INNER] JOIN\" to join\n> the table pg_subscription_rel?\n\nThanks for the comment, the v47 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm33T%3D23P-GWvy3O7cT1BfHuGV8dosAw1AVLf40MPvg2bg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 7 Sep 2022 09:54:09 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, 6 Sept 2022 at 15:25, Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Sep 6, 2022 at 9:31 AM wangw.fnst@fujitsu.com\n> <wangw.fnst@fujitsu.com> wrote:\n> >\n> > On Tues, 6 Sept 2022 at 11:14, vignesh C <vignesh21@gmail.com> wrote:\n> > > Thanks for the comments, the attached patch has the changes for the same.\n> >\n> > Thanks for updating the patch.\n> >\n> > Here is one comment for 0001 patch.\n> > 1. The query in function check_publications_origin.\n> > + appendStringInfoString(&cmd,\n> > + \"SELECT DISTINCT P.pubname AS pubname\\n\"\n> > + \"FROM pg_publication P,\\n\"\n> > + \" LATERAL pg_get_publication_tables(P.pubname) GPT\\n\"\n> > + \" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n\"\n> > + \" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n\"\n> > + \"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\");\n> >\n> > Since I found that we only use \"PS.srrelid\" in the WHERE statement by\n> > specifying \"PS.srrelid IS NOT NULL\", could we just use \"[INNER] JOIN\" to join\n> > the table pg_subscription_rel?\n> >\n>\n> I also think we can use INNER JOIN here but maybe there is a reason\n> why that is not used in the first place. If we change it in the code\n> then also let's change the same in the docs section as well.\n\nModified\n\n> Few minor points:\n> ===============\n> 1.\n> This scenario is detected and a WARNING is logged to the\n> + user, but the warning is only an indication of a potential problem; it is\n> + the user responsibility to make the necessary checks to ensure the copied\n> + data origins are really as wanted or not.\n> + </para>\n>\n> /user/user's\n\nModified\n\n> 2. How about a commit message like:\n> Raise a warning if there is a possibility of data from multiple origins.\n>\n> This commit raises a warning message for a combination of options\n> ('copy_data = true' and 'origin = none') during CREATE/ALTER subscription\n> operations if the publication tables were also replicated from other\n> publishers.\n>\n> During replication, we can skip the data from other origins as we have that\n> information in WAL but that is not possible during initial sync so we raise\n> a warning if there is such a possibility.\n\nYes, this looks good. Modified\n\nThanks for the comment, the v47 patch attached at [1] has the changes\nfor the same.\n[1] - https://www.postgresql.org/message-id/CALDaNm33T%3D23P-GWvy3O7cT1BfHuGV8dosAw1AVLf40MPvg2bg%40mail.gmail.com\n\nRegards,\nVignesh\n\n\n","msg_date":"Wed, 7 Sep 2022 09:55:16 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Sep 7, 2022 12:23 PM vignesh C <vignesh21@gmail.com> wrote:\r\n> \r\n> \r\n> Thanks for the comments, the attached v47 patch has the changes for the\r\n> same.\r\n> \r\n\r\nThanks for updating the patch.\r\n\r\nHere is a comment.\r\n\r\n+\tfor (i = 0; i < subrel_count; i++)\r\n+\t{\r\n+\t\tOid\t\t\trelid = subrel_local_oids[i];\r\n+\t\tchar\t *schemaname = get_namespace_name(get_rel_namespace(relid));\r\n+\t\tchar\t *tablename = get_rel_name(relid);\r\n+\r\n+\t\tappendStringInfo(&cmd, \"AND NOT (N.nspname = '%s' AND C.relname = '%s')\\n\",\r\n+\t\t\t\t\t\t schemaname, tablename);\r\n+\t}\r\n\r\nWould it be better to add \"pfree(schemaname)\" and \"pfree(tablename)\" after\r\ncalling appendStringInfo()?\r\n\r\nRegards,\r\nShi yu\r\n","msg_date":"Wed, 7 Sep 2022 07:39:23 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Sep 7, 2022 at 1:09 PM shiy.fnst@fujitsu.com\n<shiy.fnst@fujitsu.com> wrote:\n>\n> On Wed, Sep 7, 2022 12:23 PM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> >\n> > Thanks for the comments, the attached v47 patch has the changes for the\n> > same.\n> >\n>\n> Thanks for updating the patch.\n>\n> Here is a comment.\n>\n> + for (i = 0; i < subrel_count; i++)\n> + {\n> + Oid relid = subrel_local_oids[i];\n> + char *schemaname = get_namespace_name(get_rel_namespace(relid));\n> + char *tablename = get_rel_name(relid);\n> +\n> + appendStringInfo(&cmd, \"AND NOT (N.nspname = '%s' AND C.relname = '%s')\\n\",\n> + schemaname, tablename);\n> + }\n>\n> Would it be better to add \"pfree(schemaname)\" and \"pfree(tablename)\" after\n> calling appendStringInfo()?\n>\n\nNo, I don't think we need to do retail pfree of each and every\nallocation as these allocations are made in the portal context which\nwill be freed by the command end.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 7 Sep 2022 13:39:14 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Sep 7, 2022 at 9:53 AM vignesh C <vignesh21@gmail.com> wrote:\n>\n> Thanks for the comments, the attached v47 patch has the changes for the same.\n>\n\nV47-0001* looks good to me apart from below minor things. I would like\nto commit this tomorrow unless there are more comments on it.\n\nFew minor suggestions:\n==================\n1.\n+ list_free_deep(publist);\n+ pfree(pubnames->data);\n+ pfree(pubnames);\n\nI don't think we need to free this memory as it will automatically be\nreleased at end of the command (this memory is allocated in portal\ncontext). I understand there is no harm in freeing it as well but\nbetter to leave it as there doesn't appear to be any danger of leaking\nmemory for a longer time.\n\n2.\n+ res = walrcv_exec(wrconn, cmd.data, 1, tableRow);\n+ pfree(cmd.data);\n\nDon't you need to free cmd as well here? I think it is better to avoid\nfreeing it due to reasons mentioned in the previous point.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 7 Sep 2022 14:34:02 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, 7 Sept 2022 at 14:34, Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Sep 7, 2022 at 9:53 AM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > Thanks for the comments, the attached v47 patch has the changes for the same.\n> >\n>\n> V47-0001* looks good to me apart from below minor things. I would like\n> to commit this tomorrow unless there are more comments on it.\n>\n> Few minor suggestions:\n> ==================\n> 1.\n> + list_free_deep(publist);\n> + pfree(pubnames->data);\n> + pfree(pubnames);\n>\n> I don't think we need to free this memory as it will automatically be\n> released at end of the command (this memory is allocated in portal\n> context). I understand there is no harm in freeing it as well but\n> better to leave it as there doesn't appear to be any danger of leaking\n> memory for a longer time.\n\nModified\n\n> 2.\n> + res = walrcv_exec(wrconn, cmd.data, 1, tableRow);\n> + pfree(cmd.data);\n>\n> Don't you need to free cmd as well here? I think it is better to avoid\n> freeing it due to reasons mentioned in the previous point.\n\ncmd need not be freed here as we use initStringInfo for initialization\nand initStringInfo allocates memory for data member only.\n\nThe attached patch has the changes for the same.\n\nRegards,\nVignesh","msg_date":"Wed, 7 Sep 2022 17:10:14 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"Hi,\n\nThere is a buildfarm failure on mylodon at [1] because of the new test\nadded. I will analyse and share the findings for the same.\n\n[1] - https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mylodon&dt=2022-09-08%2001%3A40%3A27\n\nRegards,\nVignesh\n\nOn Wed, 7 Sept 2022 at 17:10, vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Wed, 7 Sept 2022 at 14:34, Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Wed, Sep 7, 2022 at 9:53 AM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > Thanks for the comments, the attached v47 patch has the changes for the same.\n> > >\n> >\n> > V47-0001* looks good to me apart from below minor things. I would like\n> > to commit this tomorrow unless there are more comments on it.\n> >\n> > Few minor suggestions:\n> > ==================\n> > 1.\n> > + list_free_deep(publist);\n> > + pfree(pubnames->data);\n> > + pfree(pubnames);\n> >\n> > I don't think we need to free this memory as it will automatically be\n> > released at end of the command (this memory is allocated in portal\n> > context). I understand there is no harm in freeing it as well but\n> > better to leave it as there doesn't appear to be any danger of leaking\n> > memory for a longer time.\n>\n> Modified\n>\n> > 2.\n> > + res = walrcv_exec(wrconn, cmd.data, 1, tableRow);\n> > + pfree(cmd.data);\n> >\n> > Don't you need to free cmd as well here? I think it is better to avoid\n> > freeing it due to reasons mentioned in the previous point.\n>\n> cmd need not be freed here as we use initStringInfo for initialization\n> and initStringInfo allocates memory for data member only.\n>\n> The attached patch has the changes for the same.\n>\n> Regards,\n> Vignesh\n\n\n","msg_date":"Thu, 8 Sep 2022 07:56:45 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Sep 8, 2022 at 7:57 AM vignesh C <vignesh21@gmail.com> wrote:\n>\n> There is a buildfarm failure on mylodon at [1] because of the new test\n> added. I will analyse and share the findings for the same.\n>\n> [1] - https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mylodon&dt=2022-09-08%2001%3A40%3A27\n>\n\nThe log is as below:\n\nerror running SQL: 'psql:<stdin>:1: ERROR: could not drop relation\nmapping for subscription \"tap_sub_a_b_2\"\nDETAIL: Table synchronization for relation \"tab_new\" is in progress\nand is in state \"s\".\nHINT: Use ALTER SUBSCRIPTION ... ENABLE to enable subscription if not\nalready enabled or use DROP SUBSCRIPTION ... to drop the\nsubscription.'\nwhile running 'psql -XAtq -d port=50352 host=/tmp/WMoRd6ngw2\ndbname='postgres' -f - -v ON_ERROR_STOP=1' with sql 'DROP TABLE\ntab_new' at /mnt/resource/bf/build/mylodon/HEAD/pgsql.build/../pgsql/src/test/perl/PostgreSQL/Test/Cluster.pm\nline 1860.\n### Stopping node \"node_A\" using mode immediate\n\nThis clearly indicates the problem. We can't drop the relation till it\nis marked ready in pg_subscription_rel and prior to dropping, the test\njust does $node_A->wait_for_subscription_sync($node_B, $subname_AB2);.\nThis doesn't ensure that relation is in 'ready' state as it will\nfinish even when the relation is in 'syncdone' state.\n\nSo, I think if we want to drop the tables, we need to poll for 'ready'\nstate or otherwise, anyway, this is the end of all tests and nodes\nwon't be reused, so we can remove the clean-up in the end. Any other\nideas?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 8 Sep 2022 08:23:00 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, 8 Sept 2022 at 08:23, Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Thu, Sep 8, 2022 at 7:57 AM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> > There is a buildfarm failure on mylodon at [1] because of the new test\n> > added. I will analyse and share the findings for the same.\n> >\n> > [1] - https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mylodon&dt=2022-09-08%2001%3A40%3A27\n> >\n>\n> The log is as below:\n>\n> error running SQL: 'psql:<stdin>:1: ERROR: could not drop relation\n> mapping for subscription \"tap_sub_a_b_2\"\n> DETAIL: Table synchronization for relation \"tab_new\" is in progress\n> and is in state \"s\".\n> HINT: Use ALTER SUBSCRIPTION ... ENABLE to enable subscription if not\n> already enabled or use DROP SUBSCRIPTION ... to drop the\n> subscription.'\n> while running 'psql -XAtq -d port=50352 host=/tmp/WMoRd6ngw2\n> dbname='postgres' -f - -v ON_ERROR_STOP=1' with sql 'DROP TABLE\n> tab_new' at /mnt/resource/bf/build/mylodon/HEAD/pgsql.build/../pgsql/src/test/perl/PostgreSQL/Test/Cluster.pm\n> line 1860.\n> ### Stopping node \"node_A\" using mode immediate\n>\n> This clearly indicates the problem. We can't drop the relation till it\n> is marked ready in pg_subscription_rel and prior to dropping, the test\n> just does $node_A->wait_for_subscription_sync($node_B, $subname_AB2);.\n> This doesn't ensure that relation is in 'ready' state as it will\n> finish even when the relation is in 'syncdone' state.\n\nI agree with the analysis, adding wait for ready state before dropping\nthe table approach looks good to me. I will prepare a patch for the\nsame and share it.\n\nRegards,\nVignesh\n\n\n","msg_date":"Thu, 8 Sep 2022 08:57:00 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, 8 Sept 2022 at 08:57, vignesh C <vignesh21@gmail.com> wrote:\n>\n> On Thu, 8 Sept 2022 at 08:23, Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Thu, Sep 8, 2022 at 7:57 AM vignesh C <vignesh21@gmail.com> wrote:\n> > >\n> > > There is a buildfarm failure on mylodon at [1] because of the new test\n> > > added. I will analyse and share the findings for the same.\n> > >\n> > > [1] - https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mylodon&dt=2022-09-08%2001%3A40%3A27\n> > >\n> >\n> > The log is as below:\n> >\n> > error running SQL: 'psql:<stdin>:1: ERROR: could not drop relation\n> > mapping for subscription \"tap_sub_a_b_2\"\n> > DETAIL: Table synchronization for relation \"tab_new\" is in progress\n> > and is in state \"s\".\n> > HINT: Use ALTER SUBSCRIPTION ... ENABLE to enable subscription if not\n> > already enabled or use DROP SUBSCRIPTION ... to drop the\n> > subscription.'\n> > while running 'psql -XAtq -d port=50352 host=/tmp/WMoRd6ngw2\n> > dbname='postgres' -f - -v ON_ERROR_STOP=1' with sql 'DROP TABLE\n> > tab_new' at /mnt/resource/bf/build/mylodon/HEAD/pgsql.build/../pgsql/src/test/perl/PostgreSQL/Test/Cluster.pm\n> > line 1860.\n> > ### Stopping node \"node_A\" using mode immediate\n> >\n> > This clearly indicates the problem. We can't drop the relation till it\n> > is marked ready in pg_subscription_rel and prior to dropping, the test\n> > just does $node_A->wait_for_subscription_sync($node_B, $subname_AB2);.\n> > This doesn't ensure that relation is in 'ready' state as it will\n> > finish even when the relation is in 'syncdone' state.\n>\n> I agree with the analysis, adding wait for ready state before dropping\n> the table approach looks good to me. I will prepare a patch for the\n> same and share it.\n\nThe attached patch has the changes to handle the same.\n\nRegards,\nVignesh","msg_date":"Thu, 8 Sep 2022 09:31:50 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\n>\n>\n> The attached patch has the changes to handle the same.\n>\n\nPushed. I am not completely sure whether we want the remaining\ndocumentation patch in this thread in its current form or by modifying\nit. Johnathan has shown some interest in it. I feel you can start a\nseparate thread for it to see if there is any interest in the same and\nclose the CF entry for this work.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 9 Sep 2022 11:12:02 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\n> >\n> >\n> > The attached patch has the changes to handle the same.\n> >\n>\n> Pushed. I am not completely sure whether we want the remaining\n> documentation patch in this thread in its current form or by modifying\n> it. Johnathan has shown some interest in it. I feel you can start a\n> separate thread for it to see if there is any interest in the same and\n> close the CF entry for this work.\n\nThanks for pushing the patch. I have closed this entry in commitfest.\nI will wait for some more time and see the response regarding the\ndocumentation patch and then start a new thread if required.\n\nRegards,\nVignesh\n\n\n","msg_date":"Mon, 12 Sep 2022 10:53:08 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":true,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On 9/12/22 1:23 AM, vignesh C wrote:\r\n> On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\r\n>>\r\n>> On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\r\n>>>\r\n>>>\r\n>>> The attached patch has the changes to handle the same.\r\n>>>\r\n>>\r\n>> Pushed. I am not completely sure whether we want the remaining\r\n>> documentation patch in this thread in its current form or by modifying\r\n>> it. Johnathan has shown some interest in it. I feel you can start a\r\n>> separate thread for it to see if there is any interest in the same and\r\n>> close the CF entry for this work.\r\n> \r\n> Thanks for pushing the patch. I have closed this entry in commitfest.\r\n> I will wait for some more time and see the response regarding the\r\n> documentation patch and then start a new thread if required.\r\n\r\nI've been testing this patch in advancing of working on the \r\ndocumentation and came across a behavior I wanted to note. Specifically, \r\nI am hitting a deadlock while trying to synchronous replicate between \r\nthe two instances at any `synchronous_commit` level above `local`.\r\n\r\nHere is my set up. I have two instances, \"A\" and \"B\".\r\n\r\nOn A and B, run:\r\n\r\n CREATE TABLE sync (id int PRIMARY KEY, info float);\r\n CREATE PUBLICATION sync FOR TABLE sync;\r\n\r\nOn A, run:\r\n\r\n CREATE SUBSCRIPTION sync\r\n CONNECTION 'connstr-to-B'\r\n PUBLICATION sync\r\n WITH (\r\n streaming=true, copy_data=false,\r\n origin=none, synchronous_commit='on');\r\n\r\nOn B, run:\r\n\r\n CREATE SUBSCRIPTION sync\r\n CONNECTION 'connstr-to-A'\r\n PUBLICATION sync\r\n WITH (\r\n streaming=true, copy_data=false,\r\n origin=none, synchronous_commit='on');\r\n\r\nOn A and B, run:\r\n\r\n ALTER SYSTEM SET synchronous_standby_names TO 'sync';\r\n SELECT pg_reload_conf();\r\n\r\nVerify on A and B that pg_stat_replication.sync_state is set to \"sync\"\r\n\r\n SELECT application_name, sync_state = 'sync' AS is_sync\r\n FROM pg_stat_replication\r\n WHERE application_name = 'sync';\r\n\r\nThe next to commands should be run simultaneously on A and B:\r\n\r\n-- run this on A\r\nINSERT INTO sync\r\nSELECT x, random() FROM generate_series(1,2000000, 2) x;\r\n\r\n-- run this on B\r\nINSERT INTO sync\r\nSELECT x, random() FROM generate_series(2,2000000, 2) x;\r\n\r\nThis consistently created the deadlock in my testing.\r\n\r\nDiscussing with Masahiko off-list, this is due to a deadlock from 4 \r\nprocesses: the walsenders on A and B, and the apply workers on A and B. \r\nThe walsenders are waiting for feedback from the apply workers, and the \r\napply workers are waiting for the walsenders to synchronize (I may be \r\noversimplifying).\r\n\r\nHe suggested I try the above example instead with `synchronous_commit` \r\nset to `local`. In this case, I verified that there is no more deadlock, \r\nbut he informed me that we would not be able to use cascading \r\nsynchronous replication when \"origin=none\".\r\n\r\nIf we decide that this is a documentation issue, I'd suggest we improve \r\nthe guidance around using `synchronous_commit`[1] on the CREATE \r\nSUBSCRIPTION page, as the GUC page[2] warns against using `local`:\r\n\r\n\"The setting local causes commits to wait for local flush to disk, but \r\nnot for replication. This is usually not desirable when synchronous \r\nreplication is in use, but is provided for completeness.\"\r\n\r\nThanks,\r\n\r\nJonathan\r\n\r\n[1] https://www.postgresql.org/docs/devel/sql-createsubscription.html\r\n[2] \r\nhttps://www.postgresql.org/docs/devel/runtime-config-wal.html#GUC-SYNCHRONOUS-COMMIT","msg_date":"Mon, 9 Jan 2023 21:43:15 -0500","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jan 10, 2023 at 8:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> On 9/12/22 1:23 AM, vignesh C wrote:\n> > On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > Thanks for pushing the patch. I have closed this entry in commitfest.\n> > I will wait for some more time and see the response regarding the\n> > documentation patch and then start a new thread if required.\n>\n> I've been testing this patch in advancing of working on the\n> documentation and came across a behavior I wanted to note. Specifically,\n> I am hitting a deadlock while trying to synchronous replicate between\n> the two instances at any `synchronous_commit` level above `local`.\n>\n> Here is my set up. I have two instances, \"A\" and \"B\".\n>\n> On A and B, run:\n>\n> CREATE TABLE sync (id int PRIMARY KEY, info float);\n> CREATE PUBLICATION sync FOR TABLE sync;\n>\n> On A, run:\n>\n> CREATE SUBSCRIPTION sync\n> CONNECTION 'connstr-to-B'\n> PUBLICATION sync\n> WITH (\n> streaming=true, copy_data=false,\n> origin=none, synchronous_commit='on');\n>\n> On B, run:\n>\n> CREATE SUBSCRIPTION sync\n> CONNECTION 'connstr-to-A'\n> PUBLICATION sync\n> WITH (\n> streaming=true, copy_data=false,\n> origin=none, synchronous_commit='on');\n>\n> On A and B, run:\n>\n> ALTER SYSTEM SET synchronous_standby_names TO 'sync';\n> SELECT pg_reload_conf();\n>\n> Verify on A and B that pg_stat_replication.sync_state is set to \"sync\"\n>\n> SELECT application_name, sync_state = 'sync' AS is_sync\n> FROM pg_stat_replication\n> WHERE application_name = 'sync';\n>\n> The next to commands should be run simultaneously on A and B:\n>\n> -- run this on A\n> INSERT INTO sync\n> SELECT x, random() FROM generate_series(1,2000000, 2) x;\n>\n> -- run this on B\n> INSERT INTO sync\n> SELECT x, random() FROM generate_series(2,2000000, 2) x;\n>\n> This consistently created the deadlock in my testing.\n>\n> Discussing with Masahiko off-list, this is due to a deadlock from 4\n> processes: the walsenders on A and B, and the apply workers on A and B.\n> The walsenders are waiting for feedback from the apply workers, and the\n> apply workers are waiting for the walsenders to synchronize (I may be\n> oversimplifying).\n>\n> He suggested I try the above example instead with `synchronous_commit`\n> set to `local`. In this case, I verified that there is no more deadlock,\n> but he informed me that we would not be able to use cascading\n> synchronous replication when \"origin=none\".\n>\n\nThis has nothing to do with the origin feature. I mean this should\nhappen with origin=any or even in PG15 without using origin at all.\nAm, I missing something? One related point to note is that in physical\nreplication cascading replication is asynchronous. See docs [1]\n(Cascading replication is currently asynchronous....)\n\n> If we decide that this is a documentation issue, I'd suggest we improve\n> the guidance around using `synchronous_commit`[1] on the CREATE\n> SUBSCRIPTION page, as the GUC page[2] warns against using `local`:\n>\n\nYeah, but on Create Subscription page, we have mentioned that it is\nsafe to use off for logical replication. One can use local or higher\nfor reducing the latency for COMMIT when synchronous replication is\nused in the publisher. Won't using 'local' while creating subscription\nwould suffice the need to consistently replicate the data? I mean it\nis equivalent to somebody using levels greater than local in case of\nphysical replication. I think in the case of physical replication, we\nwon't wait for standby to replicate to another node before sending a\nresponse, so why to wait in the case of logical replication? If this\nunderstanding is correct, then probably it is sufficient to support\n'local' for a subscription.\n\n[1] - https://www.postgresql.org/docs/devel/warm-standby.html\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 10 Jan 2023 20:47:13 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On 1/10/23 10:17 AM, Amit Kapila wrote:\r\n> On Tue, Jan 10, 2023 at 8:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\r\n\r\n>> This consistently created the deadlock in my testing.\r\n>>\r\n>> Discussing with Masahiko off-list, this is due to a deadlock from 4\r\n>> processes: the walsenders on A and B, and the apply workers on A and B.\r\n>> The walsenders are waiting for feedback from the apply workers, and the\r\n>> apply workers are waiting for the walsenders to synchronize (I may be\r\n>> oversimplifying).\r\n>>\r\n>> He suggested I try the above example instead with `synchronous_commit`\r\n>> set to `local`. In this case, I verified that there is no more deadlock,\r\n>> but he informed me that we would not be able to use cascading\r\n>> synchronous replication when \"origin=none\".\r\n>>\r\n> \r\n> This has nothing to do with the origin feature. I mean this should\r\n> happen with origin=any or even in PG15 without using origin at all.\r\n> Am, I missing something? One related point to note is that in physical\r\n> replication cascading replication is asynchronous. See docs [1]\r\n> (Cascading replication is currently asynchronous....)\r\n\r\nThis is not directly related to the origin feature, but the origin \r\nfeature makes it apparent. There is a common use-case where people want \r\nto replicate data synchronously between two tables, and this is enabled \r\nby the \"origin\" feature.\r\n\r\nTo be clear, my bigger concern is that it's fairly easy for users to \r\ncreate a deadlock situation based on how they set \"synchronous_commit\" \r\nwith the origin feature -- this is the main reason why I brought it up. \r\nI'm less concerned about the \"cascading\" case, though I want to try out \r\nsync rep between 3 instances to see what happens.\r\n\r\n>> If we decide that this is a documentation issue, I'd suggest we improve\r\n>> the guidance around using `synchronous_commit`[1] on the CREATE\r\n>> SUBSCRIPTION page, as the GUC page[2] warns against using `local`:\r\n>>\r\n> \r\n> Yeah, but on Create Subscription page, we have mentioned that it is\r\n> safe to use off for logical replication.\r\n\r\nWhile I think you can infer that it's \"safe\" for synchronous \r\nreplication, I don't think it's clear.\r\n\r\nWe say it's \"safe to use `off` for logical replication\", but provide a \r\nlengthy explanation around synchronous logical replication that \r\nconcludes that it's \"advantageous to set synchronous_commit to local or \r\nhigher\" but does not address safety. The first line in the explanation \r\nof the parameter links to the `synchronous_commit` GUC which \r\nspecifically advises against using \"local\" for synchronous replication.\r\n\r\nAdditionally, because we say \"local\" or higher, we increase the risk of \r\nthe aforementioned in HEAD with the origin feature.\r\n\r\nI know I'm hammering on this point a bit, but it feels like this is \r\nrelatively easy to misconfigure with the upcoming \"origin\" change (I did \r\nso myself from reading the devel docs) and we should ensure we guide our \r\nusers appropriately.\r\n\r\n> One can use local or higher\r\n> for reducing the latency for COMMIT when synchronous replication is\r\n> used in the publisher. Won't using 'local' while creating subscription\r\n> would suffice the need to consistently replicate the data? I mean it\r\n> is equivalent to somebody using levels greater than local in case of\r\n> physical replication. I think in the case of physical replication, we\r\n> won't wait for standby to replicate to another node before sending a\r\n> response, so why to wait in the case of logical replication? If this\r\n> understanding is correct, then probably it is sufficient to support\r\n> 'local' for a subscription.\r\n\r\nI think this is a good explanation. We should incorporate some version \r\nof this into the docs, and add some clarity around the use of \r\n`synchronous_commit` option in `CREATE SUBSCRIPTION` in particular with \r\nthe origin feature. I can make an attempt at this.\r\n\r\nPerhaps another question: based on this, should we disallow setting \r\nvalues of `synchronous_commit` greater than `local` when using \r\n\"origin=none\"? That might be too strict, but maybe we should warn around \r\nthe risk of deadlock either in the logs or in the docs.\r\n\r\nThanks,\r\n\r\nJonathan","msg_date":"Tue, 10 Jan 2023 12:54:28 -0500","msg_from":"\"Jonathan S. Katz\" <jkatz@postgresql.org>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Jan 10, 2023 at 11:24 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> On 1/10/23 10:17 AM, Amit Kapila wrote:\n> > On Tue, Jan 10, 2023 at 8:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\n>\n> > One can use local or higher\n> > for reducing the latency for COMMIT when synchronous replication is\n> > used in the publisher. Won't using 'local' while creating subscription\n> > would suffice the need to consistently replicate the data? I mean it\n> > is equivalent to somebody using levels greater than local in case of\n> > physical replication. I think in the case of physical replication, we\n> > won't wait for standby to replicate to another node before sending a\n> > response, so why to wait in the case of logical replication? If this\n> > understanding is correct, then probably it is sufficient to support\n> > 'local' for a subscription.\n>\n> I think this is a good explanation. We should incorporate some version\n> of this into the docs, and add some clarity around the use of\n> `synchronous_commit` option in `CREATE SUBSCRIPTION` in particular with\n> the origin feature. I can make an attempt at this.\n>\n\nOkay, thanks!\n\n> Perhaps another question: based on this, should we disallow setting\n> values of `synchronous_commit` greater than `local` when using\n> \"origin=none\"?\n>\n\nI think it should be done irrespective of the value of origin because\nit can create a deadlock in those cases as well. I think one idea as\nyou suggest is to block levels higher than local and the other is to\nmake it behave like 'local' even if the level is higher than 'local'\nwhich would be somewhat similar to our physical replication.\n\n> That might be too strict, but maybe we should warn around\n> the risk of deadlock either in the logs or in the docs.\n>\n\nYeah, we can mention it in docs as well. We can probably document as\npart of the docs work for this thread but it would be better to start\na separate thread to fix this behavior by either of the above two\napproaches.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 11 Jan 2023 07:46:42 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On 12.09.22 07:23, vignesh C wrote:\n> On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\n>>\n>> On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\n>>>\n>>>\n>>> The attached patch has the changes to handle the same.\n>>>\n>>\n>> Pushed. I am not completely sure whether we want the remaining\n>> documentation patch in this thread in its current form or by modifying\n>> it. Johnathan has shown some interest in it. I feel you can start a\n>> separate thread for it to see if there is any interest in the same and\n>> close the CF entry for this work.\n> \n> Thanks for pushing the patch. I have closed this entry in commitfest.\n> I will wait for some more time and see the response regarding the\n> documentation patch and then start a new thread if required.\n\nThis patch added the following error message:\n\nerrdetail_plural(\"Subscribed publication %s is subscribing to other \npublications.\",\n\"Subscribed publications %s are subscribing to other publications.\",\nlist_length(publist), pubnames->data),\n\nBut in PostgreSQL, a publication cannot subscribe to a publication, so \nthis is not giving accurate information. Apparently, what it is trying \nto say is that\n\nThe subscription that you are creating subscribes to publications that \ncontain tables that are written to by other subscriptions.\n\nCan we get to a more accurate wording like this?\n\nThere is also a translatability issue there, in the way the publication \nlist is pasted into the message.\n\nIs the list of affected publications really that interesting? I wonder \nwhether the list of affected tables might be more relevant?\n\n\n\n","msg_date":"Tue, 8 Aug 2023 10:20:42 +0200","msg_from":"Peter Eisentraut <peter@eisentraut.org>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\n>\n> On 12.09.22 07:23, vignesh C wrote:\n> > On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >>\n> >> On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\n> >>>\n> >>>\n> >>> The attached patch has the changes to handle the same.\n> >>>\n> >>\n> >> Pushed. I am not completely sure whether we want the remaining\n> >> documentation patch in this thread in its current form or by modifying\n> >> it. Johnathan has shown some interest in it. I feel you can start a\n> >> separate thread for it to see if there is any interest in the same and\n> >> close the CF entry for this work.\n> >\n> > Thanks for pushing the patch. I have closed this entry in commitfest.\n> > I will wait for some more time and see the response regarding the\n> > documentation patch and then start a new thread if required.\n>\n> This patch added the following error message:\n>\n> errdetail_plural(\"Subscribed publication %s is subscribing to other\n> publications.\",\n> \"Subscribed publications %s are subscribing to other publications.\",\n> list_length(publist), pubnames->data),\n>\n> But in PostgreSQL, a publication cannot subscribe to a publication, so\n> this is not giving accurate information. Apparently, what it is trying\n> to say is that\n>\n> The subscription that you are creating subscribes to publications that\n> contain tables that are written to by other subscriptions.\n>\n> Can we get to a more accurate wording like this?\n>\n\n+1 for changing the message as per your suggestion.\n\n> There is also a translatability issue there, in the way the publication\n> list is pasted into the message.\n>\n> Is the list of affected publications really that interesting? I wonder\n> whether the list of affected tables might be more relevant?\n>\n\nIn that case, we need to specify both schema name and table name in\nthat case. I guess the list could be very long and not sure what to do\nfor schema publications ( Create Publication ... For Schema).\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 8 Aug 2023 14:22:43 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Tue, Aug 8, 2023 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\n> >\n> > This patch added the following error message:\n> >\n> > errdetail_plural(\"Subscribed publication %s is subscribing to other\n> > publications.\",\n> > \"Subscribed publications %s are subscribing to other publications.\",\n> > list_length(publist), pubnames->data),\n> >\n> > But in PostgreSQL, a publication cannot subscribe to a publication, so\n> > this is not giving accurate information. Apparently, what it is trying\n> > to say is that\n> >\n> > The subscription that you are creating subscribes to publications that\n> > contain tables that are written to by other subscriptions.\n> >\n> > Can we get to a more accurate wording like this?\n> >\n>\n> +1 for changing the message as per your suggestion.\n>\n\nPSA a patch to change this message text. The message now has wording\nsimilar to the suggestion.\n\n> > There is also a translatability issue there, in the way the publication\n> > list is pasted into the message.\n> >\n\nThe name/list substitution is now done within parentheses, which AFAIK\nwill be enough to eliminate any translation ambiguities.\n\n> > Is the list of affected publications really that interesting? I wonder\n> > whether the list of affected tables might be more relevant?\n> >\n>\n> In that case, we need to specify both schema name and table name in\n> that case. I guess the list could be very long and not sure what to do\n> for schema publications ( Create Publication ... For Schema).\n\nRight, IIUC that was the reason for not choosing to show the tables in\nthe original patch -- i.e. the list might easily be very long with\n100s or 1000s of tables it, and so inappropriate to substitute in the\nmessage. OTOH, showing only problematic publications is a short list\nbut it is still more useful than showing nothing (e.g. there other\npublications of the subscription might be OK so those ones are not\nlisted)\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia","msg_date":"Wed, 9 Aug 2023 12:50:10 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 9, 2023 at 8:20 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Tue, Aug 8, 2023 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\n> > >\n> > > This patch added the following error message:\n> > >\n> > > errdetail_plural(\"Subscribed publication %s is subscribing to other\n> > > publications.\",\n> > > \"Subscribed publications %s are subscribing to other publications.\",\n> > > list_length(publist), pubnames->data),\n> > >\n> > > But in PostgreSQL, a publication cannot subscribe to a publication, so\n> > > this is not giving accurate information. Apparently, what it is trying\n> > > to say is that\n> > >\n> > > The subscription that you are creating subscribes to publications that\n> > > contain tables that are written to by other subscriptions.\n> > >\n> > > Can we get to a more accurate wording like this?\n> > >\n> >\n> > +1 for changing the message as per your suggestion.\n> >\n>\n> PSA a patch to change this message text. The message now has wording\n> similar to the suggestion.\n>\n> > > There is also a translatability issue there, in the way the publication\n> > > list is pasted into the message.\n> > >\n>\n> The name/list substitution is now done within parentheses, which AFAIK\n> will be enough to eliminate any translation ambiguities.\n>\n\nA similar instance as below uses \\\"%s\\\" instead. So, isn't using the\nsame a better idea?\nerrdetail_plural(\"LDAP search for filter \\\"%s\\\" on server \\\"%s\\\"\nreturned %d entry.\",\n \"LDAP search for filter \\\"%s\\\" on server \\\"%s\\\" returned %d entries.\",\n\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 9 Aug 2023 10:14:05 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 9, 2023 at 2:44 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Aug 9, 2023 at 8:20 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > On Tue, Aug 8, 2023 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > >\n> > > On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\n> > > >\n> > > > This patch added the following error message:\n> > > >\n> > > > errdetail_plural(\"Subscribed publication %s is subscribing to other\n> > > > publications.\",\n> > > > \"Subscribed publications %s are subscribing to other publications.\",\n> > > > list_length(publist), pubnames->data),\n> > > >\n> > > > But in PostgreSQL, a publication cannot subscribe to a publication, so\n> > > > this is not giving accurate information. Apparently, what it is trying\n> > > > to say is that\n> > > >\n> > > > The subscription that you are creating subscribes to publications that\n> > > > contain tables that are written to by other subscriptions.\n> > > >\n> > > > Can we get to a more accurate wording like this?\n> > > >\n> > >\n> > > +1 for changing the message as per your suggestion.\n> > >\n> >\n> > PSA a patch to change this message text. The message now has wording\n> > similar to the suggestion.\n> >\n> > > > There is also a translatability issue there, in the way the publication\n> > > > list is pasted into the message.\n> > > >\n> >\n> > The name/list substitution is now done within parentheses, which AFAIK\n> > will be enough to eliminate any translation ambiguities.\n> >\n>\n> A similar instance as below uses \\\"%s\\\" instead. So, isn't using the\n> same a better idea?\n> errdetail_plural(\"LDAP search for filter \\\"%s\\\" on server \\\"%s\\\"\n> returned %d entry.\",\n> \"LDAP search for filter \\\"%s\\\" on server \\\"%s\\\" returned %d entries.\",\n>\n>\n\nHmm. I don't see any similarity -- there the plural is for the %d, not\nfor a list of string items. And in our case the publication name (or\nlist of names) are already quoted by the function returning that list,\nso quoting them again doesn't really make sense.\n\nExample output using the patch look like this:\n\nSINGLULAR\ntest_sub=# create subscription sub_test connection 'dbname=test_pub'\npublication pub_all_at_pub with(origin=NONE);\nWARNING: subscription \"sub_test\" requested copy_data with origin =\nNONE but might copy data that had a different origin\nDETAIL: The subscription that you are creating has a publication\n(\"pub_all_at_pub\") containing tables written to by other\nsubscriptions.\nHINT: Verify that initial data copied from the publisher tables did\nnot come from other origins.\nNOTICE: created replication slot \"sub_test\" on publisher\nCREATE SUBSCRIPTION\n\nPLURAL\ntest_sub=# create subscription sub_test3 connection 'dbname=test_pub'\npublication pub_all_at_pub,pub_s1_at_pub,pub_s2_at_pub\nwith(origin=NONE);\nWARNING: subscription \"sub_test3\" requested copy_data with origin =\nNONE but might copy data that had a different origin\nDETAIL: The subscription that you are creating has publications\n(\"pub_s1_at_pub\", \"pub_all_at_pub\") containing tables written to by\nother subscriptions.\nHINT: Verify that initial data copied from the publisher tables did\nnot come from other origins.\nNOTICE: created replication slot \"sub_test3\" on publisher\nCREATE SUBSCRIPTION\n\n~\n\nI thought above looked fine but if PeterE also does not like the ()\nthen the message can be rearranged slightly like below to put the\noffending publications at the end of the message, like this:\n\nDETAIL: The subscription that you are creating has the following\npublications containing tables written to by other subscriptions:\n\"pub_s1_at_pub\", \"pub_all_at_pub\"\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia\n\n\n","msg_date":"Wed, 9 Aug 2023 15:29:31 +1000","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On Wed, Aug 9, 2023 at 10:59 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> Example output using the patch look like this:\n>\n> SINGLULAR\n> test_sub=# create subscription sub_test connection 'dbname=test_pub'\n> publication pub_all_at_pub with(origin=NONE);\n> WARNING: subscription \"sub_test\" requested copy_data with origin =\n> NONE but might copy data that had a different origin\n> DETAIL: The subscription that you are creating has a publication\n> (\"pub_all_at_pub\") containing tables written to by other\n> subscriptions.\n> HINT: Verify that initial data copied from the publisher tables did\n> not come from other origins.\n> NOTICE: created replication slot \"sub_test\" on publisher\n> CREATE SUBSCRIPTION\n>\n> PLURAL\n> test_sub=# create subscription sub_test3 connection 'dbname=test_pub'\n> publication pub_all_at_pub,pub_s1_at_pub,pub_s2_at_pub\n> with(origin=NONE);\n> WARNING: subscription \"sub_test3\" requested copy_data with origin =\n> NONE but might copy data that had a different origin\n> DETAIL: The subscription that you are creating has publications\n> (\"pub_s1_at_pub\", \"pub_all_at_pub\") containing tables written to by\n> other subscriptions.\n> HINT: Verify that initial data copied from the publisher tables did\n> not come from other origins.\n> NOTICE: created replication slot \"sub_test3\" on publisher\n> CREATE SUBSCRIPTION\n>\n> ~\n>\n> I thought above looked fine but if PeterE also does not like the ()\n> then the message can be rearranged slightly like below to put the\n> offending publications at the end of the message, like this:\n>\n\nOkay, I didn't know strings were already quoted but not sure if Round\nBrackets () exactly will address the translatability concern.\n\n> DETAIL: The subscription that you are creating has the following\n> publications containing tables written to by other subscriptions:\n> \"pub_s1_at_pub\", \"pub_all_at_pub\"\n>\n\nFair enough. Peter E., do let us know what you think makes sense here?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 9 Aug 2023 11:58:23 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"},{"msg_contents":"On 09.08.23 04:50, Peter Smith wrote:\n> On Tue, Aug 8, 2023 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>>\n>> On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\n>>>\n>>> This patch added the following error message:\n>>>\n>>> errdetail_plural(\"Subscribed publication %s is subscribing to other\n>>> publications.\",\n>>> \"Subscribed publications %s are subscribing to other publications.\",\n>>> list_length(publist), pubnames->data),\n>>>\n>>> But in PostgreSQL, a publication cannot subscribe to a publication, so\n>>> this is not giving accurate information. Apparently, what it is trying\n>>> to say is that\n>>>\n>>> The subscription that you are creating subscribes to publications that\n>>> contain tables that are written to by other subscriptions.\n>>>\n>>> Can we get to a more accurate wording like this?\n>>>\n>>\n>> +1 for changing the message as per your suggestion.\n>>\n> \n> PSA a patch to change this message text. The message now has wording\n> similar to the suggestion.\n\ncommitted, thanks\n\n\n\n","msg_date":"Wed, 16 Aug 2023 15:50:09 +0200","msg_from":"Peter Eisentraut <peter@eisentraut.org>","msg_from_op":false,"msg_subject":"Re: Handle infinite recursion in logical replication setup"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nIn logical replication, currently Walsender sends the data that is\\ngenerated locally and the data that are replicated from other\\ninstances. This results in infinite recursion in circular logical\\nreplication setup.\\nHere the user is trying to have a 2-way replication setup with node 1\\npublishing data to node2 and node2 publishing data to node1, so that\\nthe user can perform dml operations from any node, it can act as a\\n2-way multi master replication setup.\\n\\nThis problem can be reproduced with the following steps:\\n-- Instance 1\\ncreate publication pub1 for table t1;\\ncreate table t1(c1 int);\\n\\n-- Instance 2\\ncreate table t1(c1 int);\\ncreate publication pub2 for table t1;\\ncreate subscription sub1 CONNECTION 'dbname=postgres port=5432'\\npublication pub1;\\n\\n-- Instance 1\\ncreate subscription sub2 CONNECTION 'dbname=postgres port=5433'\\npublication pub2; insert into t1 values(10);\\n\\nIn this scenario, the Walsender in publisher pub1 sends data to the\\napply worker in subscriber sub1, the apply worker in sub1 maps the\\ndata to local tables and applies the individual changes as they are\\nreceived. Then the Walsender in publisher pub2 sends data to the apply\\nworker in subscriber sub2, the apply worker in sub2 maps the data to\\nlocal tables and applies the individual changes as they are received.\\nThis process repeats infinitely.\\n\\nCurrently we do not differentiate if the data is locally generated\\ndata, or a replicated data and we send both the data which causes\\ninfinite recursion.\\n\\nWe could see that the record count has increased significantly within sometime:\\nselect count(*) from t1;\\n count\\n--------------\\n 4000000\\n(1 row)\\n\\nIf the table had primary key constraint, we could notice that the\\nfirst insert is successful and when the same insert is sent back, the\\ninsert fails because of constraint error:\\n2022-02-23 09:28:43.592 IST [14743] ERROR: duplicate key value\\nviolates unique constraint \\\"t1_pkey\\\"\\n2022-02-23 09:28:43.592 IST [14743] DETAIL: Key (c1)=(10) already exists.\\n2022-02-23 09:28:43.592 IST [14743] CONTEXT: processing remote data\\nduring \\\"INSERT\\\" for replication target relation \\\"public.t1\\\" in\\ntransaction 727 at 2022-02-23 09:28:43.406738+05:30\\n2022-02-23 09:28:43.593 IST [14678] LOG: background worker \\\"logical\\nreplication worker\\\" (PID 14743) exited with exit code 1\\n2022-02-23 09:28:48.608 IST [14745] LOG: logical replication apply\\nworker for subscription \\\"sub2\\\" has started\\n2022-02-23 09:28:48.624 IST [14745] ERROR: duplicate key value\\nviolates unique constraint \\\"t1_pkey\\\"\\n2022-02-23 09:28:48.624 IST [14745] DETAIL: Key (c1)=(10) already exists.\\n2022-02-23 09:28:48.624 IST [14745] CONTEXT: processing remote data\\nduring \\\"INSERT\\\" for replication target relation \\\"public.t1\\\" in\\ntransaction 727 at 2022-02-23 09:28:43.406738+05:30\\n2022-02-23 09:28:48.626 IST [14678] LOG: background worker \\\"logical\\nreplication worker\\\" (PID 14745) exited with exit code 1\\n\\nThe same problem can occur in any circular node setup like 3 nodes,\\n4node etc like: a) node1 publishing to node2 b) node2 publishing to\\nnode3 c) node3 publishing back to node1.\\n\\nHere there are two problems for the user: a) incremental\\nsynchronization of table sending both local data and replicated data\\nby walsender b) Table synchronization of table using copy command\\nsending both local data and replicated data\\n\\nFor the first problem \\\"Incremental synchronization of table by\\nWalsender\\\" can be solved by:\\nCurrently the locally generated data does not have replication origin\\nassociated and the data that has originated from another instance will\\nhave a replication origin associated. We could use this information to\\ndifferentiate locally generated data and replicated data and send only\\nthe locally generated data. This \\\"only_local\\\" could be provided as an\\noption while subscription is created:\\nex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\\nPUBLICATION pub1 with (only_local = on);\\n\\nI have attached a basic patch for this, if the idea is accepted, I\\nwill work further to test more scenarios, add documentation, and test\\nand post an updated patch.\\nFor the second problem, Table synchronization of table including local\\ndata and replicated data using copy command.\\n\\nLet us consider the following scenario:\\na) node1 publishing to node2 b) node2 publishing to node1. Here in\\nthis case node1 will have replicated data from node2 and vice versa.\\n\\nIn the above if user wants to include node3 to subscribe data from\\nnode2. Users will have to create a subscription in node3 to get the\\ndata from node2. During table synchronization we send the complete\\ntable data from node2 to node3. Node2 will have local data from node2\\nand also replicated data from node1. Currently we don't have an option\\nto differentiate between the locally generated data and replicated\\ndata in the heap which will cause infinite recursion as described\\nabove.\\n\\nTo handle this if user has specified only_local option, we could throw\\na warning or error out while creating subscription in this case, we\\ncould have a column srreplicateddata in pg_subscription_rel which\\ncould indicate if the table has any replicated data or not:\\npostgres=# select * from pg_subscription_rel;\\n srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\\n---------+---------+------------+-----------+------------------\\n 16389 | 16384 | r | 0/14A4640 | t\\n 16389 | 16385 | r | 0/14A4690 | f\\n(1 row)\\n\\nIn the above example, srreplicateddata with true indicates, tabel t1\\nwhose relid is 16384 has replicated data and the other row having\\nsrreplicateddata as false indicates table t2 whose relid is 16385\\ndoes not have replicated data.\\nWhen creating a new subscription, the subscriber will connect to the\\npublisher and check if the relation has replicated data by checking\\nsrreplicateddata in pg_subscription_rel table.\\nIf the table has any replicated data, log a warning or error for this.\\n\\nAlso, we could document the steps on how to handle the initial sync like:\\na) Complete the ongoing transactions on this table in the replication\\nsetup nodes i.e. node1 and node2 in the above case, so that the table\\ndata is consistent, b) Once there are no ongoing transaction, Copy the\\ntable data using copy command from any one of the nodes, c) create\\nsubscription with copy_data option as off d) Perform further\\ntransactions on the table e) All the further transactions performed\\nwill be handled by the walsender which will take care of skipping\\nreplicated data and sending only the local data. i.e. node2 will send\\nthe locally generated data to node3.\\n\\nI'm not sure if there is any other better way to handle this. If there\\nis a better way, we could handle it accordingly.\\nThoughts?\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 23 Feb 2022 11:59:02 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi Vignesh,\\r\\n\\r\\n> In logical replication, currently Walsender sends the data that is\\r\\n> generated locally and the data that are replicated from other\\r\\n> instances. This results in infinite recursion in circular logical\\r\\n> replication setup.\\r\\n\\r\\nThank you for good explanation. I understand that this fix can be used\\r\\nfor a bidirectional replication.\\r\\n\\r\\n> Here there are two problems for the user: a) incremental\\r\\n> synchronization of table sending both local data and replicated data\\r\\n> by walsender b) Table synchronization of table using copy command\\r\\n> sending both local data and replicated data\\r\\n\\r\\nSo you wanted to solve these two problem and currently focused on\\r\\nthe first one, right? We can check one by one.\\r\\n\\r\\n> For the first problem \\\"Incremental synchronization of table by\\r\\n> Walsender\\\" can be solved by:\\r\\n> Currently the locally generated data does not have replication origin\\r\\n> associated and the data that has originated from another instance will\\r\\n> have a replication origin associated. We could use this information to\\r\\n> differentiate locally generated data and replicated data and send only\\r\\n> the locally generated data. This \\\"only_local\\\" could be provided as an\\r\\n> option while subscription is created:\\r\\n> ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\\r\\n> PUBLICATION pub1 with (only_local = on);\\r\\n\\r\\nSounds good, but I cannot distinguish whether the assumption will keep.\\r\\n\\r\\nI played with your patch, but it could not be applied to current master.\\r\\nI tested from bd74c40 and I confirmed infinite loop was not appeared.\\r\\n\\r\\nlocal_only could not be set from ALTER SUBSCRIPTION command.\\r\\nIs it expected?\\r\\n\\r\\nBest Regards,\\r\\nHayato Kuroda\\r\\nFUJITSU LIMITED\\r\\n\\r\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 10:42:27 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 1, 2022 at 4:12 PM kuroda.hayato@fujitsu.com\\n<kuroda.hayato@fujitsu.com> wrote:\\n>\\n> Hi Vignesh,\\n>\\n> > In logical replication, currently Walsender sends the data that is\\n> > generated locally and the data that are replicated from other\\n> > instances. This results in infinite recursion in circular logical\\n> > replication setup.\\n>\\n> Thank you for good explanation. I understand that this fix can be used\\n> for a bidirectional replication.\\n\\nOnce these issues are resolved, it can be used for bi-directional\\nlogical replication.\\n\\n> > Here there are two problems for the user: a) incremental\\n> > synchronization of table sending both local data and replicated data\\n> > by walsender b) Table synchronization of table using copy command\\n> > sending both local data and replicated data\\n>\\n> So you wanted to solve these two problem and currently focused on\\n> the first one, right? We can check one by one.\\n\\nYes.\\n\\n> > For the first problem \\\"Incremental synchronization of table by\\n> > Walsender\\\" can be solved by:\\n> > Currently the locally generated data does not have replication origin\\n> > associated and the data that has originated from another instance will\\n> > have a replication origin associated. We could use this information to\\n> > differentiate locally generated data and replicated data and send only\\n> > the locally generated data. This \\\"only_local\\\" could be provided as an\\n> > option while subscription is created:\\n> > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\\n> > PUBLICATION pub1 with (only_local = on);\\n>\\n> Sounds good, but I cannot distinguish whether the assumption will keep.\\n\\nReplication origin is created by the apply worker and it will be used\\nfor all the transactions received from the walsender. I feel the\\nreplication origin will be present always.\\n\\n> I played with your patch, but it could not be applied to current master.\\n> I tested from bd74c40 and I confirmed infinite loop was not appeared.\\n\\nI will post an updated version for this soon.\\n\\n> local_only could not be set from ALTER SUBSCRIPTION command.\\n> Is it expected?\\n\\nI wanted to get the opinion from others too just to make sure the\\napproach is right. I will fix this including the documentation, test,\\netc in the later versions.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 08:56:57 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n...\\n...\\n> I have attached a basic patch for this, if the idea is accepted, I\\n> will work further to test more scenarios, add documentation, and test\\n> and post an updated patch.\\n> For the second problem, Table synchronization of table including local\\n> data and replicated data using copy command.\\n>\\n> Let us consider the following scenario:\\n> a) node1 publishing to node2 b) node2 publishing to node1. Here in\\n> this case node1 will have replicated data from node2 and vice versa.\\n>\\n> In the above if user wants to include node3 to subscribe data from\\n> node2. Users will have to create a subscription in node3 to get the\\n> data from node2. During table synchronization we send the complete\\n> table data from node2 to node3. Node2 will have local data from node2\\n> and also replicated data from node1. Currently we don't have an option\\n> to differentiate between the locally generated data and replicated\\n> data in the heap which will cause infinite recursion as described\\n> above.\\n>\\n> To handle this if user has specified only_local option, we could throw\\n> a warning or error out while creating subscription in this case, we\\n> could have a column srreplicateddata in pg_subscription_rel which\\n> could indicate if the table has any replicated data or not:\\n> postgres=# select * from pg_subscription_rel;\\n> srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\\n> ---------+---------+------------+-----------+------------------\\n> 16389 | 16384 | r | 0/14A4640 | t\\n> 16389 | 16385 | r | 0/14A4690 | f\\n> (1 row)\\n>\\n> In the above example, srreplicateddata with true indicates, tabel t1\\n> whose relid is 16384 has replicated data and the other row having\\n> srreplicateddata as false indicates table t2 whose relid is 16385\\n> does not have replicated data.\\n> When creating a new subscription, the subscriber will connect to the\\n> publisher and check if the relation has replicated data by checking\\n> srreplicateddata in pg_subscription_rel table.\\n> If the table has any replicated data, log a warning or error for this.\\n>\\n\\nIf you want to give the error in this case, then I think we need to\\nprovide an option to the user to allow copy. One possibility could be\\nto extend existing copy_data option as 'false', 'true', 'force'. For\\n'false', there shouldn't be any change, for 'true', if 'only_local'\\noption is also set and the new column indicates replicated data then\\ngive an error, for 'force', we won't give an error even if the\\nconditions as mentioned for 'true' case are met, rather we will allow\\ncopy in this case.\\n\\n> Also, we could document the steps on how to handle the initial sync like:\\n> a) Complete the ongoing transactions on this table in the replication\\n> setup nodes i.e. node1 and node2 in the above case, so that the table\\n> data is consistent, b) Once there are no ongoing transaction, Copy the\\n> table data using copy command from any one of the nodes, c) create\\n> subscription with copy_data option as off d) Perform further\\n> transactions on the table e) All the further transactions performed\\n> will be handled by the walsender which will take care of skipping\\n> replicated data and sending only the local data. i.e. node2 will send\\n> the locally generated data to node3.\\n>\\n> I'm not sure if there is any other better way to handle this.\\n>\\n\\nI could think of the below options for users to set up bi-directional\\nreplication for the same table.\\n\\nOption-1:\\nThere is no pre-existing data in the tables that are going to\\nparticipate in bi-directional replication. In such a case, Users can\\ncreate pub/sub (with only_local option as proposed by you) on both\\nnodes before starting any writes on tables. This will allow\\nbi-directional replication for the required tables. Now, if the user\\nwants one of the nodes to join at a later point, then the strategy in\\nOption-2/3 could be used.\\n\\nOption-2:\\nOne of the nodes (say node-1) has some pre-existing data and another\\nnode (say node-2) doesn't have any pre-existing data. In this case,\\nthe user can set up pub/sub (with only_local and copy_data as 'false'\\noptions) for node-1 first before any of the operations on node-2.\\nThen, it can set up pub/sub on node-2. This will allow bi-directional\\nreplication for the required tables.\\n\\nOption-3:\\nBoth the nodes have some pre-existing data. I think the easiest option\\ncould be to truncate data on one of the nodes and set up pub/sub on\\nboth nodes. See, one way to achieve it among two nodes as below:\\n\\nNode-1:\\nTable t1 has data\\n1, 2, 3, 4\\n\\nPublication for t1, pub1: Create Publication pub1 For Table t1;\\n\\nNode-2:\\nTable t1 has data\\n5, 6, 7, 8\\nPublication for t1, pub1_2: Create Publication pub1_2 For Table t1;\\n\\nNow, Create Subscription for pub1 on node1: Create Subscription sub1_2\\nConnection '<node-1 details>' Publication pub1 WITH (only_local =\\ntrue);\\n\\nNode-1:\\nBegin;\\n# Disallow truncates to be published\\nAlter Publication pub1 Set (publish = 'insert, update, delete');\\nTruncate t1;\\nCreate Subscription sub1 Connection '<node-2 details>' Publication\\npub1 WITH (only_local = true, copy_data = 'force');\\nAlter Publication pub1 Set (publish = 'insert, update, delete, truncate');\\nCommit;\\n\\nI think this will allow the bi-directional replication between two\\nnodes. In this scheme, the user needs to manually perform some steps\\nincluding truncate of the table on one of the nodes which she might or\\nmight not like but at least there will be a way to set up a\\nbi-directional replication on two nodes for same table operations\\nwhich is not possible now.\\n\\nI think one can even imagine using and extending this functionality so\\nthat users don't need to perform TRUNCATE on one of the nodes. Say, in\\nthe above case for tablesync phase, we make both nodes to start a\\ntransaction, create a slot on another node (with USE_SNAPSHOT option),\\nand then allow copy from another node. I think it will be important to\\nallow copy on each node once the slots are created and the initial\\nsnapshot is established.\\n\\nFor more than two nodes, I think we can suggest having either of the\\noption-1 or 2 for setup. But, there could be other ways as well\\ndepending on how the user wants to do the setup.\\n\\nThoughts?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 15:58:55 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 1, 2022 at 4:12 PM kuroda.hayato@fujitsu.com\\n<kuroda.hayato@fujitsu.com> wrote:\\n>\\n> Hi Vignesh,\\n>\\n> > In logical replication, currently Walsender sends the data that is\\n> > generated locally and the data that are replicated from other\\n> > instances. This results in infinite recursion in circular logical\\n> > replication setup.\\n>\\n> Thank you for good explanation. I understand that this fix can be used\\n> for a bidirectional replication.\\n>\\n> > Here there are two problems for the user: a) incremental\\n> > synchronization of table sending both local data and replicated data\\n> > by walsender b) Table synchronization of table using copy command\\n> > sending both local data and replicated data\\n>\\n> So you wanted to solve these two problem and currently focused on\\n> the first one, right? We can check one by one.\\n>\\n> > For the first problem \\\"Incremental synchronization of table by\\n> > Walsender\\\" can be solved by:\\n> > Currently the locally generated data does not have replication origin\\n> > associated and the data that has originated from another instance will\\n> > have a replication origin associated. We could use this information to\\n> > differentiate locally generated data and replicated data and send only\\n> > the locally generated data. This \\\"only_local\\\" could be provided as an\\n> > option while subscription is created:\\n> > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\\n> > PUBLICATION pub1 with (only_local = on);\\n>\\n> Sounds good, but I cannot distinguish whether the assumption will keep.\\n>\\n> I played with your patch, but it could not be applied to current master.\\n> I tested from bd74c40 and I confirmed infinite loop was not appeared.\\nRebased the patch on top of head\\n\\n> local_only could not be set from ALTER SUBSCRIPTION command.\\n> Is it expected?\\nModified\\n\\nThanks for the comments, the attached patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 2 Mar 2022 21:42:11 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"FYI, the v2 patch did not apply to HEAD\\n\\n[postgres@CentOS7-x64 oss_postgres_misc]$ git apply\\n../patches_misc/v2-0001-Skip-replication-of-non-local-data.patch\\n--verbose\\n...\\nerror: patch failed: src/backend/replication/slotfuncs.c:231\\nerror: src/backend/replication/slotfuncs.c: patch does not apply\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia.\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 14:37:02 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\\n> Here there are two problems for the user: a) incremental\\n> synchronization of table sending both local data and replicated data\\n> by walsender b) Table synchronization of table using copy command\\n> sending both local data and replicated data\\n>\\n> For the first problem \\\"Incremental synchronization of table by\\n> Walsender\\\" can be solved by:\\n> Currently the locally generated data does not have replication origin\\n> associated and the data that has originated from another instance will\\n> have a replication origin associated. We could use this information to\\n> differentiate locally generated data and replicated data and send only\\n> the locally generated data. This \\\"only_local\\\" could be provided as an\\n> option while subscription is created:\\n> ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\\n> PUBLICATION pub1 with (only_local = on);\\n\\nI haven't yet gone through the patch, but I have a question about the\\nidea. Suppose I want to set up a logical replication like,\\nnode1->node2->node3->node1. So how would I create the subscriber at\\nnode1? only_local=on or off?. I mean on node1, I want the changes\\nfrom node3 which are generated on node3 or which are replicated from\\nnode2 but I do not want changes that are replicated from node1 itself?\\n So if I set only_local=on then node1 will not get the changes\\nreplicated from node2, is that right? and If I set only_local=off then\\nit will create the infinite loop again? So how are we protecting\\nagainst this case?\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 09:25:33 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 9:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>\\n> On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\\n> > Here there are two problems for the user: a) incremental\\n> > synchronization of table sending both local data and replicated data\\n> > by walsender b) Table synchronization of table using copy command\\n> > sending both local data and replicated data\\n> >\\n> > For the first problem \\\"Incremental synchronization of table by\\n> > Walsender\\\" can be solved by:\\n> > Currently the locally generated data does not have replication origin\\n> > associated and the data that has originated from another instance will\\n> > have a replication origin associated. We could use this information to\\n> > differentiate locally generated data and replicated data and send only\\n> > the locally generated data. This \\\"only_local\\\" could be provided as an\\n> > option while subscription is created:\\n> > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\\n> > PUBLICATION pub1 with (only_local = on);\\n>\\n> I haven't yet gone through the patch, but I have a question about the\\n> idea. Suppose I want to set up a logical replication like,\\n> node1->node2->node3->node1. So how would I create the subscriber at\\n> node1? only_local=on or off?. I mean on node1, I want the changes\\n> from node3 which are generated on node3 or which are replicated from\\n> node2 but I do not want changes that are replicated from node1 itself?\\n> So if I set only_local=on then node1 will not get the changes\\n> replicated from node2, is that right? and If I set only_local=off then\\n> it will create the infinite loop again? So how are we protecting\\n> against this case?\\n>\\n\\nIn the above topology if you want local changes from both node3 and\\nnode2 then I think the way to get that would be you have to create two\\nsubscriptions on node1. The first one points to node2 (with\\nonly_local=off) and the second one points to node3 (with only_local\\n=off). Will that address your case or am I missing something?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 10:05:09 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 10:05 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 9:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >\\n> > On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\\n> > > Here there are two problems for the user: a) incremental\\n> > > synchronization of table sending both local data and replicated data\\n> > > by walsender b) Table synchronization of table using copy command\\n> > > sending both local data and replicated data\\n> > >\\n> > > For the first problem \\\"Incremental synchronization of table by\\n> > > Walsender\\\" can be solved by:\\n> > > Currently the locally generated data does not have replication origin\\n> > > associated and the data that has originated from another instance will\\n> > > have a replication origin associated. We could use this information to\\n> > > differentiate locally generated data and replicated data and send only\\n> > > the locally generated data. This \\\"only_local\\\" could be provided as an\\n> > > option while subscription is created:\\n> > > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\\n> > > PUBLICATION pub1 with (only_local = on);\\n> >\\n> > I haven't yet gone through the patch, but I have a question about the\\n> > idea. Suppose I want to set up a logical replication like,\\n> > node1->node2->node3->node1. So how would I create the subscriber at\\n> > node1? only_local=on or off?. I mean on node1, I want the changes\\n> > from node3 which are generated on node3 or which are replicated from\\n> > node2 but I do not want changes that are replicated from node1 itself?\\n> > So if I set only_local=on then node1 will not get the changes\\n> > replicated from node2, is that right? and If I set only_local=off then\\n> > it will create the infinite loop again? So how are we protecting\\n> > against this case?\\n> >\\n>\\n> In the above topology if you want local changes from both node3 and\\n> node2 then I think the way to get that would be you have to create two\\n> subscriptions on node1. The first one points to node2 (with\\n> only_local=off) and the second one points to node3 (with only_local\\n> =off).\\n>\\n\\nSorry, I intend to say 'only_local=on' at both places in my previous email.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 10:15:46 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi Vignesh, I also have not looked at the patch yet, but I have what\\nseems like a very fundamental (and possibly dumb) question...\\n\\nBasically, I do not understand the choice of syntax for setting things up.\\n\\nIMO that \\\"only-local\\\" option sounds very similar to the other\\nPUBLICATION (\\\"publish\\\") options which decide the kinds of things that\\nwill be published. So it feels more natural for me to think of the\\npublisher as being the one to decide what will be published.\\n\\ne.g.\\n\\noption 1:\\nCREATE PUBLICATION p1 FOR TABLE t1;\\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n\\noption 2:\\nCREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n\\n~~\\n\\nIIUC the patch is using option 1. My first impression was it feels\\nback-to-front for the SUBSCRIPTION telling the PUBLICATION what to\\npublish.\\n\\nSo, why does the patch use syntax option 1?\\n\\n------\\nKind Regards,\\nPeter Smith\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 15:56:25 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 3:56 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Hi Vignesh, I also have not looked at the patch yet, but I have what\\n> seems like a very fundamental (and possibly dumb) question...\\n>\\n> Basically, I do not understand the choice of syntax for setting things up.\\n>\\n> IMO that \\\"only-local\\\" option sounds very similar to the other\\n> PUBLICATION (\\\"publish\\\") options which decide the kinds of things that\\n> will be published. So it feels more natural for me to think of the\\n> publisher as being the one to decide what will be published.\\n>\\n> e.g.\\n>\\n> option 1:\\n> CREATE PUBLICATION p1 FOR TABLE t1;\\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n>\\n> option 2:\\n> CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n>\\n\\nSorry, I mean to write WITH.\\n\\noption 2:\\nCREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n\\n> ~~\\n>\\n> IIUC the patch is using option 1. My first impression was it feels\\n> back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\\n> publish.\\n>\\n> So, why does the patch use syntax option 1?\\n>\\n> ------\\n> Kind Regards,\\n> Peter Smith\\n> Fujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 16:06:53 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Hi Vignesh, I also have not looked at the patch yet, but I have what\\n> seems like a very fundamental (and possibly dumb) question...\\n>\\n> Basically, I do not understand the choice of syntax for setting things up.\\n>\\n> IMO that \\\"only-local\\\" option sounds very similar to the other\\n> PUBLICATION (\\\"publish\\\") options which decide the kinds of things that\\n> will be published. So it feels more natural for me to think of the\\n> publisher as being the one to decide what will be published.\\n>\\n> e.g.\\n>\\n> option 1:\\n> CREATE PUBLICATION p1 FOR TABLE t1;\\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n>\\n> option 2:\\n> CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n>\\n> ~~\\n>\\n> IIUC the patch is using option 1. My first impression was it feels\\n> back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\\n> publish.\\n>\\n> So, why does the patch use syntax option 1?\\n\\nI felt the advantage with keeping it at the subscription side is that,\\nthe subscriber from one node can subscribe with only_local option on\\nand a different subscriber from a different node can subscribe with\\nonly_local option as off. This might not be possible with having the\\noption at publisher side. Having it at the subscriber side might give\\nmore flexibility for the user.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 10:50:35 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 1, 2022 at 4:12 PM kuroda.hayato@fujitsu.com\\n<kuroda.hayato@fujitsu.com> wrote:\\n>\\n> Hi Vignesh,\\n>\\n> > In logical replication, currently Walsender sends the data that is\\n> > generated locally and the data that are replicated from other\\n> > instances. This results in infinite recursion in circular logical\\n> > replication setup.\\n>\\n> Thank you for good explanation. I understand that this fix can be used\\n> for a bidirectional replication.\\n>\\n> > Here there are two problems for the user: a) incremental\\n> > synchronization of table sending both local data and replicated data\\n> > by walsender b) Table synchronization of table using copy command\\n> > sending both local data and replicated data\\n>\\n> So you wanted to solve these two problem and currently focused on\\n> the first one, right? We can check one by one.\\n>\\n> > For the first problem \\\"Incremental synchronization of table by\\n> > Walsender\\\" can be solved by:\\n> > Currently the locally generated data does not have replication origin\\n> > associated and the data that has originated from another instance will\\n> > have a replication origin associated. We could use this information to\\n> > differentiate locally generated data and replicated data and send only\\n> > the locally generated data. This \\\"only_local\\\" could be provided as an\\n> > option while subscription is created:\\n> > ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=5433'\\n> > PUBLICATION pub1 with (only_local = on);\\n>\\n> Sounds good, but I cannot distinguish whether the assumption will keep.\\n>\\n> I played with your patch, but it could not be applied to current master.\\n> I tested from bd74c40 and I confirmed infinite loop was not appeared.\\n>\\n> local_only could not be set from ALTER SUBSCRIPTION command.\\n> Is it expected?\\n\\nI felt changing only_local option might be useful for the user while\\nmodifying the subscription like setting it with a different set of\\npublications. Changes for this are included in the v2 patch attached\\nat [1].\\n[2] - https://www.postgresql.org/message-id/CALDaNm0WSo5369pr2eN1obTGBeiJU9cQdF6Ju1sC4hMQNy5BfQ%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 11:14:31 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Dear Peter,\\r\\n\\r\\n> > So, why does the patch use syntax option 1?\\r\\n\\r\\nIMU it might be useful for the following case.\\r\\n\\r\\nAssuming that multi-master configuration with node1, node2.\\r\\nNode1 has a publication pub1 and a subscription sub2, node2 has pub2 and sub1.\\r\\n\\r\\nFrom that situation, please consider that new node node3 is added\\r\\nthat subscribe some changes from node2.\\r\\n\\r\\nIf the feature is introduced as option1, new publication must be defined in node2.\\r\\nIf that is introduced as option2, however, maybe pub2 can be reused.\\r\\ni.e. multiple declaration of publications can be avoided.\\r\\n\\r\\nBest Regards,\\r\\nHayato Kuroda\\r\\nFUJITSU LIMITED\\r\\n\\r\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 06:12:50 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Hi Vignesh, I also have not looked at the patch yet, but I have what\\n> > seems like a very fundamental (and possibly dumb) question...\\n> >\\n> > Basically, I do not understand the choice of syntax for setting things up.\\n> >\\n> > IMO that \\\"only-local\\\" option sounds very similar to the other\\n> > PUBLICATION (\\\"publish\\\") options which decide the kinds of things that\\n> > will be published. So it feels more natural for me to think of the\\n> > publisher as being the one to decide what will be published.\\n> >\\n> > e.g.\\n> >\\n> > option 1:\\n> > CREATE PUBLICATION p1 FOR TABLE t1;\\n> > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n> >\\n> > option 2:\\n> > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\\n> > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n> >\\n> > ~~\\n> >\\n> > IIUC the patch is using option 1. My first impression was it feels\\n> > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\\n> > publish.\\n> >\\n> > So, why does the patch use syntax option 1?\\n>\\n> I felt the advantage with keeping it at the subscription side is that,\\n> the subscriber from one node can subscribe with only_local option on\\n> and a different subscriber from a different node can subscribe with\\n> only_local option as off. This might not be possible with having the\\n> option at publisher side. Having it at the subscriber side might give\\n> more flexibility for the user.\\n>\\n\\nOK. Option 2 needs two publications for that scenario. IMO it's more\\nintuitive this way, but maybe you wanted to avoid the extra\\npublications?\\n\\nnode0:\\nCREATE PUBLICATION p1 FOR TABLE t1;\\nCREATE PUBLICATION p1_local FOR TABLE t1 WITH (publish = 'only_local');\\n\\nnode1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1_local;\\nnode2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1;\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 17:14:48 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 5:12 PM kuroda.hayato@fujitsu.com\\n<kuroda.hayato@fujitsu.com> wrote:\\n>\\n> Dear Peter,\\n>\\n> > > So, why does the patch use syntax option 1?\\n>\\n> IMU it might be useful for the following case.\\n>\\n> Assuming that multi-master configuration with node1, node2.\\n> Node1 has a publication pub1 and a subscription sub2, node2 has pub2 and sub1.\\n>\\n> From that situation, please consider that new node node3 is added\\n> that subscribe some changes from node2.\\n>\\n> If the feature is introduced as option1, new publication must be defined in node2.\\n> If that is introduced as option2, however, maybe pub2 can be reused.\\n> i.e. multiple declaration of publications can be avoided.\\n>\\n\\nYes. Thanks for the example. I had the same observation in my last post [1]\\n\\n------\\n[1] https://www.postgresql.org/message-id/CAHut%2BPtRxiQR_4UFLNThg-NNRV447FvwtcR-BvqMzjyMJXKwfw%40mail.gmail.com\\n\\nKind Regards,\\nPeter Smith\\nFujitsu Australia.\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 17:20:03 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 11:45 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > >\\n> > > Hi Vignesh, I also have not looked at the patch yet, but I have what\\n> > > seems like a very fundamental (and possibly dumb) question...\\n> > >\\n> > > Basically, I do not understand the choice of syntax for setting things up.\\n> > >\\n> > > IMO that \\\"only-local\\\" option sounds very similar to the other\\n> > > PUBLICATION (\\\"publish\\\") options which decide the kinds of things that\\n> > > will be published. So it feels more natural for me to think of the\\n> > > publisher as being the one to decide what will be published.\\n> > >\\n> > > e.g.\\n> > >\\n> > > option 1:\\n> > > CREATE PUBLICATION p1 FOR TABLE t1;\\n> > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n> > >\\n> > > option 2:\\n> > > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\\n> > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n> > >\\n> > > ~~\\n> > >\\n> > > IIUC the patch is using option 1. My first impression was it feels\\n> > > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\\n> > > publish.\\n> > >\\n> > > So, why does the patch use syntax option 1?\\n> >\\n> > I felt the advantage with keeping it at the subscription side is that,\\n> > the subscriber from one node can subscribe with only_local option on\\n> > and a different subscriber from a different node can subscribe with\\n> > only_local option as off. This might not be possible with having the\\n> > option at publisher side. Having it at the subscriber side might give\\n> > more flexibility for the user.\\n> >\\n>\\n> OK. Option 2 needs two publications for that scenario. IMO it's more\\n> intuitive this way, but maybe you wanted to avoid the extra\\n> publications?\\n\\nYes, I wanted to avoid the extra publication creation that you pointed\\nout. Option 1 can handle this scenario without creating the extra\\npublications:\\nnode0: CREATE PUBLICATION p1 FOR TABLE t1;\\nnode1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = on);\\nnode2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = off);\\n\\nI'm ok with both the approaches, now that this scenario can be handled\\nby using both the options. i.e providing only_local option as an\\noption while creating publication or providing only_local option as an\\noption while creating subscription as Peter has pointed out at [1].\\noption 1:\\nCREATE PUBLICATION p1 FOR TABLE t1;\\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n\\noption 2:\\nCREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\\nCREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n\\nShall we get a few opinions on this and take it in that direction?\\n\\n[1] - https://www.postgresql.org/message-id/CAHut%2BPsAWaETh9VMymbBfMrqiE1KuqMq%2BwpBg0s7eMzwLATr%2Bw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 12:47:43 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Dear Vignesh,\\r\\n\\r\\n> I felt changing only_local option might be useful for the user while\\r\\n> modifying the subscription like setting it with a different set of\\r\\n> publications. Changes for this are included in the v2 patch attached\\r\\n> at [1].\\r\\n\\r\\n+1, thanks. I'll post if I notice something to say.\\r\\n\\r\\n> Shall we get a few opinions on this and take it in that direction?\\r\\n\\r\\nI prefer subscriber-option, but I also think both are reasonable.\\r\\n+1 about asking other reviewers.\\r\\n\\r\\nBest Regards,\\r\\nHayato Kuroda\\r\\nFUJITSU LIMITED\\r\\n\\r\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 07:37:46 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 10:15 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > I haven't yet gone through the patch, but I have a question about the\\n> > > idea. Suppose I want to set up a logical replication like,\\n> > > node1->node2->node3->node1. So how would I create the subscriber at\\n> > > node1? only_local=on or off?. I mean on node1, I want the changes\\n> > > from node3 which are generated on node3 or which are replicated from\\n> > > node2 but I do not want changes that are replicated from node1 itself?\\n> > > So if I set only_local=on then node1 will not get the changes\\n> > > replicated from node2, is that right? and If I set only_local=off then\\n> > > it will create the infinite loop again? So how are we protecting\\n> > > against this case?\\n> > >\\n> >\\n> > In the above topology if you want local changes from both node3 and\\n> > node2 then I think the way to get that would be you have to create two\\n> > subscriptions on node1. The first one points to node2 (with\\n> > only_local=off) and the second one points to node3 (with only_local\\n> > =off).\\n> >\\n>\\n> Sorry, I intend to say 'only_local=on' at both places in my previous email.\\n\\nHmm okay, so for this topology we will have to connect node1 directly\\nto node2 as well as to node3 but can not cascade the changes. I was\\nwondering can it be done without using the extra connection between\\nnode2 to node1? I mean instead of making this a boolean flag that\\nwhether we want local change or remote change, can't we control the\\nchanges based on the origin id? Such that node1 will get the local\\nchanges of node3 but with using the same subscription it will get\\nchanges from node3 which are originated from node2 but it will not\\nreceive the changes which are originated from node1.\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 13:10:44 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 6:17 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 11:45 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > > >\\n> > > > Hi Vignesh, I also have not looked at the patch yet, but I have what\\n> > > > seems like a very fundamental (and possibly dumb) question...\\n> > > >\\n> > > > Basically, I do not understand the choice of syntax for setting things up.\\n> > > >\\n> > > > IMO that \\\"only-local\\\" option sounds very similar to the other\\n> > > > PUBLICATION (\\\"publish\\\") options which decide the kinds of things that\\n> > > > will be published. So it feels more natural for me to think of the\\n> > > > publisher as being the one to decide what will be published.\\n> > > >\\n> > > > e.g.\\n> > > >\\n> > > > option 1:\\n> > > > CREATE PUBLICATION p1 FOR TABLE t1;\\n> > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n> > > >\\n> > > > option 2:\\n> > > > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\\n> > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n> > > >\\n> > > > ~~\\n> > > >\\n> > > > IIUC the patch is using option 1. My first impression was it feels\\n> > > > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\\n> > > > publish.\\n> > > >\\n> > > > So, why does the patch use syntax option 1?\\n> > >\\n> > > I felt the advantage with keeping it at the subscription side is that,\\n> > > the subscriber from one node can subscribe with only_local option on\\n> > > and a different subscriber from a different node can subscribe with\\n> > > only_local option as off. This might not be possible with having the\\n> > > option at publisher side. Having it at the subscriber side might give\\n> > > more flexibility for the user.\\n> > >\\n> >\\n> > OK. Option 2 needs two publications for that scenario. IMO it's more\\n> > intuitive this way, but maybe you wanted to avoid the extra\\n> > publications?\\n>\\n> Yes, I wanted to avoid the extra publication creation that you pointed\\n> out. Option 1 can handle this scenario without creating the extra\\n> publications:\\n> node0: CREATE PUBLICATION p1 FOR TABLE t1;\\n> node1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = on);\\n> node2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = off);\\n>\\n> I'm ok with both the approaches, now that this scenario can be handled\\n> by using both the options. i.e providing only_local option as an\\n> option while creating publication or providing only_local option as an\\n> option while creating subscription as Peter has pointed out at [1].\\n> option 1:\\n> CREATE PUBLICATION p1 FOR TABLE t1;\\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n>\\n> option 2:\\n> CREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n>\\n> Shall we get a few opinions on this and take it in that direction?\\n>\\n> [1] - https://www.postgresql.org/message-id/CAHut%2BPsAWaETh9VMymbBfMrqiE1KuqMq%2BwpBg0s7eMzwLATr%2Bw%40mail.gmail.com\\n>\\n> Regards,\\n> Vignesh\\n\\nBTW here is a counter-example to your scenario from earlier.\\n\\nLet's say I have a publication p1 and p2 and want to subscribe to p1\\nwith only_local=true, and p2 with only_local = false;\\n\\nUsing the current OPtion 1 syntax you cannot do this with a single\\nsubscription because the option is tied to the subscription.\\nBut using syntax Option 2 you may be able to do it.\\n\\nOption 1:\\nCREATE PUBLICATION p1 FOR TABLE t1;\\nCREATE PUBLICATION p2 FOR TABLE t2;\\nCREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 WITH (local_only = true);\\nCREATE SUBSCRIPTION s2 ... FOR PUBLICATION p1 WITH (local_only = false);\\n\\nOption 2:\\nCREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'local_only');\\nCREATE PUBLICATION p2 FOR TABLE t2;\\nCREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1, p2;\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 19:14:39 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 12:48 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 11:45 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > > >\\n> > > > Hi Vignesh, I also have not looked at the patch yet, but I have what\\n> > > > seems like a very fundamental (and possibly dumb) question...\\n> > > >\\n> > > > Basically, I do not understand the choice of syntax for setting things up.\\n> > > >\\n> > > > IMO that \\\"only-local\\\" option sounds very similar to the other\\n> > > > PUBLICATION (\\\"publish\\\") options which decide the kinds of things that\\n> > > > will be published. So it feels more natural for me to think of the\\n> > > > publisher as being the one to decide what will be published.\\n> > > >\\n> > > > e.g.\\n> > > >\\n> > > > option 1:\\n> > > > CREATE PUBLICATION p1 FOR TABLE t1;\\n> > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n> > > >\\n> > > > option 2:\\n> > > > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\\n> > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n> > > >\\n> > > > ~~\\n> > > >\\n> > > > IIUC the patch is using option 1. My first impression was it feels\\n> > > > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\\n> > > > publish.\\n> > > >\\n> > > > So, why does the patch use syntax option 1?\\n> > >\\n> > > I felt the advantage with keeping it at the subscription side is that,\\n> > > the subscriber from one node can subscribe with only_local option on\\n> > > and a different subscriber from a different node can subscribe with\\n> > > only_local option as off. This might not be possible with having the\\n> > > option at publisher side. Having it at the subscriber side might give\\n> > > more flexibility for the user.\\n> > >\\n> >\\n> > OK. Option 2 needs two publications for that scenario. IMO it's more\\n> > intuitive this way, but maybe you wanted to avoid the extra\\n> > publications?\\n>\\n> Yes, I wanted to avoid the extra publication creation that you pointed\\n> out. Option 1 can handle this scenario without creating the extra\\n> publications:\\n> node0: CREATE PUBLICATION p1 FOR TABLE t1;\\n> node1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = on);\\n> node2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = off);\\n>\\n> I'm ok with both the approaches, now that this scenario can be handled\\n> by using both the options. i.e providing only_local option as an\\n> option while creating publication or providing only_local option as an\\n> option while creating subscription as Peter has pointed out at [1].\\n> option 1:\\n> CREATE PUBLICATION p1 FOR TABLE t1;\\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n>\\n> option 2:\\n> CREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\\n> CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n>\\n> Shall we get a few opinions on this and take it in that direction?\\n>\\n\\nAllowing multiple publications will lead to a lot of duplicate data in\\ncatalogs like pg_publication_rel, so, I don't see how option-2 can be\\nbeneficial?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 14:24:43 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi Vignesh,\\n\\nHere are some review comments for patch v2.\\n\\n======\\n\\n1. Question about syntax\\n\\nI already posted some questions about why the syntax is on the CREATE\\nSUBSCRCRIBER side.\\nIMO \\\"local_only\\\" is a publisher option, so it seemed more natural to\\nme for it to be specified as a \\\"publish\\\" option.\\n\\nRef [1] my original question + suggestion for Option 2\\nRef [2] some other examples of subscribing to multiple-publishers\\n\\nAnyway, +1 to see what other people think.\\n\\n~~~\\n\\n2. ALTER\\n\\n(related also to the question about syntax)\\n\\nIf subscribing to multiple publications then ALTER is going to change\\nthe 'local_only' for all of them, which might not be what you want\\n(??)\\n\\n~~~\\n\\n3. subscription_parameter\\n\\n(related also to the question about syntax)\\n\\nCREATE SUBSCRIPTION subscription_name\\n CONNECTION 'conninfo'\\n PUBLICATION publication_name [, ...]\\n [ WITH ( subscription_parameter [= value] [, ... ] ) ]\\n\\n~\\n\\nThat WITH is for *subscription* options, not the publication options.\\n\\nSo IMO 'local_only' intuitively seems like \\\"local\\\" means local where\\nthe subscriber is.\\n\\nSo, if the Option 1 syntax is chosen (see comment #1) then I think the\\noption name maybe should change to be something more like\\n'publish_local_only' or something similar to be more clear what local\\nactually means.\\n\\n~~~\\n\\n4. contrib/test_decoding/test_decoding.c\\n\\n@@ -484,6 +487,16 @@ pg_decode_filter(LogicalDecodingContext *ctx,\\n return false;\\n }\\n\\n+static bool\\n+pg_decode_filter_remotedata(LogicalDecodingContext *ctx,\\n+ RepOriginId origin_id)\\n+{\\n+ TestDecodingData *data = ctx->output_plugin_private;\\n+\\n+ if (data->only_local && origin_id != InvalidRepOriginId)\\n+ return true;\\n+ return false;\\n+}\\n\\n4a. Maybe needs function comment.\\n\\n4b. Missing blank line following this function\\n\\n~~~\\n\\n5. General - please check all of the patch.\\n\\nThere seems inconsistency with the member names, local variable names,\\nparameter names etc. There are all variations of:\\n\\n- only_local\\n- onlylocaldata\\n- onlylocal_data\\n- etc\\n\\nPlease try using the same name everywhere for everything if possible.\\n\\n~~~\\n\\n6. src/backend/replication/logical/decode.c - FilterRemoteOriginData\\n\\n@@ -585,7 +594,8 @@ logicalmsg_decode(LogicalDecodingContext *ctx,\\nXLogRecordBuffer *buf)\\n message = (xl_logical_message *) XLogRecGetData(r);\\n\\n if (message->dbId != ctx->slot->data.database ||\\n- FilterByOrigin(ctx, origin_id))\\n+ FilterByOrigin(ctx, origin_id) ||\\n+ FilterRemoteOriginData(ctx, origin_id))\\n return;\\n\\nI noticed that every call to FilterRemoteOriginData has an associated\\npreceding call to FilterByOrigin. It might be worth just combining the\\nlogic into FilterByOrigin. Then none of that calling code (9 x places)\\nwould need to change at all.\\n\\n~~~\\n\\n7. src/backend/replication/logical/logical.c - CreateInitDecodingContext\\n\\n@@ -451,6 +453,8 @@ CreateInitDecodingContext(const char *plugin,\\n */\\n ctx->twophase &= slot->data.two_phase;\\n\\n+ ctx->onlylocal_data &= slot->data.onlylocal_data;\\n\\nThe equivalent 'twophase' option had a big comment. Probably this new\\noption should also have a similar comment?\\n\\n~~~\\n\\n8. src/backend/replication/logical/logical.c - filter_remotedata_cb_wrapper\\n\\n+bool\\n+filter_remotedata_cb_wrapper(LogicalDecodingContext *ctx,\\n+ RepOriginId origin_id)\\n+{\\n+ LogicalErrorCallbackState state;\\n+ ErrorContextCallback errcallback;\\n+ bool ret;\\n+\\n+ Assert(!ctx->fast_forward);\\n+\\n+ /* Push callback + info on the error context stack */\\n+ state.ctx = ctx;\\n+ state.callback_name = \\\"filter_remoteorigin\\\";\\n\\nThere is no consistency between the function and the name:\\n\\n\\\"filter_remoteorigin\\\" versus filter_remotedata_cb.\\n\\nA similar inconsistency for this is elsewhere. See review comment #9\\n\\n~~~\\n\\n9. src/backend/replication/pgoutput/pgoutput.c\\n\\n@@ -215,6 +217,7 @@ _PG_output_plugin_init(OutputPluginCallbacks *cb)\\n cb->commit_prepared_cb = pgoutput_commit_prepared_txn;\\n cb->rollback_prepared_cb = pgoutput_rollback_prepared_txn;\\n cb->filter_by_origin_cb = pgoutput_origin_filter;\\n+ cb->filter_remotedata_cb = pgoutput_remoteorigin_filter;\\n\\nInconsistent names for the member and function.\\n\\nfilter_remotedata_cb VS pgoutput_remoteorigin_filter.\\n\\n~~~\\n\\n10. src/backend/replication/pgoutput/pgoutput.c\\n\\n@@ -1450,6 +1465,16 @@ pgoutput_origin_filter(LogicalDecodingContext *ctx,\\n return false;\\n }\\n\\n+static bool\\n+pgoutput_remoteorigin_filter(LogicalDecodingContext *ctx,\\n+ RepOriginId origin_id)\\n+{\\n+ PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\\n+\\n+ if (data->onlylocal_data && origin_id != InvalidRepOriginId)\\n+ return true;\\n+ return false;\\n+}\\n /*\\n * Shutdown the output plugin.\\n *\\n\\n10a. Add a function comment.\\n\\n10b. Missing blank line after the function\\n\\n~~~\\n\\n11. src/backend/replication/slotfuncs.c - pg_create_logical_replication_slot\\n\\n@@ -171,6 +174,7 @@ pg_create_logical_replication_slot(PG_FUNCTION_ARGS)\\n Name plugin = PG_GETARG_NAME(1);\\n bool temporary = PG_GETARG_BOOL(2);\\n bool two_phase = PG_GETARG_BOOL(3);\\n+ bool onlylocal_data = PG_GETARG_BOOL(4);\\n Datum result;\\n TupleDesc tupdesc;\\n HeapTuple tuple;\\n\\n\\nWon't there be some PG Docs needing to be updated now there is another\\nparameter?\\n\\n~~~\\n\\n12. src/include/catalog/pg_proc.dat - pg_get_replication_slots\\n\\nI did not see any update for pg_get_replication_slots, but you added\\nthe 4th parameter elsewhere. Is something missing here?\\n\\n~~~\\n\\n13. src/include/replication/logical.h\\n\\n@@ -99,6 +99,8 @@ typedef struct LogicalDecodingContext\\n */\\n bool twophase_opt_given;\\n\\n+ bool onlylocal_data;\\n+\\n\\nI think the new member needs some comment.\\n\\n~~~\\n\\n14. src/include/replication/walreceiver.h\\n\\n@@ -183,6 +183,7 @@ typedef struct\\n bool streaming; /* Streaming of large transactions */\\n bool twophase; /* Streaming of two-phase transactions at\\n * prepare time */\\n+ bool onlylocal_data;\\n } logical;\\n } proto;\\n } WalRcvStreamOptions;\\n\\nI think the new member needs some comment.\\n\\n~~~\\n\\n15. src/test/regress/sql/subscription.sql\\n\\nALTER SUBSCRIPTION test missing?\\n\\n------\\n[1] https://www.postgresql.org/message-id/CAHut%2BPsAWaETh9VMymbBfMrqiE1KuqMq%2BwpBg0s7eMzwLATr%2Bw%40mail.gmail.com\\n[2] https://www.postgresql.org/message-id/CAHut%2BPvQonJd5epJBM0Yfh1499mL9kTL9a%3DGrMhvnL6Ok05zqw%40mail.gmail.com\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 19:57:47 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 1:11 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 10:15 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > > I haven't yet gone through the patch, but I have a question about the\\n> > > > idea. Suppose I want to set up a logical replication like,\\n> > > > node1->node2->node3->node1. So how would I create the subscriber at\\n> > > > node1? only_local=on or off?. I mean on node1, I want the changes\\n> > > > from node3 which are generated on node3 or which are replicated from\\n> > > > node2 but I do not want changes that are replicated from node1 itself?\\n> > > > So if I set only_local=on then node1 will not get the changes\\n> > > > replicated from node2, is that right? and If I set only_local=off then\\n> > > > it will create the infinite loop again? So how are we protecting\\n> > > > against this case?\\n> > > >\\n> > >\\n> > > In the above topology if you want local changes from both node3 and\\n> > > node2 then I think the way to get that would be you have to create two\\n> > > subscriptions on node1. The first one points to node2 (with\\n> > > only_local=off) and the second one points to node3 (with only_local\\n> > > =off).\\n> > >\\n> >\\n> > Sorry, I intend to say 'only_local=on' at both places in my previous email.\\n>\\n> Hmm okay, so for this topology we will have to connect node1 directly\\n> to node2 as well as to node3 but can not cascade the changes. I was\\n> wondering can it be done without using the extra connection between\\n> node2 to node1? I mean instead of making this a boolean flag that\\n> whether we want local change or remote change, can't we control the\\n> changes based on the origin id? Such that node1 will get the local\\n> changes of node3 but with using the same subscription it will get\\n> changes from node3 which are originated from node2 but it will not\\n> receive the changes which are originated from node1.\\n>\\n\\nGood point. I think we can provide that as an additional option to\\ngive more flexibility but we won't be able to use it for initial sync\\nwhere we can't differentiate between data from different origins.\\nAlso, I think as origins are internally generated, we may need some\\nbetter way to expose it to users so that they can specify it as an\\noption. Isn't it better to provide first some simple way like a\\nboolean option so that users have some way to replicate the same table\\ndata among different nodes without causing an infinite loop and then\\nextend it as you are suggesting or may be in some other ways as well?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 15:01:15 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 3:01 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 1:11 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >\\n> > On Mon, Mar 7, 2022 at 10:15 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > > > I haven't yet gone through the patch, but I have a question about the\\n> > > > > idea. Suppose I want to set up a logical replication like,\\n> > > > > node1->node2->node3->node1. So how would I create the subscriber at\\n> > > > > node1? only_local=on or off?. I mean on node1, I want the changes\\n> > > > > from node3 which are generated on node3 or which are replicated from\\n> > > > > node2 but I do not want changes that are replicated from node1 itself?\\n> > > > > So if I set only_local=on then node1 will not get the changes\\n> > > > > replicated from node2, is that right? and If I set only_local=off then\\n> > > > > it will create the infinite loop again? So how are we protecting\\n> > > > > against this case?\\n> > > > >\\n> > > >\\n> > > > In the above topology if you want local changes from both node3 and\\n> > > > node2 then I think the way to get that would be you have to create two\\n> > > > subscriptions on node1. The first one points to node2 (with\\n> > > > only_local=off) and the second one points to node3 (with only_local\\n> > > > =off).\\n> > > >\\n> > >\\n> > > Sorry, I intend to say 'only_local=on' at both places in my previous email.\\n> >\\n> > Hmm okay, so for this topology we will have to connect node1 directly\\n> > to node2 as well as to node3 but can not cascade the changes. I was\\n> > wondering can it be done without using the extra connection between\\n> > node2 to node1? I mean instead of making this a boolean flag that\\n> > whether we want local change or remote change, can't we control the\\n> > changes based on the origin id? Such that node1 will get the local\\n> > changes of node3 but with using the same subscription it will get\\n> > changes from node3 which are originated from node2 but it will not\\n> > receive the changes which are originated from node1.\\n> >\\n>\\n> Good point. I think we can provide that as an additional option to\\n> give more flexibility but we won't be able to use it for initial sync\\n> where we can't differentiate between data from different origins.\\n> Also, I think as origins are internally generated, we may need some\\n> better way to expose it to users so that they can specify it as an\\n> option. Isn't it better to provide first some simple way like a\\n> boolean option so that users have some way to replicate the same table\\n> data among different nodes without causing an infinite loop and then\\n> extend it as you are suggesting or may be in some other ways as well?\\n\\nYeah, that makes sense that first we provide some simple mechanism to\\nenable it and we can extend it later.\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 15:25:18 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 1:45 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 6:17 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Mon, Mar 7, 2022 at 11:45 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > >\\n> > > On Mon, Mar 7, 2022 at 4:20 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > > >\\n> > > > On Mon, Mar 7, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > > > >\\n> > > > > Hi Vignesh, I also have not looked at the patch yet, but I have what\\n> > > > > seems like a very fundamental (and possibly dumb) question...\\n> > > > >\\n> > > > > Basically, I do not understand the choice of syntax for setting things up.\\n> > > > >\\n> > > > > IMO that \\\"only-local\\\" option sounds very similar to the other\\n> > > > > PUBLICATION (\\\"publish\\\") options which decide the kinds of things that\\n> > > > > will be published. So it feels more natural for me to think of the\\n> > > > > publisher as being the one to decide what will be published.\\n> > > > >\\n> > > > > e.g.\\n> > > > >\\n> > > > > option 1:\\n> > > > > CREATE PUBLICATION p1 FOR TABLE t1;\\n> > > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n> > > > >\\n> > > > > option 2:\\n> > > > > CREATE PUBLICATION p1 FOR TABLE t1 WEHRE (publish = 'only_local');\\n> > > > > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n> > > > >\\n> > > > > ~~\\n> > > > >\\n> > > > > IIUC the patch is using option 1. My first impression was it feels\\n> > > > > back-to-front for the SUBSCRIPTION telling the PUBLICATION what to\\n> > > > > publish.\\n> > > > >\\n> > > > > So, why does the patch use syntax option 1?\\n> > > >\\n> > > > I felt the advantage with keeping it at the subscription side is that,\\n> > > > the subscriber from one node can subscribe with only_local option on\\n> > > > and a different subscriber from a different node can subscribe with\\n> > > > only_local option as off. This might not be possible with having the\\n> > > > option at publisher side. Having it at the subscriber side might give\\n> > > > more flexibility for the user.\\n> > > >\\n> > >\\n> > > OK. Option 2 needs two publications for that scenario. IMO it's more\\n> > > intuitive this way, but maybe you wanted to avoid the extra\\n> > > publications?\\n> >\\n> > Yes, I wanted to avoid the extra publication creation that you pointed\\n> > out. Option 1 can handle this scenario without creating the extra\\n> > publications:\\n> > node0: CREATE PUBLICATION p1 FOR TABLE t1;\\n> > node1: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = on);\\n> > node2: CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 with (only_local = off);\\n> >\\n> > I'm ok with both the approaches, now that this scenario can be handled\\n> > by using both the options. i.e providing only_local option as an\\n> > option while creating publication or providing only_local option as an\\n> > option while creating subscription as Peter has pointed out at [1].\\n> > option 1:\\n> > CREATE PUBLICATION p1 FOR TABLE t1;\\n> > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1 WITH (only_local = true);\\n> >\\n> > option 2:\\n> > CREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'only_local');\\n> > CREATE SUBSCRITION s1 ... FOR PUBLICATION p1;\\n> >\\n> > Shall we get a few opinions on this and take it in that direction?\\n> >\\n> > [1] - https://www.postgresql.org/message-id/CAHut%2BPsAWaETh9VMymbBfMrqiE1KuqMq%2BwpBg0s7eMzwLATr%2Bw%40mail.gmail.com\\n> >\\n> > Regards,\\n> > Vignesh\\n>\\n> BTW here is a counter-example to your scenario from earlier.\\n>\\n> Let's say I have a publication p1 and p2 and want to subscribe to p1\\n> with only_local=true, and p2 with only_local = false;\\n>\\n> Using the current OPtion 1 syntax you cannot do this with a single\\n> subscription because the option is tied to the subscription.\\n> But using syntax Option 2 you may be able to do it.\\n>\\n> Option 1:\\n> CREATE PUBLICATION p1 FOR TABLE t1;\\n> CREATE PUBLICATION p2 FOR TABLE t2;\\n> CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1 WITH (local_only = true);\\n> CREATE SUBSCRIPTION s2 ... FOR PUBLICATION p1 WITH (local_only = false);\\n>\\n> Option 2:\\n> CREATE PUBLICATION p1 FOR TABLE t1 WITH (publish = 'local_only');\\n> CREATE PUBLICATION p2 FOR TABLE t2;\\n> CREATE SUBSCRIPTION s1 ... FOR PUBLICATION p1, p2;\\n\\nI felt having multiple publications will create duplicate entries in\\nthe system table, Amit also has pointed this at [1]. Also enhancing\\nthis approach to support filtering based on replication origin which\\nis suggested by dilip at [2] is also on the client side and also the\\ninitial check to handle the copy_data specified by Amit at [3] will be\\ndone by the client side. Based on the above I feel the existing\\napproach is better. I might be missing something here.\\n\\n[1] - https://www.postgresql.org/message-id/CAA4eK1LgCVv8u-fOsMPbGC96sWXhT3EKOBAeFW3g84otjStztw%40mail.gmail.com\\n[2] - https://www.postgresql.org/message-id/CAFiTN-tKbjHDjAFNnqRoR8u1B%2Bfs0wunGz%3D3wp0iU-sUaxZJTQ%40mail.gmail.com\\n[3] - https://www.postgresql.org/message-id/CAA4eK1%2Bco2cd8a6okgUD_pcFEHcc7mVc0k_RE2%3D6ahyv3WPRMg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 16:23:17 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi Vignesh,\\nI agree with Peter's comment that the changes to\\nFilterRemoteOriginData() should be part of FilterByOrigin()\\n\\nFurther, I wonder why \\\"onlylocal_data\\\" is a replication slot's\\nproperty. A replication slot tracks the progress of replication and it\\nmay be used by different receivers with different options. I could\\nstart one receiver which wants only local data, say using\\n\\\"pg_logical_slot_get_changes\\\" and later start another receiver which\\nfetches all the data starting from where the first receiver left. This\\noption prevents such flexibility.\\n\\nAs discussed earlier in the thread, local_only can be property of\\npublication or subscription, depending upon the use case, but I can't\\nsee any reason that it should be tied to a replication slot.\\n\\nI have a similar question for \\\"two_phase\\\" but the ship has sailed and\\nprobably it makes some sense there which I don't know.\\n\\nAs for publication vs subscription, I think both are useful cases.\\n1. It will be a publication's property, if we want the node to not\\npublish any data that it receives from other nodes for a given set of\\ntables.\\n2. It will be the subscription's property, if we want the subscription\\nto decide whether it wants to fetch the data changed on only upstream\\nor other nodes as well.\\n\\nMaybe we want to add it to both and use the stricter of both. If a\\npublication has local_only, it publishes only local changes. If the\\nsubscription has local only then WAL sender sends only local changes.\\n\\n--\\nBest Wishes,\\nAshutosh\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 17:00:56 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 5:01 PM Ashutosh Bapat\\n<ashutosh.bapat.oss@gmail.com> wrote:\\n>\\n> Hi Vignesh,\\n> I agree with Peter's comment that the changes to\\n> FilterRemoteOriginData() should be part of FilterByOrigin()\\n>\\n> Further, I wonder why \\\"onlylocal_data\\\" is a replication slot's\\n> property. A replication slot tracks the progress of replication and it\\n> may be used by different receivers with different options. I could\\n> start one receiver which wants only local data, say using\\n> \\\"pg_logical_slot_get_changes\\\" and later start another receiver which\\n> fetches all the data starting from where the first receiver left. This\\n> option prevents such flexibility.\\n>\\n> As discussed earlier in the thread, local_only can be property of\\n> publication or subscription, depending upon the use case, but I can't\\n> see any reason that it should be tied to a replication slot.\\n>\\n\\nI thought it should be similar to 'streaming' option of subscription\\nbut may be Vignesh has some other reason which makes it different.\\n\\n> I have a similar question for \\\"two_phase\\\" but the ship has sailed and\\n> probably it makes some sense there which I don't know.\\n>\\n\\ntwo_phase is different from some of the other subscription options\\nlike 'streaming' such that it can be enabled only at the time of slot\\nand subscription creation, we can't change/specify it via\\npg_logical_slot_get_changes. This is to avoid the case where we won't\\nknow at the time of the commit prepared whether the prepare for the\\ntransaction has already been sent. For the same reason, we need to\\nalso know the 'two_phase_at' information.\\n\\n> As for publication vs subscription, I think both are useful cases.\\n> 1. It will be a publication's property, if we want the node to not\\n> publish any data that it receives from other nodes for a given set of\\n> tables.\\n> 2. It will be the subscription's property, if we want the subscription\\n> to decide whether it wants to fetch the data changed on only upstream\\n> or other nodes as well.\\n>\\n\\nI think it could be useful to allow it via both publication and\\nsubscription but I guess it is better to provide it via one way\\ninitially just to keep things simple and give users some way to deal\\nwith such cases. I would prefer to allow it via subscription initially\\nfor the reasons specified by Vignesh in his previous email [1]. Now,\\nif we think those all are ignorable things and it is more important to\\nallow this option first by publication or we must allow it via both\\npublication and subscription then it makes sense to change it.\\n\\n\\n[1] - https://www.postgresql.org/message-id/CALDaNm3jkotRhKfCqu5CXOf36_yiiW_cYE5%3DbG%3Dj6N3gOWJkqw%40mail.gmail.com\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 17:50:59 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 5:51 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 5:01 PM Ashutosh Bapat\\n> <ashutosh.bapat.oss@gmail.com> wrote:\\n> >\\n> > Hi Vignesh,\\n> > I agree with Peter's comment that the changes to\\n> > FilterRemoteOriginData() should be part of FilterByOrigin()\\n> >\\n> > Further, I wonder why \\\"onlylocal_data\\\" is a replication slot's\\n> > property. A replication slot tracks the progress of replication and it\\n> > may be used by different receivers with different options. I could\\n> > start one receiver which wants only local data, say using\\n> > \\\"pg_logical_slot_get_changes\\\" and later start another receiver which\\n> > fetches all the data starting from where the first receiver left. This\\n> > option prevents such flexibility.\\n> >\\n> > As discussed earlier in the thread, local_only can be property of\\n> > publication or subscription, depending upon the use case, but I can't\\n> > see any reason that it should be tied to a replication slot.\\n> >\\n>\\n> I thought it should be similar to 'streaming' option of subscription\\n> but may be Vignesh has some other reason which makes it different.\\n\\nYes, this can be removed from the replication slot. It is my mistake\\nthat I have made while making the code similar to two-phase, I'm\\nworking on making the changes for this. I will fix and post an updated\\nversion for this.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 21:57:03 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"IIUC the new option may be implemented subscriber-side and/or\\npublisher-side and/or both, and the subscriber-side option may be\\n\\\"enhanced\\\" in future to prevent cycles. And probably there are more\\nfeatures I don't know about or that have not yet been thought of.\\n\\n~~\\n\\nEven if the plan is only to implement just one part now and then add\\nmore later, I think there still should be some consideration for what\\nyou expect all possible future options to look like, because that may\\naffect current implementation choices.\\n\\nThe point is:\\n\\n- we should take care so don't accidentally end up with an option that\\nturned out to be inconsistent looking on the subscriber-side /\\npublisher-side.\\n\\n- we should try to avoid accidentally painting ourselves into a corner\\n(e.g. stuck with a boolean option that cannot be enhanced later on)\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Mar 2022 16:00:45 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 8, 2022 at 10:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> IIUC the new option may be implemented subscriber-side and/or\\n> publisher-side and/or both, and the subscriber-side option may be\\n> \\\"enhanced\\\" in future to prevent cycles. And probably there are more\\n> features I don't know about or that have not yet been thought of.\\n>\\n> ~~\\n>\\n> Even if the plan is only to implement just one part now and then add\\n> more later, I think there still should be some consideration for what\\n> you expect all possible future options to look like, because that may\\n> affect current implementation choices.\\n>\\n> The point is:\\n>\\n> - we should take care so don't accidentally end up with an option that\\n> turned out to be inconsistent looking on the subscriber-side /\\n> publisher-side.\\n>\\n> - we should try to avoid accidentally painting ourselves into a corner\\n> (e.g. stuck with a boolean option that cannot be enhanced later on)\\n>\\n\\nAgreed. I think it is important to see we shouldn't do something which\\nis not extendable in future or paint us in the corner. But, as of now,\\nwith the current proposal, the main thing we should consider is\\nwhether exposing the boolean option is okay or shall we do something\\ndifferent so that we can extend it later to specific origin ids?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Mar 2022 10:51:08 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 8, 2022 at 4:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Mar 8, 2022 at 10:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > IIUC the new option may be implemented subscriber-side and/or\\n> > publisher-side and/or both, and the subscriber-side option may be\\n> > \\\"enhanced\\\" in future to prevent cycles. And probably there are more\\n> > features I don't know about or that have not yet been thought of.\\n> >\\n> > ~~\\n> >\\n> > Even if the plan is only to implement just one part now and then add\\n> > more later, I think there still should be some consideration for what\\n> > you expect all possible future options to look like, because that may\\n> > affect current implementation choices.\\n> >\\n> > The point is:\\n> >\\n> > - we should take care so don't accidentally end up with an option that\\n> > turned out to be inconsistent looking on the subscriber-side /\\n> > publisher-side.\\n> >\\n> > - we should try to avoid accidentally painting ourselves into a corner\\n> > (e.g. stuck with a boolean option that cannot be enhanced later on)\\n> >\\n>\\n> Agreed. I think it is important to see we shouldn't do something which\\n> is not extendable in future or paint us in the corner. But, as of now,\\n> with the current proposal, the main thing we should consider is\\n> whether exposing the boolean option is okay or shall we do something\\n> different so that we can extend it later to specific origin ids?\\n>\\n\\nI was wondering, assuming later there is an enhancement to detect and\\nprevent cycles using origin ids, then what really is the purpose of\\nthe option?\\n\\nI can imagine if a cycle happens should probably log some one-time\\nWARNING (just to bring it to the user's attention in case it was\\ncaused by some accidental misconfiguration) but apart from that why\\nwould this need to be an option at all - i.e. is there a use-case\\nwhere a user would NOT want to prevent a recursive error?\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Mar 2022 17:46:37 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 8, 2022 at 12:17 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Tue, Mar 8, 2022 at 4:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Tue, Mar 8, 2022 at 10:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > >\\n> > > IIUC the new option may be implemented subscriber-side and/or\\n> > > publisher-side and/or both, and the subscriber-side option may be\\n> > > \\\"enhanced\\\" in future to prevent cycles. And probably there are more\\n> > > features I don't know about or that have not yet been thought of.\\n> > >\\n> > > ~~\\n> > >\\n> > > Even if the plan is only to implement just one part now and then add\\n> > > more later, I think there still should be some consideration for what\\n> > > you expect all possible future options to look like, because that may\\n> > > affect current implementation choices.\\n> > >\\n> > > The point is:\\n> > >\\n> > > - we should take care so don't accidentally end up with an option that\\n> > > turned out to be inconsistent looking on the subscriber-side /\\n> > > publisher-side.\\n> > >\\n> > > - we should try to avoid accidentally painting ourselves into a corner\\n> > > (e.g. stuck with a boolean option that cannot be enhanced later on)\\n> > >\\n> >\\n> > Agreed. I think it is important to see we shouldn't do something which\\n> > is not extendable in future or paint us in the corner. But, as of now,\\n> > with the current proposal, the main thing we should consider is\\n> > whether exposing the boolean option is okay or shall we do something\\n> > different so that we can extend it later to specific origin ids?\\n> >\\n>\\n> I was wondering, assuming later there is an enhancement to detect and\\n> prevent cycles using origin ids, then what really is the purpose of\\n> the option?\\n>\\n\\nWe can't use origin ids during the initial sync as we can't\\ndifferentiate between data from local or remote node. Also, how will\\nwe distinguish between different origins? AFAIU, all the changes\\napplied on any particular node use the same origin.\\n\\n> I can imagine if a cycle happens should probably log some one-time\\n> WARNING (just to bring it to the user's attention in case it was\\n> caused by some accidental misconfiguration) but apart from that why\\n> would this need to be an option at all - i.e. is there a use-case\\n> where a user would NOT want to prevent a recursive error?\\n>\\n\\nI think there will be plenty of such setups where there won't be a\\nneed for any such option like where users won't need bi-directional\\nreplication for the same table.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Mar 2022 13:24:47 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 2:28 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Hi Vignesh,\\n>\\n> Here are some review comments for patch v2.\\n>\\n> ======\\n>\\n> 1. Question about syntax\\n>\\n> I already posted some questions about why the syntax is on the CREATE\\n> SUBSCRCRIBER side.\\n> IMO \\\"local_only\\\" is a publisher option, so it seemed more natural to\\n> me for it to be specified as a \\\"publish\\\" option.\\n>\\n> Ref [1] my original question + suggestion for Option 2\\n> Ref [2] some other examples of subscribing to multiple-publishers\\n>\\n> Anyway, +1 to see what other people think.\\n>\\n\\nI feel we can support it in the subscriber side first and then extend\\nit to the publisher side as being discussed in [1]. I have retained it\\nas it is.\\n\\n> ~~~\\n>\\n> 2. ALTER\\n>\\n> (related also to the question about syntax)\\n>\\n> If subscribing to multiple publications then ALTER is going to change\\n> the 'local_only' for all of them, which might not be what you want\\n> (??)\\n>\\n\\n I feel we can support it in the subscriber side first and then extend\\nit to the publisher side as being discussed in [1]. When it is\\nextended to the publisher side, it will get handled. I have retained\\nit as it is.\\n\\n> ~~~\\n>\\n> 3. subscription_parameter\\n>\\n> (related also to the question about syntax)\\n>\\n> CREATE SUBSCRIPTION subscription_name\\n> CONNECTION 'conninfo'\\n> PUBLICATION publication_name [, ...]\\n> [ WITH ( subscription_parameter [= value] [, ... ] ) ]\\n>\\n> ~\\n>\\n> That WITH is for *subscription* options, not the publication options.\\n>\\n> So IMO 'local_only' intuitively seems like \\\"local\\\" means local where\\n> the subscriber is.\\n>\\n> So, if the Option 1 syntax is chosen (see comment #1) then I think the\\n> option name maybe should change to be something more like\\n> 'publish_local_only' or something similar to be more clear what local\\n> actually means.\\n>\\n\\nChanged it to publish_local_only\\n\\n> ~~~\\n>\\n> 4. contrib/test_decoding/test_decoding.c\\n>\\n> @@ -484,6 +487,16 @@ pg_decode_filter(LogicalDecodingContext *ctx,\\n> return false;\\n> }\\n>\\n> +static bool\\n> +pg_decode_filter_remotedata(LogicalDecodingContext *ctx,\\n> + RepOriginId origin_id)\\n> +{\\n> + TestDecodingData *data = ctx->output_plugin_private;\\n> +\\n> + if (data->only_local && origin_id != InvalidRepOriginId)\\n> + return true;\\n> + return false;\\n> +}\\n>\\n> 4a. Maybe needs function comment.\\n\\nModified\\n\\n> 4b. Missing blank line following this function\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. General - please check all of the patch.\\n>\\n> There seems inconsistency with the member names, local variable names,\\n> parameter names etc. There are all variations of:\\n>\\n> - only_local\\n> - onlylocaldata\\n> - onlylocal_data\\n> - etc\\n>\\n> Please try using the same name everywhere for everything if possible.\\n>\\n\\nI have changed it to only_local wherever possible.\\n\\n> ~~~\\n>\\n> 6. src/backend/replication/logical/decode.c - FilterRemoteOriginData\\n>\\n> @@ -585,7 +594,8 @@ logicalmsg_decode(LogicalDecodingContext *ctx,\\n> XLogRecordBuffer *buf)\\n> message = (xl_logical_message *) XLogRecGetData(r);\\n>\\n> if (message->dbId != ctx->slot->data.database ||\\n> - FilterByOrigin(ctx, origin_id))\\n> + FilterByOrigin(ctx, origin_id) ||\\n> + FilterRemoteOriginData(ctx, origin_id))\\n> return;\\n>\\n> I noticed that every call to FilterRemoteOriginData has an associated\\n> preceding call to FilterByOrigin. It might be worth just combining the\\n> logic into FilterByOrigin. Then none of that calling code (9 x places)\\n> would need to change at all.\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. src/backend/replication/logical/logical.c - CreateInitDecodingContext\\n>\\n> @@ -451,6 +453,8 @@ CreateInitDecodingContext(const char *plugin,\\n> */\\n> ctx->twophase &= slot->data.two_phase;\\n>\\n> + ctx->onlylocal_data &= slot->data.onlylocal_data;\\n>\\n> The equivalent 'twophase' option had a big comment. Probably this new\\n> option should also have a similar comment?\\n\\nThese change is not required anymore, the comment no more applies. I\\nhave not made any change for this.\\n\\n> ~~~\\n>\\n> 8. src/backend/replication/logical/logical.c - filter_remotedata_cb_wrapper\\n>\\n> +bool\\n> +filter_remotedata_cb_wrapper(LogicalDecodingContext *ctx,\\n> + RepOriginId origin_id)\\n> +{\\n> + LogicalErrorCallbackState state;\\n> + ErrorContextCallback errcallback;\\n> + bool ret;\\n> +\\n> + Assert(!ctx->fast_forward);\\n> +\\n> + /* Push callback + info on the error context stack */\\n> + state.ctx = ctx;\\n> + state.callback_name = \\\"filter_remoteorigin\\\";\\n>\\n> There is no consistency between the function and the name:\\n>\\n> \\\"filter_remoteorigin\\\" versus filter_remotedata_cb.\\n>\\n> A similar inconsistency for this is elsewhere. See review comment #9\\n\\nModified it to filter_remote_origin_cb_wrapper and changed to\\n*_remotedata_* to *_remote_origin_*\\n\\n> ~~~\\n>\\n> 9. src/backend/replication/pgoutput/pgoutput.c\\n>\\n> @@ -215,6 +217,7 @@ _PG_output_plugin_init(OutputPluginCallbacks *cb)\\n> cb->commit_prepared_cb = pgoutput_commit_prepared_txn;\\n> cb->rollback_prepared_cb = pgoutput_rollback_prepared_txn;\\n> cb->filter_by_origin_cb = pgoutput_origin_filter;\\n> + cb->filter_remotedata_cb = pgoutput_remoteorigin_filter;\\n>\\n> Inconsistent names for the member and function.\\n>\\n> filter_remotedata_cb VS pgoutput_remoteorigin_filter.\\n\\nModified it to filter_remote_origin_cb and changed to *_remotedata_*\\nto *_remote_origin_*\\n\\n> ~~~\\n>\\n> 10. src/backend/replication/pgoutput/pgoutput.c\\n>\\n> @@ -1450,6 +1465,16 @@ pgoutput_origin_filter(LogicalDecodingContext *ctx,\\n> return false;\\n> }\\n>\\n> +static bool\\n> +pgoutput_remoteorigin_filter(LogicalDecodingContext *ctx,\\n> + RepOriginId origin_id)\\n> +{\\n> + PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\\n> +\\n> + if (data->onlylocal_data && origin_id != InvalidRepOriginId)\\n> + return true;\\n> + return false;\\n> +}\\n> /*\\n> * Shutdown the output plugin.\\n> *\\n>\\n> 10a. Add a function comment.\\n\\nModified\\n\\n> 10b. Missing blank line after the function\\n\\nModified\\n\\n> ~~~\\n>\\n> 11. src/backend/replication/slotfuncs.c - pg_create_logical_replication_slot\\n>\\n> @@ -171,6 +174,7 @@ pg_create_logical_replication_slot(PG_FUNCTION_ARGS)\\n> Name plugin = PG_GETARG_NAME(1);\\n> bool temporary = PG_GETARG_BOOL(2);\\n> bool two_phase = PG_GETARG_BOOL(3);\\n> + bool onlylocal_data = PG_GETARG_BOOL(4);\\n> Datum result;\\n> TupleDesc tupdesc;\\n> HeapTuple tuple;\\n>\\n>\\n> Won't there be some PG Docs needing to be updated now there is another\\n> parameter?\\n\\nThis change is not required anymore, the comment no longer applies. I\\nhave not made any changes for this.\\n\\n> ~~~\\n>\\n> 12. src/include/catalog/pg_proc.dat - pg_get_replication_slots\\n>\\n> I did not see any update for pg_get_replication_slots, but you added\\n> the 4th parameter elsewhere. Is something missing here?\\n\\nThis change is not required anymore, the comment no longer applies. I\\nhave not made any changes for this.\\n\\n> ~~~\\n>\\n> 13. src/include/replication/logical.h\\n>\\n> @@ -99,6 +99,8 @@ typedef struct LogicalDecodingContext\\n> */\\n> bool twophase_opt_given;\\n>\\n> + bool onlylocal_data;\\n> +\\n>\\n> I think the new member needs some comment.\\n\\nModified\\n\\n> ~~~\\n>\\n> 14. src/include/replication/walreceiver.h\\n>\\n> @@ -183,6 +183,7 @@ typedef struct\\n> bool streaming; /* Streaming of large transactions */\\n> bool twophase; /* Streaming of two-phase transactions at\\n> * prepare time */\\n> + bool onlylocal_data;\\n> } logical;\\n> } proto;\\n> } WalRcvStreamOptions;\\n>\\n> I think the new member needs some comment.\\n\\nModified\\n\\n> ~~~\\n>\\n> 15. src/test/regress/sql/subscription.sql\\n>\\n> ALTER SUBSCRIPTION test missing?\\n\\nAdded\\n\\nThanks for the comments, the attached v3 patch has the fixes for the same.\\n[1] - https://www.postgresql.org/message-id/CAA4eK1%2BMtz%2BStvNNtTg9%3D9BTq8%3DpMu-V5i4yWqs%3DKJUh0Z_L4g%40mail.gmail.com\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 8 Mar 2022 20:08:00 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 9:57 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 5:51 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Mon, Mar 7, 2022 at 5:01 PM Ashutosh Bapat\\n> > <ashutosh.bapat.oss@gmail.com> wrote:\\n> > >\\n> > > Hi Vignesh,\\n> > > I agree with Peter's comment that the changes to\\n> > > FilterRemoteOriginData() should be part of FilterByOrigin()\\n> > >\\n> > > Further, I wonder why \\\"onlylocal_data\\\" is a replication slot's\\n> > > property. A replication slot tracks the progress of replication and it\\n> > > may be used by different receivers with different options. I could\\n> > > start one receiver which wants only local data, say using\\n> > > \\\"pg_logical_slot_get_changes\\\" and later start another receiver which\\n> > > fetches all the data starting from where the first receiver left. This\\n> > > option prevents such flexibility.\\n> > >\\n> > > As discussed earlier in the thread, local_only can be property of\\n> > > publication or subscription, depending upon the use case, but I can't\\n> > > see any reason that it should be tied to a replication slot.\\n> > >\\n> >\\n> > I thought it should be similar to 'streaming' option of subscription\\n> > but may be Vignesh has some other reason which makes it different.\\n>\\n> Yes, this can be removed from the replication slot. It is my mistake\\n> that I have made while making the code similar to two-phase, I'm\\n> working on making the changes for this. I will fix and post an updated\\n> version for this.\\n\\nI have made the changes for this, the changes for this are available\\nin the v3 patch attached at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm0JcV-7iQZhyy3kehnWTy6x%3Dz%2BsX6u6Df%2B%2By8z33pz%2BBw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Mar 2022 20:16:24 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi Vegnesh,\\r\\n\\r\\nWhile considering about second problem, I was very confusing about it.\\r\\nI'm happy if you answer my question.\\r\\n\\r\\n> To handle this if user has specified only_local option, we could throw\\r\\n> a warning or error out while creating subscription in this case, we\\r\\n> could have a column srreplicateddata in pg_subscription_rel which\\r\\n> could indicate if the table has any replicated data or not:\\r\\n> postgres=# select * from pg_subscription_rel;\\r\\n> srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\\r\\n> ---------+---------+------------+-----------+------------------\\r\\n> 16389 | 16384 | r | 0/14A4640 | t\\r\\n> 16389 | 16385 | r | 0/14A4690 | f\\r\\n> (1 row)\\r\\n> In the above example, srreplicateddata with true indicates, tabel t1\\r\\n> whose relid is 16384 has replicated data and the other row having\\r\\n> srreplicateddata as false indicates table t2 whose relid is 16385\\r\\n> does not have replicated data.\\r\\n> When creating a new subscription, the subscriber will connect to the\\r\\n> publisher and check if the relation has replicated data by checking\\r\\n> srreplicateddata in pg_subscription_rel table.\\r\\n> If the table has any replicated data, log a warning or error for this.\\r\\n\\r\\nIIUC srreplicateddata represents whether the subscribed data is not\\r\\ngenerated from the publisher, but another node.\\r\\nMy first impression was that the name 'srreplicateddata' is not friendly\\r\\nbecause all subscribed data is replicated from publisher.\\r\\nAlso I was not sure how value of the column was set.\\r\\nIIUC a filtering by replication origins is done in publisher node\\r\\nand subscriber node cannot know\\r\\nwhether some data are really filtered or not.\\r\\nIf we distinguish by subscriber option publish_local_only,\\r\\nit cannot reproduce your example because same subscriber have different 'srreplicateddata'.\\r\\n\\r\\nBest Regards,\\r\\nHayato Kuroda\\r\\nFUJITSU LIMITED\\r\\n\\r\\n\",\n \"msg_date\": \"Fri, 11 Mar 2022 10:58:25 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 3:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Feb 23, 2022 at 11:59 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> ...\\n> ...\\n> > I have attached a basic patch for this, if the idea is accepted, I\\n> > will work further to test more scenarios, add documentation, and test\\n> > and post an updated patch.\\n> > For the second problem, Table synchronization of table including local\\n> > data and replicated data using copy command.\\n> >\\n> > Let us consider the following scenario:\\n> > a) node1 publishing to node2 b) node2 publishing to node1. Here in\\n> > this case node1 will have replicated data from node2 and vice versa.\\n> >\\n> > In the above if user wants to include node3 to subscribe data from\\n> > node2. Users will have to create a subscription in node3 to get the\\n> > data from node2. During table synchronization we send the complete\\n> > table data from node2 to node3. Node2 will have local data from node2\\n> > and also replicated data from node1. Currently we don't have an option\\n> > to differentiate between the locally generated data and replicated\\n> > data in the heap which will cause infinite recursion as described\\n> > above.\\n> >\\n> > To handle this if user has specified only_local option, we could throw\\n> > a warning or error out while creating subscription in this case, we\\n> > could have a column srreplicateddata in pg_subscription_rel which\\n> > could indicate if the table has any replicated data or not:\\n> > postgres=# select * from pg_subscription_rel;\\n> > srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\\n> > ---------+---------+------------+-----------+------------------\\n> > 16389 | 16384 | r | 0/14A4640 | t\\n> > 16389 | 16385 | r | 0/14A4690 | f\\n> > (1 row)\\n> >\\n> > In the above example, srreplicateddata with true indicates, tabel t1\\n> > whose relid is 16384 has replicated data and the other row having\\n> > srreplicateddata as false indicates table t2 whose relid is 16385\\n> > does not have replicated data.\\n> > When creating a new subscription, the subscriber will connect to the\\n> > publisher and check if the relation has replicated data by checking\\n> > srreplicateddata in pg_subscription_rel table.\\n> > If the table has any replicated data, log a warning or error for this.\\n> >\\n>\\n> If you want to give the error in this case, then I think we need to\\n> provide an option to the user to allow copy. One possibility could be\\n> to extend existing copy_data option as 'false', 'true', 'force'. For\\n> 'false', there shouldn't be any change, for 'true', if 'only_local'\\n> option is also set and the new column indicates replicated data then\\n> give an error, for 'force', we won't give an error even if the\\n> conditions as mentioned for 'true' case are met, rather we will allow\\n> copy in this case.\\n\\nWhen a subscription is created with publish_local_only and copy_data,\\nit will connect to the publisher and check if the published tables\\nhave also been subscribed from other nodes by checking if the entry is\\npresent in pg_subscription_rel and throw an error if present. The\\nattached v4-0002-Support-force-option-for-copy_data-check-and-thro.patch\\nhas the implementation for the same.\\nThoughts?\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Sat, 12 Mar 2022 10:46:19 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 11, 2022 at 4:28 PM kuroda.hayato@fujitsu.com\\n<kuroda.hayato@fujitsu.com> wrote:\\n>\\n> Hi Vegnesh,\\n>\\n> While considering about second problem, I was very confusing about it.\\n> I'm happy if you answer my question.\\n>\\n> > To handle this if user has specified only_local option, we could throw\\n> > a warning or error out while creating subscription in this case, we\\n> > could have a column srreplicateddata in pg_subscription_rel which\\n> > could indicate if the table has any replicated data or not:\\n> > postgres=# select * from pg_subscription_rel;\\n> > srsubid | srrelid | srsubstate | srsublsn | srreplicateddata\\n> > ---------+---------+------------+-----------+------------------\\n> > 16389 | 16384 | r | 0/14A4640 | t\\n> > 16389 | 16385 | r | 0/14A4690 | f\\n> > (1 row)\\n> > In the above example, srreplicateddata with true indicates, tabel t1\\n> > whose relid is 16384 has replicated data and the other row having\\n> > srreplicateddata as false indicates table t2 whose relid is 16385\\n> > does not have replicated data.\\n> > When creating a new subscription, the subscriber will connect to the\\n> > publisher and check if the relation has replicated data by checking\\n> > srreplicateddata in pg_subscription_rel table.\\n> > If the table has any replicated data, log a warning or error for this.\\n>\\n> IIUC srreplicateddata represents whether the subscribed data is not\\n> generated from the publisher, but another node.\\n> My first impression was that the name 'srreplicateddata' is not friendly\\n> because all subscribed data is replicated from publisher.\\n> Also I was not sure how value of the column was set.\\n> IIUC a filtering by replication origins is done in publisher node\\n> and subscriber node cannot know\\n> whether some data are really filtered or not.\\n> If we distinguish by subscriber option publish_local_only,\\n> it cannot reproduce your example because same subscriber have different 'srreplicateddata'.\\n\\nLet's consider an existing Multi master logical replication setup\\nbetween Node1 and Node2 that is created using the following steps:\\na) Node1 - Publication publishing employee table - pub1\\nb) Node2 - Subscription subscribing from publication pub1 with\\npublish_local_only - sub1_pub1_node1\\nc) Node2 - Publication publishing employee table - pub2\\nd) Node1 - Subscription subscribing from publication pub2 with\\npublish_local_only - sub2_pub2_node2\\n\\nTo create a subscription in node3, we will be using the following steps:\\na) Node2 - Publication publishing employee table. - pub3\\nb) Node3 - Subscription subscribing from publication in Node2 with\\npublish_local_only - sub3_pub3_node2\\n\\nWhen we create a subscription in Node3, Node3 will connect to\\nNode2(this will not be done in Node3) and check if the employee table\\nis present in pg_subscription_rel, in our case Node2 will have\\nemployee table present in pg_subscription_rel (sub1_pub1_node1\\nsubscribing to employee table from pub1 in Node1). As employee table\\nis being subscribed in node2 from node1, we will throw an error like\\nbelow:\\npostgres=# create subscription sub2 CONNECTION 'dbname =postgres port\\n= 9999' publication pub2 with (publish_local_only=on);\\nERROR: CREATE/ALTER SUBSCRIPTION with publish_local_only and\\ncopy_data as true is not allowed when the publisher might have\\nreplicated data, table:public.t1 might have replicated data in the\\npublisher\\nHINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\n\\nI was initially planning to add srreplicateddata field but I have\\nchanged it slightly to keep the design simple. Now we just check if\\nthe relation is present in pg_subscription_rel and throw an error if\\ncopy_data and publish_local_only option is specified. The changes for\\nthe same are available at [1].\\n\\n[1] - https://www.postgresql.org/message-id/CALDaNm0V%2B%3Db%3DCeZJNAAUO2PmSXH5QzNX3jADXb-0hGO_jVj0vA%40mail.gmail.com\\nThoughts?\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Sat, 12 Mar 2022 12:25:27 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Dear Vignesh,\\r\\n\\r\\nThank you for updating your patch!\\r\\n\\r\\n> Let's consider an existing Multi master logical replication setup\\r\\n> between Node1 and Node2 that is created using the following steps:\\r\\n> a) Node1 - Publication publishing employee table - pub1\\r\\n> b) Node2 - Subscription subscribing from publication pub1 with\\r\\n> publish_local_only - sub1_pub1_node1\\r\\n> c) Node2 - Publication publishing employee table - pub2\\r\\n> d) Node1 - Subscription subscribing from publication pub2 with\\r\\n> publish_local_only - sub2_pub2_node2\\r\\n> \\r\\n> To create a subscription in node3, we will be using the following steps:\\r\\n> a) Node2 - Publication publishing employee table. - pub3\\r\\n> b) Node3 - Subscription subscribing from publication in Node2 with\\r\\n> publish_local_only - sub3_pub3_node2\\r\\n> \\r\\n> When we create a subscription in Node3, Node3 will connect to\\r\\n> Node2(this will not be done in Node3) and check if the employee table\\r\\n> is present in pg_subscription_rel, in our case Node2 will have\\r\\n> employee table present in pg_subscription_rel (sub1_pub1_node1\\r\\n> subscribing to employee table from pub1 in Node1). As employee table\\r\\n> is being subscribed in node2 from node1, we will throw an error like\\r\\n> below:\\r\\n> postgres=# create subscription sub2 CONNECTION 'dbname =postgres port\\r\\n> = 9999' publication pub2 with (publish_local_only=on);\\r\\n> ERROR: CREATE/ALTER SUBSCRIPTION with publish_local_only and\\r\\n> copy_data as true is not allowed when the publisher might have\\r\\n> replicated data, table:public.t1 might have replicated data in the\\r\\n> publisher\\r\\n> HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\r\\n\\r\\nThanks for kind explanation. \\r\\nI read above and your doc in 0002, and I put some comments.\\r\\n\\r\\n1. alter_subscription.sgml\\r\\n\\r\\n```\\r\\n- <term><literal>copy_data</literal> (<type>boolean</type>)</term>\\r\\n+ <term><literal>copy_data</literal> (<type>boolean</type> | <literal>force</literal>)</term>\\r\\n```\\r\\n\\r\\nI thought that it should be written as enum. For example, huge_pages GUC parameter\\r\\ncan accept {on, off, try}, and it has been written as enum.\\r\\n\\r\\n2. create_subscription.sgml\\r\\n\\r\\n```\\r\\n- <term><literal>copy_data</literal> (<type>boolean</type>)</term>\\r\\n+ <term><literal>copy_data</literal> (<type>boolean</type> | <literal>force</literal>)</term>\\r\\n```\\r\\n\\r\\nSame as above.\\r\\n\\r\\n3. create_subscription.sgml\\r\\n\\r\\n```\\r\\n+\\r\\n+ <para>\\r\\n+ If the publication tables were also subscribing data in the publisher\\r\\n+ from other publishers, it will affect the\\r\\n+ <command>CREATE SUBSCRIPTION</command> based on the value specified\\r\\n+ for <literal>publish_local_only</literal> option. Refer to the\\r\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\r\\n+ </para>\\r\\n```\\r\\n\\r\\nI seeked docs, but the words \\\" publication tables \\\" have not seen.\\r\\nHow about \\\"tables in the publication\\\"?\\r\\n\\r\\n4. create_subscription.sgml - about your example \\r\\n\\r\\nIn the first section, we should describe about 2-nodes case more detail\\r\\nlike Amit mentioned in [1]. I thought that Option-3 can be resolved by defining\\r\\nsubscriptions in both nodes with publish_local_only = true and copy_data = force.\\r\\n\\r\\n> I was initially planning to add srreplicateddata field but I have\\r\\n> changed it slightly to keep the design simple. Now we just check if\\r\\n> the relation is present in pg_subscription_rel and throw an error if\\r\\n> copy_data and publish_local_only option is specified. The changes for\\r\\n> the same are available at [1].\\r\\n> \\r\\n> [1] -\\r\\n> https://www.postgresql.org/message-id/CALDaNm0V%2B%3Db%3DCeZJNAAU\\r\\n> O2PmSXH5QzNX3jADXb-0hGO_jVj0vA%40mail.gmail.com\\r\\n> Thoughts?\\r\\n\\r\\nActually I doubted that the column is really needed, so it is good.\\r\\n\\r\\nBest Regards,\\r\\nHayato Kuroda\\r\\nFUJITSU LIMITED\\r\\n\\r\\n\",\n \"msg_date\": \"Tue, 15 Mar 2022 01:39:49 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Dear Vignesh,\\r\\n\\r\\n> Thanks for kind explanation.\\r\\n> I read above and your doc in 0002, and I put some comments.\\r\\n\\r\\nI forgot a comment about 0002 doc.\\r\\n\\r\\n5. create_subscription.sgml - about your example\\r\\n\\r\\nThree possibilities were listed in the doc,\\r\\nbut I was not sure about b) case.\\r\\nIn the situation Node1 and Node2 have already become multi-master,\\r\\nand data has already synced at that time.\\r\\nIf so, how do we realize that \\\"there is data present only in one Node\\\"?\\r\\nCase a) and c) seem reasonable.\\r\\n\\r\\nBest Regards,\\r\\nHayato Kuroda\\r\\nFUJITSU LIMITED\\r\\n\\r\\n\",\n \"msg_date\": \"Tue, 15 Mar 2022 04:25:21 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 15, 2022 at 7:09 AM kuroda.hayato@fujitsu.com\\n<kuroda.hayato@fujitsu.com> wrote:\\n>\\n> Dear Vignesh,\\n>\\n> Thank you for updating your patch!\\n>\\n> > Let's consider an existing Multi master logical replication setup\\n> > between Node1 and Node2 that is created using the following steps:\\n> > a) Node1 - Publication publishing employee table - pub1\\n> > b) Node2 - Subscription subscribing from publication pub1 with\\n> > publish_local_only - sub1_pub1_node1\\n> > c) Node2 - Publication publishing employee table - pub2\\n> > d) Node1 - Subscription subscribing from publication pub2 with\\n> > publish_local_only - sub2_pub2_node2\\n> >\\n> > To create a subscription in node3, we will be using the following steps:\\n> > a) Node2 - Publication publishing employee table. - pub3\\n> > b) Node3 - Subscription subscribing from publication in Node2 with\\n> > publish_local_only - sub3_pub3_node2\\n> >\\n> > When we create a subscription in Node3, Node3 will connect to\\n> > Node2(this will not be done in Node3) and check if the employee table\\n> > is present in pg_subscription_rel, in our case Node2 will have\\n> > employee table present in pg_subscription_rel (sub1_pub1_node1\\n> > subscribing to employee table from pub1 in Node1). As employee table\\n> > is being subscribed in node2 from node1, we will throw an error like\\n> > below:\\n> > postgres=# create subscription sub2 CONNECTION 'dbname =postgres port\\n> > = 9999' publication pub2 with (publish_local_only=on);\\n> > ERROR: CREATE/ALTER SUBSCRIPTION with publish_local_only and\\n> > copy_data as true is not allowed when the publisher might have\\n> > replicated data, table:public.t1 might have replicated data in the\\n> > publisher\\n> > HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\n>\\n> Thanks for kind explanation.\\n> I read above and your doc in 0002, and I put some comments.\\n>\\n> 1. alter_subscription.sgml\\n>\\n> ```\\n> - <term><literal>copy_data</literal> (<type>boolean</type>)</term>\\n> + <term><literal>copy_data</literal> (<type>boolean</type> | <literal>force</literal>)</term>\\n> ```\\n>\\n> I thought that it should be written as enum. For example, huge_pages GUC parameter\\n> can accept {on, off, try}, and it has been written as enum.\\n\\nModified\\n\\n> 2. create_subscription.sgml\\n>\\n> ```\\n> - <term><literal>copy_data</literal> (<type>boolean</type>)</term>\\n> + <term><literal>copy_data</literal> (<type>boolean</type> | <literal>force</literal>)</term>\\n> ```\\n>\\n> Same as above.\\n\\nModified\\n\\n> 3. create_subscription.sgml\\n>\\n> ```\\n> +\\n> + <para>\\n> + If the publication tables were also subscribing data in the publisher\\n> + from other publishers, it will affect the\\n> + <command>CREATE SUBSCRIPTION</command> based on the value specified\\n> + for <literal>publish_local_only</literal> option. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n> + </para>\\n> ```\\n>\\n> I seeked docs, but the words \\\" publication tables \\\" have not seen.\\n> How about \\\"tables in the publication\\\"?\\n\\nModified\\n\\n> 4. create_subscription.sgml - about your example\\n>\\n> In the first section, we should describe about 2-nodes case more detail\\n> like Amit mentioned in [1]. I thought that Option-3 can be resolved by defining\\n> subscriptions in both nodes with publish_local_only = true and copy_data = force.\\n\\nI thought existing information is enough because we have mentioned\\nthat node1 and node2 have bidirectional replication setup done and\\nboth the table data will be replicated and synchronized as and when\\nthe DML operations are happening. In option-3 we need to create a\\nsubscription with copy_data as force to one node and copy_data as\\nfalse to another node because both nodes will be having the same data,\\ncopying the data just from one of the nodes should be enough.\\n\\nThanks for the comments, the attached v5 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 16 Mar 2022 11:15:09 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 15, 2022 at 9:55 AM kuroda.hayato@fujitsu.com\\n<kuroda.hayato@fujitsu.com> wrote:\\n>\\n> Dear Vignesh,\\n>\\n> > Thanks for kind explanation.\\n> > I read above and your doc in 0002, and I put some comments.\\n>\\n> I forgot a comment about 0002 doc.\\n>\\n> 5. create_subscription.sgml - about your example\\n>\\n> Three possibilities were listed in the doc,\\n> but I was not sure about b) case.\\n> In the situation Node1 and Node2 have already become multi-master,\\n> and data has already synced at that time.\\n> If so, how do we realize that \\\"there is data present only in one Node\\\"?\\n> Case a) and c) seem reasonable.\\n\\nYour point is valid, modified.\\n\\nThe changes for the same are available int the v5 patch available at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 11:17:16 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi Vignesh,\\nSome review comments on 0001\\n\\nOn Wed, Mar 16, 2022 at 11:17 AM vignesh C <vignesh21@gmail.com> wrote:\\n\\n>\\n> The changes for the same are available int the v5 patch available at [1].\\n> [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\\n>\\n\\n cb->truncate_cb = pg_decode_truncate;\\n cb->commit_cb = pg_decode_commit_txn;\\n cb->filter_by_origin_cb = pg_decode_filter;\\n+ cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\\n\\nWhy do we need a new hook? Can we use filter_by_origin_cb? Also it looks like\\nimplementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\\nunless my eyes are deceiving me.\\n\\n- <literal>binary</literal>, <literal>streaming</literal>, and\\n- <literal>disable_on_error</literal>.\\n+ <literal>binary</literal>, <literal>streaming</literal>,\\n+ <literal>disable_on_error</literal> and\\n+ <literal>publish_local_only</literal>.\\n\\n\\\"publish_local_only\\\" as a \\\"subscription\\\" option looks odd. Should it be\\n\\\"subscribe_local_only\\\"?\\n\\n\\n+ <varlistentry>\\n+ <term><literal>publish_local_only</literal>\\n(<type>boolean</type>)</term>\\n+ <listitem>\\n+ <para>\\n+ Specifies whether the subscription should subscribe only to the\\n+ locally generated changes or subscribe to both the locally generated\\n+ changes and the replicated changes that was generated from other\\n+ nodes in the publisher. The default is <literal>false</literal>.\\n\\nThis description to uses verb \\\"subscribe\\\" instead of \\\"publish\\\".\\n\\n@@ -936,6 +951,13 @@ AlterSubscription(ParseState *pstate,\\nAlterSubscriptionStmt *stmt,\\n = true;\\n }\\n\\n+ if (IsSet(opts.specified_opts, SUBOPT_PUBLISH_LOCAL_ONLY))\\n+ {\\n+ values[Anum_pg_subscription_sublocal - 1] =\\n+ BoolGetDatum(opts.streaming);\\n\\nshould this be opts.sublocal?\\n\\n cb->commit_prepared_cb = pgoutput_commit_prepared_txn;\\n cb->rollback_prepared_cb = pgoutput_rollback_prepared_txn;\\n cb->filter_by_origin_cb = pgoutput_origin_filter;\\n+ cb->filter_remote_origin_cb = pgoutput_remote_origin_filter;\\n\\npgoutput_origin_filter just returns false now. I think we should just enhance\\nthat function and reuse the callback, instead of adding a new callback.\\n\\n--- a/src/include/replication/logical.h\\n+++ b/src/include/replication/logical.h\\n@@ -99,6 +99,8 @@ typedef struct LogicalDecodingContext\\n */\\n bool twophase_opt_given;\\n\\n+ bool only_local; /* publish only locally generated data */\\n+\\n\\nIf we get rid of the new callback, we won't need this new member as well.\\nAnyway the filtering happens within the output plugin. There is nothing that\\ncore is required to do here.\\n\\n--- a/src/include/replication/walreceiver.h\\n+++ b/src/include/replication/walreceiver.h\\n@@ -183,6 +183,7 @@ typedef struct\\n bool streaming; /* Streaming of large transactions */\\n bool twophase; /* Streaming of two-phase transactions at\\n * prepare time */\\n+ bool only_local; /* publish only locally generated data */\\n\\nAre we using this anywhere. I couldn't spot any usage of this member. I might\\nbe missing something though.\\n\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 19:22:12 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2022 at 7:40 PM Ashutosh Bapat\\n<ashutosh.bapat.oss@gmail.com> wrote:\\n> >\\n> > The changes for the same are available int the v5 patch available at [1].\\n> > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\\n> >\\n>\\n> cb->truncate_cb = pg_decode_truncate;\\n> cb->commit_cb = pg_decode_commit_txn;\\n> cb->filter_by_origin_cb = pg_decode_filter;\\n> + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\\n>\\n> Why do we need a new hook? Can we use filter_by_origin_cb?\\n>\\n\\nI also think there is no need for a new hook in this case but I might\\nalso be missing something that Vignesh has in his mind.\\n\\n> Also it looks like\\n> implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\\n> unless my eyes are deceiving me.\\n>\\n> - <literal>binary</literal>, <literal>streaming</literal>, and\\n> - <literal>disable_on_error</literal>.\\n> + <literal>binary</literal>, <literal>streaming</literal>,\\n> + <literal>disable_on_error</literal> and\\n> + <literal>publish_local_only</literal>.\\n>\\n> \\\"publish_local_only\\\" as a \\\"subscription\\\" option looks odd. Should it be\\n> \\\"subscribe_local_only\\\"?\\n>\\n\\nI also think \\\"subscribe_local_only\\\" fits better here.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Mar 2022 09:14:53 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 31, 2022 at 9:14 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Mar 30, 2022 at 7:40 PM Ashutosh Bapat\\n> <ashutosh.bapat.oss@gmail.com> wrote:\\n> > >\\n> > > The changes for the same are available int the v5 patch available at [1].\\n> > > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\\n> > >\\n> >\\n> > cb->truncate_cb = pg_decode_truncate;\\n> > cb->commit_cb = pg_decode_commit_txn;\\n> > cb->filter_by_origin_cb = pg_decode_filter;\\n> > + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\\n> >\\n> > Why do we need a new hook? Can we use filter_by_origin_cb?\\n> >\\n>\\n> I also think there is no need for a new hook in this case but I might\\n> also be missing something that Vignesh has in his mind.\\n>\\n> > Also it looks like\\n> > implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\\n> > unless my eyes are deceiving me.\\n> >\\n> > - <literal>binary</literal>, <literal>streaming</literal>, and\\n> > - <literal>disable_on_error</literal>.\\n> > + <literal>binary</literal>, <literal>streaming</literal>,\\n> > + <literal>disable_on_error</literal> and\\n> > + <literal>publish_local_only</literal>.\\n> >\\n> > \\\"publish_local_only\\\" as a \\\"subscription\\\" option looks odd. Should it be\\n> > \\\"subscribe_local_only\\\"?\\n> >\\n>\\n> I also think \\\"subscribe_local_only\\\" fits better here.\\n>\\n\\nAbout 0002 patch,\\nCommit message:\\n------\\nIf a subscription is created to Node1 from Node3 with publish_local_only\\nand copy_data as ON, then throw an error so that user can handle\\ncreation of subscription with table having consistent data.\\n------\\n\\nDo you want to refer to Node2 instead of Node3 here as Node3 doesn't\\nmake sense in the description?\\n\\nAlso, you haven't explained anywhere in the patch why\\n'publish_local_only' (or whatever we call it) won't work for initial\\nsync? IIUC, it is because we can identify origin changes only based on\\nWAL and initial sync directly copies data from the heap. Am, I missing\\nsomething here?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Mar 2022 14:51:54 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2022 at 7:22 PM Ashutosh Bapat\\n<ashutosh.bapat.oss@gmail.com> wrote:\\n>\\n> Hi Vignesh,\\n> Some review comments on 0001\\n>\\n> On Wed, Mar 16, 2022 at 11:17 AM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> >\\n> > The changes for the same are available int the v5 patch available at [1].\\n> > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\\n> >\\n>\\n> cb->truncate_cb = pg_decode_truncate;\\n> cb->commit_cb = pg_decode_commit_txn;\\n> cb->filter_by_origin_cb = pg_decode_filter;\\n> + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\\n>\\n> Why do we need a new hook? Can we use filter_by_origin_cb? Also it looks like\\n> implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\\n> unless my eyes are deceiving me.\\n\\nI have used filter_by_origin_cb for the implementation, removed\\nfilter_remote_origin_cb\\n\\n> - <literal>binary</literal>, <literal>streaming</literal>, and\\n> - <literal>disable_on_error</literal>.\\n> + <literal>binary</literal>, <literal>streaming</literal>,\\n> + <literal>disable_on_error</literal> and\\n> + <literal>publish_local_only</literal>.\\n>\\n> \\\"publish_local_only\\\" as a \\\"subscription\\\" option looks odd. Should it be\\n> \\\"subscribe_local_only\\\"?\\n\\nModified\\n\\n>\\n> + <varlistentry>\\n> + <term><literal>publish_local_only</literal>\\n> (<type>boolean</type>)</term>\\n> + <listitem>\\n> + <para>\\n> + Specifies whether the subscription should subscribe only to the\\n> + locally generated changes or subscribe to both the locally generated\\n> + changes and the replicated changes that was generated from other\\n> + nodes in the publisher. The default is <literal>false</literal>.\\n>\\n> This description to uses verb \\\"subscribe\\\" instead of \\\"publish\\\".\\n\\nModified\\n\\n> @@ -936,6 +951,13 @@ AlterSubscription(ParseState *pstate,\\n> AlterSubscriptionStmt *stmt,\\n> = true;\\n> }\\n>\\n> + if (IsSet(opts.specified_opts, SUBOPT_PUBLISH_LOCAL_ONLY))\\n> + {\\n> + values[Anum_pg_subscription_sublocal - 1] =\\n> + BoolGetDatum(opts.streaming);\\n>\\n> should this be opts.sublocal?\\n\\nYes you are right, Modified\\n\\n> cb->commit_prepared_cb = pgoutput_commit_prepared_txn;\\n> cb->rollback_prepared_cb = pgoutput_rollback_prepared_txn;\\n> cb->filter_by_origin_cb = pgoutput_origin_filter;\\n> + cb->filter_remote_origin_cb = pgoutput_remote_origin_filter;\\n>\\n> pgoutput_origin_filter just returns false now. I think we should just enhance\\n> that function and reuse the callback, instead of adding a new callback.\\n\\nModified\\n\\n> --- a/src/include/replication/logical.h\\n> +++ b/src/include/replication/logical.h\\n> @@ -99,6 +99,8 @@ typedef struct LogicalDecodingContext\\n> */\\n> bool twophase_opt_given;\\n>\\n> + bool only_local; /* publish only locally generated data */\\n> +\\n>\\n> If we get rid of the new callback, we won't need this new member as well.\\n> Anyway the filtering happens within the output plugin. There is nothing that\\n> core is required to do here.\\n\\nModified\\n\\n> --- a/src/include/replication/walreceiver.h\\n> +++ b/src/include/replication/walreceiver.h\\n> @@ -183,6 +183,7 @@ typedef struct\\n> bool streaming; /* Streaming of large transactions */\\n> bool twophase; /* Streaming of two-phase transactions at\\n> * prepare time */\\n> + bool only_local; /* publish only locally generated data */\\n>\\n> Are we using this anywhere. I couldn't spot any usage of this member. I might\\n> be missing something though.\\n\\nThis is set in ApplyWorkerMain before calling libpqrcv_startstreaming.\\nThis will be used in building the START_REPLICATION command.\\nThanks for the comments, the attached v6 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Fri, 1 Apr 2022 12:34:21 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 31, 2022 at 2:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Thu, Mar 31, 2022 at 9:14 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Wed, Mar 30, 2022 at 7:40 PM Ashutosh Bapat\\n> > <ashutosh.bapat.oss@gmail.com> wrote:\\n> > > >\\n> > > > The changes for the same are available int the v5 patch available at [1].\\n> > > > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\\n> > > >\\n> > >\\n> > > cb->truncate_cb = pg_decode_truncate;\\n> > > cb->commit_cb = pg_decode_commit_txn;\\n> > > cb->filter_by_origin_cb = pg_decode_filter;\\n> > > + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\\n> > >\\n> > > Why do we need a new hook? Can we use filter_by_origin_cb?\\n> > >\\n> >\\n> > I also think there is no need for a new hook in this case but I might\\n> > also be missing something that Vignesh has in his mind.\\n> >\\n> > > Also it looks like\\n> > > implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\\n> > > unless my eyes are deceiving me.\\n> > >\\n> > > - <literal>binary</literal>, <literal>streaming</literal>, and\\n> > > - <literal>disable_on_error</literal>.\\n> > > + <literal>binary</literal>, <literal>streaming</literal>,\\n> > > + <literal>disable_on_error</literal> and\\n> > > + <literal>publish_local_only</literal>.\\n> > >\\n> > > \\\"publish_local_only\\\" as a \\\"subscription\\\" option looks odd. Should it be\\n> > > \\\"subscribe_local_only\\\"?\\n> > >\\n> >\\n> > I also think \\\"subscribe_local_only\\\" fits better here.\\n> >\\n>\\n> About 0002 patch,\\n> Commit message:\\n> ------\\n> If a subscription is created to Node1 from Node3 with publish_local_only\\n> and copy_data as ON, then throw an error so that user can handle\\n> creation of subscription with table having consistent data.\\n> ------\\n> Do you want to refer to Node2 instead of Node3 here as Node3 doesn't\\n> make sense in the description?\\n\\nIt should be Node3, I have added more details in the commit message\\nmentioning about the scenario.\\n\\n> Also, you haven't explained anywhere in the patch why\\n> 'publish_local_only' (or whatever we call it) won't work for initial\\n> sync? IIUC, it is because we can identify origin changes only based on\\n> WAL and initial sync directly copies data from the heap. Am, I missing\\n> something here?\\n\\nI have added the explanation where we are throwing the error.\\n\\nThe v6 patch at [1] has the changes for the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm16eBx2BjLFjfFHSU4pb25HmgV--M692OPgH_91Dwn%3D2g%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Apr 2022 12:45:50 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"I didn't find this in https://commitfest.postgresql.org/37/. Is this\\nsomewhere in the commitfest?\\n\\nOn Fri, Apr 1, 2022 at 12:46 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Thu, Mar 31, 2022 at 2:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Thu, Mar 31, 2022 at 9:14 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > >\\n> > > On Wed, Mar 30, 2022 at 7:40 PM Ashutosh Bapat\\n> > > <ashutosh.bapat.oss@gmail.com> wrote:\\n> > > > >\\n> > > > > The changes for the same are available int the v5 patch available at [1].\\n> > > > > [1] - https://www.postgresql.org/message-id/CALDaNm3wCf0YcvVo%2BgHMGpupk9K6WKJxCyLUvhPC2GkPKRZUWA%40mail.gmail.com\\n> > > > >\\n> > > >\\n> > > > cb->truncate_cb = pg_decode_truncate;\\n> > > > cb->commit_cb = pg_decode_commit_txn;\\n> > > > cb->filter_by_origin_cb = pg_decode_filter;\\n> > > > + cb->filter_remote_origin_cb = pg_decode_filter_remote_origin;\\n> > > >\\n> > > > Why do we need a new hook? Can we use filter_by_origin_cb?\\n> > > >\\n> > >\\n> > > I also think there is no need for a new hook in this case but I might\\n> > > also be missing something that Vignesh has in his mind.\\n> > >\\n> > > > Also it looks like\\n> > > > implementation of pg_decode_filter and pg_decode_filter_remote_origin is same,\\n> > > > unless my eyes are deceiving me.\\n> > > >\\n> > > > - <literal>binary</literal>, <literal>streaming</literal>, and\\n> > > > - <literal>disable_on_error</literal>.\\n> > > > + <literal>binary</literal>, <literal>streaming</literal>,\\n> > > > + <literal>disable_on_error</literal> and\\n> > > > + <literal>publish_local_only</literal>.\\n> > > >\\n> > > > \\\"publish_local_only\\\" as a \\\"subscription\\\" option looks odd. Should it be\\n> > > > \\\"subscribe_local_only\\\"?\\n> > > >\\n> > >\\n> > > I also think \\\"subscribe_local_only\\\" fits better here.\\n> > >\\n> >\\n> > About 0002 patch,\\n> > Commit message:\\n> > ------\\n> > If a subscription is created to Node1 from Node3 with publish_local_only\\n> > and copy_data as ON, then throw an error so that user can handle\\n> > creation of subscription with table having consistent data.\\n> > ------\\n> > Do you want to refer to Node2 instead of Node3 here as Node3 doesn't\\n> > make sense in the description?\\n>\\n> It should be Node3, I have added more details in the commit message\\n> mentioning about the scenario.\\n>\\n> > Also, you haven't explained anywhere in the patch why\\n> > 'publish_local_only' (or whatever we call it) won't work for initial\\n> > sync? IIUC, it is because we can identify origin changes only based on\\n> > WAL and initial sync directly copies data from the heap. Am, I missing\\n> > something here?\\n>\\n> I have added the explanation where we are throwing the error.\\n>\\n> The v6 patch at [1] has the changes for the same.\\n> [1] - https://www.postgresql.org/message-id/CALDaNm16eBx2BjLFjfFHSU4pb25HmgV--M692OPgH_91Dwn%3D2g%40mail.gmail.com\\n>\\n> Regards,\\n> Vignesh\\n\\n\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Apr 2022 18:50:08 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my comments for the latest patch v6-0001.\\n\\n(I will post my v6-0002 review comments separately)\\n\\nPATCH v6-0001 comments\\n======================\\n\\n1.1 General - Option name\\n\\nI still feel like the option name is not ideal. Unfortunately, this is\\nimportant because any name change would impact lots of these patch\\nfiles and docs, struct members etc.\\n\\nIt was originally called \\\"local_only\\\", but I thought that as a\\nSUBSCRIPTION option that was confusing because \\\"local\\\" means local to\\nwhat? Really it is local to the publisher, not local to the\\nsubscriber, so that name seemed misleading.\\n\\nThen I suggested \\\"publish_local_only\\\". Although that resolved the\\nambiguity problem, other people thought it seemed odd to have the\\n\\\"publish\\\" prefix for a subscription-side option.\\n\\nSo now it is changed again to \\\"subscribe_local_only\\\" -- It's getting\\nbetter but still, it is implied that the \\\"local\\\" means local to the\\npublisher except there is nothing in the option name really to convey\\nthat meaning. IMO we here all understand the meaning of this option\\nmostly by familiarity with this discussion thread, but I think a user\\ncoming to this for the first time will still be confused by the name.\\n\\nBelow are some other name ideas. Maybe they are not improvements, but\\nit might help other people to come up with something better.\\n\\nsubscribe_publocal_only = true/false\\norigin = pub_only/any\\nadjacent_data_only = true/false\\ndirect_subscriptions_only = true/false\\n...\\n\\n\\n(FYI, the remainder of these review comments will assume the option is\\nstill called \\\"subscribe_local_only\\\")\\n\\n~~~\\n\\n1.2 General - inconsistent members and args\\n\\nIMO the struct members and args should also be named for close\\nconsistency with whatever the option name is.\\n\\nCurrently the option is called \\\"subscription_local_only\\\". So I think\\nthe members/args would be better to be called \\\"local_only\\\" instead of\\n\\\"only_local\\\".\\n\\n~~~\\n\\n1.3 Commit message - wrong option name\\n\\nThe commit message refers to the option name as \\\"publish_local_only\\\"\\ninstead of the option name that is currently implemented.\\n\\n~~~\\n\\n1.4 Commit message - wording\\n\\nThe wording seems a bit off. Below is suggested simpler wording which\\nI AFAIK conveys the same information.\\n\\nBEFORE\\nAdd an option publish_local_only which will subscribe only to the locally\\ngenerated data in the publisher node. If subscriber is created with this\\noption, publisher will skip publishing the data that was subscribed\\nfrom other nodes. It can be created using following syntax:\\nex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\\nPUBLICATION pub1 with (publish_local_only = on);\\n\\nSUGGESTION\\nThis patch adds a new SUBSCRIPTION boolean option\\n\\\"subscribe_local_only\\\". The default is false. When a SUBSCRIPTION is\\ncreated with this option enabled, the publisher will only publish data\\nthat originated at the publisher node.\\nUsage:\\nCREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\\nPUBLICATION pub1 with (subscribe_local_only = true);\\n\\n~~~\\n\\n1.5 doc/src/sgml/ref/create_subscription.sgml - \\\"generated\\\" changes.\\n\\n+ <para>\\n+ Specifies whether the subscription will request the publisher to send\\n+ locally generated changes or both the locally generated changes and\\n+ the replicated changes that was generated from other nodes. The\\n+ default is <literal>false</literal>.\\n+ </para>\\n\\nFor some reason, it seemed a bit strange to me to use the term\\n\\\"generated\\\" changes. Maybe better to refer to the origin of changes?\\n\\nSUGGESTION\\nSpecifies whether the publisher should send only changes that\\noriginated locally at the publisher node, or send any publisher node\\nchanges regardless of their origin. The default is false.\\n\\n~~~\\n\\n1.6 src/backend/replication/pgoutput/pgoutput.c -\\nLOGICALREP_PROTO_TWOPHASE_VERSION_NUM\\n\\n@@ -496,6 +509,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\\nOutputPluginOptions *opt,\\n else\\n ctx->twophase_opt_given = true;\\n\\n+ if (data->only_local && data->protocol_version <\\nLOGICALREP_PROTO_TWOPHASE_VERSION_NUM)\\n+ ereport(ERROR,\\n+ (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n+ errmsg(\\\"requested proto_version=%d does not support\\nsubscribe_local_only, need %d or higher\\\",\\n+ data->protocol_version, LOGICALREP_PROTO_TWOPHASE_VERSION_NUM)));\\n\\nI thought this code should not be using\\nLOGICALREP_PROTO_TWOPHASE_VERSION_NUM. Shouldn't there be some newly\\nintroduced constant like LOGICALREP_PROTO_LOCALONLY_VERSION_NUM which\\nyou will use here?\\n\\n~~~\\n\\n1.7 src/bin/pg_dump/pg_dump.c - 150000\\n\\n@@ -4451,11 +4452,13 @@ getSubscriptions(Archive *fout)\\n if (fout->remoteVersion >= 150000)\\n appendPQExpBufferStr(query,\\n \\\" s.subtwophasestate,\\\\n\\\"\\n- \\\" s.subdisableonerr\\\\n\\\");\\n+ \\\" s.subdisableonerr,\\\\n\\\"\\n+ \\\" s.sublocal\\\\n\\\");\\n else\\n appendPQExpBuffer(query,\\n \\\" '%c' AS subtwophasestate,\\\\n\\\"\\n- \\\" false AS subdisableonerr\\\\n\\\",\\n+ \\\" false AS subdisableonerr,\\\\n\\\"\\n+ \\\" false AS s.sublocal\\\\n\\\",\\n LOGICALREP_TWOPHASE_STATE_DISABLED);\\n\\nI think this local_only feature is unlikely to get into the PG15\\nrelease, so this code should be split out into a separate condition\\nbecause later will need to change to say >= 160000.\\n\\n~~~\\n\\n1.8 src/bin/pg_dump/pg_dump.c - dumpSubscription\\n\\n@@ -4585,6 +4591,9 @@ dumpSubscription(Archive *fout, const\\nSubscriptionInfo *subinfo)\\n if (strcmp(subinfo->subdisableonerr, \\\"t\\\") == 0)\\n appendPQExpBufferStr(query, \\\", disable_on_error = true\\\");\\n\\n+ if (strcmp(subinfo->sublocal, \\\"f\\\") != 0)\\n+ appendPQExpBufferStr(query, \\\", subscribe_local_only = on\\\");\\n+\\n\\nI felt it is more natural to say \\\"if it is true set to true\\\", instead\\nof \\\"if it is not false set to on\\\".\\n\\nSUGGESTION\\nif (strcmp(subinfo->sublocal, \\\"t\\\") == 0)\\nappendPQExpBufferStr(query, \\\", subscribe_local_only = true\\\");\\n\\n~~~\\n\\n1.9 src/bin/psql/describe.c - 150000\\n\\n@@ -6318,9 +6318,11 @@ describeSubscriptions(const char *pattern, bool verbose)\\n if (pset.sversion >= 150000)\\n appendPQExpBuffer(&buf,\\n \\\", subtwophasestate AS \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n- \\\", subdisableonerr AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n+ \\\", subdisableonerr AS \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n+ \\\", sublocal AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n gettext_noop(\\\"Two phase commit\\\"),\\n- gettext_noop(\\\"Disable on error\\\"));\\n+ gettext_noop(\\\"Disable on error\\\"),\\n+ gettext_noop(\\\"Only local\\\"));\\n\\nI think this local_only feature is unlikely to get into the PG15\\nrelease, so this code should be split out into a separate condition\\nbecause later will need to change to say >= 160000.\\n\\n~~~\\n\\n1.10 src/bin/psql/describe.c - describeSubscriptions column\\n\\n@@ -6318,9 +6318,11 @@ describeSubscriptions(const char *pattern, bool verbose)\\n if (pset.sversion >= 150000)\\n appendPQExpBuffer(&buf,\\n \\\", subtwophasestate AS \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n- \\\", subdisableonerr AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n+ \\\", subdisableonerr AS \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n+ \\\", sublocal AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n gettext_noop(\\\"Two phase commit\\\"),\\n- gettext_noop(\\\"Disable on error\\\"));\\n+ gettext_noop(\\\"Disable on error\\\"),\\n+ gettext_noop(\\\"Only local\\\"));\\n\\nI think the column name here should be more consistent with the option\\nname. e.g. it should be \\\"Local only\\\", not \\\"Only local\\\".\\n\\n~~~\\n\\n1.11 src/bin/psql/tab-complete.c - whitespace\\n\\n@@ -3167,7 +3167,7 @@ psql_completion(const char *text, int start, int end)\\n /* Complete \\\"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\\\" */\\n else if (HeadMatches(\\\"CREATE\\\", \\\"SUBSCRIPTION\\\") && TailMatches(\\\"WITH\\\", \\\"(\\\"))\\n COMPLETE_WITH(\\\"binary\\\", \\\"connect\\\", \\\"copy_data\\\", \\\"create_slot\\\",\\n- \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n+ \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\", \\\"subscribe_local_only\\\",\\n\\nThe patch accidentally added a space char before the \\\"slot_name\\\".\\n\\n~~~\\n\\n1.12 src/include/replication/walreceiver.h - \\\"generated\\\"\\n\\n@@ -183,6 +183,7 @@ typedef struct\\n bool streaming; /* Streaming of large transactions */\\n bool twophase; /* Streaming of two-phase transactions at\\n * prepare time */\\n+ bool only_local; /* publish only locally generated data */\\n\\nThis is a similar review comment as #1.5 about saying the word \\\"generated\\\".\\nMaybe there is another way to word this?\\n\\n~~~\\n\\n1.13 src/test/regress/sql/subscription.sql - missing test case\\n\\nIsn't there a missing test case for ensuring that the new option is boolean?\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 10:51:20 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Apr 4, 2022 at 6:50 PM Ashutosh Bapat\\n<ashutosh.bapat.oss@gmail.com> wrote:\\n>\\n> I didn't find this in https://commitfest.postgresql.org/37/. Is this\\n> somewhere in the commitfest?\\n\\nI missed adding it earlier, I have added it to commitfest today:\\nhttps://commitfest.postgresql.org/38/3610/\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 07:24:37 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my comments for the latest patch v6-0001.\\n>\\n> (I will post my v6-0002 review comments separately)\\n>\\n> PATCH v6-0001 comments\\n> ======================\\n>\\n> 1.1 General - Option name\\n>\\n> I still feel like the option name is not ideal. Unfortunately, this is\\n> important because any name change would impact lots of these patch\\n> files and docs, struct members etc.\\n>\\n> It was originally called \\\"local_only\\\", but I thought that as a\\n> SUBSCRIPTION option that was confusing because \\\"local\\\" means local to\\n> what? Really it is local to the publisher, not local to the\\n> subscriber, so that name seemed misleading.\\n>\\n> Then I suggested \\\"publish_local_only\\\". Although that resolved the\\n> ambiguity problem, other people thought it seemed odd to have the\\n> \\\"publish\\\" prefix for a subscription-side option.\\n>\\n> So now it is changed again to \\\"subscribe_local_only\\\" -- It's getting\\n> better but still, it is implied that the \\\"local\\\" means local to the\\n> publisher except there is nothing in the option name really to convey\\n> that meaning. IMO we here all understand the meaning of this option\\n> mostly by familiarity with this discussion thread, but I think a user\\n> coming to this for the first time will still be confused by the name.\\n>\\n> Below are some other name ideas. Maybe they are not improvements, but\\n> it might help other people to come up with something better.\\n>\\n> subscribe_publocal_only = true/false\\n> origin = pub_only/any\\n> adjacent_data_only = true/false\\n> direct_subscriptions_only = true/false\\n> ...\\n>\\n\\nI think local_only with a description should be okay considering other\\ncurrent options. Some of the options like slot_name, binary, etc. are\\npublisher-specific which becomes clear only after reading the\\ndescription. I am not sure adding pub*/sub* to it is much helpful. So,\\n+1 to a simple boolean like 'local_only', 'data_local', or something\\nlike that.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 08:06:18 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my comments for the latest patch v6-0002.\\n\\nPATCH v6-0002 comments\\n======================\\n\\n2.1 General - should this be an independent patch?\\n\\nIn many ways, I think most of this patch is unrelated to the other\\n\\\"local_only\\\" patch (v6-0001).\\n\\nFor example, IIUC even in the current HEAD, we could consider it to be\\na user error if multiple SUBSCRIPTIONS or multiple PUBLICATIONS of the\\nsame SUBSCRIPTION are replicating to the same TABLE on the same node\\nand using \\\"copy_data = on\\\".\\n\\nSo I think it would be ok to throw an ERROR if such a copy_data clash\\nis detected, and then the user will have to change to use \\\"copy_data =\\noff\\\" for some/all of them to avoid data duplication.\\n\\nThe \\\"local_only\\\" option only adds some small logic to this new ERROR,\\nbut it's not really a prerequisite at all.\\n\\ne.g. this whole ERROR part of the patch can be a separate thread.\\n\\n~~~\\n\\n2.2 General - can we remove the \\\"force\\\" enum?\\n\\nNow, because I consider the clashing \\\"copy_data = on\\\" ERROR to be a\\nuser error, I think that is something that the user can already take\\ncare of themselves just using the copy_data = off.\\n\\nI did not really like the modifying of the \\\"copy_data\\\" option from\\njust boolean to some kind of hybrid boolean + \\\"force\\\".\\n\\na) It smells a bit off to me. IMO replication is supposed to end up\\nwith the same (replicated) data on the standby machine but this\\n\\\"force\\\" mode seems to be just helping the user to break that concept\\nand say - \\\"I know what I'm doing, and I don't care if I get lots of\\nduplicated data in the replica table - just let me do it\\\"...\\n\\nb) It also somehow feels like the new \\\"force\\\" was introduced mostly to\\nmake the code ERROR handling implementation simpler, rather than to\\nmake the life of the end-user better. Yes, if force is removed maybe\\nthe copy-clash-detection-code will need to be internally quite more\\ncomplex than it is now, but that is how it should be, instead of\\nputting extra burden on the user (e.g. by complicating the PG docs and\\ngiving them yet more alternatives to configure). I think any clashing\\ncopy_data options really is a user error, but also I think the current\\nboolean copy_data true/false already gives the user a way to fix it.\\n\\nc) Actually, your new error hint messages are similar to my\\nperspective: They already say \\\"Use CREATE/ALTER SUBSCRIPTION with\\ncopy_data = off or force\\\". All I am saying is remove the \\\"force\\\", and\\nthe user can still fix the problem just by using \\\"copy_data = off\\\"\\nappropriately.\\n\\n======\\n\\nSo (from above) I am not much in favour of the copy_data becoming a\\nhybrid enum and using \\\"force\\\", yet that is what most of this patch is\\nimplementing. Anyway, the remainder of my review comments here are for\\nthe code in its current form. Maybe if \\\"force\\\" can be removed most of\\nthe following comments end up being redundant.\\n\\n======\\n\\n2.3 Commit message - wording\\n\\nThis message is difficult to understand.\\n\\nI think that the long sentence \\\"Now when user is trying...\\\" can be\\nbroken into more manageable parts.\\n\\nThis part \\\"and throw an error so that user can handle the initial copy\\ndata.\\\" also seemed a bit vague.\\n\\n~~~\\n\\n2.4 Commit message - more functions\\n\\n\\\"This patch does couple of things:\\\"\\n\\nIIUC, there seems a third thing implemented by this patch but not\\ndescribed by the comment. I think it also adds support for ALTER\\nSUBSCRIPTION SET PUBLICATION WITH (subscribe_local_only)\\n\\n~~~\\n\\n2.5 doc/src/sgml/ref/create_subscription.sgml - wording\\n\\n@@ -161,6 +161,13 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n the replicated changes that was generated from other nodes. The\\n default is <literal>false</literal>.\\n </para>\\n+ <para>\\n+ If the tables in the publication were also subscribing to the data in\\n+ the publisher from other publishers, it will affect the\\n+ <command>CREATE SUBSCRIPTION</command> based on the value specified\\n+ for <literal>copy_data</literal> option. Refer to the\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n+ </para>\\n\\nIs there is a simpler way to express all that?\\n\\nSUGGESTION\\nThere is some interation between the option \\\"subscribe_local_only\\\" and\\noption \\\"copy_data\\\". Refer to the <xref\\nlinkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n\\n~~~\\n\\n2.6 doc/src/sgml/ref/create_subscription.sgml - whitespace\\n\\n@@ -213,18 +220,28 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n </varlistentry>\\n\\n <varlistentry>\\n- <term><literal>copy_data</literal> (<type>boolean</type>)</term>\\n+ <term><literal>copy_data</literal> (<type>enum</type>)</term>\\n <listitem>\\n <para>\\n Specifies whether to copy pre-existing data in the publications\\n- that are being subscribed to when the replication starts.\\n- The default is <literal>true</literal>.\\n+ that are being subscribed to when the replication starts. This\\n+ parameter may be either <literal>true</literal>,\\n+ <literal>false</literal> or <literal>force</literal>. The default is\\n+ <literal>true</literal>.\\n\\nThat last line has trailing whitespace.\\n\\n~~~\\n\\n2.7 doc/src/sgml/ref/create_subscription.sgml - wording\\n\\n+ <para>\\n+ If the tables in the publication were also subscribing to the data in\\n+ the publisher from other publishers, it will affect the\\n+ <command>CREATE SUBSCRIPTION</command> based on the value specified\\n+ for <literal>subscribe_local_only</literal> option. Refer to the\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n+ </para>\\n\\nThis is similar to review comment #2.5 which I thought could be\\nwritten in a simpler way.\\n\\n~~~\\n\\n2.8 doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -375,6 +392,42 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n can have non-existent publications.\\n </para>\\n\\n+ <para>\\n+ Let's consider an existing Multi master logical replication setup between\\n+ Node1 and Node2 that is created using the following steps:\\n+ a) Node1 - Publication publishing employee table.\\n+ b) Node2 - Subscription subscribing from publication pub1 with\\n+ <literal>subscribe_local_only</literal>.\\n+ c) Node2 - Publication publishing employee table.\\n+ d) Node1 - Subscription subscribing from publication pub2 with\\n+ <literal>subscribe_local_only</literal>.\\n+ Now when user is trying to add another node Node3 to the above Multi master\\n+ logical replication setup, user will have to create one subscription\\n+ subscribing from Node1 and another subscription subscribing from Node2 in\\n+ Node3 using <literal>subscribe_local_only</literal> option and\\n+ <literal>copy_data</literal> as <literal>true</literal>, while\\n+ the subscription is created, server will identify that Node2 is subscribing\\n+ from Node1 and Node1 is subscribing from Node2 and throw an error like:\\n+ <programlisting>\\n+postgres=# CREATE SUBSCRIPTION mysub CONNECTION 'host=192.168.1.50\\nport=5432 user=foo dbname=foodb'\\n+ PUBLICATION mypublication, insert_only with\\n(subscribe_local_only=on);\\n+ERROR: CREATE/ALTER SUBSCRIPTION with subscribe_local_only and\\ncopy_data as true is not allowed when the publisher might have\\nreplicated data, table:public.t1 might have replicated data in the\\npublisher\\n+HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\n+ </programlisting>\\n+ In this scenario user can solve this based on one of the 2 possibilities,\\n+ a) If there are no data present in Node1 and Node2, then the user can create\\n+ the subscriptions to Node1 and Node2 with\\n+ <literal>subscribe_local_only</literal> as <literal>true</literal> and\\n+ <literal>copy_data</literal> as <literal>false</literal>. b) If the data is\\n+ present, then the user can create subscription with\\n+ <literal>copy_data</literal> as <literal>force</literal> on Node1 and\\n+ <literal>copy_data</literal> as <literal>false</literal> on Node2, before\\n+ allowing any operations on the respective tables of Node1 and Node2, in this\\n+ case <literal>copy_data</literal> is <literal>false</literal> on Node2\\n+ because the data will be replicated to each other and available on both the\\n+ nodes.\\n+ </para>\\n+\\n\\nThat is a large slab of text in the Notes, so not very easy to digest it.\\n\\nI'm not sure what to suggest for this -\\n- Perhaps the a,b,c,d should all be \\\"lists\\\" so it renders differently?\\n- It almost seems like too much information to be included in this\\n\\\"Notes\\\" section, Maybe it needs its own full page in PG Docs \\\"Logical\\nReplication\\\" to discuss this topic.\\n\\n~~~\\n\\n2.9 src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_VALID\\n\\n@@ -69,6 +69,18 @@\\n /* check if the 'val' has 'bits' set */\\n #define IsSet(val, bits) (((val) & (bits)) == (bits))\\n\\n+#define IS_COPY_DATA_VALID(copy_data) (copy_data != COPY_DATA_OFF)\\n\\nThe macro seems misnamed because \\\"off\\\" is also \\\"valid\\\". It seems like\\nit should be called something different like IS_COPY_DATA, or\\nIS_COPY_DATA_ON_OR_FORCE, etc.\\n\\n~~~\\n\\n2.10 src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n\\n+ /*\\n+ * The set of strings accepted here should match up with\\n+ * the grammar's opt_boolean_or_string production.\\n+ */\\n+ if (pg_strcasecmp(sval, \\\"true\\\") == 0)\\n+ return COPY_DATA_ON;\\n+ if (pg_strcasecmp(sval, \\\"false\\\") == 0)\\n+ return COPY_DATA_OFF;\\n+ if (pg_strcasecmp(sval, \\\"on\\\") == 0)\\n+ return COPY_DATA_ON;\\n+ if (pg_strcasecmp(sval, \\\"off\\\") == 0)\\n+ return COPY_DATA_OFF;\\n+ if (pg_strcasecmp(sval, \\\"force\\\") == 0)\\n+ return COPY_DATA_FORCE;\\n\\nMaybe combine the return for true/on and false/off?\\n\\n~~~\\n\\n2.11 src/backend/commands/subscriptioncmds.c - fetch_table_list\\n\\n+ appendStringInfoString(&cmd,\\n+ \\\"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\\nPS.srsubstate as replicated\\\\n\\\"\\n+ \\\"FROM pg_publication P,\\\\n\\\"\\n+ \\\"LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n+ \\\"LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n+ \\\"pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n+ \\\"WHERE C.oid = GPT.relid AND P.pubname in (\\\");\\n+\\n\\nThat blank line is not needed (it was not there previously) because\\nthe next get_publications_str is a continuation of this SQL.\\n\\n~~~\\n\\n2.12 src/backend/commands/subscriptioncmds.c - fetch_table_list comment\\n\\n+ * It is quite possible that subscriber has not yet pulled data to\\n+ * the tables, but in ideal cases the table data will be subscribed.\\n+ * Too keep the code simple it is not checked if the subscriber table\\n+ * has pulled the data or not.\\n+ */\\n\\nTypo \\\"Too\\\" -> \\\"To\\\"\\n\\n~~~\\n\\n2.13 src/backend/commands/subscriptioncmds.c - force\\n\\n+ if (copydata == COPY_DATA_ON && only_local && !slot_attisnull(slot, 3))\\n+ ereport(ERROR,\\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n+ errmsg(\\\"CREATE/ALTER SUBSCRIPTION with subscribe_local_only and\\ncopy_data as true is not allowed when the publisher might have\\nreplicated data, table:%s.%s might have replicated data in the\\npublisher\\\",\\n+ nspname, relname),\\n+ errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\\"));\\n\\nAFAIK this is the only code of this patch that is distinguishing\\nbetween \\\"force\\\" and the \\\"on\\\". As I wrote at the beginning of this post\\nI feel you can keep this ERROR (maybe the way to detect it needs to be\\nmore complex) and the user can fix it just by using \\\"copy_data = off\\\"\\nappropriately. So then copy_data does not need to be changed.\\n\\n~~~\\n\\n2.14 src/test/regress/sql/subscription.sql - missing tests\\n\\nThe new copy_data is not really an enum true/false/force like the PG\\ndocs claims.\\n\\nIt seems more like some kind of hybrid boolean+force. So if that is\\nhow it is going to be then there are missing test cases to make sure\\nthat values like \\\"on\\\"/\\\"off\\\"/\\\"0\\\"/\\\"1\\\" still are working.\\n\\n~~~\\n\\n2.15 src/test/subscription/t/032_circular.pl\\n\\n@@ -65,6 +69,25 @@ $node_A->safe_psql('postgres', \\\"\\n PUBLICATION tap_pub_B\\n WITH (subscribe_local_only = on, copy_data = off)\\\");\\n\\n+($result, $stdout, $stderr) = $node_A->psql('postgres', \\\"\\n+ CREATE SUBSCRIPTION tap_sub_A1\\n+ CONNECTION '$node_B_connstr application_name=$appname_A'\\n+ PUBLICATION tap_pub_B\\n+ WITH (subscribe_local_only = on, copy_data = on)\\\");\\n+like(\\n+ $stderr,\\n+ qr/ERROR: CREATE\\\\/ALTER SUBSCRIPTION with\\nsubscribe_local_only and copy_data as true is not allowed when the\\npublisher might have replicated data/,\\n+ \\\"Create subscription with subscribe_local_only and copy_data\\nhaving replicated table in publisher\\\");\\n+\\n+$node_A->safe_psql('postgres', \\\"\\n+ CREATE SUBSCRIPTION tap_sub_A2\\n+ CONNECTION '$node_B_connstr application_name=$appname_A'\\n+ PUBLICATION tap_pub_B\\n+ WITH (subscribe_local_only = on, copy_data = force)\\\");\\n+\\n+$node_A->safe_psql('postgres', \\\"\\n+ DROP SUBSCRIPTION tap_sub_A2\\\");\\n+\\n\\nMaybe underneath it is the same, but from the outside, this looks like\\na slightly different scenario from what is mentioned everywhere else\\nin the patch.\\n\\nI think it would be better to create a new Node_C (aka Node3) so then\\nthe TAP test can use the same example that you give in the commit\\nmessage and the PG docs notes.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 12:47:04 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2022 at 12:36 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Here are my comments for the latest patch v6-0001.\\n> >\\n> > (I will post my v6-0002 review comments separately)\\n> >\\n> > PATCH v6-0001 comments\\n> > ======================\\n> >\\n> > 1.1 General - Option name\\n> >\\n> > I still feel like the option name is not ideal. Unfortunately, this is\\n> > important because any name change would impact lots of these patch\\n> > files and docs, struct members etc.\\n> >\\n> > It was originally called \\\"local_only\\\", but I thought that as a\\n> > SUBSCRIPTION option that was confusing because \\\"local\\\" means local to\\n> > what? Really it is local to the publisher, not local to the\\n> > subscriber, so that name seemed misleading.\\n> >\\n> > Then I suggested \\\"publish_local_only\\\". Although that resolved the\\n> > ambiguity problem, other people thought it seemed odd to have the\\n> > \\\"publish\\\" prefix for a subscription-side option.\\n> >\\n> > So now it is changed again to \\\"subscribe_local_only\\\" -- It's getting\\n> > better but still, it is implied that the \\\"local\\\" means local to the\\n> > publisher except there is nothing in the option name really to convey\\n> > that meaning. IMO we here all understand the meaning of this option\\n> > mostly by familiarity with this discussion thread, but I think a user\\n> > coming to this for the first time will still be confused by the name.\\n> >\\n> > Below are some other name ideas. Maybe they are not improvements, but\\n> > it might help other people to come up with something better.\\n> >\\n> > subscribe_publocal_only = true/false\\n> > origin = pub_only/any\\n> > adjacent_data_only = true/false\\n> > direct_subscriptions_only = true/false\\n> > ...\\n> >\\n>\\n> I think local_only with a description should be okay considering other\\n> current options. Some of the options like slot_name, binary, etc. are\\n> publisher-specific which becomes clear only after reading the\\n> description. I am not sure adding pub*/sub* to it is much helpful. So,\\n> +1 to a simple boolean like 'local_only', 'data_local', or something\\n> like that.\\n\\n+1 to use a simple boolean option.\\n\\nThanks for the examples of the other options which are also really\\npublisher-specific. Seen in that light, and with a clear description,\\nprobably the original \\\"local_only\\\" was fine after all.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 13:09:01 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2022 at 8:17 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my comments for the latest patch v6-0002.\\n>\\n> PATCH v6-0002 comments\\n> ======================\\n>\\n> 2.1 General - should this be an independent patch?\\n>\\n\\nI think both the patches are dependent and might get committed\\ntogether if the concept proved to be useful and doesn't have flaws.\\n\\n> In many ways, I think most of this patch is unrelated to the other\\n> \\\"local_only\\\" patch (v6-0001).\\n>\\n> For example, IIUC even in the current HEAD, we could consider it to be\\n> a user error if multiple SUBSCRIPTIONS or multiple PUBLICATIONS of the\\n> same SUBSCRIPTION are replicating to the same TABLE on the same node\\n> and using \\\"copy_data = on\\\".\\n>\\n> So I think it would be ok to throw an ERROR if such a copy_data clash\\n> is detected, and then the user will have to change to use \\\"copy_data =\\n> off\\\" for some/all of them to avoid data duplication.\\n>\\n> The \\\"local_only\\\" option only adds some small logic to this new ERROR,\\n> but it's not really a prerequisite at all.\\n>\\n> e.g. this whole ERROR part of the patch can be a separate thread.\\n>\\n> ~~~\\n>\\n> 2.2 General - can we remove the \\\"force\\\" enum?\\n>\\n> Now, because I consider the clashing \\\"copy_data = on\\\" ERROR to be a\\n> user error, I think that is something that the user can already take\\n> care of themselves just using the copy_data = off.\\n>\\n> I did not really like the modifying of the \\\"copy_data\\\" option from\\n> just boolean to some kind of hybrid boolean + \\\"force\\\".\\n>\\n> a) It smells a bit off to me. IMO replication is supposed to end up\\n> with the same (replicated) data on the standby machine but this\\n> \\\"force\\\" mode seems to be just helping the user to break that concept\\n> and say - \\\"I know what I'm doing, and I don't care if I get lots of\\n> duplicated data in the replica table - just let me do it\\\"...\\n>\\n> b) It also somehow feels like the new \\\"force\\\" was introduced mostly to\\n> make the code ERROR handling implementation simpler, rather than to\\n> make the life of the end-user better. Yes, if force is removed maybe\\n> the copy-clash-detection-code will need to be internally quite more\\n> complex than it is now, but that is how it should be, instead of\\n> putting extra burden on the user (e.g. by complicating the PG docs and\\n> giving them yet more alternatives to configure). I think any clashing\\n> copy_data options really is a user error, but also I think the current\\n> boolean copy_data true/false already gives the user a way to fix it.\\n>\\n> c) Actually, your new error hint messages are similar to my\\n> perspective: They already say \\\"Use CREATE/ALTER SUBSCRIPTION with\\n> copy_data = off or force\\\". All I am saying is remove the \\\"force\\\", and\\n> the user can still fix the problem just by using \\\"copy_data = off\\\"\\n> appropriately.\\n>\\n\\nHow will it work if there is more than one table? copy_data = \\\"off\\\"\\nmeans it won't copy any of the other tables in the corresponding\\npublication(s). I think it is primarily to give the user some option\\nwhere she understands the logical replication setup better and\\nunderstands that this won't be a problem.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 11:45:50 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my comments for the latest patch v6-0001.\\n>\\n> (I will post my v6-0002 review comments separately)\\n>\\n> PATCH v6-0001 comments\\n> ======================\\n>\\n> 1.1 General - Option name\\n>\\n> I still feel like the option name is not ideal. Unfortunately, this is\\n> important because any name change would impact lots of these patch\\n> files and docs, struct members etc.\\n>\\n> It was originally called \\\"local_only\\\", but I thought that as a\\n> SUBSCRIPTION option that was confusing because \\\"local\\\" means local to\\n> what? Really it is local to the publisher, not local to the\\n> subscriber, so that name seemed misleading.\\n>\\n> Then I suggested \\\"publish_local_only\\\". Although that resolved the\\n> ambiguity problem, other people thought it seemed odd to have the\\n> \\\"publish\\\" prefix for a subscription-side option.\\n>\\n> So now it is changed again to \\\"subscribe_local_only\\\" -- It's getting\\n> better but still, it is implied that the \\\"local\\\" means local to the\\n> publisher except there is nothing in the option name really to convey\\n> that meaning. IMO we here all understand the meaning of this option\\n> mostly by familiarity with this discussion thread, but I think a user\\n> coming to this for the first time will still be confused by the name.\\n>\\n> Below are some other name ideas. Maybe they are not improvements, but\\n> it might help other people to come up with something better.\\n>\\n> subscribe_publocal_only = true/false\\n> origin = pub_only/any\\n> adjacent_data_only = true/false\\n> direct_subscriptions_only = true/false\\n> ...\\n\\nFWIW, The subscriber wants \\\"changes originated on publisher\\\". From\\nthat angle origin = publisher/any looks attractive. It also leaves\\nopen the possibility that the subscriber may ask changes from a set of\\norigins or even non-publisher origins.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 19:06:18 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2022 at 7:06 PM Ashutosh Bapat\\n<ashutosh.bapat.oss@gmail.com> wrote:\\n>\\n> On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Below are some other name ideas. Maybe they are not improvements, but\\n> > it might help other people to come up with something better.\\n> >\\n> > subscribe_publocal_only = true/false\\n> > origin = pub_only/any\\n> > adjacent_data_only = true/false\\n> > direct_subscriptions_only = true/false\\n> > ...\\n>\\n> FWIW, The subscriber wants \\\"changes originated on publisher\\\". From\\n> that angle origin = publisher/any looks attractive. It also leaves\\n> open the possibility that the subscriber may ask changes from a set of\\n> origins or even non-publisher origins.\\n>\\n\\nSo, how are you imagining extending it for multiple origins? I think\\nwe can have multiple origins data on a node when there are multiple\\nsubscriptions pointing to the different or same node. The origin names\\nare internally generated names but are present in\\npg_replication_origin. So, are we going to ask the user to check that\\ntable and specify it as an option? But, note, that the user may not be\\nable to immediately recognize which origin data is useful for her.\\nSurely, with some help or advice in docs, she may be able to identify\\nit but that doesn't seem so straightforward. The other point is that\\nsuch an option won't be applicable for initial sync as we can't\\ndifferentiate between data from a local or remote node.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Apr 2022 09:36:52 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"FYI, here is a test script that is using the current patch (v6) to\\ndemonstrate a way to share table data between different numbers of\\nnodes (up to 5 of them here).\\n\\nThe script starts off with just 2-way sharing (nodes N1, N2),\\nthen expands to 3-way sharing (nodes N1, N2, N3),\\nthen 4-way sharing (nodes N1, N2, N3, N4),\\nthen 5-way sharing (nodes N1, N2, N3, N4, N5).\\n\\nAs an extra complication, for this test, all 5 nodes have different\\ninitial table data, which gets replicated to the others whenever each\\nnew node joins the existing share group.\\n\\nPSA.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\",\n \"msg_date\": \"Thu, 7 Apr 2022 14:09:23 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Dear Peter,\\r\\n\\r\\n> FYI, here is a test script that is using the current patch (v6) to\\r\\n> demonstrate a way to share table data between different numbers of\\r\\n> nodes (up to 5 of them here).\\r\\n\\r\\nThanks for sharing your script! It's very helpful for us.\\r\\n\\r\\nWhile reading your script, however, I had a question about it.\\r\\nLine 121-122, you defined subscriptions for 2-nodes cluster:\\r\\n\\r\\npsql -p $port_N1 -c \\\"create subscription sub12 connection 'port=$port_N2' publication pub2 with ($copy_force);\\\"\\r\\npsql -p $port_N2 -c \\\"create subscription sub21 connection 'port=$port_N1' publication pub1 with ($copy_force);\\\"\\r\\n\\r\\nBut I was not sure it works well.\\r\\nN2 already have shared data from N1 when subscription sub21 is created.\\r\\nDid you assume that the initial copying is not so quick and\\r\\ndata synchronization will be not done when creating sub21?\\r\\n\\r\\nBest Regards,\\r\\nHayato Kuroda\\r\\nFUJITSU LIMITED\\r\\n\\r\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 06:03:34 +0000\",\n \"msg_from\": \"\\\"kuroda.hayato@fujitsu.com\\\" <kuroda.hayato@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 7, 2022 at 4:03 PM kuroda.hayato@fujitsu.com\\n<kuroda.hayato@fujitsu.com> wrote:\\n>\\n> Dear Peter,\\n>\\n> > FYI, here is a test script that is using the current patch (v6) to\\n> > demonstrate a way to share table data between different numbers of\\n> > nodes (up to 5 of them here).\\n>\\n> Thanks for sharing your script! It's very helpful for us.\\n>\\n> While reading your script, however, I had a question about it.\\n> Line 121-122, you defined subscriptions for 2-nodes cluster:\\n>\\n> psql -p $port_N1 -c \\\"create subscription sub12 connection 'port=$port_N2' publication pub2 with ($copy_force);\\\"\\n> psql -p $port_N2 -c \\\"create subscription sub21 connection 'port=$port_N1' publication pub1 with ($copy_force);\\\"\\n>\\n> But I was not sure it works well.\\n> N2 already have shared data from N1 when subscription sub21 is created.\\n> Did you assume that the initial copying is not so quick and\\n> data synchronization will be not done when creating sub21?\\n\\nOops. Good catch.\\n\\nAlthough the 2-way test was working OK for me, I think that it worked\\nonly because of lucky timing. e.g. When I put a delay between those 2\\nsubscriptions then the 2nd one would cause the PK violation that\\nprobably you were anticipating would happen.\\n\\nI have modified the 2-way example to use the same truncate pattern as others.\\n\\nPSA the fixed test.sh script and accompanying files.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\",\n \"msg_date\": \"Thu, 7 Apr 2022 16:13:56 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi Vignesh, FYI the patch is recently broken again and is failing on cfbot [1].\\n\\n------\\n[1] http://cfbot.cputube.org/patch_38_3610.log\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Apr 2022 09:07:45 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 8, 2022 at 4:38 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Hi Vignesh, FYI the patch is recently broken again and is failing on cfbot [1].\\n>\\n> ------\\n> [1] http://cfbot.cputube.org/patch_38_3610.log\\n\\nI'm working on fixing a few review comments, I will be posting an\\nupdated version to handle this today.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Apr 2022 06:41:50 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my comments for the latest patch v6-0001.\\n>\\n> (I will post my v6-0002 review comments separately)\\n>\\n> PATCH v6-0001 comments\\n> ======================\\n>\\n> 1.1 General - Option name\\n>\\n> I still feel like the option name is not ideal. Unfortunately, this is\\n> important because any name change would impact lots of these patch\\n> files and docs, struct members etc.\\n>\\n> It was originally called \\\"local_only\\\", but I thought that as a\\n> SUBSCRIPTION option that was confusing because \\\"local\\\" means local to\\n> what? Really it is local to the publisher, not local to the\\n> subscriber, so that name seemed misleading.\\n>\\n> Then I suggested \\\"publish_local_only\\\". Although that resolved the\\n> ambiguity problem, other people thought it seemed odd to have the\\n> \\\"publish\\\" prefix for a subscription-side option.\\n>\\n> So now it is changed again to \\\"subscribe_local_only\\\" -- It's getting\\n> better but still, it is implied that the \\\"local\\\" means local to the\\n> publisher except there is nothing in the option name really to convey\\n> that meaning. IMO we here all understand the meaning of this option\\n> mostly by familiarity with this discussion thread, but I think a user\\n> coming to this for the first time will still be confused by the name.\\n>\\n> Below are some other name ideas. Maybe they are not improvements, but\\n> it might help other people to come up with something better.\\n>\\n> subscribe_publocal_only = true/false\\n> origin = pub_only/any\\n> adjacent_data_only = true/false\\n> direct_subscriptions_only = true/false\\n> ...\\n>\\n>\\n> (FYI, the remainder of these review comments will assume the option is\\n> still called \\\"subscribe_local_only\\\")\\n\\nModified to local_only\\n\\n> ~~~\\n>\\n> 1.2 General - inconsistent members and args\\n>\\n> IMO the struct members and args should also be named for close\\n> consistency with whatever the option name is.\\n>\\n> Currently the option is called \\\"subscription_local_only\\\". So I think\\n> the members/args would be better to be called \\\"local_only\\\" instead of\\n> \\\"only_local\\\".\\n\\nModified\\n\\n> ~~~\\n>\\n> 1.3 Commit message - wrong option name\\n>\\n> The commit message refers to the option name as \\\"publish_local_only\\\"\\n> instead of the option name that is currently implemented.\\n\\nModified\\n\\n> ~~~\\n>\\n> 1.4 Commit message - wording\\n>\\n> The wording seems a bit off. Below is suggested simpler wording which\\n> I AFAIK conveys the same information.\\n>\\n> BEFORE\\n> Add an option publish_local_only which will subscribe only to the locally\\n> generated data in the publisher node. If subscriber is created with this\\n> option, publisher will skip publishing the data that was subscribed\\n> from other nodes. It can be created using following syntax:\\n> ex: CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\\n> PUBLICATION pub1 with (publish_local_only = on);\\n>\\n> SUGGESTION\\n> This patch adds a new SUBSCRIPTION boolean option\\n> \\\"subscribe_local_only\\\". The default is false. When a SUBSCRIPTION is\\n> created with this option enabled, the publisher will only publish data\\n> that originated at the publisher node.\\n> Usage:\\n> CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\\n> PUBLICATION pub1 with (subscribe_local_only = true);\\n\\nModified\\n\\n> ~~~\\n>\\n> 1.5 doc/src/sgml/ref/create_subscription.sgml - \\\"generated\\\" changes.\\n>\\n> + <para>\\n> + Specifies whether the subscription will request the publisher to send\\n> + locally generated changes or both the locally generated changes and\\n> + the replicated changes that was generated from other nodes. The\\n> + default is <literal>false</literal>.\\n> + </para>\\n>\\n> For some reason, it seemed a bit strange to me to use the term\\n> \\\"generated\\\" changes. Maybe better to refer to the origin of changes?\\n>\\n> SUGGESTION\\n> Specifies whether the publisher should send only changes that\\n> originated locally at the publisher node, or send any publisher node\\n> changes regardless of their origin. The default is false.\\n\\nModified\\n\\n> ~~~\\n>\\n> 1.6 src/backend/replication/pgoutput/pgoutput.c -\\n> LOGICALREP_PROTO_TWOPHASE_VERSION_NUM\\n>\\n> @@ -496,6 +509,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\\n> OutputPluginOptions *opt,\\n> else\\n> ctx->twophase_opt_given = true;\\n>\\n> + if (data->only_local && data->protocol_version <\\n> LOGICALREP_PROTO_TWOPHASE_VERSION_NUM)\\n> + ereport(ERROR,\\n> + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n> + errmsg(\\\"requested proto_version=%d does not support\\n> subscribe_local_only, need %d or higher\\\",\\n> + data->protocol_version, LOGICALREP_PROTO_TWOPHASE_VERSION_NUM)));\\n>\\n> I thought this code should not be using\\n> LOGICALREP_PROTO_TWOPHASE_VERSION_NUM. Shouldn't there be some newly\\n> introduced constant like LOGICALREP_PROTO_LOCALONLY_VERSION_NUM which\\n> you will use here?\\n\\nModified, I have set LOGICALREP_PROTO_LOCALONLY_VERSION_NUM to same\\nvalue as LOGICALREP_PROTO_TWOPHASE_VERSION_NUM, will increment this\\nonce server version is changed.\\n\\n> ~~~\\n>\\n> 1.7 src/bin/pg_dump/pg_dump.c - 150000\\n>\\n> @@ -4451,11 +4452,13 @@ getSubscriptions(Archive *fout)\\n> if (fout->remoteVersion >= 150000)\\n> appendPQExpBufferStr(query,\\n> \\\" s.subtwophasestate,\\\\n\\\"\\n> - \\\" s.subdisableonerr\\\\n\\\");\\n> + \\\" s.subdisableonerr,\\\\n\\\"\\n> + \\\" s.sublocal\\\\n\\\");\\n> else\\n> appendPQExpBuffer(query,\\n> \\\" '%c' AS subtwophasestate,\\\\n\\\"\\n> - \\\" false AS subdisableonerr\\\\n\\\",\\n> + \\\" false AS subdisableonerr,\\\\n\\\"\\n> + \\\" false AS s.sublocal\\\\n\\\",\\n> LOGICALREP_TWOPHASE_STATE_DISABLED);\\n>\\n> I think this local_only feature is unlikely to get into the PG15\\n> release, so this code should be split out into a separate condition\\n> because later will need to change to say >= 160000.\\n\\n I have split the condition and checked it with 150000, this will be\\nchanged later to 160000 after the branch is created.\\n\\n> ~~~\\n>\\n> 1.8 src/bin/pg_dump/pg_dump.c - dumpSubscription\\n>\\n> @@ -4585,6 +4591,9 @@ dumpSubscription(Archive *fout, const\\n> SubscriptionInfo *subinfo)\\n> if (strcmp(subinfo->subdisableonerr, \\\"t\\\") == 0)\\n> appendPQExpBufferStr(query, \\\", disable_on_error = true\\\");\\n>\\n> + if (strcmp(subinfo->sublocal, \\\"f\\\") != 0)\\n> + appendPQExpBufferStr(query, \\\", subscribe_local_only = on\\\");\\n> +\\n>\\n> I felt it is more natural to say \\\"if it is true set to true\\\", instead\\n> of \\\"if it is not false set to on\\\".\\n>\\n> SUGGESTION\\n> if (strcmp(subinfo->sublocal, \\\"t\\\") == 0)\\n> appendPQExpBufferStr(query, \\\", subscribe_local_only = true\\\");\\n\\nModified\\n\\n> ~~~\\n>\\n> 1.9 src/bin/psql/describe.c - 150000\\n>\\n> @@ -6318,9 +6318,11 @@ describeSubscriptions(const char *pattern, bool verbose)\\n> if (pset.sversion >= 150000)\\n> appendPQExpBuffer(&buf,\\n> \\\", subtwophasestate AS \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n> - \\\", subdisableonerr AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n> + \\\", subdisableonerr AS \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n> + \\\", sublocal AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n> gettext_noop(\\\"Two phase commit\\\"),\\n> - gettext_noop(\\\"Disable on error\\\"));\\n> + gettext_noop(\\\"Disable on error\\\"),\\n> + gettext_noop(\\\"Only local\\\"));\\n>\\n> I think this local_only feature is unlikely to get into the PG15\\n> release, so this code should be split out into a separate condition\\n> because later will need to change to say >= 160000.\\n\\nI have split the condition and checked it with 150000, this will be\\nchanged later to 160000 after the branch is created.\\n\\n> ~~\\n>\\n> 1.10 src/bin/psql/describe.c - describeSubscriptions column\\n>\\n> @@ -6318,9 +6318,11 @@ describeSubscriptions(const char *pattern, bool verbose)\\n> if (pset.sversion >= 150000)\\n> appendPQExpBuffer(&buf,\\n> \\\", subtwophasestate AS \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n> - \\\", subdisableonerr AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n> + \\\", subdisableonerr AS \\\\\\\"%s\\\\\\\"\\\\n\\\"\\n> + \\\", sublocal AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n> gettext_noop(\\\"Two phase commit\\\"),\\n> - gettext_noop(\\\"Disable on error\\\"));\\n> + gettext_noop(\\\"Disable on error\\\"),\\n> + gettext_noop(\\\"Only local\\\"));\\n>\\n> I think the column name here should be more consistent with the option\\n> name. e.g. it should be \\\"Local only\\\", not \\\"Only local\\\".\\n\\nModified\\n\\n> ~~~\\n>\\n> 1.11 src/bin/psql/tab-complete.c - whitespace\\n>\\n> @@ -3167,7 +3167,7 @@ psql_completion(const char *text, int start, int end)\\n> /* Complete \\\"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\\\" */\\n> else if (HeadMatches(\\\"CREATE\\\", \\\"SUBSCRIPTION\\\") && TailMatches(\\\"WITH\\\", \\\"(\\\"))\\n> COMPLETE_WITH(\\\"binary\\\", \\\"connect\\\", \\\"copy_data\\\", \\\"create_slot\\\",\\n> - \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> + \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\", \\\"subscribe_local_only\\\",\\n>\\n> The patch accidentally added a space char before the \\\"slot_name\\\".\\n\\nModified\\n\\n> ~~~\\n>\\n> 1.12 src/include/replication/walreceiver.h - \\\"generated\\\"\\n>\\n> @@ -183,6 +183,7 @@ typedef struct\\n> bool streaming; /* Streaming of large transactions */\\n> bool twophase; /* Streaming of two-phase transactions at\\n> * prepare time */\\n> + bool only_local; /* publish only locally generated data */\\n>\\n> This is a similar review comment as #1.5 about saying the word \\\"generated\\\".\\n> Maybe there is another way to word this?\\n\\nModified\\n\\n> ~~~\\n>\\n> 1.13 src/test/regress/sql/subscription.sql - missing test case\\n>\\n> Isn't there a missing test case for ensuring that the new option is boolean?\\n\\nAdded test\\n\\nThanks for the comments, the attached v7 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Fri, 8 Apr 2022 13:00:04 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 5, 2022 at 8:17 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my comments for the latest patch v6-0002.\\n>\\n> PATCH v6-0002 comments\\n> ======================\\n>\\n> 2.1 General - should this be an independent patch?\\n>\\n> In many ways, I think most of this patch is unrelated to the other\\n> \\\"local_only\\\" patch (v6-0001).\\n>\\n> For example, IIUC even in the current HEAD, we could consider it to be\\n> a user error if multiple SUBSCRIPTIONS or multiple PUBLICATIONS of the\\n> same SUBSCRIPTION are replicating to the same TABLE on the same node\\n> and using \\\"copy_data = on\\\".\\n>\\n> So I think it would be ok to throw an ERROR if such a copy_data clash\\n> is detected, and then the user will have to change to use \\\"copy_data =\\n> off\\\" for some/all of them to avoid data duplication.\\n>\\n> The \\\"local_only\\\" option only adds some small logic to this new ERROR,\\n> but it's not really a prerequisite at all.\\n>\\n> e.g. this whole ERROR part of the patch can be a separate thread.\\n\\nAs Amit also pointed out, the patches have some dependency, I will\\nkeep it as it is.\\n\\n> ~~~\\n>\\n> 2.2 General - can we remove the \\\"force\\\" enum?\\n>\\n> Now, because I consider the clashing \\\"copy_data = on\\\" ERROR to be a\\n> user error, I think that is something that the user can already take\\n> care of themselves just using the copy_data = off.\\n>\\n> I did not really like the modifying of the \\\"copy_data\\\" option from\\n> just boolean to some kind of hybrid boolean + \\\"force\\\".\\n>\\n> a) It smells a bit off to me. IMO replication is supposed to end up\\n> with the same (replicated) data on the standby machine but this\\n> \\\"force\\\" mode seems to be just helping the user to break that concept\\n> and say - \\\"I know what I'm doing, and I don't care if I get lots of\\n> duplicated data in the replica table - just let me do it\\\"...\\n>\\n> b) It also somehow feels like the new \\\"force\\\" was introduced mostly to\\n> make the code ERROR handling implementation simpler, rather than to\\n> make the life of the end-user better. Yes, if force is removed maybe\\n> the copy-clash-detection-code will need to be internally quite more\\n> complex than it is now, but that is how it should be, instead of\\n> putting extra burden on the user (e.g. by complicating the PG docs and\\n> giving them yet more alternatives to configure). I think any clashing\\n> copy_data options really is a user error, but also I think the current\\n> boolean copy_data true/false already gives the user a way to fix it.\\n>\\n> c) Actually, your new error hint messages are similar to my\\n> perspective: They already say \\\"Use CREATE/ALTER SUBSCRIPTION with\\n> copy_data = off or force\\\". All I am saying is remove the \\\"force\\\", and\\n> the user can still fix the problem just by using \\\"copy_data = off\\\"\\n> appropriately.\\n\\nWhen a user is trying to add a node to bidirectional replication\\nsetup, the user will use the force option to copy the data from one of\\nthe nodes and use off to skip copying the data from other nodes. This\\noption will be used while adding nodes to bidirectional replication,\\nthe same has been documented with examples in the patch. I felt we\\nshould retain this option.\\n\\n> ======\\n>\\n> So (from above) I am not much in favour of the copy_data becoming a\\n> hybrid enum and using \\\"force\\\", yet that is what most of this patch is\\n> implementing. Anyway, the remainder of my review comments here are for\\n> the code in its current form. Maybe if \\\"force\\\" can be removed most of\\n> the following comments end up being redundant.\\n>\\n> ======\\n>\\n> 2.3 Commit message - wording\\n>\\n> This message is difficult to understand.\\n>\\n> I think that the long sentence \\\"Now when user is trying...\\\" can be\\n> broken into more manageable parts.\\n\\nSlightly modified\\n\\n> This part \\\"and throw an error so that user can handle the initial copy\\n> data.\\\" also seemed a bit vague.\\n\\nI have given the reference to the documentation section of the patch\\nfor initial data copy handling\\n\\n> ~~~\\n>\\n> 2.4 Commit message - more functions\\n>\\n> \\\"This patch does couple of things:\\\"\\n>\\n> IIUC, there seems a third thing implemented by this patch but not\\n> described by the comment. I think it also adds support for ALTER\\n> SUBSCRIPTION SET PUBLICATION WITH (subscribe_local_only)\\n\\nThis is part of 0001 patch, I felt this should not be part of 002 patch commit.\\n\\n> ~~~\\n>\\n> 2.5 doc/src/sgml/ref/create_subscription.sgml - wording\\n>\\n> @@ -161,6 +161,13 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> the replicated changes that was generated from other nodes. The\\n> default is <literal>false</literal>.\\n> </para>\\n> + <para>\\n> + If the tables in the publication were also subscribing to the data in\\n> + the publisher from other publishers, it will affect the\\n> + <command>CREATE SUBSCRIPTION</command> based on the value specified\\n> + for <literal>copy_data</literal> option. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n> + </para>\\n>\\n> Is there is a simpler way to express all that?\\n>\\n> SUGGESTION\\n> There is some interation between the option \\\"subscribe_local_only\\\" and\\n> option \\\"copy_data\\\". Refer to the <xref\\n> linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.6 doc/src/sgml/ref/create_subscription.sgml - whitespace\\n>\\n> @@ -213,18 +220,28 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> </varlistentry>\\n>\\n> <varlistentry>\\n> - <term><literal>copy_data</literal> (<type>boolean</type>)</term>\\n> + <term><literal>copy_data</literal> (<type>enum</type>)</term>\\n> <listitem>\\n> <para>\\n> Specifies whether to copy pre-existing data in the publications\\n> - that are being subscribed to when the replication starts.\\n> - The default is <literal>true</literal>.\\n> + that are being subscribed to when the replication starts. This\\n> + parameter may be either <literal>true</literal>,\\n> + <literal>false</literal> or <literal>force</literal>. The default is\\n> + <literal>true</literal>.\\n>\\n> That last line has trailing whitespace.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.7 doc/src/sgml/ref/create_subscription.sgml - wording\\n>\\n> + <para>\\n> + If the tables in the publication were also subscribing to the data in\\n> + the publisher from other publishers, it will affect the\\n> + <command>CREATE SUBSCRIPTION</command> based on the value specified\\n> + for <literal>subscribe_local_only</literal> option. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n> + </para>\\n>\\n> This is similar to review comment #2.5 which I thought could be\\n> written in a simpler way.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.8 doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -375,6 +392,42 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> can have non-existent publications.\\n> </para>\\n>\\n> + <para>\\n> + Let's consider an existing Multi master logical replication setup between\\n> + Node1 and Node2 that is created using the following steps:\\n> + a) Node1 - Publication publishing employee table.\\n> + b) Node2 - Subscription subscribing from publication pub1 with\\n> + <literal>subscribe_local_only</literal>.\\n> + c) Node2 - Publication publishing employee table.\\n> + d) Node1 - Subscription subscribing from publication pub2 with\\n> + <literal>subscribe_local_only</literal>.\\n> + Now when user is trying to add another node Node3 to the above Multi master\\n> + logical replication setup, user will have to create one subscription\\n> + subscribing from Node1 and another subscription subscribing from Node2 in\\n> + Node3 using <literal>subscribe_local_only</literal> option and\\n> + <literal>copy_data</literal> as <literal>true</literal>, while\\n> + the subscription is created, server will identify that Node2 is subscribing\\n> + from Node1 and Node1 is subscribing from Node2 and throw an error like:\\n> + <programlisting>\\n> +postgres=# CREATE SUBSCRIPTION mysub CONNECTION 'host=192.168.1.50\\n> port=5432 user=foo dbname=foodb'\\n> + PUBLICATION mypublication, insert_only with\\n> (subscribe_local_only=on);\\n> +ERROR: CREATE/ALTER SUBSCRIPTION with subscribe_local_only and\\n> copy_data as true is not allowed when the publisher might have\\n> replicated data, table:public.t1 might have replicated data in the\\n> publisher\\n> +HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\n> + </programlisting>\\n> + In this scenario user can solve this based on one of the 2 possibilities,\\n> + a) If there are no data present in Node1 and Node2, then the user can create\\n> + the subscriptions to Node1 and Node2 with\\n> + <literal>subscribe_local_only</literal> as <literal>true</literal> and\\n> + <literal>copy_data</literal> as <literal>false</literal>. b) If the data is\\n> + present, then the user can create subscription with\\n> + <literal>copy_data</literal> as <literal>force</literal> on Node1 and\\n> + <literal>copy_data</literal> as <literal>false</literal> on Node2, before\\n> + allowing any operations on the respective tables of Node1 and Node2, in this\\n> + case <literal>copy_data</literal> is <literal>false</literal> on Node2\\n> + because the data will be replicated to each other and available on both the\\n> + nodes.\\n> + </para>\\n> +\\n>\\n> That is a large slab of text in the Notes, so not very easy to digest it.\\n>\\n> I'm not sure what to suggest for this -\\n> - Perhaps the a,b,c,d should all be \\\"lists\\\" so it renders differently?\\n> - It almost seems like too much information to be included in this\\n> \\\"Notes\\\" section, Maybe it needs its own full page in PG Docs \\\"Logical\\n> Replication\\\" to discuss this topic.\\n\\nThis is moved to the new page \\\"BiDirectional logical replication Quick\\nSetup\\\" page now.\\n\\n> ~~~\\n>\\n> 2.9 src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_VALID\\n>\\n> @@ -69,6 +69,18 @@\\n> /* check if the 'val' has 'bits' set */\\n> #define IsSet(val, bits) (((val) & (bits)) == (bits))\\n>\\n> +#define IS_COPY_DATA_VALID(copy_data) (copy_data != COPY_DATA_OFF)\\n>\\n> The macro seems misnamed because \\\"off\\\" is also \\\"valid\\\". It seems like\\n> it should be called something different like IS_COPY_DATA, or\\n> IS_COPY_DATA_ON_OR_FORCE, etc.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.10 src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n>\\n> + /*\\n> + * The set of strings accepted here should match up with\\n> + * the grammar's opt_boolean_or_string production.\\n> + */\\n> + if (pg_strcasecmp(sval, \\\"true\\\") == 0)\\n> + return COPY_DATA_ON;\\n> + if (pg_strcasecmp(sval, \\\"false\\\") == 0)\\n> + return COPY_DATA_OFF;\\n> + if (pg_strcasecmp(sval, \\\"on\\\") == 0)\\n> + return COPY_DATA_ON;\\n> + if (pg_strcasecmp(sval, \\\"off\\\") == 0)\\n> + return COPY_DATA_OFF;\\n> + if (pg_strcasecmp(sval, \\\"force\\\") == 0)\\n> + return COPY_DATA_FORCE;\\n>\\n> Maybe combine the return for true/on and false/off?\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.11 src/backend/commands/subscriptioncmds.c - fetch_table_list\\n>\\n> + appendStringInfoString(&cmd,\\n> + \\\"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\\n> PS.srsubstate as replicated\\\\n\\\"\\n> + \\\"FROM pg_publication P,\\\\n\\\"\\n> + \\\"LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n> + \\\"LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n> + \\\"pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n> + \\\"WHERE C.oid = GPT.relid AND P.pubname in (\\\");\\n> +\\n>\\n> That blank line is not needed (it was not there previously) because\\n> the next get_publications_str is a continuation of this SQL.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.12 src/backend/commands/subscriptioncmds.c - fetch_table_list comment\\n>\\n> + * It is quite possible that subscriber has not yet pulled data to\\n> + * the tables, but in ideal cases the table data will be subscribed.\\n> + * Too keep the code simple it is not checked if the subscriber table\\n> + * has pulled the data or not.\\n> + */\\n>\\n> Typo \\\"Too\\\" -> \\\"To\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.13 src/backend/commands/subscriptioncmds.c - force\\n>\\n> + if (copydata == COPY_DATA_ON && only_local && !slot_attisnull(slot, 3))\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"CREATE/ALTER SUBSCRIPTION with subscribe_local_only and\\n> copy_data as true is not allowed when the publisher might have\\n> replicated data, table:%s.%s might have replicated data in the\\n> publisher\\\",\\n> + nspname, relname),\\n> + errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\\"));\\n>\\n> AFAIK this is the only code of this patch that is distinguishing\\n> between \\\"force\\\" and the \\\"on\\\". As I wrote at the beginning of this post\\n> I feel you can keep this ERROR (maybe the way to detect it needs to be\\n> more complex) and the user can fix it just by using \\\"copy_data = off\\\"\\n> appropriately. So then copy_data does not need to be changed.\\n\\nWhen a user is trying to add a node to bidirectional replication\\nsetup, the user will use the force option to copy the data from one of\\nthe nodes and use off to skip copying the data from other nodes. This\\noption will be used while adding nodes to bidirectional replication,\\nthe same has been documented with examples in the patch. I felt we\\nshould retain this option.\\n\\n> ~~~\\n>\\n> 2.14 src/test/regress/sql/subscription.sql - missing tests\\n>\\n> The new copy_data is not really an enum true/false/force like the PG\\n> docs claims.\\n>\\n> It seems more like some kind of hybrid boolean+force. So if that is\\n> how it is going to be then there are missing test cases to make sure\\n> that values like \\\"on\\\"/\\\"off\\\"/\\\"0\\\"/\\\"1\\\" still are working.\\n\\nAdded tests\\n\\n> ~~~\\n>\\n> 2.15 src/test/subscription/t/032_circular.pl\\n>\\n> @@ -65,6 +69,25 @@ $node_A->safe_psql('postgres', \\\"\\n> PUBLICATION tap_pub_B\\n> WITH (subscribe_local_only = on, copy_data = off)\\\");\\n>\\n> +($result, $stdout, $stderr) = $node_A->psql('postgres', \\\"\\n> + CREATE SUBSCRIPTION tap_sub_A1\\n> + CONNECTION '$node_B_connstr application_name=$appname_A'\\n> + PUBLICATION tap_pub_B\\n> + WITH (subscribe_local_only = on, copy_data = on)\\\");\\n> +like(\\n> + $stderr,\\n> + qr/ERROR: CREATE\\\\/ALTER SUBSCRIPTION with\\n> subscribe_local_only and copy_data as true is not allowed when the\\n> publisher might have replicated data/,\\n> + \\\"Create subscription with subscribe_local_only and copy_data\\n> having replicated table in publisher\\\");\\n> +\\n> +$node_A->safe_psql('postgres', \\\"\\n> + CREATE SUBSCRIPTION tap_sub_A2\\n> + CONNECTION '$node_B_connstr application_name=$appname_A'\\n> + PUBLICATION tap_pub_B\\n> + WITH (subscribe_local_only = on, copy_data = force)\\\");\\n> +\\n> +$node_A->safe_psql('postgres', \\\"\\n> + DROP SUBSCRIPTION tap_sub_A2\\\");\\n> +\\n>\\n> Maybe underneath it is the same, but from the outside, this looks like\\n> a slightly different scenario from what is mentioned everywhere else\\n> in the patch.\\n>\\n> I think it would be better to create a new Node_C (aka Node3) so then\\n> the TAP test can use the same example that you give in the commit\\n> message and the PG docs notes.\\n\\nModified the tests to include Node_C.\\n\\nThanks for the comments, the v7 patch attached at [1] has the fixes for the same\\n[1] - https://www.postgresql.org/message-id/CALDaNm1ei%3DrRwCBKWtUu8b5OsS6FFcvaxg9h0oXcjgFn8GoZnQ%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Apr 2022 13:10:05 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 7, 2022 at 2:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> FYI, here is a test script that is using the current patch (v6) to\\n> demonstrate a way to share table data between different numbers of\\n> nodes (up to 5 of them here).\\n>\\n> The script starts off with just 2-way sharing (nodes N1, N2),\\n> then expands to 3-way sharing (nodes N1, N2, N3),\\n> then 4-way sharing (nodes N1, N2, N3, N4),\\n> then 5-way sharing (nodes N1, N2, N3, N4, N5).\\n>\\n> As an extra complication, for this test, all 5 nodes have different\\n> initial table data, which gets replicated to the others whenever each\\n> new node joins the existing share group.\\n>\\n> PSA.\\n>\\n\\n\\nHi Vignesh. I had some problems getting the above test script working.\\nIt was OK up until I tried to join the 5th node (N5) to the existing 4\\nnodes. The ERROR was manifesting itself strangely because it appeared\\nthat there was an index violation in the pg_subscription_rel catalog\\neven though AFAIK the N5 did not have any entries in it.\\n\\n\\ne.g.\\n2022-04-07 09:13:28.361 AEST [24237] ERROR: duplicate key value\\nviolates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n2022-04-07 09:13:28.361 AEST [24237] DETAIL: Key (srrelid,\\nsrsubid)=(16384, 16393) already exists.\\n2022-04-07 09:13:28.361 AEST [24237] STATEMENT: create subscription\\nsub51 connection 'port=7651' publication pub1 with\\n(subscribe_local_only=true,copy_data=force);\\n2022-04-07 09:13:28.380 AEST [24237] ERROR: duplicate key value\\nviolates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n2022-04-07 09:13:28.380 AEST [24237] DETAIL: Key (srrelid,\\nsrsubid)=(16384, 16394) already exists.\\n2022-04-07 09:13:28.380 AEST [24237] STATEMENT: create subscription\\nsub52 connection 'port=7652' publication pub2 with\\n(subscribe_local_only=true,copy_data=false);\\n2022-04-07 09:13:28.405 AEST [24237] ERROR: duplicate key value\\nviolates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n2022-04-07 09:13:28.405 AEST [24237] DETAIL: Key (srrelid,\\nsrsubid)=(16384, 16395) already exists.\\n2022-04-07 09:13:28.405 AEST [24237] STATEMENT: create subscription\\nsub53 connection 'port=7653' publication pub3 with\\n(subscribe_local_only=true,copy_data=false);\\n2022-04-07 09:13:28.425 AEST [24237] ERROR: duplicate key value\\nviolates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n2022-04-07 09:13:28.425 AEST [24237] DETAIL: Key (srrelid,\\nsrsubid)=(16384, 16396) already exists.\\n2022-04-07 09:13:28.425 AEST [24237] STATEMENT: create subscription\\nsub54 connection 'port=7654' publication pub4 with\\n(subscribe_local_only=true,copy_data=false);\\n2022-04-07 09:17:52.472 AEST [25852] ERROR: duplicate key value\\nviolates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n2022-04-07 09:17:52.472 AEST [25852] DETAIL: Key (srrelid,\\nsrsubid)=(16384, 16397) already exists.\\n2022-04-07 09:17:52.472 AEST [25852] STATEMENT: create subscription\\nsub51 connection 'port=7651' publication pub1;\\n\\n~~~\\n\\nWhen I debugged this it seemed like each of the CREAT SUBSCRIPTION was\\ntrying to make a double-entry, because the fetch_tables (your patch\\nv6-0002 modified SQL of this) was retuning the same table 2x.\\n\\n(gdb) bt\\n#0 errfinish (filename=0xbc1057 \\\"nbtinsert.c\\\", lineno=671,\\nfuncname=0xbc25e0 <__func__.15798> \\\"_bt_check_unique\\\") at elog.c:510\\n#1 0x0000000000526d83 in _bt_check_unique (rel=0x7f654219c2a0,\\ninsertstate=0x7ffd9629ddd0, heapRel=0x7f65421b0e28,\\ncheckUnique=UNIQUE_CHECK_YES, is_unique=0x7ffd9629de01,\\nspeculativeToken=0x7ffd9629ddcc) at nbtinsert.c:664\\n#2 0x0000000000526157 in _bt_doinsert (rel=0x7f654219c2a0,\\nitup=0x19ea8e8, checkUnique=UNIQUE_CHECK_YES, indexUnchanged=false,\\nheapRel=0x7f65421b0e28) at nbtinsert.c:208\\n#3 0x000000000053450e in btinsert (rel=0x7f654219c2a0,\\nvalues=0x7ffd9629df10, isnull=0x7ffd9629def0, ht_ctid=0x19ea894,\\nheapRel=0x7f65421b0e28, checkUnique=UNIQUE_CHECK_YES,\\nindexUnchanged=false, indexInfo=0x19dea80) at nbtree.c:201\\n#4 0x00000000005213b6 in index_insert (indexRelation=0x7f654219c2a0,\\nvalues=0x7ffd9629df10, isnull=0x7ffd9629def0, heap_t_ctid=0x19ea894,\\nheapRelation=0x7f65421b0e28, checkUnique=UNIQUE_CHECK_YES,\\nindexUnchanged=false, indexInfo=0x19dea80) at indexam.c:193\\n#5 0x00000000005c81d5 in CatalogIndexInsert (indstate=0x19de540,\\nheapTuple=0x19ea890) at indexing.c:158\\n#6 0x00000000005c8325 in CatalogTupleInsert (heapRel=0x7f65421b0e28,\\ntup=0x19ea890) at indexing.c:231\\n#7 0x00000000005f0170 in AddSubscriptionRelState (subid=16400,\\nrelid=16384, state=105 'i', sublsn=0) at pg_subscription.c:315\\n#8 0x00000000006d6fa5 in CreateSubscription (pstate=0x1942dc0,\\nstmt=0x191f6a0, isTopLevel=true) at subscriptioncmds.c:767\\n\\n~~\\n\\nAside: All this was happening when I did not have enough logical\\nreplication workers configured. (There were WARNINGS in the logfile\\nthat I had not noticed).\\nWhen I fix the configuration then all these other problems went away!\\n\\n~~\\n\\nSo to summarize, I'm not sure if the fetch_tables still has some\\npotential problem lurking or not, but I feel that the SQL in that\\nfunction maybe needs a closer look to ensure it is always impossible\\nto return the same table multiple times.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia.\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Apr 2022 14:56:12 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 12, 2022 at 10:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Thu, Apr 7, 2022 at 2:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > FYI, here is a test script that is using the current patch (v6) to\\n> > demonstrate a way to share table data between different numbers of\\n> > nodes (up to 5 of them here).\\n> >\\n> > The script starts off with just 2-way sharing (nodes N1, N2),\\n> > then expands to 3-way sharing (nodes N1, N2, N3),\\n> > then 4-way sharing (nodes N1, N2, N3, N4),\\n> > then 5-way sharing (nodes N1, N2, N3, N4, N5).\\n> >\\n> > As an extra complication, for this test, all 5 nodes have different\\n> > initial table data, which gets replicated to the others whenever each\\n> > new node joins the existing share group.\\n> >\\n> > PSA.\\n> >\\n>\\n>\\n> Hi Vignesh. I had some problems getting the above test script working.\\n> It was OK up until I tried to join the 5th node (N5) to the existing 4\\n> nodes. The ERROR was manifesting itself strangely because it appeared\\n> that there was an index violation in the pg_subscription_rel catalog\\n> even though AFAIK the N5 did not have any entries in it.\\n>\\n>\\n> e.g.\\n> 2022-04-07 09:13:28.361 AEST [24237] ERROR: duplicate key value\\n> violates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n> 2022-04-07 09:13:28.361 AEST [24237] DETAIL: Key (srrelid,\\n> srsubid)=(16384, 16393) already exists.\\n> 2022-04-07 09:13:28.361 AEST [24237] STATEMENT: create subscription\\n> sub51 connection 'port=7651' publication pub1 with\\n> (subscribe_local_only=true,copy_data=force);\\n> 2022-04-07 09:13:28.380 AEST [24237] ERROR: duplicate key value\\n> violates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n> 2022-04-07 09:13:28.380 AEST [24237] DETAIL: Key (srrelid,\\n> srsubid)=(16384, 16394) already exists.\\n> 2022-04-07 09:13:28.380 AEST [24237] STATEMENT: create subscription\\n> sub52 connection 'port=7652' publication pub2 with\\n> (subscribe_local_only=true,copy_data=false);\\n> 2022-04-07 09:13:28.405 AEST [24237] ERROR: duplicate key value\\n> violates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n> 2022-04-07 09:13:28.405 AEST [24237] DETAIL: Key (srrelid,\\n> srsubid)=(16384, 16395) already exists.\\n> 2022-04-07 09:13:28.405 AEST [24237] STATEMENT: create subscription\\n> sub53 connection 'port=7653' publication pub3 with\\n> (subscribe_local_only=true,copy_data=false);\\n> 2022-04-07 09:13:28.425 AEST [24237] ERROR: duplicate key value\\n> violates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n> 2022-04-07 09:13:28.425 AEST [24237] DETAIL: Key (srrelid,\\n> srsubid)=(16384, 16396) already exists.\\n> 2022-04-07 09:13:28.425 AEST [24237] STATEMENT: create subscription\\n> sub54 connection 'port=7654' publication pub4 with\\n> (subscribe_local_only=true,copy_data=false);\\n> 2022-04-07 09:17:52.472 AEST [25852] ERROR: duplicate key value\\n> violates unique constraint \\\"pg_subscription_rel_srrelid_srsubid_index\\\"\\n> 2022-04-07 09:17:52.472 AEST [25852] DETAIL: Key (srrelid,\\n> srsubid)=(16384, 16397) already exists.\\n> 2022-04-07 09:17:52.472 AEST [25852] STATEMENT: create subscription\\n> sub51 connection 'port=7651' publication pub1;\\n>\\n> ~~~\\n>\\n> When I debugged this it seemed like each of the CREAT SUBSCRIPTION was\\n> trying to make a double-entry, because the fetch_tables (your patch\\n> v6-0002 modified SQL of this) was retuning the same table 2x.\\n>\\n> (gdb) bt\\n> #0 errfinish (filename=0xbc1057 \\\"nbtinsert.c\\\", lineno=671,\\n> funcname=0xbc25e0 <__func__.15798> \\\"_bt_check_unique\\\") at elog.c:510\\n> #1 0x0000000000526d83 in _bt_check_unique (rel=0x7f654219c2a0,\\n> insertstate=0x7ffd9629ddd0, heapRel=0x7f65421b0e28,\\n> checkUnique=UNIQUE_CHECK_YES, is_unique=0x7ffd9629de01,\\n> speculativeToken=0x7ffd9629ddcc) at nbtinsert.c:664\\n> #2 0x0000000000526157 in _bt_doinsert (rel=0x7f654219c2a0,\\n> itup=0x19ea8e8, checkUnique=UNIQUE_CHECK_YES, indexUnchanged=false,\\n> heapRel=0x7f65421b0e28) at nbtinsert.c:208\\n> #3 0x000000000053450e in btinsert (rel=0x7f654219c2a0,\\n> values=0x7ffd9629df10, isnull=0x7ffd9629def0, ht_ctid=0x19ea894,\\n> heapRel=0x7f65421b0e28, checkUnique=UNIQUE_CHECK_YES,\\n> indexUnchanged=false, indexInfo=0x19dea80) at nbtree.c:201\\n> #4 0x00000000005213b6 in index_insert (indexRelation=0x7f654219c2a0,\\n> values=0x7ffd9629df10, isnull=0x7ffd9629def0, heap_t_ctid=0x19ea894,\\n> heapRelation=0x7f65421b0e28, checkUnique=UNIQUE_CHECK_YES,\\n> indexUnchanged=false, indexInfo=0x19dea80) at indexam.c:193\\n> #5 0x00000000005c81d5 in CatalogIndexInsert (indstate=0x19de540,\\n> heapTuple=0x19ea890) at indexing.c:158\\n> #6 0x00000000005c8325 in CatalogTupleInsert (heapRel=0x7f65421b0e28,\\n> tup=0x19ea890) at indexing.c:231\\n> #7 0x00000000005f0170 in AddSubscriptionRelState (subid=16400,\\n> relid=16384, state=105 'i', sublsn=0) at pg_subscription.c:315\\n> #8 0x00000000006d6fa5 in CreateSubscription (pstate=0x1942dc0,\\n> stmt=0x191f6a0, isTopLevel=true) at subscriptioncmds.c:767\\n>\\n> ~~\\n>\\n> Aside: All this was happening when I did not have enough logical\\n> replication workers configured. (There were WARNINGS in the logfile\\n> that I had not noticed).\\n> When I fix the configuration then all these other problems went away!\\n>\\n> ~~\\n>\\n> So to summarize, I'm not sure if the fetch_tables still has some\\n> potential problem lurking or not, but I feel that the SQL in that\\n> function maybe needs a closer look to ensure it is always impossible\\n> to return the same table multiple times.\\n\\nI have earlier used a distinct of srsubstate, but there is a\\npossibility that when multiple subscriptions are being created there\\ncan be multiple entries because the srsubstate can be different like i\\nand r. I have changed srsubstate to srrelid to get the unique values.\\nThe attached v8 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 12 Apr 2022 12:22:22 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for v8-0001.\\n\\n======\\n\\n1. Commit message\\n\\nCREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\\nPUBLICATION pub1 with (local_only = true);\\n\\nThe spaces in the CONNECTION string are a bit strange.\\n\\n~~~\\n\\n2. src/backend/catalog/pg_subscription.\\n\\n@@ -71,6 +71,7 @@ GetSubscription(Oid subid, bool missing_ok)\\n sub->stream = subform->substream;\\n sub->twophasestate = subform->subtwophasestate;\\n sub->disableonerr = subform->subdisableonerr;\\n+ sub->local_only = subform->sublocal;\\n\\nMaybe call it subform->sublocalonly;\\n\\n~~~\\n\\n3. src/backend/catalog/system_views.sql\\n\\n@@ -1290,8 +1290,8 @@ REVOKE ALL ON pg_replication_origin_status FROM public;\\n -- All columns of pg_subscription except subconninfo are publicly readable.\\n REVOKE ALL ON pg_subscription FROM public;\\n GRANT SELECT (oid, subdbid, subskiplsn, subname, subowner, subenabled,\\n- subbinary, substream, subtwophasestate,\\nsubdisableonerr, subslotname,\\n- subsynccommit, subpublications)\\n+ subbinary, substream, subtwophasestate, subdisableonerr,\\n+ sublocal, subslotname, subsynccommit, subpublications)\\n\\nMaybe call the column sublocalonly\\n\\n~~~\\n\\n4. .../libpqwalreceiver/libpqwalreceiver.c\\n\\n@@ -453,6 +453,10 @@ libpqrcv_startstreaming(WalReceiverConn *conn,\\n PQserverVersion(conn->streamConn) >= 150000)\\n appendStringInfoString(&cmd, \\\", two_phase 'on'\\\");\\n\\n+ if (options->proto.logical.local_only &&\\n+ PQserverVersion(conn->streamConn) >= 150000)\\n+ appendStringInfoString(&cmd, \\\", local_only 'on'\\\");\\n\\nAdd a FIXME comment here as a reminder that this condition needs to\\nchange for PG16.\\n\\n~~~\\n\\n5. src/bin/pg_dump/pg_dump.c\\n\\n@@ -4361,6 +4361,7 @@ getSubscriptions(Archive *fout)\\n int i_subsynccommit;\\n int i_subpublications;\\n int i_subbinary;\\n+ int i_sublocal;\\n int i,\\n ntups;\\n\\nMaybe call this member i_sublocalonly;\\n\\n~~~\\n\\n6. src/bin/pg_dump/pg_dump.c\\n+ if (fout->remoteVersion >= 150000)\\n+ appendPQExpBufferStr(query, \\\" s.sublocal\\\\n\\\");\\n+ else\\n+ appendPQExpBufferStr(query, \\\" false AS sublocal\\\\n\\\");\\n+\\n\\n6a. Add a FIXME comment as a reminder that this condition needs to\\nchange for PG16.\\n\\n6b. Change the column name to sublocalonly.\\n\\n~~~\\n\\n7. src/bin/pg_dump/pg_dump.h\\n\\n@@ -661,6 +661,7 @@ typedef struct _SubscriptionInfo\\n char *subdisableonerr;\\n char *subsynccommit;\\n char *subpublications;\\n+ char *sublocal;\\n } SubscriptionInfo;\\n\\nChange the member name to sublocalonly\\n\\n~~~\\n\\n8. src/bin/psql/describe.c\\n\\n@@ -6241,6 +6241,12 @@ describeSubscriptions(const char *pattern, bool verbose)\\n gettext_noop(\\\"Two phase commit\\\"),\\n gettext_noop(\\\"Disable on error\\\"));\\n\\n+ /* local_only is supported only in v15 and higher */\\n+ if (pset.sversion >= 150000)\\n+ appendPQExpBuffer(&buf,\\n+ \\\", sublocal AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n+ gettext_noop(\\\"Local only\\\"));\\n\\n7a. The comment is wrong to mention v15.\\n\\n7b. Add a FIXME comment as a reminder that this condition needs to\\nchange for PG16.\\n\\n7c. Change the column name to sublocalonly.\\n\\n~~~\\n\\n9. src/include/catalog/pg_subscription.h\\n\\n@@ -70,6 +70,8 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\\nBKI_SHARED_RELATION BKI_ROW\\n\\n bool substream; /* Stream in-progress transactions. */\\n\\n+ bool sublocal; /* skip copying of remote origin data */\\n\\nChange the member name to sublocalonly\\n\\n~~~\\n\\n10. src/include/replication/logicalproto.h\\n\\n@@ -32,12 +32,17 @@\\n *\\n * LOGICALREP_PROTO_TWOPHASE_VERSION_NUM is the minimum protocol version with\\n * support for two-phase commit decoding (at prepare time). Introduced in PG15.\\n+ *\\n+ * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\\n+ * support for sending only locally originated data from the publisher.\\n+ * Introduced in PG16.\\n\\nAdd a FIXME comment here as a reminder that the proto version number\\nneeds to be bumped to 4 in PG16.\\n\\n~~~\\n\\n11. src/test/subscription/t/032_circular.pl\\n\\nPerhaps there should be another test using a third \\\"Node_C\\\" which\\npublishes some data to Node_B. Then you can ensure that by using\\nlocal_only (when Node_A is subscribing to Node_A) that nothing from\\nthe Node_C can find its way onto Node_A. But then the test name\\n\\\"circular\\\" is a bit misleading. Maybe it should be \\\"032_localonly\\\"?\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Apr 2022 18:22:59 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 14, 2022 at 1:53 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for v8-0001.\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> CREATE SUBSCRIPTION sub1 CONNECTION 'dbname =postgres port=9999'\\n> PUBLICATION pub1 with (local_only = true);\\n>\\n> The spaces in the CONNECTION string are a bit strange.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2. src/backend/catalog/pg_subscription.\\n>\\n> @@ -71,6 +71,7 @@ GetSubscription(Oid subid, bool missing_ok)\\n> sub->stream = subform->substream;\\n> sub->twophasestate = subform->subtwophasestate;\\n> sub->disableonerr = subform->subdisableonerr;\\n> + sub->local_only = subform->sublocal;\\n>\\n> Maybe call it subform->sublocalonly;\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. src/backend/catalog/system_views.sql\\n>\\n> @@ -1290,8 +1290,8 @@ REVOKE ALL ON pg_replication_origin_status FROM public;\\n> -- All columns of pg_subscription except subconninfo are publicly readable.\\n> REVOKE ALL ON pg_subscription FROM public;\\n> GRANT SELECT (oid, subdbid, subskiplsn, subname, subowner, subenabled,\\n> - subbinary, substream, subtwophasestate,\\n> subdisableonerr, subslotname,\\n> - subsynccommit, subpublications)\\n> + subbinary, substream, subtwophasestate, subdisableonerr,\\n> + sublocal, subslotname, subsynccommit, subpublications)\\n>\\n> Maybe call the column sublocalonly\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. .../libpqwalreceiver/libpqwalreceiver.c\\n>\\n> @@ -453,6 +453,10 @@ libpqrcv_startstreaming(WalReceiverConn *conn,\\n> PQserverVersion(conn->streamConn) >= 150000)\\n> appendStringInfoString(&cmd, \\\", two_phase 'on'\\\");\\n>\\n> + if (options->proto.logical.local_only &&\\n> + PQserverVersion(conn->streamConn) >= 150000)\\n> + appendStringInfoString(&cmd, \\\", local_only 'on'\\\");\\n>\\n> Add a FIXME comment here as a reminder that this condition needs to\\n> change for PG16.\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. src/bin/pg_dump/pg_dump.c\\n>\\n> @@ -4361,6 +4361,7 @@ getSubscriptions(Archive *fout)\\n> int i_subsynccommit;\\n> int i_subpublications;\\n> int i_subbinary;\\n> + int i_sublocal;\\n> int i,\\n> ntups;\\n>\\n> Maybe call this member i_sublocalonly;\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 6. src/bin/pg_dump/pg_dump.c\\n> + if (fout->remoteVersion >= 150000)\\n> + appendPQExpBufferStr(query, \\\" s.sublocal\\\\n\\\");\\n> + else\\n> + appendPQExpBufferStr(query, \\\" false AS sublocal\\\\n\\\");\\n> +\\n>\\n> 6a. Add a FIXME comment as a reminder that this condition needs to\\n> change for PG16.\\n>\\n\\nModified\\n\\n> 6b. Change the column name to sublocalonly.\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. src/bin/pg_dump/pg_dump.h\\n>\\n> @@ -661,6 +661,7 @@ typedef struct _SubscriptionInfo\\n> char *subdisableonerr;\\n> char *subsynccommit;\\n> char *subpublications;\\n> + char *sublocal;\\n> } SubscriptionInfo;\\n>\\n> Change the member name to sublocalonly\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 8. src/bin/psql/describe.c\\n>\\n> @@ -6241,6 +6241,12 @@ describeSubscriptions(const char *pattern, bool verbose)\\n> gettext_noop(\\\"Two phase commit\\\"),\\n> gettext_noop(\\\"Disable on error\\\"));\\n>\\n> + /* local_only is supported only in v15 and higher */\\n> + if (pset.sversion >= 150000)\\n> + appendPQExpBuffer(&buf,\\n> + \\\", sublocal AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n> + gettext_noop(\\\"Local only\\\"));\\n>\\n> 7a. The comment is wrong to mention v15.\\n>\\n\\nI have removed this comment and added a FIXME. I will add it in once\\nversion change is done to avoid confusion.\\n\\n> 7b. Add a FIXME comment as a reminder that this condition needs to\\n> change for PG16.\\n>\\n\\nModified\\n\\n> 7c. Change the column name to sublocalonly.\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 9. src/include/catalog/pg_subscription.h\\n>\\n> @@ -70,6 +70,8 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\\n> BKI_SHARED_RELATION BKI_ROW\\n>\\n> bool substream; /* Stream in-progress transactions. */\\n>\\n> + bool sublocal; /* skip copying of remote origin data */\\n>\\n> Change the member name to sublocalonly\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 10. src/include/replication/logicalproto.h\\n>\\n> @@ -32,12 +32,17 @@\\n> *\\n> * LOGICALREP_PROTO_TWOPHASE_VERSION_NUM is the minimum protocol version with\\n> * support for two-phase commit decoding (at prepare time). Introduced in PG15.\\n> + *\\n> + * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\\n> + * support for sending only locally originated data from the publisher.\\n> + * Introduced in PG16.\\n>\\n> Add a FIXME comment here as a reminder that the proto version number\\n> needs to be bumped to 4 in PG16.\\n>\\n\\nModified\\n\\n> ~~~\\n>\\n> 11. src/test/subscription/t/032_circular.pl\\n>\\n> Perhaps there should be another test using a third \\\"Node_C\\\" which\\n> publishes some data to Node_B. Then you can ensure that by using\\n> local_only (when Node_A is subscribing to Node_A) that nothing from\\n> the Node_C can find its way onto Node_A. But then the test name\\n> \\\"circular\\\" is a bit misleading. Maybe it should be \\\"032_localonly\\\"?\\n\\nAdded the test and changed the file.\\n\\nThanks for the comments, attached v9 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Sat, 16 Apr 2022 19:48:22 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"I checked the latest v9-0001 patch. Below are my review comments.\\n\\nOther than these few trivial comments this 0001 patch looks good to me.\\n\\n~~~\\n\\n1. src/backend/replication/pgoutput/pgoutput.c - whitespace\\n\\n@@ -1696,6 +1714,10 @@ static bool\\n pgoutput_origin_filter(LogicalDecodingContext *ctx,\\n RepOriginId origin_id)\\n {\\n+ PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\\n+\\n+ if (data->local_only && origin_id != InvalidRepOriginId)\\n+ return true;\\n return false;\\n }\\n\\nSuggest to add a blank line after the return true;\\n\\n~~~\\n\\n2. src/bin/psql/tab-complete.c - not alphabetical\\n\\n@@ -1874,7 +1874,7 @@ psql_completion(const char *text, int start, int end)\\n COMPLETE_WITH(\\\"(\\\", \\\"PUBLICATION\\\");\\n /* ALTER SUBSCRIPTION <name> SET ( */\\n else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\nTailMatches(\\\"SET\\\", \\\"(\\\"))\\n- COMPLETE_WITH(\\\"binary\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n\\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n+ COMPLETE_WITH(\\\"binary\\\", \\\"slot_name\\\", \\\"streaming\\\", \\\"local_only\\\",\\n\\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n\\n2a. AFAIK the code intended that these options be listed in\\nalphabetical order (I think the recent addition of disable_on_error is\\nalso wrong here). So \\\"local_only\\\" should be moved.\\n\\n@@ -3156,7 +3156,7 @@ psql_completion(const char *text, int start, int end)\\n /* Complete \\\"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\\\" */\\n else if (HeadMatches(\\\"CREATE\\\", \\\"SUBSCRIPTION\\\") && TailMatches(\\\"WITH\\\", \\\"(\\\"))\\n COMPLETE_WITH(\\\"binary\\\", \\\"connect\\\", \\\"copy_data\\\", \\\"create_slot\\\",\\n- \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n+ \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\", \\\"local_only\\\",\\n \\\"synchronous_commit\\\", \\\"two_phase\\\", \\\"disable_on_error\\\");\\n\\n2b. ditto\\n\\n~~~\\n\\n3. src/test/subscription/t/032_localonly.pl - wrong message\\n\\n+$node_C->wait_for_catchup($appname_B2);\\n+$node_B->wait_for_catchup($appname_A);\\n+\\n+$result = $node_B->safe_psql('postgres', \\\"SELECT * FROM tab_full;\\\");\\n+is( $result, qq(11\\n+12\\n+13),\\n+ 'Inserted successfully without leading to infinite recursion in\\ncircular replication setup'\\n+);\\n+\\n+# check that the data published from node_C to node_B is not sent to node_A\\n+$result = $node_A->safe_psql('postgres', \\\"SELECT * FROM tab_full;\\\");\\n+is( $result, qq(11\\n+12),\\n+ 'Inserted successfully without leading to infinite recursion in\\ncircular replication setup'\\n+);\\n+\\n\\nThe new test looked good, but the cut/paste text message ('Inserted\\nsuccessfully without leading to infinite recursion in circular\\nreplication setup') maybe needs changing because there is nothing\\nreally \\\"circular\\\" about this test case.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia.\\n\\n\\n\",\n \"msg_date\": \"Tue, 19 Apr 2022 12:59:16 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Below are my review comments for the v9-0002 patch (except I did not\\nyet look at the TAP tests).\\n\\n~~~\\n\\n1. General comment - describe.c\\n\\nI wondered why the copy_data enum value is not displayed by the psql\\n\\\\drs+ command. Should it be?\\n\\n~~~\\n\\n2. General comment - SUBOPT_LOCAL_ONLY\\n\\n@@ -1134,7 +1204,7 @@ AlterSubscription(ParseState *pstate,\\nAlterSubscriptionStmt *stmt,\\n\\n case ALTER_SUBSCRIPTION_SET_PUBLICATION:\\n {\\n- supported_opts = SUBOPT_COPY_DATA | SUBOPT_REFRESH;\\n+ supported_opts = SUBOPT_COPY_DATA | SUBOPT_REFRESH | SUBOPT_LOCAL_ONLY;\\n parse_subscription_options(pstate, stmt->options,\\n supported_opts, &opts);\\n\\n@@ -1236,7 +1308,8 @@ AlterSubscription(ParseState *pstate,\\nAlterSubscriptionStmt *stmt,\\n errmsg(\\\"ALTER SUBSCRIPTION ... REFRESH is not allowed for disabled\\nsubscriptions\\\")));\\n\\n parse_subscription_options(pstate, stmt->options,\\n- SUBOPT_COPY_DATA, &opts);\\n+ SUBOPT_COPY_DATA | SUBOPT_LOCAL_ONLY,\\n+ &opts);\\n\\nI noticed there is some new code that appears to be setting the\\nSUBOT_LOCAL_ONLY as a supported option. Shouldn't those changes belong\\nin the patch 0001 instead of in this patch 0002?\\n\\n~~~\\n\\n3. Commit message - wording\\n\\nCURRENT\\na) Node1 - Publication publishing employee table.\\nb) Node2 - Subscription subscribing from publication pub1 with\\nlocal_only.\\nc) Node2 - Publication publishing employee table.\\nd) Node1 - Subscription subscribing from publication pub2 with\\nlocal_only.\\n\\nSUGGESTION (I think below has the same meaning but is simpler)\\na) Node1 - PUBLICATION pub1 for the employee table\\nb) Node2 - SUBSCRIPTION from pub1 with local_only=true\\nc) Node2 - PUBLICATION pub2 for the employee table\\nd) Node1 - SUBSCRIPTION from pub2 with local_only=true\\n\\n~~~\\n\\n4. Commit message - meaning?\\n\\nCURRENT\\nNow when user is trying to add another node Node3 to the\\nabove Multi master logical replication setup:\\na) user will have to create one subscription subscribing from Node1 to\\nNode3\\nb) user wil have to create another subscription subscribing from\\nNode2 to Node3 using local_only option and copy_data as true.\\n\\nWhile the subscription is created, server will identify that Node2 is\\nsubscribing from Node1 and Node1 is subscribing from Node2 and throw an\\nerror so that user can handle the initial copy data.\\n\\n~\\n\\nThe wording above confused me. Can you clarify it?\\ne.g.\\na) What exactly was the user hoping to achieve here?\\nb) Are the user steps doing something deliberately wrong just so you\\ncan describe later that an error gets thrown?\\n\\n~~~\\n\\n5. doc/src/sgml/logical-replication.sgml - how to get here?\\n\\nI didn’t see any easy way to get to this page. (No cross refs from anywhere?)\\n\\n~~~\\n\\n6. doc/src/sgml/logical-replication.sgml - section levels\\n\\nI think the section levels may be a bit messed up. e.g. The HTML\\nrendering of sections looks a bit confused. Maybe this is same as my\\nreview comment #13.\\n\\n~~\\n\\n7. doc/src/sgml/logical-replication.sgml - headings\\n\\n<title>Setting Bidirection logical replication between two nodes:</title>\\n\\n7a. Maybe better just to have a simpler main heading like\\n\\\"Bidirectional logical replication\\\".\\n\\n7b. Don't put \\\":\\\" in the title.\\n\\n~~~\\n\\n8. doc/src/sgml/logical-replication.sgml - missing intro\\n\\nIMO this page needs some sort of introduction/blurb instead of leaping\\nstraight into examples without any preamble text to give context.\\n\\n~~~\\n\\n9. doc/src/sgml/logical-replication.sgml - bullets\\n\\nSuggest removing all the bullets from the example steps. (I had\\nsomething similar a while ago for the \\\"Row Filter\\\" page but\\neventually, they all had to be removed).\\n\\n~~~\\n\\n10. doc/src/sgml/logical-replication.sgml - SQL too long\\n\\nMany of the example commands are much too long, and for me, they are\\ngiving scroll bars when rendered. It would be better if they can be\\nwrapped in appropriate places so easier to read (and no resulting\\nscroll bars).\\n\\n~~~\\n\\n11. doc/src/sgml/logical-replication.sgml - add the psql prompt\\n\\nIMO it would also be easier to understand the examples if you show the\\npsql prompt. Then you can easily know the node context without having\\nto describe it in the text so often.\\n\\ne.g.\\n\\n+ <para>\\n+ Create the subscription in node2 to subscribe the changes from node1:\\n+<programlisting>\\n+CREATE SUBSCRIPTION sub_node1_node2 ...\\n\\nSUGGGESTION\\n+ <para>\\n+ Create the subscription in node2 to subscribe the changes from node1\\n+<programlisting>\\n+node_2=# CREATE SUBSCRIPTION sub_node1_node2 ...\\n\\n~~~\\n\\n12. doc/src/sgml/logical-replication.sgml - typo\\n\\n+ <para>\\n+ Now the BiDirectional logical replication setup is complete between node1\\n\\ntypo \\\"BiDirectional\\\"\\n\\n~~~\\n\\n13. doc/src/sgml/logical-replication.sgml - deep levels\\n\\nThe section levels are very deep, but > 3 will not appear in the table\\nof contents when rendered. Maybe you can rearrange to raise them all\\nup one level, then IMO the TOC will work better and the whole page\\nwill be easier to read.\\n\\n~~~\\n\\n14. doc/src/sgml/logical-replication.sgml - unnecessarily complex?\\n\\n+<programlisting>\\n+# Truncate the table data but do not replicate the truncate.\\n+ALTER PUBLICATION pub_node3 SET (publish='insert,update,delete');\\n+TRUNCATE t1;\\n+\\n+CREATE SUBSCRIPTION sub_node3_node1 CONNECTION 'dbname=foo host=node1\\nuser=repuser' PUBLICATION pub_node1 WITH (copy_data = force,\\nlocal_only = on);\\n+\\n+CREATE SUBSCRIPTION sub_node3_node2 CONNECTION 'dbname=foo host=node2\\nuser=repuser' PUBLICATION pub_node2 WITH (copy_data = off, local_only\\n= on);\\n+\\n+# Include truncate operations from now\\n+ALTER PUBLICATION pub_node3 SET (publish='insert,update,delete,truncate');\\n+</programlisting>\\n\\nIs it really necessary for those CREATE SUBSCRIPTION to be placed\\nwhere they are? I felt it would be less confusing if you just do the\\nTRUNCATE between the 2x ALTER PUBLICATION... and then do those CREATE\\nSUBSCRIPTION separately later.\\n\\n~~~\\n\\n14. doc/src/sgml/logical-replication.sgml - force?\\n\\nThere seems no documentation anywhere that says what is the\\npurpose/meaning of the copy_data enum \\\"force\\\".\\n\\n~~~\\n\\n15. doc/src/sgml/ref/alter_subscription.sgml - force?\\n\\nMeanings of copy_data true/false are self-evident but should something\\nhere be explaining what is the meaning of \\\"force\\\"? Or a reference to\\nsome place that explains it?\\n\\n~~~\\n\\n16. doc/src/sgml/ref/create_subscription.sgml - local_only missing notes?\\n\\n@@ -161,6 +161,11 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n publisher node changes regardless of their origin. The default is\\n <literal>false</literal>.\\n </para>\\n+ <para>\\n+ There is some interation between the \\\"local_only\\\" option and\\n+ \\\"copy_data\\\" option. Refer to the\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n+ </para>\\n </listitem>\\n </varlistentry>\\n\\n16a. Typo \\\"interation\\\"\\n\\n16b. But where in the referenced notes does it actually say anything about this?\\n\\n~~~\\n\\n17. doc/src/sgml/ref/create_subscription.sgml - force?\\n\\nThe notes about the copy_data do not say anything about what the\\nvalues mean. Specifically, there seems nothing that describes what is\\nthe meaning of \\\"force\\\".\\n\\n~~~\\n\\n18. doc/src/sgml/ref/create_subscription.sgml - copy_data missing notes?\\n\\n+\\n+ <para>\\n+ There is some interation between the \\\"local_only\\\" option and\\n+ \\\"copy_data\\\" option. Refer to the\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n+ </para>\\n\\n18a. Typo \\\"interation\\\"\\n\\n18b. But where in the referenced notes does it actually say anything about this?\\n\\n~~~\\n\\n19. doc/src/sgml/ref/create_subscription.sgml - xref to the new page\\n\\nShouldn't there be an xref on this page somewhere to your other new\\n\\\"Bidirectional\\\" docs page?\\n\\n~~~\\n\\n20. src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_ON_OR_FORCE\\n\\n@@ -69,6 +69,18 @@\\n /* check if the 'val' has 'bits' set */\\n #define IsSet(val, bits) (((val) & (bits)) == (bits))\\n\\n+#define IS_COPY_DATA_ON_OR_FORCE(copy_data) (copy_data != COPY_DATA_OFF)\\n\\nMaybe this would be better as a static inline function?\\n\\n~~~\\n\\n21. src/backend/commands/subscriptioncmds.c - comment\\n\\n+/*\\n+ * Represents whether copy_data option is specified with on, off or force.\\n+ */\\n+typedef enum CopyData\\n+{\\n+ COPY_DATA_OFF,\\n+ COPY_DATA_ON,\\n+ COPY_DATA_FORCE\\n+} CopyData;\\n\\nI felt it might be better if the comment described these enums in the\\nsame order they are defined.\\n\\nE.g. \\\"Represents whether copy_data option is specified as off/on, or force.\\\"\\n\\n~~~\\n\\n22. src/backend/commands/subscriptioncmds.c - CreateSubscription bug?\\n\\n@@ -720,7 +788,8 @@ CreateSubscription(ParseState *pstate,\\nCreateSubscriptionStmt *stmt,\\n * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\\n * PUBLICATION to work.\\n */\\n- if (opts.twophase && !opts.copy_data && tables != NIL)\\n+ if (opts.twophase && IS_COPY_DATA_ON_OR_FORCE(opts.copy_data)\\n+ && tables != NIL)\\n twophase_enabled = true;\\nIs that a bug? It does not seem to match the previous code. Should\\nthat IS_COPY_DATA_ON_OR_FORCE be \\\"not\\\" ?\\n\\n~~~\\n\\n23. src/backend/commands/subscriptioncmds.c - long errmsg\\n\\n+ if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\\n+ ereport(ERROR,\\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n+ errmsg(\\\"CREATE/ALTER SUBSCRIPTION with local_only and copy_data as\\ntrue is not allowed when the publisher might have replicated data,\\ntable:%s.%s might have replicated data in the publisher\\\",\\n+ nspname, relname),\\n+ errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\\"));\\n+\\n\\nThe errmsg seems way too long for the source code. Can you use string\\nconcatenation or continuation chars to wrap the message over multiple\\nlines?\\n\\n~~~\\n\\n24. src/test/regress/sql/subscription.sql - typo\\n\\n@@ -66,6 +69,15 @@ ALTER SUBSCRIPTION regress_testsub4 SET (local_only = false);\\n DROP SUBSCRIPTION regress_testsub3;\\n DROP SUBSCRIPTION regress_testsub4;\\n\\n+-- ok - valid coy_data options\\n\\nTypo \\\"coy_data\\\". (it looks like this typo is not caused by this patch,\\nbut I think this patch should fix it anyhow).\\n\\n~~~\\n\\n25. src/test/regress/sql/subscription.sql - test order\\n\\nThe new tests are OK but IMO they could be re-ordered so then they\\nwill be more consistent for the positive and negative tests.\\n\\nCURRENT\\n\\\"true/force/on/1\\\" and \\\"off/false/0\\\"\\n\\nSUGGEST\\n\\\"true/on/1/force\\\" and \\\"false/off/0\\\"\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 20 Apr 2022 11:56:08 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Below are my review comments for the v9-0002 patch (TAP test part only).\\n\\n(The order of my comments is a bit muddled because I jumped around in\\nthe file a bit while reviewing it).\\n\\n======\\n\\n1. create_subscription - missing comment.\\n\\n+sub create_subscription\\n+{\\n+ my ($node_subscriber, $node_publisher, $sub_name, $node_connstr,\\n+ $application_name, $pub_name, $copy_data_val)\\n+ = @_;\\n\\nMaybe add a comment for this subroutine and describe the expected parameters.\\n\\n~~\\n\\n2. create_subscription - hides the options\\n\\nIMO the \\\"create_subscription\\\" subroutine is hiding too many of the\\ndetails, so now it is less clear (from the caller's POV) what the test\\nis really doing. Perhaps the options could be passed more explicitly\\nto this subroutine.\\n\\ne.g. Instead of\\n\\ncreate_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\\n $appname_B1, 'tap_pub_A', 'on');\\n\\nperhaps explicitly set the WITH options like:\\n\\ncreate_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\\n $appname_B1, 'tap_pub_A', 'local_only = on, copy_data = on');\\n\\n~~~\\n\\n3. the application names are confusing\\n\\n+my $appname_A1 = 'tap_sub_A_B';\\n+my $appname_A2 = 'tap_sub_A_C';\\n+my $appname_B1 = 'tap_sub_B_A';\\n+my $appname_B2 = 'tap_sub_B_C';\\n+my $appname_C1 = 'tap_sub_C_A';\\n+my $appname_C2 = 'tap_sub_C_B';\\n\\nI found the application names using '1' and '2' to be a bit confusing.\\ni.e it was unnecessarily hard to associate them (in my head) with\\ntheir relevant subscriptions. IMO it would be easier to name them\\nusing letters like below:\\n\\nSUGGESTED\\nmy $appname_AB = 'tap_sub_A_B';\\nmy $appname_AC = 'tap_sub_A_C';\\nmy $appname_BA = 'tap_sub_B_A';\\nmy $appname_BC = 'tap_sub_B_C';\\nmy $appname_CA = 'tap_sub_C_A';\\nmy $appname_CB = 'tap_sub_C_B';\\n\\n~~~\\n\\n4. create_subscription - passing the $appname 2x.\\n\\n+create_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\\n+ $appname_B1, 'tap_pub_A', 'on');\\n\\n\\nIt seemed confusing that the $app_name is passed twice.\\n\\nIMO should rename all those $appname_XX vars (see previous comment) to\\nbe $subname_XX, and just pass that create_subscription instead.\\n\\nmy $subname_AB = 'tap_sub_A_B';\\nmy $subname_AC = 'tap_sub_A_C';\\nmy $subname_BA = 'tap_sub_B_A';\\nmy $subname_BC = 'tap_sub_B_C';\\nmy $subname_CA = 'tap_sub_C_A';\\nmy $subname_CB = 'tap_sub_C_B';\\n\\nThen create_subscription subroutine should have one less argument.\\nJust add a comment saying that the appname is always assigned the same\\nvalue as the subname.\\n\\n~~~\\n\\n5. cleanup part - seems a bit misleading\\n\\n+# cleanup\\n+$node_B->safe_psql(\\n+ 'postgres', \\\"\\n+ DROP SUBSCRIPTION $appname_B2\\\");\\n+$node_C->safe_psql(\\n+ 'postgres', \\\"\\n+ DELETE FROM tab_full\\\");\\n+$node_B->safe_psql(\\n+ 'postgres', \\\"\\n+ DELETE FROM tab_full where a = 13\\\");\\n\\nIs that comment misleading? IIUC this is not really cleaning up\\neverything. It is just a cleanup of the previous Node_C test part\\nisn't it?\\n\\n~~~\\n\\n6. Error case (when copy_data is true)\\n\\n+# Error when creating susbcription with local_only and copy_data as true when\\n+# the publisher has replicated data\\n\\n6a. Typo \\\"susbcription\\\"\\n\\n6b. That comment maybe needs some more explanation - eg. say that\\nsince Node_A is already subscribing to Node_B so when Node_B makes\\nanother subscription to Node_A the copy doesn't really know if the\\ndata really originated from Node_A or not...\\n\\n6c. Maybe the comment needed to be more like ############## style to\\ndenote this (and the one that follows) is a separate test case.\\n\\n~~~\\n\\n7. Error case (when copy_data is force)\\n\\n+# Creating subscription with local_only and copy_data as force should be\\n+# successful when the publisher has replicated data\\n\\n7a. Same typo \\\"subscription\\\"\\n\\n~~~\\n\\n8. Add 3rd node when the existing node has some data\\n\\n8a. Maybe that comment needs to be expanded more. This isn't just\\n\\\"adding a node\\\" - you are joining it to the others as another\\nbi-directional participant in a 3-way group. And the comment should\\nclarify exactly which nodes have the initial data. Both the existing 2\\nyou are joining to? The new one only? All of them?\\n\\n8b. IMO is would be much clearer to do SELECT from all the nodes at\\nthe start of this test just to re-confirm what is all the initial data\\non these nodes before joining the 3rd node.\\n\\nNOTE - These same review comments apply to the other test combinations too\\n- Add 3rd node when the existing node has no data\\n- Add 3rd node when the new node has some data\\n\\n~~~\\n\\n9. Inserted data\\n\\n+# insert some data in all the nodes\\n+$node_A->safe_psql('postgres', \\\"INSERT INTO tab_full VALUES (13);\\\");\\n+$node_B->safe_psql('postgres', \\\"INSERT INTO tab_full VALUES (21);\\\");\\n+$node_C->safe_psql('postgres', \\\"INSERT INTO tab_full VALUES (31);\\\");\\n\\nThey seemed strange values (13, 21, 31) to add. I couldn't work out\\nthe \\\"meaning\\\" of them.\\n\\nWouldn't values like 13, 23, 33 make more sense (e.g pattern is 10 x\\nnode# + something)\\n\\n~~~\\n\\n10. verify_data($node_A, $node_B, $node_C);\\n\\nAll the expected values are too buried in this subroutine which makes\\nit hard to read. I think it would be easier to read (from the caller's\\nPOV) if you can pass the *expected* values as another arg into the\\nverify_data subroutine.\\n\\ne.g. is something like this possible?\\nverify_data($node_A, $node_B, $node_C, [11,12,13])\\n\\n~~~\\n\\n11. ALTER for ignoring 'truncate'\\n\\n+$node_C->safe_psql('postgres',\\n+ \\\"ALTER PUBLICATION tap_pub_C SET (publish='insert,update,delete');\\\");\\n+\\n+$node_C->safe_psql('postgres', \\\"TRUNCATE tab_full\\\");\\n+\\n+create_subscription($node_C, $node_A, $appname_C1, $node_A_connstr,\\n+ $appname_C1, 'tap_pub_A', 'force');\\n+create_subscription($node_C, $node_B, $appname_C2, $node_B_connstr,\\n+ $appname_C2, 'tap_pub_B', 'off');\\n+\\n+#include truncates now\\n+$node_C->safe_psql('postgres',\\n+ \\\"ALTER PUBLICATION tap_pub_C SET\\n(publish='insert,update,delete,truncate');\\\"\\n+);\\n\\nDo those create_subscription calls need to be encapsulated by the\\nALTER like that? I did not think so. It looks a bit more complex if\\ndone this way.\\n\\n~~~\\n\\n12. clean_subscriber_contents - misnamed?\\n\\nThis subroutine feels a bit misnamed. It seems to be doing lots of\\nthings like detaching the Node_C and deleting all table data from all\\nnodes. That all seems quite different from just \\\"clean subscriber\\ncontents\\\".\\n\\n~~~\\n\\n13. table initial data?\\n\\n+clean_subscriber_contents($node_A, $node_B, $node_C);\\n+\\n+##########################################################################\\n+# Add 3rd node when the new node has some data\\n+##########################################################################\\n\\nBut does this test case *really* have some data? I am not so sure.\\nDoesn't the preceding \\\"clean_subscriber_contents\\\" call remove all the\\ndata that might have been there? That is why I think all the tests out\\nto have SELECT (previous comment #8) so they can re-confirm what data\\nis really in those tables before doing each test part.\\n\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 20 Apr 2022 15:49:24 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 19, 2022 at 8:29 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> I checked the latest v9-0001 patch. Below are my review comments.\\n>\\n> Other than these few trivial comments this 0001 patch looks good to me.\\n>\\n> ~~~\\n>\\n> 1. src/backend/replication/pgoutput/pgoutput.c - whitespace\\n>\\n> @@ -1696,6 +1714,10 @@ static bool\\n> pgoutput_origin_filter(LogicalDecodingContext *ctx,\\n> RepOriginId origin_id)\\n> {\\n> + PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\\n> +\\n> + if (data->local_only && origin_id != InvalidRepOriginId)\\n> + return true;\\n> return false;\\n> }\\n>\\n> Suggest to add a blank line after the return true;\\n\\nModified\\n\\n> ~~~\\n>\\n> 2. src/bin/psql/tab-complete.c - not alphabetical\\n>\\n> @@ -1874,7 +1874,7 @@ psql_completion(const char *text, int start, int end)\\n> COMPLETE_WITH(\\\"(\\\", \\\"PUBLICATION\\\");\\n> /* ALTER SUBSCRIPTION <name> SET ( */\\n> else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\n> TailMatches(\\\"SET\\\", \\\"(\\\"))\\n> - COMPLETE_WITH(\\\"binary\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n> + COMPLETE_WITH(\\\"binary\\\", \\\"slot_name\\\", \\\"streaming\\\", \\\"local_only\\\",\\n> \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n>\\n> 2a. AFAIK the code intended that these options be listed in\\n> alphabetical order (I think the recent addition of disable_on_error is\\n> also wrong here). So \\\"local_only\\\" should be moved.\\n\\nModified\\n\\n> @@ -3156,7 +3156,7 @@ psql_completion(const char *text, int start, int end)\\n> /* Complete \\\"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\\\" */\\n> else if (HeadMatches(\\\"CREATE\\\", \\\"SUBSCRIPTION\\\") && TailMatches(\\\"WITH\\\", \\\"(\\\"))\\n> COMPLETE_WITH(\\\"binary\\\", \\\"connect\\\", \\\"copy_data\\\", \\\"create_slot\\\",\\n> - \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> + \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\", \\\"local_only\\\",\\n> \\\"synchronous_commit\\\", \\\"two_phase\\\", \\\"disable_on_error\\\");\\n>\\n> 2b. ditto\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. src/test/subscription/t/032_localonly.pl - wrong message\\n>\\n> +$node_C->wait_for_catchup($appname_B2);\\n> +$node_B->wait_for_catchup($appname_A);\\n> +\\n> +$result = $node_B->safe_psql('postgres', \\\"SELECT * FROM tab_full;\\\");\\n> +is( $result, qq(11\\n> +12\\n> +13),\\n> + 'Inserted successfully without leading to infinite recursion in\\n> circular replication setup'\\n> +);\\n> +\\n> +# check that the data published from node_C to node_B is not sent to node_A\\n> +$result = $node_A->safe_psql('postgres', \\\"SELECT * FROM tab_full;\\\");\\n> +is( $result, qq(11\\n> +12),\\n> + 'Inserted successfully without leading to infinite recursion in\\n> circular replication setup'\\n> +);\\n> +\\n>\\n> The new test looked good, but the cut/paste text message ('Inserted\\n> successfully without leading to infinite recursion in circular\\n> replication setup') maybe needs changing because there is nothing\\n> really \\\"circular\\\" about this test case.\\n\\nModified\\n\\nAttached v10 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Fri, 22 Apr 2022 21:45:25 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 20, 2022 at 7:26 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Below are my review comments for the v9-0002 patch (except I did not\\n> yet look at the TAP tests).\\n>\\n> ~~~\\n>\\n> 1. General comment - describe.c\\n>\\n> I wondered why the copy_data enum value is not displayed by the psql\\n> \\\\drs+ command. Should it be?\\n>\\n\\nI noticed that we generally don't display the values for the options.\\nI did not add it to keep it consistent with other options.\\n\\n> ~~~\\n>\\n> 2. General comment - SUBOPT_LOCAL_ONLY\\n>\\n> @@ -1134,7 +1204,7 @@ AlterSubscription(ParseState *pstate,\\n> AlterSubscriptionStmt *stmt,\\n>\\n> case ALTER_SUBSCRIPTION_SET_PUBLICATION:\\n> {\\n> - supported_opts = SUBOPT_COPY_DATA | SUBOPT_REFRESH;\\n> + supported_opts = SUBOPT_COPY_DATA | SUBOPT_REFRESH | SUBOPT_LOCAL_ONLY;\\n> parse_subscription_options(pstate, stmt->options,\\n> supported_opts, &opts);\\n>\\n> @@ -1236,7 +1308,8 @@ AlterSubscription(ParseState *pstate,\\n> AlterSubscriptionStmt *stmt,\\n> errmsg(\\\"ALTER SUBSCRIPTION ... REFRESH is not allowed for disabled\\n> subscriptions\\\")));\\n>\\n> parse_subscription_options(pstate, stmt->options,\\n> - SUBOPT_COPY_DATA, &opts);\\n> + SUBOPT_COPY_DATA | SUBOPT_LOCAL_ONLY,\\n> + &opts);\\n>\\n> I noticed there is some new code that appears to be setting the\\n> SUBOT_LOCAL_ONLY as a supported option. Shouldn't those changes belong\\n> in the patch 0001 instead of in this patch 0002?\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. Commit message - wording\\n>\\n> CURRENT\\n> a) Node1 - Publication publishing employee table.\\n> b) Node2 - Subscription subscribing from publication pub1 with\\n> local_only.\\n> c) Node2 - Publication publishing employee table.\\n> d) Node1 - Subscription subscribing from publication pub2 with\\n> local_only.\\n>\\n> SUGGESTION (I think below has the same meaning but is simpler)\\n> a) Node1 - PUBLICATION pub1 for the employee table\\n> b) Node2 - SUBSCRIPTION from pub1 with local_only=true\\n> c) Node2 - PUBLICATION pub2 for the employee table\\n> d) Node1 - SUBSCRIPTION from pub2 with local_only=true\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. Commit message - meaning?\\n>\\n> CURRENT\\n> Now when user is trying to add another node Node3 to the\\n> above Multi master logical replication setup:\\n> a) user will have to create one subscription subscribing from Node1 to\\n> Node3\\n> b) user wil have to create another subscription subscribing from\\n> Node2 to Node3 using local_only option and copy_data as true.\\n>\\n> While the subscription is created, server will identify that Node2 is\\n> subscribing from Node1 and Node1 is subscribing from Node2 and throw an\\n> error so that user can handle the initial copy data.\\n>\\n> ~\\n>\\n> The wording above confused me. Can you clarify it?\\n> e.g.\\n> a) What exactly was the user hoping to achieve here?\\n> b) Are the user steps doing something deliberately wrong just so you\\n> can describe later that an error gets thrown?\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. doc/src/sgml/logical-replication.sgml - how to get here?\\n>\\n> I didn’t see any easy way to get to this page. (No cross refs from anywhere?)\\n\\nI have added a link from create subscription page\\n\\n> ~~~\\n>\\n> 6. doc/src/sgml/logical-replication.sgml - section levels\\n>\\n> I think the section levels may be a bit messed up. e.g. The HTML\\n> rendering of sections looks a bit confused. Maybe this is same as my\\n> review comment #13.\\n\\nModified\\n\\n> ~~\\n>\\n> 7. doc/src/sgml/logical-replication.sgml - headings\\n>\\n> <title>Setting Bidirection logical replication between two nodes:</title>\\n>\\n> 7a. Maybe better just to have a simpler main heading like\\n> \\\"Bidirectional logical replication\\\".\\n\\nModified\\n\\n> 7b. Don't put \\\":\\\" in the title.\\n\\nModified\\n\\n> ~~~\\n>\\n> 8. doc/src/sgml/logical-replication.sgml - missing intro\\n>\\n> IMO this page needs some sort of introduction/blurb instead of leaping\\n> straight into examples without any preamble text to give context.\\n\\nModified\\n\\n> ~~~\\n>\\n> 9. doc/src/sgml/logical-replication.sgml - bullets\\n>\\n> Suggest removing all the bullets from the example steps. (I had\\n> something similar a while ago for the \\\"Row Filter\\\" page but\\n> eventually, they all had to be removed).\\n\\nModified\\n\\n> ~~~\\n>\\n> 10. doc/src/sgml/logical-replication.sgml - SQL too long\\n>\\n> Many of the example commands are much too long, and for me, they are\\n> giving scroll bars when rendered. It would be better if they can be\\n> wrapped in appropriate places so easier to read (and no resulting\\n> scroll bars).\\n\\nModified\\n\\n> ~~~\\n>\\n> 11. doc/src/sgml/logical-replication.sgml - add the psql prompt\\n>\\n> IMO it would also be easier to understand the examples if you show the\\n> psql prompt. Then you can easily know the node context without having\\n> to describe it in the text so often.\\n>\\n> e.g.\\n>\\n> + <para>\\n> + Create the subscription in node2 to subscribe the changes from node1:\\n> +<programlisting>\\n> +CREATE SUBSCRIPTION sub_node1_node2 ...\\n>\\n> SUGGGESTION\\n> + <para>\\n> + Create the subscription in node2 to subscribe the changes from node1\\n> +<programlisting>\\n> +node_2=# CREATE SUBSCRIPTION sub_node1_node2 ...\\n\\nModified\\n\\n> ~~~\\n>\\n> 12. doc/src/sgml/logical-replication.sgml - typo\\n>\\n> + <para>\\n> + Now the BiDirectional logical replication setup is complete between node1\\n>\\n> typo \\\"BiDirectional\\\"\\n\\nModified to bidirectional\\n\\n> ~~~\\n>\\n> 13. doc/src/sgml/logical-replication.sgml - deep levels\\n>\\n> The section levels are very deep, but > 3 will not appear in the table\\n> of contents when rendered. Maybe you can rearrange to raise them all\\n> up one level, then IMO the TOC will work better and the whole page\\n> will be easier to read.\\n\\nModified\\n\\n> ~~~\\n>\\n> 14. doc/src/sgml/logical-replication.sgml - unnecessarily complex?\\n>\\n> +<programlisting>\\n> +# Truncate the table data but do not replicate the truncate.\\n> +ALTER PUBLICATION pub_node3 SET (publish='insert,update,delete');\\n> +TRUNCATE t1;\\n> +\\n> +CREATE SUBSCRIPTION sub_node3_node1 CONNECTION 'dbname=foo host=node1\\n> user=repuser' PUBLICATION pub_node1 WITH (copy_data = force,\\n> local_only = on);\\n> +\\n> +CREATE SUBSCRIPTION sub_node3_node2 CONNECTION 'dbname=foo host=node2\\n> user=repuser' PUBLICATION pub_node2 WITH (copy_data = off, local_only\\n> = on);\\n> +\\n> +# Include truncate operations from now\\n> +ALTER PUBLICATION pub_node3 SET (publish='insert,update,delete,truncate');\\n> +</programlisting>\\n>\\n> Is it really necessary for those CREATE SUBSCRIPTION to be placed\\n> where they are? I felt it would be less confusing if you just do the\\n> TRUNCATE between the 2x ALTER PUBLICATION... and then do those CREATE\\n> SUBSCRIPTION separately later.\\n\\nModified\\n\\n> ~~~\\n>\\n> 14. doc/src/sgml/logical-replication.sgml - force?\\n>\\n> There seems no documentation anywhere that says what is the\\n> purpose/meaning of the copy_data enum \\\"force\\\".\\n\\nThis is added in create subscription notes\\n\\n> ~~~\\n>\\n> 15. doc/src/sgml/ref/alter_subscription.sgml - force?\\n>\\n> Meanings of copy_data true/false are self-evident but should something\\n> here be explaining what is the meaning of \\\"force\\\"? Or a reference to\\n> some place that explains it?\\n\\nThis is mentioned in create subscription notes, added a reference to\\ncreate subscription notes.\\n\\n> ~~~\\n>\\n> 16. doc/src/sgml/ref/create_subscription.sgml - local_only missing notes?\\n>\\n> @@ -161,6 +161,11 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> publisher node changes regardless of their origin. The default is\\n> <literal>false</literal>.\\n> </para>\\n> + <para>\\n> + There is some interation between the \\\"local_only\\\" option and\\n> + \\\"copy_data\\\" option. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n> + </para>\\n> </listitem>\\n> </varlistentry>\\n>\\n> 16a. Typo \\\"interation\\\"\\n\\nModified\\n\\n> 16b. But where in the referenced notes does it actually say anything about this?\\n\\nModified\\n\\n> ~~~\\n>\\n> 17. doc/src/sgml/ref/create_subscription.sgml - force?\\n>\\n> The notes about the copy_data do not say anything about what the\\n> values mean. Specifically, there seems nothing that describes what is\\n> the meaning of \\\"force\\\".\\n\\nI have mentioned this in the notes section.\\n\\n> ~~~\\n>\\n> 18. doc/src/sgml/ref/create_subscription.sgml - copy_data missing notes?\\n>\\n> +\\n> + <para>\\n> + There is some interation between the \\\"local_only\\\" option and\\n> + \\\"copy_data\\\" option. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n> + </para>\\n>\\n> 18a. Typo \\\"interation\\\"\\n\\nModified\\n\\n> 18b. But where in the referenced notes does it actually say anything about this?\\n\\nAdded\\n\\n> ~~~\\n>\\n> 19. doc/src/sgml/ref/create_subscription.sgml - xref to the new page\\n>\\n> Shouldn't there be an xref on this page somewhere to your other new\\n> \\\"Bidirectional\\\" docs page?\\n\\nAdded reference\\n\\n> ~~~\\n>\\n> 20. src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_ON_OR_FORCE\\n>\\n> @@ -69,6 +69,18 @@\\n> /* check if the 'val' has 'bits' set */\\n> #define IsSet(val, bits) (((val) & (bits)) == (bits))\\n>\\n> +#define IS_COPY_DATA_ON_OR_FORCE(copy_data) (copy_data != COPY_DATA_OFF)\\n>\\n> Maybe this would be better as a static inline function?\\n\\n What is the advantage of doing this change? I have not changed this\\nas the macro usage is fine.\\nThoughts?\\n\\n> ~~~\\n>\\n> 21. src/backend/commands/subscriptioncmds.c - comment\\n>\\n> +/*\\n> + * Represents whether copy_data option is specified with on, off or force.\\n> + */\\n> +typedef enum CopyData\\n> +{\\n> + COPY_DATA_OFF,\\n> + COPY_DATA_ON,\\n> + COPY_DATA_FORCE\\n> +} CopyData;\\n>\\n> I felt it might be better if the comment described these enums in the\\n> same order they are defined.\\n>\\n> E.g. \\\"Represents whether copy_data option is specified as off/on, or force.\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 22. src/backend/commands/subscriptioncmds.c - CreateSubscription bug?\\n>\\n> @@ -720,7 +788,8 @@ CreateSubscription(ParseState *pstate,\\n> CreateSubscriptionStmt *stmt,\\n> * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\\n> * PUBLICATION to work.\\n> */\\n> - if (opts.twophase && !opts.copy_data && tables != NIL)\\n> + if (opts.twophase && IS_COPY_DATA_ON_OR_FORCE(opts.copy_data)\\n> + && tables != NIL)\\n> twophase_enabled = true;\\n> Is that a bug? It does not seem to match the previous code. Should\\n> that IS_COPY_DATA_ON_OR_FORCE be \\\"not\\\" ?\\n\\nModified\\n\\n> ~~~\\n>\\n> 23. src/backend/commands/subscriptioncmds.c - long errmsg\\n>\\n> + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"CREATE/ALTER SUBSCRIPTION with local_only and copy_data as\\n> true is not allowed when the publisher might have replicated data,\\n> table:%s.%s might have replicated data in the publisher\\\",\\n> + nspname, relname),\\n> + errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\\"));\\n> +\\n>\\n> The errmsg seems way too long for the source code. Can you use string\\n> concatenation or continuation chars to wrap the message over multiple\\n> lines?\\n\\nI had seen that the long error message elsewhere also is in a single\\nline. I think we should keep it as it is to maintain the coding\\nstandard.\\nThoughts?\\n\\n> ~~~\\n>\\n> 24. src/test/regress/sql/subscription.sql - typo\\n>\\n> @@ -66,6 +69,15 @@ ALTER SUBSCRIPTION regress_testsub4 SET (local_only = false);\\n> DROP SUBSCRIPTION regress_testsub3;\\n> DROP SUBSCRIPTION regress_testsub4;\\n>\\n> +-- ok - valid coy_data options\\n>\\n> Typo \\\"coy_data\\\". (it looks like this typo is not caused by this patch,\\n> but I think this patch should fix it anyhow).\\n\\nModified\\n\\n> ~~~\\n>\\n> 25. src/test/regress/sql/subscription.sql - test order\\n>\\n> The new tests are OK but IMO they could be re-ordered so then they\\n> will be more consistent for the positive and negative tests.\\n>\\n> CURRENT\\n> \\\"true/force/on/1\\\" and \\\"off/false/0\\\"\\n>\\n> SUGGEST\\n> \\\"true/on/1/force\\\" and \\\"false/off/0\\\"\\n\\nModified\\n\\nThanks for the comments, the v10 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm0PmOz71O6ofhZkB0rts5Ak2HUhMuuMQoViH_LAXTBeBw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 22 Apr 2022 21:53:48 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 20, 2022 at 11:19 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Below are my review comments for the v9-0002 patch (TAP test part only).\\n>\\n> (The order of my comments is a bit muddled because I jumped around in\\n> the file a bit while reviewing it).\\n>\\n> ======\\n>\\n> 1. create_subscription - missing comment.\\n>\\n> +sub create_subscription\\n> +{\\n> + my ($node_subscriber, $node_publisher, $sub_name, $node_connstr,\\n> + $application_name, $pub_name, $copy_data_val)\\n> + = @_;\\n>\\n> Maybe add a comment for this subroutine and describe the expected parameters.\\n\\nAdded\\n\\n> ~~\\n>\\n> 2. create_subscription - hides the options\\n>\\n> IMO the \\\"create_subscription\\\" subroutine is hiding too many of the\\n> details, so now it is less clear (from the caller's POV) what the test\\n> is really doing. Perhaps the options could be passed more explicitly\\n> to this subroutine.\\n>\\n> e.g. Instead of\\n>\\n> create_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\\n> $appname_B1, 'tap_pub_A', 'on');\\n>\\n> perhaps explicitly set the WITH options like:\\n>\\n> create_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\\n> $appname_B1, 'tap_pub_A', 'local_only = on, copy_data = on');\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. the application names are confusing\\n>\\n> +my $appname_A1 = 'tap_sub_A_B';\\n> +my $appname_A2 = 'tap_sub_A_C';\\n> +my $appname_B1 = 'tap_sub_B_A';\\n> +my $appname_B2 = 'tap_sub_B_C';\\n> +my $appname_C1 = 'tap_sub_C_A';\\n> +my $appname_C2 = 'tap_sub_C_B';\\n>\\n> I found the application names using '1' and '2' to be a bit confusing.\\n> i.e it was unnecessarily hard to associate them (in my head) with\\n> their relevant subscriptions. IMO it would be easier to name them\\n> using letters like below:\\n>\\n> SUGGESTED\\n> my $appname_AB = 'tap_sub_A_B';\\n> my $appname_AC = 'tap_sub_A_C';\\n> my $appname_BA = 'tap_sub_B_A';\\n> my $appname_BC = 'tap_sub_B_C';\\n> my $appname_CA = 'tap_sub_C_A';\\n> my $appname_CB = 'tap_sub_C_B';\\n\\nRemoved appname and used subscription names for application name\\n\\n> ~~~\\n>\\n> 4. create_subscription - passing the $appname 2x.\\n>\\n> +create_subscription($node_B, $node_A, $appname_B1, $node_A_connstr,\\n> + $appname_B1, 'tap_pub_A', 'on');\\n>\\n>\\n> It seemed confusing that the $app_name is passed twice.\\n>\\n> IMO should rename all those $appname_XX vars (see previous comment) to\\n> be $subname_XX, and just pass that create_subscription instead.\\n>\\n> my $subname_AB = 'tap_sub_A_B';\\n> my $subname_AC = 'tap_sub_A_C';\\n> my $subname_BA = 'tap_sub_B_A';\\n> my $subname_BC = 'tap_sub_B_C';\\n> my $subname_CA = 'tap_sub_C_A';\\n> my $subname_CB = 'tap_sub_C_B';\\n>\\n> Then create_subscription subroutine should have one less argument.\\n> Just add a comment saying that the appname is always assigned the same\\n> value as the subname.\\n\\nModifeid\\n\\n> ~~~\\n>\\n> 5. cleanup part - seems a bit misleading\\n>\\n> +# cleanup\\n> +$node_B->safe_psql(\\n> + 'postgres', \\\"\\n> + DROP SUBSCRIPTION $appname_B2\\\");\\n> +$node_C->safe_psql(\\n> + 'postgres', \\\"\\n> + DELETE FROM tab_full\\\");\\n> +$node_B->safe_psql(\\n> + 'postgres', \\\"\\n> + DELETE FROM tab_full where a = 13\\\");\\n>\\n> Is that comment misleading? IIUC this is not really cleaning up\\n> everything. It is just a cleanup of the previous Node_C test part\\n> isn't it?\\n\\nModified to clear the operations done by this test\\n\\n> ~~~\\n>\\n> 6. Error case (when copy_data is true)\\n>\\n> +# Error when creating susbcription with local_only and copy_data as true when\\n> +# the publisher has replicated data\\n>\\n> 6a. Typo \\\"susbcription\\\"\\n\\nModified\\n\\n> 6b. That comment maybe needs some more explanation - eg. say that\\n> since Node_A is already subscribing to Node_B so when Node_B makes\\n> another subscription to Node_A the copy doesn't really know if the\\n> data really originated from Node_A or not...\\n\\nSlightly reworded and modified\\n\\n> 6c. Maybe the comment needed to be more like ############## style to\\n> denote this (and the one that follows) is a separate test case.\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. Error case (when copy_data is force)\\n>\\n> +# Creating subscription with local_only and copy_data as force should be\\n> +# successful when the publisher has replicated data\\n>\\n> 7a. Same typo \\\"subscription\\\"\\n\\nI felt subscription is correct in this case, made no change for this\\n\\n> ~~~\\n>\\n> 8. Add 3rd node when the existing node has some data\\n>\\n> 8a. Maybe that comment needs to be expanded more. This isn't just\\n> \\\"adding a node\\\" - you are joining it to the others as another\\n> bi-directional participant in a 3-way group. And the comment should\\n> clarify exactly which nodes have the initial data. Both the existing 2\\n> you are joining to? The new one only? All of them?\\n\\nModified\\n\\n> 8b. IMO is would be much clearer to do SELECT from all the nodes at\\n> the start of this test just to re-confirm what is all the initial data\\n> on these nodes before joining the 3rd node.\\n\\nModified\\n\\n> NOTE - These same review comments apply to the other test combinations too\\n> - Add 3rd node when the existing node has no data\\n> - Add 3rd node when the new node has some data\\n\\nModified\\n\\n> ~~~\\n>\\n> 9. Inserted data\\n>\\n> +# insert some data in all the nodes\\n> +$node_A->safe_psql('postgres', \\\"INSERT INTO tab_full VALUES (13);\\\");\\n> +$node_B->safe_psql('postgres', \\\"INSERT INTO tab_full VALUES (21);\\\");\\n> +$node_C->safe_psql('postgres', \\\"INSERT INTO tab_full VALUES (31);\\\");\\n>\\n> They seemed strange values (13, 21, 31) to add. I couldn't work out\\n> the \\\"meaning\\\" of them.\\n>\\n> Wouldn't values like 13, 23, 33 make more sense (e.g pattern is 10 x\\n> node# + something)\\n\\nModified\\n\\n> ~~~\\n>\\n> 10. verify_data($node_A, $node_B, $node_C);\\n>\\n> All the expected values are too buried in this subroutine which makes\\n> it hard to read. I think it would be easier to read (from the caller's\\n> POV) if you can pass the *expected* values as another arg into the\\n> verify_data subroutine.\\n>\\n> e.g. is something like this possible?\\n> verify_data($node_A, $node_B, $node_C, [11,12,13])\\n\\nModified\\n\\n> ~~~\\n>\\n> 11. ALTER for ignoring 'truncate'\\n>\\n> +$node_C->safe_psql('postgres',\\n> + \\\"ALTER PUBLICATION tap_pub_C SET (publish='insert,update,delete');\\\");\\n> +\\n> +$node_C->safe_psql('postgres', \\\"TRUNCATE tab_full\\\");\\n> +\\n> +create_subscription($node_C, $node_A, $appname_C1, $node_A_connstr,\\n> + $appname_C1, 'tap_pub_A', 'force');\\n> +create_subscription($node_C, $node_B, $appname_C2, $node_B_connstr,\\n> + $appname_C2, 'tap_pub_B', 'off');\\n> +\\n> +#include truncates now\\n> +$node_C->safe_psql('postgres',\\n> + \\\"ALTER PUBLICATION tap_pub_C SET\\n> (publish='insert,update,delete,truncate');\\\"\\n> +);\\n>\\n> Do those create_subscription calls need to be encapsulated by the\\n> ALTER like that? I did not think so. It looks a bit more complex if\\n> done this way.\\n\\nModified\\n\\n> ~~~\\n>\\n> 12. clean_subscriber_contents - misnamed?\\n>\\n> This subroutine feels a bit misnamed. It seems to be doing lots of\\n> things like detaching the Node_C and deleting all table data from all\\n> nodes. That all seems quite different from just \\\"clean subscriber\\n> contents\\\".\\n\\nchanged it to detach_node_clean_table_data\\n\\n> ~~~\\n>\\n> 13. table initial data?\\n>\\n> +clean_subscriber_contents($node_A, $node_B, $node_C);\\n> +\\n> +##########################################################################\\n> +# Add 3rd node when the new node has some data\\n> +##########################################################################\\n>\\n> But does this test case *really* have some data? I am not so sure.\\n> Doesn't the preceding \\\"clean_subscriber_contents\\\" call remove all the\\n> data that might have been there? That is why I think all the tests out\\n> to have SELECT (previous comment #8) so they can re-confirm what data\\n> is really in those tables before doing each test part.\\n\\nModified\\n\\nThanks for the comments, the v10 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm0PmOz71O6ofhZkB0rts5Ak2HUhMuuMQoViH_LAXTBeBw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 22 Apr 2022 21:58:22 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for v10-0001 and v10-0002.\\n\\n(I did not yet look at the v10-0002 TAP tests. I will check those\\nseparately later).\\n\\n=================\\nv10-0001 comments\\n=================\\n\\n1.1 src/include/catalog/pg_subscription.h\\n\\n@@ -70,6 +70,8 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\\nBKI_SHARED_RELATION BKI_ROW\\n\\n bool substream; /* Stream in-progress transactions. */\\n\\n+ bool sublocalonly; /* skip copying of remote origin data */\\n+\\n char subtwophasestate; /* Stream two-phase transactions */\\n\\n bool subdisableonerr; /* True if a worker error should cause the\\n\\n1.1.a Comment should start uppercase.\\n\\n1.1.b Perhaps it is better to say \\\"Skip replication of\\\" instead of\\n\\\"Skip copying\\\" ?\\n\\n~~~\\n\\n1.2 src/include/catalog/pg_subscription.h\\n\\n@@ -110,6 +112,7 @@ typedef struct Subscription\\n bool binary; /* Indicates if the subscription wants data in\\n * binary format */\\n bool stream; /* Allow streaming in-progress transactions. */\\n+ bool local_only; /* Skip copying of remote origin data */\\n char twophasestate; /* Allow streaming two-phase transactions */\\n\\nSame comment as #1.1\\n\\n=================\\nv10-0002 comments\\n=================\\n\\n2.1 Commit message\\n\\nb) user wil have to create another subscription subscribing from\\nNode2 to Node3 using local_only option and copy_data as true.\\n\\nTypo: \\\"wil\\\"\\n\\n~~~\\n\\n2.2 Commit message\\n\\nHere when user has specified local_only 'on' which indicates that the\\npublisher should only replicate the changes that are generated locally, but in\\nthis case since the publisher node is also subscribing data from other nodes,\\nthe publisher node can have remotely originated data, so throw an error in this\\ncase to prevent remotely generated data being replicated to the subscriber. If\\nuser still intends to continue with the operation user can specify copy_data\\nas 'force' and proceed.\\n\\nSUGGESTED (minor rewording)\\nIn the scenario above the user has specified local_only 'on' (which\\nindicates that the publisher should only replicate the changes that\\nare generated locally), but in this case, the publisher node is also\\nsubscribing data from other nodes, so the publisher node may have\\nremotely originated data. We throw an error, in this case, to draw\\nattention to there being possible remote data. If the user still\\nwishes to continue with the operation user can specify copy_data as\\n'force' and proceed.\\n\\n~~~\\n\\n2.3 doc/src/sgml/logical-replication.sgml\\n\\n+ <para>\\n+ Bidirectional replication is useful in creating multi master database\\n+ which helps in performing read/write operations from any of the nodes.\\n+ Setting up bidirectional logical replication between two nodes requires\\n+ creation of the publication in all the nodes, creating subscription in\\n+ each of the nodes that subcribes to data from all the nodes. The steps\\n+ to create two node bidirectional replication is listed below:\\n+ </para>\\n\\n2.3.a Typo: \\\"subcribes\\\"\\n\\n2.3.b Wording: \\\"creating subscription” -> \\\"and creating subscriptions...\\\"\\n\\n2.3.c Wording: \\\"The steps to create two node bidirectional replication\\nis listed below:\\\" -> \\\"The steps to create a two-node bidirectional\\nreplication are given below:\\\"\\n\\n~~~\\n\\n2.4 doc/src/sgml/logical-replication.sgml\\n\\n+<programlisting>\\n+node2=# CREATE SUBSCRIPTION sub_node1_node2\\n+node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\\n+node2=# PUBLICATION pub_node1\\n+node2=# WITH (copy_data = off, local_only = on);\\n+CREATE SUBSCRIPTION\\n+</programlisting>\\n\\nI am not sure if those psql continuation prompts are right. Shouldn't\\nthey be \\\"node2-#\\\" instead of \\\"node2=#\\\"?\\n\\n~~~\\n\\n2.5 doc/src/sgml/logical-replication.sgml\\n\\n+<programlisting>\\n+node2=# CREATE SUBSCRIPTION sub_node1_node2\\n+node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\\n+node2=# PUBLICATION pub_node1\\n+node2=# WITH (copy_data = off, local_only = on);\\n+CREATE SUBSCRIPTION\\n+</programlisting>\\n+ </para>\\n\\nIIUC the closing </para> should be on the same line as the\\n</programlisting>. I recall there was some recent github push about\\nthis sometime in the last month - maybe you can check to confirm it.\\n\\n~~~\\n\\n2.6 doc/src/sgml/logical-replication.sgml\\n\\n+ <para>\\n+ Create the subscription in node2 to subscribe the changes from node1:\\n+<programlisting>\\n+node2=# CREATE SUBSCRIPTION sub_node1_node2\\n\\nIMO the naming convention here is backwards/confusing. E.g. The\\n\\\"pub_node1\\\" is the publisher at node1. So similarly, I think the\\nsubscriber at node2 should be called \\\"sub_node2_node1\\\" (not\\n\\\"sub_node1_node2\\\")\\n\\n~~~\\n\\n2.7 doc/src/sgml/logical-replication.sgml\\n\\n+ <para>\\n+ Adding a new node node3 to the existing node1 and node2 requires setting\\n+ up subscription in node1 and node2 to replicate the data from node3 and\\n+ setting up subscription in node3 to replicate data from node1 and node2.\\n+ The steps for the same is listed below:\\n+ </para>\\n\\nThere are several sections that say \\\"The steps for the same is listed\\nbelow:\\\". The sentence for all of them seemed redundant to me.\\n\\n~~~\\n\\n2.8 doc/src/sgml/logical-replication.sgml\\n\\n+ <para>\\n+ Adding a new node node3 to the existing node1 and node2 when data is\\n+ present in existing nodes node1 and node2 needs similar steps, only change\\n+ required here is that node3 should create subscription with copy_data as\\n+ force to one of the existing nodes to receive the existing data during\\n+ initial data synchronization. The steps for the same is listed below:\\n+ </para>\\n\\nI thought it should be 2 sentences. So \\\"needs similar steps, only\\nchange required here\\\" -> \\\"needs similar steps. The only change\\nrequired here...\\\"\\n\\n~~~\\n\\n2.9 doc/src/sgml/logical-replication.sgml\\n\\nI think every time you say option names or values like \\\"copy_data\\\" in\\nthe text they should be using <literal> font.\\n\\n~~~\\n\\n2.10 doc/src/sgml/logical-replication.sgml\\n\\n+ <para>\\n+ Adding a new node node3 to the existing node1 and node2 when data is\\n+ present in the new node node3 needs similar steps, few changes are\\n+ required here to get the existing data from node3 to node1 and node2 and\\n+ later cleaning up of data in node3 before synchronization of all the data\\n+ from the existing nodes. The steps for the same is listed below:\\n+ </para>\\n\\nI thought it should be 2 sentences. So \\\"needs similar steps, few\\nchanges are required here\\\" -> \\\"needs similar steps. A few changes are\\nrequired here...\\\"\\n\\n~~~\\n\\n2.11 doc/src/sgml/logical-replication.sgml\\n\\nAll the text that says \\\"create subscription\\\" and \\\"create publication\\\"\\nmaybe should be change to \\\"create a subscription\\\" and \\\"create a\\npublication\\\" etc.\\n\\n~~~\\n\\n2.12 doc/src/sgml/ref/alter_subscription.sgml - copy_data\\n\\n+ <para>\\n+ There is some interaction between the \\\"local_only\\\" option and\\n+ \\\"copy_data\\\" option. Refer to the\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n+ details and usage of force for copy_data option.\\n </para>\\n\\n2.12.a Maybe options should not be in quotes like that. I think they\\nshould be <literal> font.\\n\\n2.12.b It is a bit misleading because there is no \\\"Notes\\\" section here\\non this page. Maybe it should say refer to the CREATE SUBSCRIPTION\\nNotes.\\n\\n2.12.c The last copy_data should also be <literal> font.\\n\\n~~~\\n\\n2.13 doc/src/sgml/ref/create_subscription.sgml - local_only\\n\\n </para>\\n+ <para>\\n+ There is some interaction between the \\\"local_only\\\" option and\\n+ \\\"copy_data\\\" option. Refer to the\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n+ </para>\\n </listitem>\\n\\n2.13.a Maybe options should not be in quotes like that. I think they\\nshould be <literal> font.\\n\\n2.13.b The last copy_data should also be <literal> font.\\n\\n~~~\\n\\n2.14 doc/src/sgml/ref/create_subscription.sgml - copy_data\\n\\n\\n <para>\\n Specifies whether to copy pre-existing data in the publications\\n- that are being subscribed to when the replication starts.\\n- The default is <literal>true</literal>.\\n+ that are being subscribed to when the replication starts. This\\n+ parameter may be either <literal>true</literal>,\\n+ <literal>false</literal> or <literal>force</literal>. The default is\\n+ <literal>true</literal>.\\n </para>\\n\\n2.14.a Maybe options should not be in quotes like that. I think they\\nshould be <literal> font.\\n\\n2.14.b The last copy_data should also be <literal> font.\\n\\n~~~\\n\\n2.15 doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -374,6 +388,16 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n can have non-existent publications.\\n </para>\\n\\n+ <para>\\n+ If subscription is created with local_only as 'on' and copy_data as 'on', it\\n+ will check if the publisher tables are being subscribed to any other\\n+ publisher and throw an error to prevent inconsistent data in the\\n+ subscription. User can continue with the copy operation without throwing any\\n+ error in this case by specifying copy_data as 'force'. Refer to the\\n+ <xref linkend=\\\"bidirectional-logical-replication\\\"/> on how\\n+ copy_data and local_only can be used in bidirectional replication.\\n+ </para>\\n\\n2.15.a I think all those mentions of copy_data and local_only and\\ntheir values should be in <literal> font instead of in quotes.\\n\\n2.15.b Wording: \\\"User can\\\" -> \\\"The user can\\\"\\n\\n~~~\\n\\n2.16 src/backend/commands/subscriptioncmds.c - macro\\n\\n> 20. src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_ON_OR_FORCE\\n>\\n> @@ -69,6 +69,18 @@\\n> /* check if the 'val' has 'bits' set */\\n> #define IsSet(val, bits) (((val) & (bits)) == (bits))\\n>\\n> +#define IS_COPY_DATA_ON_OR_FORCE(copy_data) (copy_data != COPY_DATA_OFF)\\n>\\n> Maybe this would be better as a static inline function?\\n\\nVignesh: What is the advantage of doing this change? I have not changed this\\nas the macro usage is fine. Thoughts?\\n\\nOriginally I was going to suggest the macro should use extra parens\\nlike ((copy_data) != COPY_DATA_OFF), but then I thought if it was a\\nfunction then it would have enum type-checking which would be better.\\nIf you want to keep the macro then please make the parens change.\\n\\n~~~\\n\\n2.17 src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n\\n+ /*\\n+ * The set of strings accepted here should match up with\\n+ * the grammar's opt_boolean_or_string production.\\n+ */\\n+ if (pg_strcasecmp(sval, \\\"true\\\") == 0 ||\\n+ pg_strcasecmp(sval, \\\"on\\\") == 0)\\n+ return COPY_DATA_ON;\\n+ if (pg_strcasecmp(sval, \\\"false\\\") == 0 ||\\n+ pg_strcasecmp(sval, \\\"off\\\") == 0)\\n+ return COPY_DATA_OFF;\\n+ if (pg_strcasecmp(sval, \\\"force\\\") == 0)\\n+ return COPY_DATA_FORCE;\\n\\nI think you can change the order of these to be off/on/force, so then\\nthe order is consistent with the T_Integer case.\\n\\n~~~\\n\\n2.18 src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n\\n+ ereport(ERROR,\\n+ errcode(ERRCODE_SYNTAX_ERROR),\\n+ errmsg(\\\"%s requires a boolean or \\\\\\\"force\\\\\\\"\\\", def->defname));\\n+ return COPY_DATA_OFF; /*\\nkeep compiler quiet */\\n\\nExcessive comment indent? Has pg_indent been run?\\n\\n~~~\\n\\n2.19 src/backend/commands/subscriptioncmds.c - AlterSubscription\\n\\n@@ -1236,7 +1306,8 @@ AlterSubscription(ParseState *pstate,\\nAlterSubscriptionStmt *stmt,\\n errmsg(\\\"ALTER SUBSCRIPTION ... REFRESH is not allowed for disabled\\nsubscriptions\\\")));\\n\\n parse_subscription_options(pstate, stmt->options,\\n- SUBOPT_COPY_DATA, &opts);\\n+ SUBOPT_COPY_DATA,\\n+ &opts);\\n\\nThis is a formatting change only. Maybe it does not belong in this\\npatch unless it is the result of pg_indent.\\n\\n~~~\\n\\n2.20 src/backend/commands/subscriptioncmds.c - fetch_table_list\\n\\n> 23. src/backend/commands/subscriptioncmds.c - long errmsg\\n>\\n> + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"CREATE/ALTER SUBSCRIPTION with local_only and copy_data as\\n> true is not allowed when the publisher might have replicated data,\\n> table:%s.%s might have replicated data in the publisher\\\",\\n> + nspname, relname),\\n> + errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\\"));\\n> +\\n>\\n> The errmsg seems way too long for the source code. Can you use string\\n> concatenation or continuation chars to wrap the message over multiple\\n> lines?\\n\\nVignesh: I had seen that the long error message elsewhere also is in a\\nsingle line. I think we should keep it as it is to maintain the coding\\nstandard. Thoughts?\\n\\nOK, if you say it is already common practice then it's fine by me to\\nleave it as-is.\\n\\n~~~\\n\\n2.21 src/test/regress/expected/subscription.out - make check\\n\\nmake check fails.\\n 1 of 214 tests failed.\\n\\n2.21.a It looks like maybe you did not update the expected ordering of\\nsome of the tests, after some minor adjustments in subscriprion.sql in\\nv10. So the expected output needs to be fixed in the patch.\\n\\n2.21.b. Suggest adding this patch to CF so that the cfbot can pick up\\nsuch test problems earlier.\\n\\n\\n------\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 27 Apr 2022 15:45:28 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for v10-0002 (TAP tests part only)\\n\\nFIle: src/test/subscription/t/032_localonly.pl\\n\\n======\\n\\n1.\\n\\n+# Detach node C and clean the table contents.\\n+sub detach_node_clean_table_data\\n+{\\n\\n1a. Maybe say \\\"Detach node C from the node-group of (A, B, C) and\\nclean the table contents from all nodes\\\"\\n\\n1b. Actually wondered do you need to TRUNCATE from both A and B (maybe\\nonly truncate 1 is OK since those nodes are still using MMC). OTOH\\nmaybe your explicit way makes the test simpler.\\n\\n~~~\\n\\n2.\\n\\n+# Subroutine for verify the data is replicated successfully.\\n+sub verify_data\\n+{\\n\\n2a. Typo: \\\"for verify\\\" -> \\\"to verify\\\"\\n\\n2b. The messages in this function maybe are not very appropriate. They\\nsay 'Inserted successfully without leading to infinite recursion in\\ncircular replication setup', but really the function is only testing\\nall the data is the same as 'expected'. So it could be the result of\\nany operation - not just Insert.\\n\\n~~~\\n\\n3. SELECT ORDER BY?\\n\\n$result = $node_B->safe_psql('postgres', \\\"SELECT * FROM tab_full;\\\");\\nis($result, qq(11\\n12\\n13),\\n 'Node_C data replicated to Node_B'\\n);\\n\\nI am not sure are these OK like this or if *every* SELECT use ORDER BY\\nto make sure the data is in the same qq expected order? There are\\nmultiple cases like this.\\n\\n(BTW, I think this comment needs to be applied for the v10-0001 patch,\\nmaybe not v10-0002).\\n\\n~~~\\n\\n4.\\n\\n+###############################################################################\\n+# Specifying local_only 'on' which indicates that the publisher should only\\n+# replicate the changes that are generated locally from node_B, but in\\n+# this case since the node_B is also subscribing data from node_A, node_B can\\n+# have data originated from node_A, so throw an error in this case to prevent\\n+# node_A data being replicated to the node_C.\\n+###############################################################################\\n\\nThere is something wrong with the description because there is no\\n\\\"node_C\\\" in this test. You are just creating a 2nd subscription on\\nnode A.\\n\\n~~\\n\\n5.\\n\\n+($result, $stdout, $stderr) = $node_A->psql(\\n+ 'postgres', \\\"\\n+ CREATE SUBSCRIPTION tap_sub_A3\\n+ CONNECTION '$node_B_connstr application_name=$subname_AB'\\n+ PUBLICATION tap_pub_B\\n+ WITH (local_only = on, copy_data = on)\\\");\\n+like(\\n\\nIt seemed strange to call this 2nd subscription \\\"tap_sub_A3\\\". Wouldn't\\nit be better to call it \\\"tap_sub_A2\\\"?\\n\\n~~~\\n\\n6.\\n\\n+###############################################################################\\n+# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\\n+# replication setup when the existing nodes (node_A & node_B) has no data and\\n+# the new node (node_C) some pre-existing data.\\n+###############################################################################\\n+$node_C->safe_psql('postgres', \\\"INSERT INTO tab_full VALUES (31);\\\");\\n+\\n+$result = $node_A->safe_psql('postgres', \\\"SELECT * FROM tab_full order by 1;\\\");\\n+is( $result, qq(), 'Check existing data');\\n+\\n+$result = $node_B->safe_psql('postgres', \\\"SELECT * FROM tab_full order by 1;\\\");\\n+is( $result, qq(), 'Check existing data');\\n+\\n+$result = $node_C->safe_psql('postgres', \\\"SELECT * FROM tab_full order by 1;\\\");\\n+is($result, qq(31), 'Check existing data');\\n+\\n+create_subscription($node_A, $node_C, $subname_AC, $node_C_connstr,\\n+ 'tap_pub_C', 'copy_data = force, local_only = on');\\n+create_subscription($node_B, $node_C, $subname_BC, $node_C_connstr,\\n+ 'tap_pub_C', 'copy_data = force, local_only = on');\\n+\\n\\nBecause the Node_C does not yet have any subscriptions aren't these\\ncases where you didn't really need to use \\\"force\\\"?\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 28 Apr 2022 12:27:15 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 27, 2022 at 11:15 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for v10-0001 and v10-0002.\\n>\\n> (I did not yet look at the v10-0002 TAP tests. I will check those\\n> separately later).\\n>\\n> =================\\n> v10-0001 comments\\n> =================\\n>\\n> 1.1 src/include/catalog/pg_subscription.h\\n>\\n> @@ -70,6 +70,8 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\\n> BKI_SHARED_RELATION BKI_ROW\\n>\\n> bool substream; /* Stream in-progress transactions. */\\n>\\n> + bool sublocalonly; /* skip copying of remote origin data */\\n> +\\n> char subtwophasestate; /* Stream two-phase transactions */\\n>\\n> bool subdisableonerr; /* True if a worker error should cause the\\n>\\n> 1.1.a Comment should start uppercase.\\n\\nModified\\n\\n> 1.1.b Perhaps it is better to say \\\"Skip replication of\\\" instead of\\n> \\\"Skip copying\\\" ?\\n\\nModified\\n\\n> ~~~\\n>\\n> 1.2 src/include/catalog/pg_subscription.h\\n>\\n> @@ -110,6 +112,7 @@ typedef struct Subscription\\n> bool binary; /* Indicates if the subscription wants data in\\n> * binary format */\\n> bool stream; /* Allow streaming in-progress transactions. */\\n> + bool local_only; /* Skip copying of remote origin data */\\n> char twophasestate; /* Allow streaming two-phase transactions */\\n>\\n> Same comment as #1.1\\n\\nModified\\n\\n> =================\\n> v10-0002 comments\\n> =================\\n>\\n> 2.1 Commit message\\n>\\n> b) user wil have to create another subscription subscribing from\\n> Node2 to Node3 using local_only option and copy_data as true.\\n>\\n> Typo: \\\"wil\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.2 Commit message\\n>\\n> Here when user has specified local_only 'on' which indicates that the\\n> publisher should only replicate the changes that are generated locally, but in\\n> this case since the publisher node is also subscribing data from other nodes,\\n> the publisher node can have remotely originated data, so throw an error in this\\n> case to prevent remotely generated data being replicated to the subscriber. If\\n> user still intends to continue with the operation user can specify copy_data\\n> as 'force' and proceed.\\n>\\n> SUGGESTED (minor rewording)\\n> In the scenario above the user has specified local_only 'on' (which\\n> indicates that the publisher should only replicate the changes that\\n> are generated locally), but in this case, the publisher node is also\\n> subscribing data from other nodes, so the publisher node may have\\n> remotely originated data. We throw an error, in this case, to draw\\n> attention to there being possible remote data. If the user still\\n> wishes to continue with the operation user can specify copy_data as\\n> 'force' and proceed.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.3 doc/src/sgml/logical-replication.sgml\\n>\\n> + <para>\\n> + Bidirectional replication is useful in creating multi master database\\n> + which helps in performing read/write operations from any of the nodes.\\n> + Setting up bidirectional logical replication between two nodes requires\\n> + creation of the publication in all the nodes, creating subscription in\\n> + each of the nodes that subcribes to data from all the nodes. The steps\\n> + to create two node bidirectional replication is listed below:\\n> + </para>\\n>\\n> 2.3.a Typo: \\\"subcribes\\\"\\n\\nModified\\n\\n> 2.3.b Wording: \\\"creating subscription” -> \\\"and creating subscriptions...\\\"\\n\\nModified\\n\\n> 2.3.c Wording: \\\"The steps to create two node bidirectional replication\\n> is listed below:\\\" -> \\\"The steps to create a two-node bidirectional\\n> replication are given below:\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.4 doc/src/sgml/logical-replication.sgml\\n>\\n> +<programlisting>\\n> +node2=# CREATE SUBSCRIPTION sub_node1_node2\\n> +node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\\n> +node2=# PUBLICATION pub_node1\\n> +node2=# WITH (copy_data = off, local_only = on);\\n> +CREATE SUBSCRIPTION\\n> +</programlisting>\\n>\\n> I am not sure if those psql continuation prompts are right. Shouldn't\\n> they be \\\"node2-#\\\" instead of \\\"node2=#\\\"?\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.5 doc/src/sgml/logical-replication.sgml\\n>\\n> +<programlisting>\\n> +node2=# CREATE SUBSCRIPTION sub_node1_node2\\n> +node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\\n> +node2=# PUBLICATION pub_node1\\n> +node2=# WITH (copy_data = off, local_only = on);\\n> +CREATE SUBSCRIPTION\\n> +</programlisting>\\n> + </para>\\n>\\n> IIUC the closing </para> should be on the same line as the\\n> </programlisting>. I recall there was some recent github push about\\n> this sometime in the last month - maybe you can check to confirm it.\\n\\n I have seen the programlisting across multiple places and noticed\\nthat there is no such guideline being followed. I have not made any\\nchange for this comment.\\n\\n> ~~~\\n>\\n> 2.6 doc/src/sgml/logical-replication.sgml\\n>\\n> + <para>\\n> + Create the subscription in node2 to subscribe the changes from node1:\\n> +<programlisting>\\n> +node2=# CREATE SUBSCRIPTION sub_node1_node2\\n>\\n> IMO the naming convention here is backwards/confusing. E.g. The\\n> \\\"pub_node1\\\" is the publisher at node1. So similarly, I think the\\n> subscriber at node2 should be called \\\"sub_node2_node1\\\" (not\\n> \\\"sub_node1_node2\\\")\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.7 doc/src/sgml/logical-replication.sgml\\n>\\n> + <para>\\n> + Adding a new node node3 to the existing node1 and node2 requires setting\\n> + up subscription in node1 and node2 to replicate the data from node3 and\\n> + setting up subscription in node3 to replicate data from node1 and node2.\\n> + The steps for the same is listed below:\\n> + </para>\\n>\\n> There are several sections that say \\\"The steps for the same is listed\\n> below:\\\". The sentence for all of them seemed redundant to me.\\n\\nRemoved it.\\n\\n> ~~~\\n>\\n> 2.8 doc/src/sgml/logical-replication.sgml\\n>\\n> + <para>\\n> + Adding a new node node3 to the existing node1 and node2 when data is\\n> + present in existing nodes node1 and node2 needs similar steps, only change\\n> + required here is that node3 should create subscription with copy_data as\\n> + force to one of the existing nodes to receive the existing data during\\n> + initial data synchronization. The steps for the same is listed below:\\n> + </para>\\n>\\n> I thought it should be 2 sentences. So \\\"needs similar steps, only\\n> change required here\\\" -> \\\"needs similar steps. The only change\\n> required here...\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.9 doc/src/sgml/logical-replication.sgml\\n>\\n> I think every time you say option names or values like \\\"copy_data\\\" in\\n> the text they should be using <literal> font.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.10 doc/src/sgml/logical-replication.sgml\\n>\\n> + <para>\\n> + Adding a new node node3 to the existing node1 and node2 when data is\\n> + present in the new node node3 needs similar steps, few changes are\\n> + required here to get the existing data from node3 to node1 and node2 and\\n> + later cleaning up of data in node3 before synchronization of all the data\\n> + from the existing nodes. The steps for the same is listed below:\\n> + </para>\\n>\\n> I thought it should be 2 sentences. So \\\"needs similar steps, few\\n> changes are required here\\\" -> \\\"needs similar steps. A few changes are\\n> required here...\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.11 doc/src/sgml/logical-replication.sgml\\n>\\n> All the text that says \\\"create subscription\\\" and \\\"create publication\\\"\\n> maybe should be change to \\\"create a subscription\\\" and \\\"create a\\n> publication\\\" etc.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.12 doc/src/sgml/ref/alter_subscription.sgml - copy_data\\n>\\n> + <para>\\n> + There is some interaction between the \\\"local_only\\\" option and\\n> + \\\"copy_data\\\" option. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n> + details and usage of force for copy_data option.\\n> </para>\\n>\\n> 2.12.a Maybe options should not be in quotes like that. I think they\\n> should be <literal> font.\\n\\nModified\\n\\n> 2.12.b It is a bit misleading because there is no \\\"Notes\\\" section here\\n> on this page. Maybe it should say refer to the CREATE SUBSCRIPTION\\n> Notes.\\n\\nModified\\n\\n> 2.12.c The last copy_data should also be <literal> font.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.13 doc/src/sgml/ref/create_subscription.sgml - local_only\\n>\\n> </para>\\n> + <para>\\n> + There is some interaction between the \\\"local_only\\\" option and\\n> + \\\"copy_data\\\" option. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n> + </para>\\n> </listitem>\\n>\\n> 2.13.a Maybe options should not be in quotes like that. I think they\\n> should be <literal> font.\\n\\nModified\\n\\n> 2.13.b The last copy_data should also be <literal> font.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.14 doc/src/sgml/ref/create_subscription.sgml - copy_data\\n>\\n>\\n> <para>\\n> Specifies whether to copy pre-existing data in the publications\\n> - that are being subscribed to when the replication starts.\\n> - The default is <literal>true</literal>.\\n> + that are being subscribed to when the replication starts. This\\n> + parameter may be either <literal>true</literal>,\\n> + <literal>false</literal> or <literal>force</literal>. The default is\\n> + <literal>true</literal>.\\n> </para>\\n>\\n> 2.14.a Maybe options should not be in quotes like that. I think they\\n> should be <literal> font.\\n\\nModified\\n\\n> 2.14.b The last copy_data should also be <literal> font.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.15 doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -374,6 +388,16 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> can have non-existent publications.\\n> </para>\\n>\\n> + <para>\\n> + If subscription is created with local_only as 'on' and copy_data as 'on', it\\n> + will check if the publisher tables are being subscribed to any other\\n> + publisher and throw an error to prevent inconsistent data in the\\n> + subscription. User can continue with the copy operation without throwing any\\n> + error in this case by specifying copy_data as 'force'. Refer to the\\n> + <xref linkend=\\\"bidirectional-logical-replication\\\"/> on how\\n> + copy_data and local_only can be used in bidirectional replication.\\n> + </para>\\n>\\n> 2.15.a I think all those mentions of copy_data and local_only and\\n> their values should be in <literal> font instead of in quotes.\\n\\nModified\\n\\n> 2.15.b Wording: \\\"User can\\\" -> \\\"The user can\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.16 src/backend/commands/subscriptioncmds.c - macro\\n>\\n> > 20. src/backend/commands/subscriptioncmds.c - IS_COPY_DATA_ON_OR_FORCE\\n> >\\n> > @@ -69,6 +69,18 @@\\n> > /* check if the 'val' has 'bits' set */\\n> > #define IsSet(val, bits) (((val) & (bits)) == (bits))\\n> >\\n> > +#define IS_COPY_DATA_ON_OR_FORCE(copy_data) (copy_data != COPY_DATA_OFF)\\n> >\\n> > Maybe this would be better as a static inline function?\\n>\\n> Vignesh: What is the advantage of doing this change? I have not changed this\\n> as the macro usage is fine. Thoughts?\\n>\\n> Originally I was going to suggest the macro should use extra parens\\n> like ((copy_data) != COPY_DATA_OFF), but then I thought if it was a\\n> function then it would have enum type-checking which would be better.\\n> If you want to keep the macro then please make the parens change.\\n\\nModified to include parenthesis.\\n\\n> ~~~\\n>\\n> 2.17 src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n>\\n> + /*\\n> + * The set of strings accepted here should match up with\\n> + * the grammar's opt_boolean_or_string production.\\n> + */\\n> + if (pg_strcasecmp(sval, \\\"true\\\") == 0 ||\\n> + pg_strcasecmp(sval, \\\"on\\\") == 0)\\n> + return COPY_DATA_ON;\\n> + if (pg_strcasecmp(sval, \\\"false\\\") == 0 ||\\n> + pg_strcasecmp(sval, \\\"off\\\") == 0)\\n> + return COPY_DATA_OFF;\\n> + if (pg_strcasecmp(sval, \\\"force\\\") == 0)\\n> + return COPY_DATA_FORCE;\\n>\\n> I think you can change the order of these to be off/on/force, so then\\n> the order is consistent with the T_Integer case.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.18 src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n>\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_SYNTAX_ERROR),\\n> + errmsg(\\\"%s requires a boolean or \\\\\\\"force\\\\\\\"\\\", def->defname));\\n> + return COPY_DATA_OFF; /*\\n> keep compiler quiet */\\n>\\n> Excessive comment indent? Has pg_indent been run?\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.19 src/backend/commands/subscriptioncmds.c - AlterSubscription\\n>\\n> @@ -1236,7 +1306,8 @@ AlterSubscription(ParseState *pstate,\\n> AlterSubscriptionStmt *stmt,\\n> errmsg(\\\"ALTER SUBSCRIPTION ... REFRESH is not allowed for disabled\\n> subscriptions\\\")));\\n>\\n> parse_subscription_options(pstate, stmt->options,\\n> - SUBOPT_COPY_DATA, &opts);\\n> + SUBOPT_COPY_DATA,\\n> + &opts);\\n>\\n> This is a formatting change only. Maybe it does not belong in this\\n> patch unless it is the result of pg_indent.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.20 src/backend/commands/subscriptioncmds.c - fetch_table_list\\n>\\n> > 23. src/backend/commands/subscriptioncmds.c - long errmsg\\n> >\\n> > + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\\n> > + ereport(ERROR,\\n> > + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> > + errmsg(\\\"CREATE/ALTER SUBSCRIPTION with local_only and copy_data as\\n> > true is not allowed when the publisher might have replicated data,\\n> > table:%s.%s might have replicated data in the publisher\\\",\\n> > + nspname, relname),\\n> > + errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\\"));\\n> > +\\n> >\\n> > The errmsg seems way too long for the source code. Can you use string\\n> > concatenation or continuation chars to wrap the message over multiple\\n> > lines?\\n>\\n> Vignesh: I had seen that the long error message elsewhere also is in a\\n> single line. I think we should keep it as it is to maintain the coding\\n> standard. Thoughts?\\n>\\n> OK, if you say it is already common practice then it's fine by me to\\n> leave it as-is.\\n\\nOk, no change done.\\n\\n> ~~~\\n>\\n> 2.21 src/test/regress/expected/subscription.out - make check\\n>\\n> make check fails.\\n> 1 of 214 tests failed.\\n>\\n> 2.21.a It looks like maybe you did not update the expected ordering of\\n> some of the tests, after some minor adjustments in subscriprion.sql in\\n> v10. So the expected output needs to be fixed in the patch.\\n\\nModified\\n\\n> 2.21.b. Suggest adding this patch to CF so that the cfbot can pick up\\n> such test problems earlier.\\n\\nCF entry added\\n\\nThanks for the comments, the attached v11 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Thu, 28 Apr 2022 22:34:12 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 28, 2022 at 7:57 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for v10-0002 (TAP tests part only)\\n>\\n> FIle: src/test/subscription/t/032_localonly.pl\\n>\\n> ======\\n>\\n> 1.\\n>\\n> +# Detach node C and clean the table contents.\\n> +sub detach_node_clean_table_data\\n> +{\\n>\\n> 1a. Maybe say \\\"Detach node C from the node-group of (A, B, C) and\\n> clean the table contents from all nodes\\\"\\n\\nModified\\n\\n> 1b. Actually wondered do you need to TRUNCATE from both A and B (maybe\\n> only truncate 1 is OK since those nodes are still using MMC). OTOH\\n> maybe your explicit way makes the test simpler.\\n\\ncleaning the data in all nodes to keep the test simpler\\n\\n> ~~~\\n>\\n> 2.\\n>\\n> +# Subroutine for verify the data is replicated successfully.\\n> +sub verify_data\\n> +{\\n>\\n> 2a. Typo: \\\"for verify\\\" -> \\\"to verify\\\"\\n\\nModified\\n\\n> 2b. The messages in this function maybe are not very appropriate. They\\n> say 'Inserted successfully without leading to infinite recursion in\\n> circular replication setup', but really the function is only testing\\n> all the data is the same as 'expected'. So it could be the result of\\n> any operation - not just Insert.\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. SELECT ORDER BY?\\n>\\n> $result = $node_B->safe_psql('postgres', \\\"SELECT * FROM tab_full;\\\");\\n> is($result, qq(11\\n> 12\\n> 13),\\n> 'Node_C data replicated to Node_B'\\n> );\\n>\\n> I am not sure are these OK like this or if *every* SELECT use ORDER BY\\n> to make sure the data is in the same qq expected order? There are\\n> multiple cases like this.\\n>\\n> (BTW, I think this comment needs to be applied for the v10-0001 patch,\\n> maybe not v10-0002).\\n\\nModified\\n\\n> ~~~\\n>\\n> 4.\\n>\\n> +###############################################################################\\n> +# Specifying local_only 'on' which indicates that the publisher should only\\n> +# replicate the changes that are generated locally from node_B, but in\\n> +# this case since the node_B is also subscribing data from node_A, node_B can\\n> +# have data originated from node_A, so throw an error in this case to prevent\\n> +# node_A data being replicated to the node_C.\\n> +###############################################################################\\n>\\n> There is something wrong with the description because there is no\\n> \\\"node_C\\\" in this test. You are just creating a 2nd subscription on\\n> node A.\\n\\nModified\\n\\n> ~~\\n>\\n> 5.\\n>\\n> +($result, $stdout, $stderr) = $node_A->psql(\\n> + 'postgres', \\\"\\n> + CREATE SUBSCRIPTION tap_sub_A3\\n> + CONNECTION '$node_B_connstr application_name=$subname_AB'\\n> + PUBLICATION tap_pub_B\\n> + WITH (local_only = on, copy_data = on)\\\");\\n> +like(\\n>\\n> It seemed strange to call this 2nd subscription \\\"tap_sub_A3\\\". Wouldn't\\n> it be better to call it \\\"tap_sub_A2\\\"?\\n\\nModified\\n\\n> ~~~\\n>\\n> 6.\\n>\\n> +###############################################################################\\n> +# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\\n> +# replication setup when the existing nodes (node_A & node_B) has no data and\\n> +# the new node (node_C) some pre-existing data.\\n> +###############################################################################\\n> +$node_C->safe_psql('postgres', \\\"INSERT INTO tab_full VALUES (31);\\\");\\n> +\\n> +$result = $node_A->safe_psql('postgres', \\\"SELECT * FROM tab_full order by 1;\\\");\\n> +is( $result, qq(), 'Check existing data');\\n> +\\n> +$result = $node_B->safe_psql('postgres', \\\"SELECT * FROM tab_full order by 1;\\\");\\n> +is( $result, qq(), 'Check existing data');\\n> +\\n> +$result = $node_C->safe_psql('postgres', \\\"SELECT * FROM tab_full order by 1;\\\");\\n> +is($result, qq(31), 'Check existing data');\\n> +\\n> +create_subscription($node_A, $node_C, $subname_AC, $node_C_connstr,\\n> + 'tap_pub_C', 'copy_data = force, local_only = on');\\n> +create_subscription($node_B, $node_C, $subname_BC, $node_C_connstr,\\n> + 'tap_pub_C', 'copy_data = force, local_only = on');\\n> +\\n>\\n> Because the Node_C does not yet have any subscriptions aren't these\\n> cases where you didn't really need to use \\\"force\\\"?\\n\\nModified\\n\\nThe v11 patch attached at [1] has the fixes for the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm2hRyXc4DKPLnBgem_LHBmy%3DWXFEsm9-xhckye1YXcpPA%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Thu, 28 Apr 2022 22:37:22 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for the v11* patches.\\n\\n======\\n\\nv11-0001 - no more comments. LGTM\\n\\n======\\n\\nV11-0002\\n\\n1. doc/src/sgml/logical-replication.sgml\\n\\n+ <para>\\n+ Bidirectional replication is useful in creating multi master database\\n+ which helps in performing read/write operations from any of the nodes.\\n+ Setting up bidirectional logical replication between two nodes requires\\n+ creation of the publication in all the nodes, creating subscriptions in\\n+ each of the nodes that subscribes to data from all the nodes. The steps\\n+ to create a two-node bidirectional replication are given below:\\n+ </para>\\n\\nWording: \\\"creating multi master database\\\" -> \\\"creating a multi-master database\\\"\\nWording: \\\"creation of the publication in all the nodes\\\" -> \\\"creation\\nof a publication in all the nodes\\\".\\n\\n~~~\\n\\n2. doc/src/sgml/logical-replication.sgml\\n\\n> +<programlisting>\\n> +node2=# CREATE SUBSCRIPTION sub_node1_node2\\n> +node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\\n> +node2=# PUBLICATION pub_node1\\n> +node2=# WITH (copy_data = off, local_only = on);\\n> +CREATE SUBSCRIPTION\\n> +</programlisting>\\n> + </para>\\n>\\n> IIUC the closing </para> should be on the same line as the\\n> </programlisting>. I recall there was some recent github push about\\n> this sometime in the last month - maybe you can check to confirm it.\\n\\nVignesh: I have seen the programlisting across multiple places and noticed\\nthat there is no such guideline being followed. I have not made any\\nchange for this comment.\\n\\nFYI – I found the push [1] by PeterE that I was referring to so. I\\nthought we perhaps should follow this, even if not all other code is\\ndoing so. But you can choose.\\n\\n~~~\\n\\n3. doc/src/sgml/logical-replication.sgml\\n\\n+ <para>\\n+ Create a subscription in <literal>node3</literal> to subscribe the changes\\n+ from <literal>node1</literal> and <literal>node2</literal>, here\\n+ <literal>copy_data</literal> is specified as <literal>force</literal> when\\n+ creating a subscription to <literal>node1</literal> so that the existing\\n+ table data is copied during initial sync:\\n+<programlisting>\\n+node3=# CREATE SUBSCRIPTION\\n+node3-# sub_node3_node1 CONNECTION 'dbname=foo host=node1 user=repuser'\\n+node3-# PUBLICATION pub_node1\\n+node3-# WITH (copy_data = force, local_only = on);\\n+CREATE SUBSCRIPTION\\n+node3=# CREATE SUBSCRIPTION\\n+node3-# sub_node3_node2 CONNECTION 'dbname=foo host=node2 user=repuser'\\n+node3-# PUBLICATION pub_node2\\n+node3-# WITH (copy_data = off, local_only = on);\\n+CREATE SUBSCRIPTION\\n+</programlisting>\\n\\nI think this part should be split into 2 separate program listings\\neach for the different subscriptions. Then those descriptions can be\\ndescribed separately (e.g. why one is force and the other is not).\\nThen it will also be more consistent with how you split/explained\\nsomething similar in the previous section on this page.\\n\\n------\\n[1] https://github.com/postgres/postgres/commit/d7ab2a9a3c0a2800ab36bb48d1cc97370067777e\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 29 Apr 2022 12:37:46 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 29, 2022 at 8:08 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for the v11* patches.\\n>\\n> ======\\n>\\n> v11-0001 - no more comments. LGTM\\n>\\n> ======\\n>\\n> V11-0002\\n>\\n> 1. doc/src/sgml/logical-replication.sgml\\n>\\n> + <para>\\n> + Bidirectional replication is useful in creating multi master database\\n> + which helps in performing read/write operations from any of the nodes.\\n> + Setting up bidirectional logical replication between two nodes requires\\n> + creation of the publication in all the nodes, creating subscriptions in\\n> + each of the nodes that subscribes to data from all the nodes. The steps\\n> + to create a two-node bidirectional replication are given below:\\n> + </para>\\n>\\n> Wording: \\\"creating multi master database\\\" -> \\\"creating a multi-master database\\\"\\n> Wording: \\\"creation of the publication in all the nodes\\\" -> \\\"creation\\n> of a publication in all the nodes\\\".\\n\\nModified\\n\\n> ~~~\\n>\\n> 2. doc/src/sgml/logical-replication.sgml\\n>\\n> > +<programlisting>\\n> > +node2=# CREATE SUBSCRIPTION sub_node1_node2\\n> > +node2=# CONNECTION 'dbname=foo host=node1 user=repuser'\\n> > +node2=# PUBLICATION pub_node1\\n> > +node2=# WITH (copy_data = off, local_only = on);\\n> > +CREATE SUBSCRIPTION\\n> > +</programlisting>\\n> > + </para>\\n> >\\n> > IIUC the closing </para> should be on the same line as the\\n> > </programlisting>. I recall there was some recent github push about\\n> > this sometime in the last month - maybe you can check to confirm it.\\n>\\n> Vignesh: I have seen the programlisting across multiple places and noticed\\n> that there is no such guideline being followed. I have not made any\\n> change for this comment.\\n>\\n> FYI – I found the push [1] by PeterE that I was referring to so. I\\n> thought we perhaps should follow this, even if not all other code is\\n> doing so. But you can choose.\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. doc/src/sgml/logical-replication.sgml\\n>\\n> + <para>\\n> + Create a subscription in <literal>node3</literal> to subscribe the changes\\n> + from <literal>node1</literal> and <literal>node2</literal>, here\\n> + <literal>copy_data</literal> is specified as <literal>force</literal> when\\n> + creating a subscription to <literal>node1</literal> so that the existing\\n> + table data is copied during initial sync:\\n> +<programlisting>\\n> +node3=# CREATE SUBSCRIPTION\\n> +node3-# sub_node3_node1 CONNECTION 'dbname=foo host=node1 user=repuser'\\n> +node3-# PUBLICATION pub_node1\\n> +node3-# WITH (copy_data = force, local_only = on);\\n> +CREATE SUBSCRIPTION\\n> +node3=# CREATE SUBSCRIPTION\\n> +node3-# sub_node3_node2 CONNECTION 'dbname=foo host=node2 user=repuser'\\n> +node3-# PUBLICATION pub_node2\\n> +node3-# WITH (copy_data = off, local_only = on);\\n> +CREATE SUBSCRIPTION\\n> +</programlisting>\\n>\\n> I think this part should be split into 2 separate program listings\\n> each for the different subscriptions. Then those descriptions can be\\n> described separately (e.g. why one is force and the other is not).\\n> Then it will also be more consistent with how you split/explained\\n> something similar in the previous section on this page.\\n\\nModified\\nThe attached v12 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Fri, 29 Apr 2022 16:31:41 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Thanks for updating the patches for all my prior feedback.\\n\\nFor v12* I have only minor feedback for the docs.\\n\\n======\\nV12-0001\\n\\nno comments\\n\\n======\\nV12-0002\\n\\n1. Commit message\\n\\nIn another thread using the terminology \\\"multi-master\\\" seems to be\\ncausing some problems. Maybe you should choose different wording in\\nthis commit message to avoid having the same problems?\\n\\n~~~\\n\\n2. doc/src/sgml/logical-replication.sgml\\n\\nThese are all similar minor wording suggestions to split the sentences.\\nWording: \\\", here copy_data is specified as XXX ...\\\" -> \\\". Use\\ncopy_data specified as XXX ...\\\"\\n\\nAlso:\\nWording: \\\"... to subscribe the changes from nodeXXX...\\\"\\n-> \\\"... to subscribe to the changes from nodeXXX... \\\" (add \\\"to\\\")\\n-> \\\"... to subscribe to nodeXXX.\\\" (or I preferred just remove the\\nwhole \\\"changes\\\" part)\\n\\n2a.\\nCreate a subscription in node3 to subscribe the changes from node1,\\nhere copy_data is specified as force so that the existing table data\\nis copied during initial sync:\\n\\nSUGGESTION\\nCreate a subscription in node3 to subscribe to node1. Use copy_data\\nspecified as force so that the existing table data is copied during\\ninitial sync:\\n\\n2b.\\nCreate a subscription in node1 to subscribe the changes fromnode3,\\nhere copy_data is specified as force so that the existing table data\\nis copied during initial sync:\\n\\nSUGGESTION\\nCreate a subscription in node1 to subscribe to node3. Use copy_data\\nspecified as force so that the existing table data is copied during\\ninitial sync:\\n\\n2c.\\nCreate a subscription in node2 to subscribe the changes from node3,\\nhere copy_data is specified as force so that the existing table data\\nis copied during initial sync:\\n\\nSUGGESTION\\nCreate a subscription in node2 to subscribe to node3. Use copy_data\\nspecified as force so that the existing table data is copied during\\ninitial sync:\\n\\n2d.\\nCreate a subscription in node3 to subscribe the changes from node1,\\nhere copy_data is specified as force when creating a subscription to\\nnode1 so that the existing table data is copied during initial sync:\\n\\nSUGGESTION\\nCreate a subscription in node3 to subscribe to node1. Use copy_data\\nspecified as force when creating a subscription to node1 so that the\\nexisting table data is copied during initial sync:\\n\\n2e.\\nCreate a subscription in node3 to subscribe the changes from node2,\\nhere copy_data is specified as off as the initial table data would\\nhave been copied in the earlier step:\\n\\nSUGGESTION (this one has a bit more re-wording than the others)\\nCreate a subscription in node3 to subscribe to node2. Use copy_data\\nspecified as off because the initial table data would have been\\nalready copied in the previous step:\\n\\n~~~\\n\\n3. doc/src/sgml/logical-replication.sgml\\n\\n31.11.2. Adding new node when there is no data in any of the nodes\\n31.11.3. Adding new node when data is present in the existing nodes\\n31.11.4. Adding new node when data is present in the new node\\n\\nMinor change to the above heading?\\n\\nWording: \\\"Adding new node ...\\\" -> \\\"Adding a new node ...\\\"\\n\\n------\\n[1] https://www.postgresql.org/message-id/flat/CAHut%2BPuwRAoWY9pz%3DEubps3ooQCOBFiYPU9Yi%3DVB-U%2ByORU7OA%40mail.gmail.com\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 2 May 2022 10:18:59 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, May 2, 2022 at 5:49 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Thanks for updating the patches for all my prior feedback.\\n>\\n> For v12* I have only minor feedback for the docs.\\n>\\n> ======\\n> V12-0001\\n>\\n> no comments\\n>\\n> ======\\n> V12-0002\\n>\\n> 1. Commit message\\n>\\n> In another thread using the terminology \\\"multi-master\\\" seems to be\\n> causing some problems. Maybe you should choose different wording in\\n> this commit message to avoid having the same problems?\\n\\nI felt, we can leave this to committer\\n\\n> ~~~\\n>\\n> 2. doc/src/sgml/logical-replication.sgml\\n>\\n> These are all similar minor wording suggestions to split the sentences.\\n> Wording: \\\", here copy_data is specified as XXX ...\\\" -> \\\". Use\\n> copy_data specified as XXX ...\\\"\\n>\\n> Also:\\n> Wording: \\\"... to subscribe the changes from nodeXXX...\\\"\\n> -> \\\"... to subscribe to the changes from nodeXXX... \\\" (add \\\"to\\\")\\n> -> \\\"... to subscribe to nodeXXX.\\\" (or I preferred just remove the\\n> whole \\\"changes\\\" part)\\n>\\n> 2a.\\n> Create a subscription in node3 to subscribe the changes from node1,\\n> here copy_data is specified as force so that the existing table data\\n> is copied during initial sync:\\n>\\n> SUGGESTION\\n> Create a subscription in node3 to subscribe to node1. Use copy_data\\n> specified as force so that the existing table data is copied during\\n> initial sync:\\n\\nModified\\n\\n> 2b.\\n> Create a subscription in node1 to subscribe the changes fromnode3,\\n> here copy_data is specified as force so that the existing table data\\n> is copied during initial sync:\\n>\\n> SUGGESTION\\n> Create a subscription in node1 to subscribe to node3. Use copy_data\\n> specified as force so that the existing table data is copied during\\n> initial sync:\\n\\nModified\\n\\n> 2c.\\n> Create a subscription in node2 to subscribe the changes from node3,\\n> here copy_data is specified as force so that the existing table data\\n> is copied during initial sync:\\n>\\n> SUGGESTION\\n> Create a subscription in node2 to subscribe to node3. Use copy_data\\n> specified as force so that the existing table data is copied during\\n> initial sync:\\n\\nModified\\n\\n> 2d.\\n> Create a subscription in node3 to subscribe the changes from node1,\\n> here copy_data is specified as force when creating a subscription to\\n> node1 so that the existing table data is copied during initial sync:\\n>\\n> SUGGESTION\\n> Create a subscription in node3 to subscribe to node1. Use copy_data\\n> specified as force when creating a subscription to node1 so that the\\n> existing table data is copied during initial sync:\\n\\nModified\\n\\n> 2e.\\n> Create a subscription in node3 to subscribe the changes from node2,\\n> here copy_data is specified as off as the initial table data would\\n> have been copied in the earlier step:\\n>\\n> SUGGESTION (this one has a bit more re-wording than the others)\\n> Create a subscription in node3 to subscribe to node2. Use copy_data\\n> specified as off because the initial table data would have been\\n> already copied in the previous step:\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. doc/src/sgml/logical-replication.sgml\\n>\\n> 31.11.2. Adding new node when there is no data in any of the nodes\\n> 31.11.3. Adding new node when data is present in the existing nodes\\n> 31.11.4. Adding new node when data is present in the new node\\n>\\n> Minor change to the above heading?\\n>\\n> Wording: \\\"Adding new node ...\\\" -> \\\"Adding a new node ...\\\"\\n\\nModified\\n\\nThanks for the comments, the attached v13 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 4 May 2022 12:17:00 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Thanks for making changes from all my past reviews.\\n\\n1. I checked the cfbot result is OK\\n\\n2. Both patches apply cleanly in my VM\\n\\n3. Patch changes\\nV13-0001 - No diffs; it is identical to v12-0001\\nV13-0002 - only changed for pg docs. So I rebuilt v13 docs and checked\\nthe HTML rendering. Looked OK.\\n\\nSo I have no more review comments. LGTM.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 5 May 2022 17:22:47 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On 2022-May-04, vignesh C wrote:\\n\\n> On Mon, May 2, 2022 at 5:49 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n\\n> > 1. Commit message\\n> >\\n> > In another thread using the terminology \\\"multi-master\\\" seems to be\\n> > causing some problems. Maybe you should choose different wording in\\n> > this commit message to avoid having the same problems?\\n> \\n> I felt, we can leave this to committer\\n\\nPlease don't. See\\nhttps://postgr.es/m/20200615182235.x7lch5n6kcjq4aue@alap3.anarazel.de\\n\\n-- \\nÁlvaro Herrera Breisgau, Deutschland — https://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 5 May 2022 11:11:05 +0200\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, May 5, 2022 at 2:41 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> On 2022-May-04, vignesh C wrote:\\n>\\n> > On Mon, May 2, 2022 at 5:49 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> > > 1. Commit message\\n> > >\\n> > > In another thread using the terminology \\\"multi-master\\\" seems to be\\n> > > causing some problems. Maybe you should choose different wording in\\n> > > this commit message to avoid having the same problems?\\n> >\\n> > I felt, we can leave this to committer\\n>\\n> Please don't. See\\n> https://postgr.es/m/20200615182235.x7lch5n6kcjq4aue@alap3.anarazel.de\\n\\nIn this thread changing multimaster to multi-source did not get agreed\\nand the changes did not get committed. The code still uses multimaster\\nas in [1]. I think we should keep it as \\\"multimaster\\\" as that is\\nalready being used in [1].\\n\\n[1] - https://www.postgresql.org/docs/current/different-replication-solutions.html\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Thu, 5 May 2022 16:44:14 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, May 4, 2022 at 12:17 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> Thanks for the comments, the attached v13 patch has the changes for the same.\\n>\\n\\nFew comments on v13-0001\\n======================\\n1.\\n+ *\\n+ * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\\n+ * PG16.\\n...\\n@@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\\nOutputPluginOptions *opt,\\n else\\n ctx->twophase_opt_given = true;\\n\\n+ if (data->local_only && data->protocol_version <\\nLOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\\n+ ereport(ERROR,\\n+ (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n+ errmsg(\\\"requested proto_version=%d does not support local_only, need\\n%d or higher\\\",\\n+ data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\\n\\nWhat is the need to change the protocol version for this parameter? As\\nper my understanding, we only need a different protocol version when\\nwe are sending some new message or some additional information in an\\nexisting message as we have done for the streaming/two_phase options\\nwhich doesn't seem to be the case here.\\n\\n2.\\n@@ -29,6 +29,7 @@ typedef struct PGOutputData\\n bool streaming;\\n bool messages;\\n bool two_phase;\\n+ bool local_only;\\n } PGOutputData;\\n\\nIt seems we already have a similar option named 'only_local' in\\nTestDecodingData which has the exactly same functionality. So, isn't\\nit better to name this as 'only_local' instead of 'local_only'? Also,\\nlet's add a test case for 'only_local' option of test_decoding.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 18 May 2022 10:29:30 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, May 4, 2022 2:47 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> Thanks for the comments, the attached v13 patch has the changes for the same.\\r\\n> \\r\\n\\r\\nThanks for your patch. Here are some comments on v13-0001 patch.\\r\\n\\r\\n1) \\r\\n@@ -152,6 +152,18 @@ CREATE SUBSCRIPTION <replaceable class=\\\"parameter\\\">subscription_name</replaceabl\\r\\n </listitem>\\r\\n </varlistentry>\\r\\n \\r\\n+ <varlistentry>\\r\\n+ <term><literal>local_only</literal> (<type>boolean</type>)</term>\\r\\n+ <listitem>\\r\\n+ <para>\\r\\n+ Specifies whether the subscription will request the publisher to send\\r\\n+ locally originated changes at the publisher node, or send any\\r\\n+ publisher node changes regardless of their origin. The default is\\r\\n+ <literal>false</literal>.\\r\\n+ </para>\\r\\n+ </listitem>\\r\\n+ </varlistentry>\\r\\n+\\r\\n <varlistentry>\\r\\n <term><literal>slot_name</literal> (<type>string</type>)</term>\\r\\n <listitem>\\r\\n\\r\\nI think this change should be put after \\\"The following parameters control the\\r\\nsubscription's replication behavior after it has been created\\\", thoughts?\\r\\n\\r\\n2)\\r\\nA new column \\\"sublocalonly\\\" is added to pg_subscription, so maybe we need add it\\r\\nto pg_subscription document, too. (in doc/src/sgml/catalogs.sgml)\\r\\n\\r\\n3)\\r\\n/*\\r\\n * Currently we always forward.\\r\\n */\\r\\nstatic bool\\r\\npgoutput_origin_filter(LogicalDecodingContext *ctx,\\r\\n\\t\\t\\t\\t\\t RepOriginId origin_id)\\r\\n\\r\\nShould we modify the comment of pgoutput_origin_filter()? It doesn't match the\\r\\ncode.\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Wed, 18 May 2022 08:10:42 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, May 4, 2022 at 12:17 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached v13 patch has the changes for the same.\\n> >\\n>\\n> Few comments on v13-0001\\n> ======================\\n>\\n\\nFew comments on v13-0002\\n===========================\\n1.\\nThe steps\\n+ to create a two-node bidirectional replication are given below:\\n+ </para>\\n\\nThe steps given after this will be successful only when there is no\\ndata in any of the nodes and that is not clear by reading docs.\\n\\n2.\\n+ <sect2 id=\\\"add-new-node-data-in-existing-node\\\">\\n+ <title>Adding a new node when data is present in the existing nodes</title>\\n+ <para>\\n+ Adding a new node <literal>node3</literal> to the existing\\n+ <literal>node1</literal> and <literal>node2</literal> when data is present\\n+ in existing nodes <literal>node1</literal> and <literal>node2</literal>\\n+ needs similar steps. The only change required here is that\\n+ <literal>node3</literal> should create a subscription with\\n+ <literal>copy_data = force</literal> to one of the existing nodes to\\n+ receive the existing data during initial data synchronization.\\n+ </para>\\n\\nI think the steps for these require the user to lock the required\\ntables/database (or in some other way hold the operations on required\\ntables) for node-2 till the time setup is complete, otherwise, node-3\\nmight miss some data. It seems the same is also missing in the\\nsection: \\\"Adding a new node when data is present in the new node\\\".\\n\\n3.\\n+\\n+ <sect2 id=\\\"add-node-data-present-in-new-node\\\">\\n+ <title>Adding a new node when data is present in the new node</title>\\n...\\n...\\n\\n+ <para>\\n+ Create a subscription in <literal>node1</literal> to subscribe to\\n+ <literal>node3</literal>. Use <literal>copy_data</literal> specified as\\n+ <literal>force</literal> so that the existing table data is copied during\\n+ initial sync:\\n+<programlisting>\\n+node1=# CREATE SUBSCRIPTION sub_node1_node3\\n+node1-# CONNECTION 'dbname=foo host=node3 user=repuser'\\n+node1-# PUBLICATION pub_node3\\n+node1-# WITH (copy_data = force, local_only = on);\\n+CREATE SUBSCRIPTION\\n+</programlisting></para>\\n+\\n+ <para>\\n+ Create a subscription in <literal>node2</literal> to subscribe to\\n+ <literal>node3</literal>. Use <literal>copy_data</literal> specified as\\n+ <literal>force</literal> so that the existing table data is copied during\\n+ initial sync:\\n+<programlisting>\\n+node2=# CREATE SUBSCRIPTION sub_node2_node3\\n+node2-# CONNECTION 'dbname=foo host=node3 user=repuser'\\n+node2-# PUBLICATION pub_node3\\n+node2-# WITH (copy_data = force, local_only = on);\\n+CREATE SUBSCRIPTION\\n+</programlisting></para>\\n\\nWhy do we need to use \\\"copy_data = force\\\" here? AFAIU, unless, we\\ncreate any subscription on node-3, we don't need the 'force' option.\\n\\n4. We should have a generic section to explain how users can add a new\\nnode using the new options to the existing set of nodes in all cases.\\nFor example, say when the existing set of nodes has some data and the\\nnew node also has some pre-existing data. I think the basic steps are\\nsomething like: a. create a required publication(s) on the new node.\\n(b) create subscriptions on existing nodes pointing to publication on\\nthe new node with the local_only option as true and copy_data = on.\\n(c) wait for data to be copied from the new node to existing nodes.\\n(d) Truncate the data on the new node. (e) create subscriptions\\ncorresponding to each of the existing publisher nodes on the new node\\nwith local_only as true and copy_data = force for one of the nodes.\\n(f) One needs to ensure that there is no new activity on required\\ntables/database (either by locking or in some other way) in all the\\nnodes for which copy_data option is kept as false while creating\\nsubscriptions in the previous step to avoid any data loss.\\n\\nWe also need to mention in Notes that as all operations are not\\ntransactional, user is advised to take backup of existing data to\\navoid any inconsistency.\\n\\n5.\\n * It is quite possible that subscriber has not yet pulled data to\\n+ * the tables, but in ideal cases the table data will be subscribed.\\n+ * To keep the code simple it is not checked if the subscriber table\\n+ * has pulled the data or not.\\n+ */\\n+ if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\\n\\nSorry, but I don't understand what you intend to say by the above\\ncomment. Can you please explain?\\n\\n6. I feel we can explain the case with only two nodes in the commit\\nmessage, why do we need to use a three-node case?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 18 May 2022 16:22:22 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, May 4, 2022 2:47 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> Thanks for the comments, the attached v13 patch has the changes for the same.\\r\\n>\\r\\n\\r\\nHere are some comments on v13-0002 patch.\\r\\n\\r\\n1)\\r\\n+\\t\\t * Throw an error so that the user can take care of the initial data\\r\\n+\\t\\t * copying and then create subscription with copy_data off.\\r\\n\\r\\nShould \\\"with copy_data off\\\" be changed to \\\"with copy_data off or force\\\"?\\r\\n\\r\\n2)\\r\\n\\t\\tcase ALTER_SUBSCRIPTION_ADD_PUBLICATION:\\r\\n\\t\\tcase ALTER_SUBSCRIPTION_DROP_PUBLICATION:\\r\\n...\\r\\n\\t\\t\\t\\t\\t/*\\r\\n\\t\\t\\t\\t\\t * See ALTER_SUBSCRIPTION_REFRESH for details why this is\\r\\n\\t\\t\\t\\t\\t * not allowed.\\r\\n\\t\\t\\t\\t\\t */\\r\\n\\t\\t\\t\\t\\tif (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\r\\n\\r\\nI think we need some changes here, too. Should it be modified to:\\r\\nif (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Thu, 19 May 2022 08:41:54 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, May 4, 2022 at 12:17 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached v13 patch has the changes for the same.\\n> >\\n>\\n> Few comments on v13-0001\\n> ======================\\n> 1.\\n> + *\\n> + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\\n> + * PG16.\\n> ...\\n> @@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\\n> OutputPluginOptions *opt,\\n> else\\n> ctx->twophase_opt_given = true;\\n>\\n> + if (data->local_only && data->protocol_version <\\n> LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\\n> + ereport(ERROR,\\n> + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n> + errmsg(\\\"requested proto_version=%d does not support local_only, need\\n> %d or higher\\\",\\n> + data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\\n>\\n> What is the need to change the protocol version for this parameter? As\\n> per my understanding, we only need a different protocol version when\\n> we are sending some new message or some additional information in an\\n> existing message as we have done for the streaming/two_phase options\\n> which doesn't seem to be the case here.\\n\\nModified\\n\\n> 2.\\n> @@ -29,6 +29,7 @@ typedef struct PGOutputData\\n> bool streaming;\\n> bool messages;\\n> bool two_phase;\\n> + bool local_only;\\n> } PGOutputData;\\n>\\n> It seems we already have a similar option named 'only_local' in\\n> TestDecodingData which has the exactly same functionality. So, isn't\\n> it better to name this as 'only_local' instead of 'local_only'? Also,\\n> let's add a test case for 'only_local' option of test_decoding.\\n\\nModified and added test for only_local option of test_decoding.\\n\\nThanks for the comments, the attached v14 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Fri, 20 May 2022 15:07:47 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, May 18, 2022 at 1:40 PM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Wed, May 4, 2022 2:47 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached v13 patch has the changes for the same.\\n> >\\n>\\n> Thanks for your patch. Here are some comments on v13-0001 patch.\\n>\\n> 1)\\n> @@ -152,6 +152,18 @@ CREATE SUBSCRIPTION <replaceable class=\\\"parameter\\\">subscription_name</replaceabl\\n> </listitem>\\n> </varlistentry>\\n>\\n> + <varlistentry>\\n> + <term><literal>local_only</literal> (<type>boolean</type>)</term>\\n> + <listitem>\\n> + <para>\\n> + Specifies whether the subscription will request the publisher to send\\n> + locally originated changes at the publisher node, or send any\\n> + publisher node changes regardless of their origin. The default is\\n> + <literal>false</literal>.\\n> + </para>\\n> + </listitem>\\n> + </varlistentry>\\n> +\\n> <varlistentry>\\n> <term><literal>slot_name</literal> (<type>string</type>)</term>\\n> <listitem>\\n>\\n> I think this change should be put after \\\"The following parameters control the\\n> subscription's replication behavior after it has been created\\\", thoughts?\\n\\nModified\\n\\n> 2)\\n> A new column \\\"sublocalonly\\\" is added to pg_subscription, so maybe we need add it\\n> to pg_subscription document, too. (in doc/src/sgml/catalogs.sgml)\\n\\nModified\\n\\n> 3)\\n> /*\\n> * Currently we always forward.\\n> */\\n> static bool\\n> pgoutput_origin_filter(LogicalDecodingContext *ctx,\\n> RepOriginId origin_id)\\n>\\n> Should we modify the comment of pgoutput_origin_filter()? It doesn't match the\\n> code.\\n\\nModified\\n\\nThanks for the comments, the v14 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm0xuYy35vOudVHBjov3fQ%3DjBRHJHKUUN9VarqO%3DYqtaxg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 20 May 2022 15:15:42 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, May 18, 2022 at 4:22 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Wed, May 4, 2022 at 12:17 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > Thanks for the comments, the attached v13 patch has the changes for the same.\\n> > >\\n> >\\n> > Few comments on v13-0001\\n> > ======================\\n> >\\n>\\n> Few comments on v13-0002\\n> ===========================\\n> 1.\\n> The steps\\n> + to create a two-node bidirectional replication are given below:\\n> + </para>\\n>\\n> The steps given after this will be successful only when there is no\\n> data in any of the nodes and that is not clear by reading docs.\\n\\nModified\\n\\n> 2.\\n> + <sect2 id=\\\"add-new-node-data-in-existing-node\\\">\\n> + <title>Adding a new node when data is present in the existing nodes</title>\\n> + <para>\\n> + Adding a new node <literal>node3</literal> to the existing\\n> + <literal>node1</literal> and <literal>node2</literal> when data is present\\n> + in existing nodes <literal>node1</literal> and <literal>node2</literal>\\n> + needs similar steps. The only change required here is that\\n> + <literal>node3</literal> should create a subscription with\\n> + <literal>copy_data = force</literal> to one of the existing nodes to\\n> + receive the existing data during initial data synchronization.\\n> + </para>\\n>\\n> I think the steps for these require the user to lock the required\\n> tables/database (or in some other way hold the operations on required\\n> tables) for node-2 till the time setup is complete, otherwise, node-3\\n> might miss some data. It seems the same is also missing in the\\n> section: \\\"Adding a new node when data is present in the new node\\\".\\n\\nModified. Mentioned that lock is required on new node-3 and node-2. I\\nhave mentioned node-3 also should be locked so that no operation is\\nhappened in node-3, else there can be some dml operation after\\ntruncate which is sent to node-1 and the same data is synced when\\ncreate subscription in node-3 subscribing to the publisher in node-1.\\n\\n> 3.\\n> +\\n> + <sect2 id=\\\"add-node-data-present-in-new-node\\\">\\n> + <title>Adding a new node when data is present in the new node</title>\\n> ...\\n> ...\\n>\\n> + <para>\\n> + Create a subscription in <literal>node1</literal> to subscribe to\\n> + <literal>node3</literal>. Use <literal>copy_data</literal> specified as\\n> + <literal>force</literal> so that the existing table data is copied during\\n> + initial sync:\\n> +<programlisting>\\n> +node1=# CREATE SUBSCRIPTION sub_node1_node3\\n> +node1-# CONNECTION 'dbname=foo host=node3 user=repuser'\\n> +node1-# PUBLICATION pub_node3\\n> +node1-# WITH (copy_data = force, local_only = on);\\n> +CREATE SUBSCRIPTION\\n> +</programlisting></para>\\n> +\\n> + <para>\\n> + Create a subscription in <literal>node2</literal> to subscribe to\\n> + <literal>node3</literal>. Use <literal>copy_data</literal> specified as\\n> + <literal>force</literal> so that the existing table data is copied during\\n> + initial sync:\\n> +<programlisting>\\n> +node2=# CREATE SUBSCRIPTION sub_node2_node3\\n> +node2-# CONNECTION 'dbname=foo host=node3 user=repuser'\\n> +node2-# PUBLICATION pub_node3\\n> +node2-# WITH (copy_data = force, local_only = on);\\n> +CREATE SUBSCRIPTION\\n> +</programlisting></para>\\n>\\n> Why do we need to use \\\"copy_data = force\\\" here? AFAIU, unless, we\\n> create any subscription on node-3, we don't need the 'force' option.\\n\\nModified to on.\\n\\n> 4. We should have a generic section to explain how users can add a new\\n> node using the new options to the existing set of nodes in all cases.\\n> For example, say when the existing set of nodes has some data and the\\n> new node also has some pre-existing data. I think the basic steps are\\n> something like: a. create a required publication(s) on the new node.\\n> (b) create subscriptions on existing nodes pointing to publication on\\n> the new node with the local_only option as true and copy_data = on.\\n> (c) wait for data to be copied from the new node to existing nodes.\\n> (d) Truncate the data on the new node. (e) create subscriptions\\n> corresponding to each of the existing publisher nodes on the new node\\n> with local_only as true and copy_data = force for one of the nodes.\\n> (f) One needs to ensure that there is no new activity on required\\n> tables/database (either by locking or in some other way) in all the\\n> nodes for which copy_data option is kept as false while creating\\n> subscriptions in the previous step to avoid any data loss.\\n\\nAdded\\n\\n> We also need to mention in Notes that as all operations are not\\n> transactional, user is advised to take backup of existing data to\\n> avoid any inconsistency.\\n\\nModified\\n\\n> 5.\\n> * It is quite possible that subscriber has not yet pulled data to\\n> + * the tables, but in ideal cases the table data will be subscribed.\\n> + * To keep the code simple it is not checked if the subscriber table\\n> + * has pulled the data or not.\\n> + */\\n> + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\\n>\\n> Sorry, but I don't understand what you intend to say by the above\\n> comment. Can you please explain?\\n\\nWhen the user specifies copy_data as on, we should check if the\\npublisher has the publication tables being subscribed from a remote\\npublisher. If so throw an error as it remote origin data present.\\nEx:\\nNode1 - pub1 for table t1 -- no data\\nNode2 - Sub1 subscribing to data from pub1\\nNode2 - pub2 for table t1 -- no data\\nNode3 - create subscription to Node2 with copy_data = ON\\n\\nIn this case even though the table does not have any remote origin\\ndata, as Node2 is subscribing to data from Node1, throw an error.\\nWe throw an error irrespective of data present in Node1 or not to keep\\nthe code simple.\\n\\n> 6. I feel we can explain the case with only two nodes in the commit\\n> message, why do we need to use a three-node case?\\n\\nModified\\n\\nThanks for the comments, the v14 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm0xuYy35vOudVHBjov3fQ%3DjBRHJHKUUN9VarqO%3DYqtaxg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 20 May 2022 15:30:54 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, May 19, 2022 at 2:12 PM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Wed, May 4, 2022 2:47 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached v13 patch has the changes for the same.\\n> >\\n>\\n> Here are some comments on v13-0002 patch.\\n>\\n> 1)\\n> + * Throw an error so that the user can take care of the initial data\\n> + * copying and then create subscription with copy_data off.\\n>\\n> Should \\\"with copy_data off\\\" be changed to \\\"with copy_data off or force\\\"?\\n\\nModified\\n\\n> 2)\\n> case ALTER_SUBSCRIPTION_ADD_PUBLICATION:\\n> case ALTER_SUBSCRIPTION_DROP_PUBLICATION:\\n> ...\\n> /*\\n> * See ALTER_SUBSCRIPTION_REFRESH for details why this is\\n> * not allowed.\\n> */\\n> if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\n>\\n> I think we need some changes here, too. Should it be modified to:\\n> if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\\n\\nModified\\n\\nThanks for the comments, the v14 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm0xuYy35vOudVHBjov3fQ%3DjBRHJHKUUN9VarqO%3DYqtaxg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 20 May 2022 15:32:11 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, May 20, 2022 at 3:08 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> >\\n> > Few comments on v13-0001\\n> > ======================\\n> > 1.\\n> > + *\\n> > + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\\n> > + * PG16.\\n> > ...\\n> > @@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\\n> > OutputPluginOptions *opt,\\n> > else\\n> > ctx->twophase_opt_given = true;\\n> >\\n> > + if (data->local_only && data->protocol_version <\\n> > LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\\n> > + ereport(ERROR,\\n> > + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n> > + errmsg(\\\"requested proto_version=%d does not support local_only, need\\n> > %d or higher\\\",\\n> > + data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\\n> >\\n> > What is the need to change the protocol version for this parameter? As\\n> > per my understanding, we only need a different protocol version when\\n> > we are sending some new message or some additional information in an\\n> > existing message as we have done for the streaming/two_phase options\\n> > which doesn't seem to be the case here.\\n>\\n> Modified\\n>\\n\\nIt seems you forgot to remove the comments after removing the code\\ncorresponding to the above. See below.\\n+ *\\n+ * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\\n+ * support for sending only locally originated data from the publisher.\\n+ * Introduced in PG16.\\n+ *\\n+ * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\\n+ * PG16.\\n */\\n\\nFew other comments on 0001\\n========================\\n1.\\n+ * Return true if data has originated remotely when only_local option is\\n+ * enabled, false otherwise.\\n\\nCan we slightly change the comment to:\\\"Return true if the data source\\n(origin) is remote and user has requested only local data, false\\notherwise.\\\"\\n\\n2.\\n+SELECT pg_replication_origin_session_reset();\\n+SELECT pg_drop_replication_slot('regression_slot_only_local');\\n+SELECT pg_replication_origin_drop('regress_test_decoding:\\nregression_slot_only_local');\\n\\\\ No newline at end of file\\n\\nAt the end of the file, there should be a new line.\\n\\n3.\\n+ <structfield>subonlylocal</structfield> <type>bool</type>\\n+ </para>\\n+ <para>\\n+ If true, subscription will request the publisher to send locally\\n+ originated changes at the publisher node, or send any publisher node\\n+ changes regardless of their origin\\n\\nThe case where this option is not set is not clear from the\\ndescription. Can we change the description to: \\\"If true, the\\nsubscription will request that the publisher send locally originated\\nchanges. False indicates that the publisher sends any changes\\nregardless of their origin.\\\"\\n\\n4. This new option 'subonlylocal' is placed before 'substream' in docs\\n(catalogs.sgml) and after it in structures in pg_subscription.h. I\\nsuggest adding it after 'subdisableonerr' in docs and\\npg_subscription.h. Also, adjust other places in subscriber-side code\\nto place it consistently.\\n\\n5.\\n@@ -4516,6 +4524,8 @@ getSubscriptions(Archive *fout)\\n pg_strdup(PQgetvalue(res, i, i_subtwophasestate));\\n subinfo[i].subdisableonerr =\\n pg_strdup(PQgetvalue(res, i, i_subdisableonerr));\\n+ subinfo[i].subonlylocal =\\n+ pg_strdup(PQgetvalue(res, i, i_subonlylocal));\\n\\n /* Decide whether we want to dump it */\\n selectDumpableObject(&(subinfo[i].dobj), fout);\\n@@ -4589,6 +4599,9 @@ dumpSubscription(Archive *fout, const\\nSubscriptionInfo *subinfo)\\n if (strcmp(subinfo->subdisableonerr, \\\"t\\\") == 0)\\n appendPQExpBufferStr(query, \\\", disable_on_error = true\\\");\\n\\n+ if (strcmp(subinfo->subonlylocal, \\\"t\\\") == 0)\\n+ appendPQExpBufferStr(query, \\\", only_local = true\\\");\\n+\\n if (strcmp(subinfo->subsynccommit, \\\"off\\\") != 0)\\n appendPQExpBuffer(query, \\\", synchronous_commit = %s\\\",\\nfmtId(subinfo->subsynccommit));\\n\\ndiff --git a/src/bin/pg_dump/pg_dump.h b/src/bin/pg_dump/pg_dump.h\\nindex 1d21c2906f..ddb855fd16 100644\\n--- a/src/bin/pg_dump/pg_dump.h\\n+++ b/src/bin/pg_dump/pg_dump.h\\n@@ -661,6 +661,7 @@ typedef struct _SubscriptionInfo\\n char *subdisableonerr;\\n char *subsynccommit;\\n char *subpublications;\\n+ char *subonlylocal;\\n } SubscriptionInfo;\\n\\nTo keep this part of the code consistent, I think it is better to\\nplace 'subonlylocal' after 'subdisableonerr' in SubscriptionInfo.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 23 May 2022 10:01:12 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, May 23, 2022 at 10:01 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, May 20, 2022 at 3:08 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > >\\n> > >\\n> > > Few comments on v13-0001\\n> > > ======================\\n> > > 1.\\n> > > + *\\n> > > + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\\n> > > + * PG16.\\n> > > ...\\n> > > @@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\\n> > > OutputPluginOptions *opt,\\n> > > else\\n> > > ctx->twophase_opt_given = true;\\n> > >\\n> > > + if (data->local_only && data->protocol_version <\\n> > > LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\\n> > > + ereport(ERROR,\\n> > > + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n> > > + errmsg(\\\"requested proto_version=%d does not support local_only, need\\n> > > %d or higher\\\",\\n> > > + data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\\n> > >\\n> > > What is the need to change the protocol version for this parameter? As\\n> > > per my understanding, we only need a different protocol version when\\n> > > we are sending some new message or some additional information in an\\n> > > existing message as we have done for the streaming/two_phase options\\n> > > which doesn't seem to be the case here.\\n> >\\n> > Modified\\n> >\\n>\\n> It seems you forgot to remove the comments after removing the code\\n> corresponding to the above. See below.\\n> + *\\n> + * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\\n> + * support for sending only locally originated data from the publisher.\\n> + * Introduced in PG16.\\n> + *\\n> + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\\n> + * PG16.\\n> */\\n\\nRemoved it.\\n\\n> Few other comments on 0001\\n> ========================\\n> 1.\\n> + * Return true if data has originated remotely when only_local option is\\n> + * enabled, false otherwise.\\n>\\n> Can we slightly change the comment to:\\\"Return true if the data source\\n> (origin) is remote and user has requested only local data, false\\n> otherwise.\\\"\\n\\nModified\\n\\n> 2.\\n> +SELECT pg_replication_origin_session_reset();\\n> +SELECT pg_drop_replication_slot('regression_slot_only_local');\\n> +SELECT pg_replication_origin_drop('regress_test_decoding:\\n> regression_slot_only_local');\\n> \\\\ No newline at end of file\\n>\\n> At the end of the file, there should be a new line.\\n\\nI checked the other regression sql files and found that a newline is\\nnot added at the end of the file. I did not make any change for this.\\nThoughts?\\n\\n> 3.\\n> + <structfield>subonlylocal</structfield> <type>bool</type>\\n> + </para>\\n> + <para>\\n> + If true, subscription will request the publisher to send locally\\n> + originated changes at the publisher node, or send any publisher node\\n> + changes regardless of their origin\\n>\\n> The case where this option is not set is not clear from the\\n> description. Can we change the description to: \\\"If true, the\\n> subscription will request that the publisher send locally originated\\n> changes. False indicates that the publisher sends any changes\\n> regardless of their origin.\\\"\\n\\nModified\\n\\n> 4. This new option 'subonlylocal' is placed before 'substream' in docs\\n> (catalogs.sgml) and after it in structures in pg_subscription.h. I\\n> suggest adding it after 'subdisableonerr' in docs and\\n> pg_subscription.h. Also, adjust other places in subscriber-side code\\n> to place it consistently.\\n\\nModified\\n\\n> 5.\\n> @@ -4516,6 +4524,8 @@ getSubscriptions(Archive *fout)\\n> pg_strdup(PQgetvalue(res, i, i_subtwophasestate));\\n> subinfo[i].subdisableonerr =\\n> pg_strdup(PQgetvalue(res, i, i_subdisableonerr));\\n> + subinfo[i].subonlylocal =\\n> + pg_strdup(PQgetvalue(res, i, i_subonlylocal));\\n>\\n> /* Decide whether we want to dump it */\\n> selectDumpableObject(&(subinfo[i].dobj), fout);\\n> @@ -4589,6 +4599,9 @@ dumpSubscription(Archive *fout, const\\n> SubscriptionInfo *subinfo)\\n> if (strcmp(subinfo->subdisableonerr, \\\"t\\\") == 0)\\n> appendPQExpBufferStr(query, \\\", disable_on_error = true\\\");\\n>\\n> + if (strcmp(subinfo->subonlylocal, \\\"t\\\") == 0)\\n> + appendPQExpBufferStr(query, \\\", only_local = true\\\");\\n> +\\n> if (strcmp(subinfo->subsynccommit, \\\"off\\\") != 0)\\n> appendPQExpBuffer(query, \\\", synchronous_commit = %s\\\",\\n> fmtId(subinfo->subsynccommit));\\n>\\n> diff --git a/src/bin/pg_dump/pg_dump.h b/src/bin/pg_dump/pg_dump.h\\n> index 1d21c2906f..ddb855fd16 100644\\n> --- a/src/bin/pg_dump/pg_dump.h\\n> +++ b/src/bin/pg_dump/pg_dump.h\\n> @@ -661,6 +661,7 @@ typedef struct _SubscriptionInfo\\n> char *subdisableonerr;\\n> char *subsynccommit;\\n> char *subpublications;\\n> + char *subonlylocal;\\n> } SubscriptionInfo;\\n>\\n> To keep this part of the code consistent, I think it is better to\\n> place 'subonlylocal' after 'subdisableonerr' in SubscriptionInfo.\\n\\nModified\\n\\nThanks for the comments, the attached v15 patch has the fixes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Mon, 23 May 2022 23:04:03 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, May 24, 2022 1:34 AM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> Thanks for the comments, the attached v15 patch has the fixes for the same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch. I have a comment on the document in 0002 patch.\\r\\n\\r\\n@@ -300,6 +310,11 @@ CREATE SUBSCRIPTION <replaceable class=\\\"parameter\\\">subscription_name</replaceabl\\r\\n replication from the publisher. The default is\\r\\n <literal>false</literal>.\\r\\n </para>\\r\\n+ <para>\\r\\n+ There is some interaction between the <literal>only_local</literal>\\r\\n+ option and <literal>copy_data</literal> option. Refer to the\\r\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\r\\n+ </para>\\r\\n </listitem>\\r\\n </varlistentry>\\r\\n\\r\\nThis change is related to \\\"only_local\\\" option and \\\"copy_data\\\" option, so I think\\r\\nit should be put together with \\\"only_local\\\", instead of \\\"disable_on_error\\\".\\r\\n\\r\\nBesides, I tested some cross version cases, pg_dump and describe command, the\\r\\nresults are all as expected.\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Tue, 24 May 2022 02:36:48 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, May 23, 2022 at 10:01 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, May 20, 2022 at 3:08 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Wed, May 18, 2022 at 10:29 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > >\\n> > >\\n> > > Few comments on v13-0001\\n> > > ======================\\n> > > 1.\\n> > > + *\\n> > > + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\\n> > > + * PG16.\\n> > > ...\\n> > > @@ -477,6 +489,12 @@ pgoutput_startup(LogicalDecodingContext *ctx,\\n> > > OutputPluginOptions *opt,\\n> > > else\\n> > > ctx->twophase_opt_given = true;\\n> > >\\n> > > + if (data->local_only && data->protocol_version <\\n> > > LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)\\n> > > + ereport(ERROR,\\n> > > + (errcode(ERRCODE_INVALID_PARAMETER_VALUE),\\n> > > + errmsg(\\\"requested proto_version=%d does not support local_only, need\\n> > > %d or higher\\\",\\n> > > + data->protocol_version, LOGICALREP_PROTO_LOCALONLY_VERSION_NUM)));\\n> > >\\n> > > What is the need to change the protocol version for this parameter? As\\n> > > per my understanding, we only need a different protocol version when\\n> > > we are sending some new message or some additional information in an\\n> > > existing message as we have done for the streaming/two_phase options\\n> > > which doesn't seem to be the case here.\\n> >\\n> > Modified\\n> >\\n>\\n> It seems you forgot to remove the comments after removing the code\\n> corresponding to the above. See below.\\n> + *\\n> + * LOGICALREP_PROTO_LOCALONLY_VERSION_NUM is the minimum protocol version with\\n> + * support for sending only locally originated data from the publisher.\\n> + * Introduced in PG16.\\n> + *\\n> + * FIXME: LOGICALREP_PROTO_LOCALONLY_VERSION_NUM needs to be bumped to 4 in\\n> + * PG16.\\n> */\\n>\\n> Few other comments on 0001\\n> ========================\\n> 1.\\n> + * Return true if data has originated remotely when only_local option is\\n> + * enabled, false otherwise.\\n>\\n> Can we slightly change the comment to:\\\"Return true if the data source\\n> (origin) is remote and user has requested only local data, false\\n> otherwise.\\\"\\n>\\n> 2.\\n> +SELECT pg_replication_origin_session_reset();\\n> +SELECT pg_drop_replication_slot('regression_slot_only_local');\\n> +SELECT pg_replication_origin_drop('regress_test_decoding:\\n> regression_slot_only_local');\\n> \\\\ No newline at end of file\\n>\\n> At the end of the file, there should be a new line.\\n>\\n\\nThanks for pointing this out, for some reason vim does not show these\\nnew lines at the end of the file, VS Code shows the last new line. I\\nhave added it.\\n\\nThe attached v16 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 25 May 2022 17:25:02 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, May 24, 2022 at 8:06 AM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Tue, May 24, 2022 1:34 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached v15 patch has the fixes for the same.\\n> >\\n>\\n> Thanks for updating the patch. I have a comment on the document in 0002 patch.\\n>\\n> @@ -300,6 +310,11 @@ CREATE SUBSCRIPTION <replaceable class=\\\"parameter\\\">subscription_name</replaceabl\\n> replication from the publisher. The default is\\n> <literal>false</literal>.\\n> </para>\\n> + <para>\\n> + There is some interaction between the <literal>only_local</literal>\\n> + option and <literal>copy_data</literal> option. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n> + </para>\\n> </listitem>\\n> </varlistentry>\\n>\\n> This change is related to \\\"only_local\\\" option and \\\"copy_data\\\" option, so I think\\n> it should be put together with \\\"only_local\\\", instead of \\\"disable_on_error\\\".\\n\\nModified.\\n\\nThe v16 patch attached at [1] has the changes for the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm2_Ytder-6C68ia%3Dm39cmknAhxf2KGkeNAtxt84MxMT3w%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 25 May 2022 17:26:32 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for patch v15-0001.\\n\\n======\\n\\n1. Commit message\\n\\nShould this also say new test cases were added for the test_decoding plugin?\\n\\n~~~\\n\\n2. contrib/test_decoding/sql/replorigin.sql\\n\\n@@ -119,3 +119,18 @@ SELECT data FROM\\npg_logical_slot_get_changes('regression_slot_no_lsn', NULL, NUL\\n SELECT pg_replication_origin_session_reset();\\n SELECT pg_drop_replication_slot('regression_slot_no_lsn');\\n SELECT pg_replication_origin_drop('regress_test_decoding:\\nregression_slot_no_lsn');\\n+\\n+-- Verify that remote origin data is not returned with only-local option\\n+SELECT 'init' FROM\\npg_create_logical_replication_slot('regression_slot_only_local',\\n'test_decoding');\\n+SELECT pg_replication_origin_create('regress_test_decoding:\\nregression_slot_only_local');\\n+SELECT pg_replication_origin_session_setup('regress_test_decoding:\\nregression_slot_only_local');\\n+INSERT INTO origin_tbl(data) VALUES ('only_local, commit1');\\n+-- remote origin data returned when only-local option is not set\\n+SELECT data FROM\\npg_logical_slot_get_changes('regression_slot_only_local', NULL, NULL,\\n'skip-empty-xacts', '1', 'include-xids', '0', 'only-local', '0');\\n+INSERT INTO origin_tbl(data) VALUES ('only_local, commit2');\\n+-- remote origin data not returned when only-local option is set\\n+SELECT data FROM\\npg_logical_slot_get_changes('regression_slot_only_local', NULL, NULL,\\n'skip-empty-xacts', '1', 'include-xids', '0', 'only-local', '1');\\n+-- Clean up\\n+SELECT pg_replication_origin_session_reset();\\n+SELECT pg_drop_replication_slot('regression_slot_only_local');\\n+SELECT pg_replication_origin_drop('regress_test_decoding:\\nregression_slot_only_local');\\n\\nAll the new comments should consistently start with upper case.\\n\\n~~~\\n\\n3. doc/src/sgml/catalogs.sgml\\n\\n+ <para>\\n+ If true, the subscription will request that the publisher send locally\\n+ originated changes. False indicates that the publisher sends any changes\\n+ regardless of their origin.\\n+ </para></entry>\\n\\nSUGGESTION\\nIf true, the subscription will request the publisher to only send\\nchanges that originated locally.\\n\\n~~~\\n\\n4. doc/src/sgml/ref/create_subscription.sgml\\n\\n+ <para>\\n+ Specifies whether the subscription will request the publisher to send\\n+ locally originated changes at the publisher node, or send any\\n+ publisher node changes regardless of their origin. The default is\\n+ <literal>false</literal>.\\n+ </para>\\n\\nThis wording should be more similar to the same information in catalogs.sgml\\n\\nSUGGESTION\\nSpecifies whether the subscription will request the publisher to only\\nsend changes that originated locally, or to send any changes\\nregardless of origin.\\n\\n~~~\\n\\n5. src/include/replication/walreceiver.h\\n\\n@@ -183,6 +183,7 @@ typedef struct\\n bool streaming; /* Streaming of large transactions */\\n bool twophase; /* Streaming of two-phase transactions at\\n * prepare time */\\n+ bool only_local; /* publish only locally originated data */\\n } logical;\\n } proto;\\n } WalRcvStreamOptions;\\n\\nSUGGESTION\\n/* Publish only local origin data */\\n\\n~~~\\n\\n6. src/test/subscription/t/032_onlylocal.pl - cosmetic changes\\n\\n6a.\\n+# Setup a bidirectional logical replication between Node_A & Node_B\\n\\nSUGGESTION\\n\\\"... node_A and node_B\\\"\\n\\n6b.\\n+is(1, 1, \\\"Circular replication setup is complete\\\");\\n\\nSUGGESTION (or maybe saying \\\"circular\\\" is also OK - I wasn't sure)\\n\\\"Bidirectional replication setup is complete\\\"\\n\\n6c.\\n+# check that bidirectional logical replication setup...\\n\\nStart comment sentence with upper case.\\n\\n6d.\\n+###############################################################################\\n+# check that remote data that is originated from node_C to node_B is not\\n+# published to node_A\\n+###############################################################################\\n\\nSUGGESTION\\nCheck that remote data of node_B (that originated from node_C) is not\\npublished to node_A\\n\\n6e.\\n+is($result, qq(11\\n+12\\n+13), 'Node_C data replicated to Node_B'\\n+);\\n\\nSUGGESTION for message\\n'node_C data replicated to node_B'\\n\\n6f.\\n+is($result, qq(11\\n+12), 'Remote data originated from other node is not replicated when\\nonly_local option is ON'\\n+);\\n\\nSUGGESTION for message\\n'Remote data originating from another node (not the publisher) is not\\nreplicated when only_local option is ON'\\n\\n6g.\\n\\\"Circular replication setup is complete\\\"\\n'Inserted successfully without leading to infinite recursion in\\nbidirectional replication setup'\\n'Inserted successfully without leading to infinite recursion in\\nbidirectional replication setup'\\n'Node_C data replicated to Node_B'\\n'Remote data originated from other node is not replicated when\\nonly_local option is ON'\\n\\nWhy do some of the \\\"is\\\" messages have single quotes and others have\\ndouble quotes? Should be consistent.\\n\\n~~~\\n\\n7. src/test/subscription/t/032_onlylocal.pl\\n\\n+my $appname_B2 = 'tap_sub_B2';\\n+$node_B->safe_psql(\\n+ 'postgres', \\\"\\n+ CREATE SUBSCRIPTION tap_sub_B2\\n+ CONNECTION '$node_C_connstr application_name=$appname_B2'\\n+ PUBLICATION tap_pub_C\\n+ WITH (only_local = on)\\\");\\n+\\n\\nAFAIK the \\\"WITH (only_local = on)\\\" is unnecessary here. We don't care\\nwhere the node_C data came from for this test case.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 May 2022 11:36:29 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for patch v15-0001.\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> Should this also say new test cases were added for the test_decoding plugin?\\n>\\n> ~~~\\n\\nShall we take this out as a separate test-case patch? I think it is a\\ngood idea to do a missing test case for existing functionality.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 May 2022 09:50:13 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, May 26, 2022 at 2:20 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Here are my review comments for patch v15-0001.\\n> >\\n> > ======\\n> >\\n> > 1. Commit message\\n> >\\n> > Should this also say new test cases were added for the test_decoding plugin?\\n> >\\n> > ~~~\\n>\\n> Shall we take this out as a separate test-case patch? I think it is a\\n> good idea to do a missing test case for existing functionality.\\n>\\n\\n+1. I also think this new test could be posted separately.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 May 2022 14:59:41 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n>\\n> 3. doc/src/sgml/catalogs.sgml\\n>\\n> + <para>\\n> + If true, the subscription will request that the publisher send locally\\n> + originated changes. False indicates that the publisher sends any changes\\n> + regardless of their origin.\\n> + </para></entry>\\n>\\n> SUGGESTION\\n> If true, the subscription will request the publisher to only send\\n> changes that originated locally.\\n>\\n> ~~~\\n>\\n\\nI think it is a good idea to keep the description related to the\\n'False' value as without that it is slightly unclear.\\n\\n> 4. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> + <para>\\n> + Specifies whether the subscription will request the publisher to send\\n> + locally originated changes at the publisher node, or send any\\n> + publisher node changes regardless of their origin. The default is\\n> + <literal>false</literal>.\\n> + </para>\\n>\\n> This wording should be more similar to the same information in catalogs.sgml\\n>\\n> SUGGESTION\\n> Specifies whether the subscription will request the publisher to only\\n> send changes that originated locally, or to send any changes\\n> regardless of origin.\\n>\\n> ~~~\\n>\\n\\nHere, also, let's keep the wording related to the default value.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 May 2022 10:38:42 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, May 26, 2022 at 3:08 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> >\\n> > 3. doc/src/sgml/catalogs.sgml\\n> >\\n> > + <para>\\n> > + If true, the subscription will request that the publisher send locally\\n> > + originated changes. False indicates that the publisher sends any changes\\n> > + regardless of their origin.\\n> > + </para></entry>\\n> >\\n> > SUGGESTION\\n> > If true, the subscription will request the publisher to only send\\n> > changes that originated locally.\\n> >\\n> > ~~~\\n> >\\n>\\n> I think it is a good idea to keep the description related to the\\n> 'False' value as without that it is slightly unclear.\\n>\\n> > 4. doc/src/sgml/ref/create_subscription.sgml\\n> >\\n> > + <para>\\n> > + Specifies whether the subscription will request the publisher to send\\n> > + locally originated changes at the publisher node, or send any\\n> > + publisher node changes regardless of their origin. The default is\\n> > + <literal>false</literal>.\\n> > + </para>\\n> >\\n> > This wording should be more similar to the same information in catalogs.sgml\\n> >\\n> > SUGGESTION\\n> > Specifies whether the subscription will request the publisher to only\\n> > send changes that originated locally, or to send any changes\\n> > regardless of origin.\\n> >\\n> > ~~~\\n> >\\n>\\n> Here, also, let's keep the wording related to the default value.\\n>\\n> --\\n\\nI agree. Probably I was not being clear enough with those review comments.\\n\\nMy suggestions were only meant as substitutes for the FIRST sentences\\nof those patch fragments, not the entire paragraph.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 May 2022 15:38:20 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"This post completes my review of patch v15*.\\n\\n======\\n\\n1. A general comment affecting both patches\\n\\nI realised the things defined in the subscription WITH are referred to\\nin the PG DOCS as \\\"parameters\\\" (not \\\"options\\\"). See CREATE\\nSUBSCRIPTION [1]. This is already reflected in some of my review\\ncomments for the v15-0002 below., but I'd already posted my v15-0001\\nreview comments before noticing this.\\n\\nPlease also go through the v15-0001 patch to make all the necessary\\nterminology changes (including in the v15-0001 commit message).\\n\\n=====\\n\\nPlease find below review comments for patch v15-0002.\\n\\n2. Commit message\\n\\nThis patch does couple of things:\\nchange 1) Added force option for copy_data.\\nchange 2) Check and throw an error if the publication tables were also\\nsubscribing data in the publisher from other publishers.\\n\\n--\\n\\n2a.\\n\\\"couple\\\" -> \\\"a couple\\\"\\n\\\"Added force\\\" -> \\\"Adds force ...\\\"\\n\\\"Check and throw\\\" -> \\\"Checks and throws ...\\\"\\n\\n2b.\\nIsn't \\\"Change 2)\\\" only applicable when copy_data and only_local are\\non? It should say so here.\\n\\n~~~\\n\\n3. Commit message\\n\\nnode1: Table t1 (c1 int) has data 1, 2, 3, 4\\nnode2: Table t1 (c1 int) has data 5, 6, 7, 8\\n\\n--\\n\\nI feel these kinds of replication examples are always much easier to\\nvisualize when the data somehow indicates where it came from. E.g.\\nconsider using different data in your example like:\\n\\nnode1: Table t1 (c1 int) has data 11, 12, 13, 14\\nnode2: Table t1 (c1 int) has data 21, 22, 23, 24\\n\\n~~~\\n\\n4. Commit message\\n\\nExamples in this message exceed 80 chars. Please wrap the SQL appropriately\\n\\n~~~\\n\\n5. Commit message\\n\\nAfter this the data will be something like this:\\nnode1:\\n1, 2, 3, 4, 5, 6, 7, 8\\n\\nnode2:\\n1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\\n\\nSo, you can see that data on node2 (5, 6, 7, 8) is duplicated.\\n\\n--\\n\\nYeah, but is that entirely correct though? AFAIK the scenario\\ndescribed by this example is going to recurse forever.\\n\\n~~~\\n\\n6. Commit message\\n\\nThis problem can be solved by using only_local and copy_data option as\\ngiven below\\n\\n--\\n\\n\\\"option\\\" -> \\\"parameters\\\"\\n\\n~~~\\n\\n7. Commit message\\n\\nThere are other minor review comments but it's too messy to report in\\ndetail here. I suggest checking all the \\\":\\\" (some were missing) and\\nthe letter case of/within the sentences. I also suggest\\ncutting/pasting all this text into MSWord or some grammar checker to\\ncorrect anything else reported.\\n\\n~~~\\n\\n8. Commit message\\n\\nCREATE SUBSCRIPTION sub_node2_node1 CONNECTION '<node1 details>'\\nPUBLICATION pub_node1 WITH (copy_data = force, only_local = on);\\nERROR: CREATE/ALTER SUBSCRIPTION with only_local and copy_data as\\ntrue is not allowed when the publisher might have replicated data\\n\\n--\\n\\nLooks like a typo: e.g. If copy_data is \\\"force\\\" then how is the error happening?\\n\\n~~~\\n\\n9. Commit message\\n\\nThe examples seem confused/muddled in some way IMO.\\n\\nIIUC the new ERROR is like a new integrity check for when the\\nsubscription is using the only_local = on, along with copy_data = on.\\n\\nAnd so then the copy_data = force was added as a way to override that\\nprevious error being reported.\\n\\nSo unless you know about the possible error then the \\\"force\\\" option\\nmakes no sense. So I think you need to change the order of the\\nexplanations in the commit message to describe the new error first.\\n\\n======\\n\\n10. doc/src/sgml/logical-replication.sgml\\n\\n+\\n+ <sect1 id=\\\"bidirectional-logical-replication\\\">\\n+ <title>Bidirectional logical replication</title>\\n+\\n+ <sect2 id=\\\"setting-bidirectional-replication-two-nodes\\\">\\n+ <title>Setting bidirectional replication between two nodes</title>\\n+ <para>\\n+ Bidirectional replication is useful in creating a multi-master database\\n+ which helps in performing read/write operations from any of the nodes.\\n+ Setting up bidirectional logical replication between two nodes requires\\n+ creation of a publication in all the nodes, creating subscriptions in\\n+ each of the nodes that subscribes to data from all the nodes. The steps\\n+ to create a two-node bidirectional replication when there is no data in\\n+ both the nodes are given below:\\n+ </para>\\n\\nThis description is confusing. It seems to be trying to say something\\nabout a multi-master system (e.g. you say \\\"all the nodes\\\") yet this is\\nall written inside the section about \\\"... two nodes\\\", so saying \\\"all\\nthe nodes\\\" makes no sense here.\\n\\nI think you need to split this information and put some kind of blurb\\ntext at the top level (section 31.11), and then this section (for \\\"two\\nnodes\\\") should ONLY be talking about the case when there are just TWO\\nnodes.\\n\\n~~~\\n\\n11. doc/src/sgml/logical-replication.sgml\\n\\n+ <para>\\n+ Lock the required tables in <literal>node1</literal> and\\n+ <literal>node2</literal> till the setup is completed.\\n+ </para>\\n\\n\\\"till\\\" -> \\\"until\\\" (make this same change in multiple places)\\n\\n~~~\\n\\n12. doc/src/sgml/logical-replication.sgml\\n\\n+ <sect2 id=\\\"add-new-node-data-in-existing-node\\\">\\n+ <title>Adding a new node when data is present in the existing nodes</title>\\n+ <para>\\n+ Adding a new node <literal>node3</literal> to the existing\\n+ <literal>node1</literal> and <literal>node2</literal> when data is present\\n+ in existing nodes <literal>node1</literal> and <literal>node2</literal>\\n+ needs similar steps. The only change required here is that\\n+ <literal>node3</literal> should create a subscription with\\n+ <literal>copy_data = force</literal> to one of the existing nodes to\\n+ receive the existing data during initial data synchronization.\\n+ </para>\\n\\nI think perhaps this should clarify that there is NO data in node3 for\\nthis example.\\n\\n~~~\\n\\n13. doc/src/sgml/logical-replication.sgml - section 31.11.3\\n\\n+ <para>\\n+ Lock the required tables in <literal>node2</literal> and\\n+ <literal>node3</literal> till the setup is completed.\\n+ </para>\\n\\nShould you describe why you say no lock is required for tables on\\nnode1? (Similar information is in section 31.11.4 also)\\n\\n~~~\\n\\n14. doc/src/sgml/logical-replication.sgml - section 31.11.3\\n\\n+ <para>\\n+ Create a subscription in <literal>node3</literal> to subscribe to\\n+ <literal>node2</literal>:\\n+<programlisting>\\n+node3=# CREATE SUBSCRIPTION sub_node3_node2\\n+node3-# CONNECTION 'dbname=foo host=node2 user=repuser'\\n+node3-# PUBLICATION pub_node2\\n+node3-# WITH (copy_data = off, only_local = on);\\n+CREATE SUBSCRIPTION\\n+</programlisting></para>\\n\\nUsing a similar description as in section 31.11.4 this should probably\\nalso say something like:\\n\\\"Use copy_data specified as off because the initial table data would\\nhave been already copied in the previous step\\\".\\n\\n~~~\\n\\n15. doc/src/sgml/logical-replication.sgml - section 31.11.4\\n\\n+ <sect2 id=\\\"add-node-data-present-in-new-node\\\">\\n+ <title>Adding a new node when data is present in the new node</title>\\n+ <para>\\n+ Adding a new node <literal>node3</literal> to the existing\\n+ <literal>node1</literal> and <literal>node2</literal> when data is present\\n+ in the new node <literal>node3</literal> needs similar steps. A\\nfew changes\\n+ are required here to get the existing data from <literal>node3</literal>\\n+ to <literal>node1</literal> and <literal>node2</literal> and later\\n+ cleaning up of data in <literal>node3</literal> before synchronization of\\n+ all the data from the existing nodes.\\n+ </para>\\n\\nI think perhaps this should clarify that there is NO data in node1 or\\nnode2 for this example.\\n\\n~~~\\n\\n16. doc/src/sgml/logical-replication.sgml\\n\\n+ <sect2 id=\\\"generic-steps-add-new-node\\\">\\n+ <title>Generic steps to add a new node to the existing set of nodes</title>\\n\\nI think all the steps in this section should be numbered because the\\norder is important.\\n\\n~~~\\n\\n17. doc/src/sgml/logical-replication.sgml\\n\\n+ <sect2 id=\\\"generic-steps-add-new-node\\\">\\n+ <title>Generic steps to add a new node to the existing set of nodes</title>\\n\\nIIUC the word \\\"parameter\\\" should replace the word \\\"option\\\" in all the\\ntext of this section.\\n\\n~~~\\n\\n18. doc/src/sgml/logical-replication.sgml - Notes\\n\\n+ <sect2>\\n+ <title>Notes</title>\\n+ <para>\\n+ Setting up bidirectional logical replication across nodes requires multiple\\n+ steps to be performed on various nodes, as all operations are not\\n+ transactional, user is advised to take backup of existing data to avoid any\\n+ inconsistency.\\n+ </para>\\n+ </sect2>\\n\\n18a.\\nI thought this \\\"Notes\\\" section should be rendered more like an sgml\\nnote or a warning. E.g. like the warning on this page [2].\\n\\n18b.\\nSUGGESTION (split the para into multiple sentences). e.g.\\nSetting up bidirectional logical replication across nodes requires\\nmultiple steps to be performed on various nodes. Because not all\\noperations are transactional, the user is advised to take backups.\\n\\n======\\n\\n19. doc/src/sgml/ref/alter_subscription.sgml\\n\\n+ <para>\\n+ There is some interaction between the <literal>only_local</literal>\\n+ option and <literal>copy_data</literal> option. Refer to the\\n+ <command>CREATE SUBSCRIPTION</command>\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n+ details and usage of <literal>force</literal> for\\n+ <literal>copy_data</literal> option.\\n </para>\\n\\n\\\"option\\\" -> \\\"parameter\\\" (3x)\\n\\n======\\n\\n20. doc/src/sgml/ref/create_subscription.sgml\\n\\n+ <para>\\n+ There is some interaction between the <literal>only_local</literal>\\n+ option and <literal>copy_data</literal> option. Refer to the\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n+ details and usage of <literal>force</literal> for\\n+ <literal>copy_data</literal> option.\\n+ </para>\\n\\n\\\"option\\\" -> \\\"parameter\\\" (3x)\\n\\n======\\n\\n21. src/backend/commands/subscriptioncmds.c\\n\\n@@ -69,6 +69,18 @@\\n /* check if the 'val' has 'bits' set */\\n #define IsSet(val, bits) (((val) & (bits)) == (bits))\\n\\n+#define IS_COPY_DATA_ON_OR_FORCE(copy_data) ((copy_data) != COPY_DATA_OFF)\\n+\\n+/*\\n+ * Represents whether copy_data option is specified with off, on or force.\\n+ */\\n+typedef enum CopyData\\n+{\\n+ COPY_DATA_OFF,\\n+ COPY_DATA_ON,\\n+ COPY_DATA_FORCE\\n+} CopyData;\\n+\\n\\n21a.\\n\\\"option\\\" -> \\\"parameter\\\"\\n\\n21b.\\nI think the new #define is redundant and can be removed (see the next\\nreview comment)\\n\\n21c.\\nI think you set the OFF enum value like:\\ntypedef enum CopyData\\n{\\n COPY_DATA_OFF = 0,\\n COPY_DATA_ON,\\n COPY_DATA_FORCE\\n} CopyData;\\n\\nthen it will greatly simplify things; many changes in this file will\\nbe unnecessary (see subsequent review comments)\\n\\n~~~\\n\\n22. src/backend/commands/subscriptioncmds.c\\n\\n+/*\\n+ * Validate the value specified for copy_data option.\\n+ */\\n+static CopyData\\n+DefGetCopyData(DefElem *def)\\n\\n\\\"option\\\" -> \\\"parameter\\\"\\n\\n~~~\\n\\n23. src/backend/commands/subscriptioncmds.c\\n\\n+ /*\\n+ * Allow 0, 1, \\\"true\\\", \\\"false\\\", \\\"on\\\", \\\"off\\\" or \\\"force\\\".\\n+ */\\n+ switch (nodeTag(def->arg))\\n+ {\\n+ case T_Integer:\\n+ switch (intVal(def->arg))\\n+ {\\n+ case 0:\\n+ return COPY_DATA_OFF;\\n+ case 1:\\n+ return COPY_DATA_ON;\\n+ default:\\n+ /* otherwise, error out below */\\n+ break;\\n+ }\\n+ break;\\n\\nWhat about also allowing copy_data = 2, and making it equivalent to \\\"force\\\"?\\n\\n~~~\\n\\n24. src/backend/commands/subscriptioncmds.c\\n\\n@@ -333,17 +400,17 @@ parse_subscription_options(ParseState *pstate,\\nList *stmt_options,\\n errmsg(\\\"%s and %s are mutually exclusive options\\\",\\n \\\"connect = false\\\", \\\"create_slot = true\\\")));\\n\\n- if (opts->copy_data &&\\n+ if (IS_COPY_DATA_ON_OR_FORCE(opts->copy_data) &&\\n IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\\n\\nI think this change would not be needed if you set the OFF enum = 0.\\n\\n~~~\\n25. src/backend/commands/subscriptioncmds.c\\n\\n@@ -671,13 +738,14 @@ CreateSubscription(ParseState *pstate,\\nCreateSubscriptionStmt *stmt,\\n * Set sync state based on if we were asked to do data copy or\\n * not.\\n */\\n- table_state = opts.copy_data ? SUBREL_STATE_INIT : SUBREL_STATE_READY;\\n+ table_state = IS_COPY_DATA_ON_OR_FORCE(opts.copy_data) ?\\nSUBREL_STATE_INIT : SUBREL_STATE_READY;\\n\\nI think this change would not be needed if you set the OFF enum = 0.\\n\\n~~~\\n\\n26. src/backend/commands/subscriptioncmds.c\\n\\n@@ -720,7 +788,8 @@ CreateSubscription(ParseState *pstate,\\nCreateSubscriptionStmt *stmt,\\n * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\\n * PUBLICATION to work.\\n */\\n- if (opts.twophase && !opts.copy_data && tables != NIL)\\n+ if (opts.twophase && opts.copy_data == COPY_DATA_OFF &&\\n+ tables != NIL)\\n twophase_enabled = true;\\n\\nI think this change would not be needed if you set the OFF enum = 0.\\n\\n~~~\\n\\n27. src/backend/commands/subscriptioncmds.c\\n\\n@@ -851,7 +921,7 @@ AlterSubscription_refresh(Subscription *sub, bool copy_data,\\n list_length(subrel_states), sizeof(Oid), oid_cmp))\\n {\\n AddSubscriptionRelState(sub->oid, relid,\\n- copy_data ? SUBREL_STATE_INIT : SUBREL_STATE_READY,\\n+ IS_COPY_DATA_ON_OR_FORCE(copy_data) ? SUBREL_STATE_INIT : SUBREL_STATE_READY,\\n InvalidXLogRecPtr);\\n\\nI think this change would not be needed if you set the OFF enum = 0.\\n\\n\\n~~~\\n28. src/backend/commands/subscriptioncmds.c\\n\\n@@ -1157,7 +1227,7 @@ AlterSubscription(ParseState *pstate,\\nAlterSubscriptionStmt *stmt,\\n * See ALTER_SUBSCRIPTION_REFRESH for details why this is\\n * not allowed.\\n */\\n- if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\n+ if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\\nIS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\\n ereport(ERROR,\\n (errcode(ERRCODE_SYNTAX_ERROR),\\n errmsg(\\\"ALTER SUBSCRIPTION with refresh and copy_data is not allowed\\nwhen two_phase is enabled\\\"),\\n\\nI think this change would not be needed if you set the OFF enum = 0.\\n\\n~~~\\n\\n29. src/backend/commands/subscriptioncmds.c\\n\\n@@ -1209,7 +1279,7 @@ AlterSubscription(ParseState *pstate,\\nAlterSubscriptionStmt *stmt,\\n * See ALTER_SUBSCRIPTION_REFRESH for details why this is\\n * not allowed.\\n */\\n- if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\n+ if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\\nIS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\\n ereport(ERROR,\\n (errcode(ERRCODE_SYNTAX_ERROR),\\n errmsg(\\\"ALTER SUBSCRIPTION with refresh and copy_data is not allowed\\nwhen two_phase is enabled\\\"),\\n\\nI think this change would not be needed if you set the OFF enum = 0.\\n\\n~~~\\n\\n30. src/backend/commands/subscriptioncmds.c\\n\\n@@ -1255,7 +1325,8 @@ AlterSubscription(ParseState *pstate,\\nAlterSubscriptionStmt *stmt,\\n *\\n * For more details see comments atop worker.c.\\n */\\n- if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\n+ if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\\n+ IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\\n ereport(ERROR,\\n (errcode(ERRCODE_SYNTAX_ERROR),\\n errmsg(\\\"ALTER SUBSCRIPTION ... REFRESH with copy_data is not allowed\\nwhen two_phase is enabled\\\"),\\n\\nI think this change would not be needed if you set the OFF enum = 0.\\n\\n~~~\\n\\n31. src/backend/commands/subscriptioncmds.c\\n\\n+ appendStringInfoString(&cmd,\\n+ \\\"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\\nPS.srrelid as replicated\\\\n\\\"\\n+ \\\"FROM pg_publication P,\\\\n\\\"\\n+ \\\"LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n+ \\\"LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n+ \\\"pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n+ \\\"WHERE C.oid = GPT.relid AND P.pubname in (\\\");\\n\\nuse upper case \\\"in\\\" -> \\\"IN\\\"\\n\\n======\\n\\n32. src/test/regress/sql/subscription.sql\\n\\n@@ -40,6 +40,9 @@ SET SESSION AUTHORIZATION 'regress_subscription_user';\\n\\n -- fail - invalid option combinations\\n CREATE SUBSCRIPTION regress_testsub2 CONNECTION\\n'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\\nfalse, copy_data = true);\\n+CREATE SUBSCRIPTION regress_testsub2 CONNECTION\\n'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\\nfalse, copy_data = on);\\n+CREATE SUBSCRIPTION regress_testsub2 CONNECTION\\n'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\\nfalse, copy_data = 1);\\n+CREATE SUBSCRIPTION regress_testsub2 CONNECTION\\n'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\\nfalse, copy_data = force);\\n\\n32a.\\nThe test for \\\"copy_data = 1\\\" should confirm it is the same as \\\"on\\\".\\n\\n32b.\\nShould also test copy_data = 2 (and confirm it is the same as \\\"force\\\").\\n\\n======\\n\\n33. src/test/subscription/t/032_localonly.pl - cosmetic changes\\n\\n33a.\\n+# Detach node C from the node-group of (A, B, C) and clean the table contents\\n+# from all nodes.\\n\\nSUGGESTION\\nDetach node_C from the node-group of (node_A, node_B, node_C) and ...\\n\\n33b.\\n+# Subroutine to create subscription and wait till the initial sync is\\ncompleted.\\n\\n\\\"till\\\" -> \\\"until\\\"\\n\\n33c.\\n+# Subroutine to create subscription and wait till the initial sync is\\ncompleted.\\n+# Subroutine expects subscriber node, publisher node, subscription name,\\n+# destination connection string, publication name and the subscription with\\n+# options to be passed as input parameters.\\n+sub create_subscription\\n+{\\n+ my ($node_subscriber, $node_publisher, $sub_name, $node_connstr,\\n+ $pub_name, $with_options)\\n+ = @_;\\n\\n\\\"subscription with options\\\" => \\\"subscription parameters\\\"\\n\\\"$with_options\\\" -> \\\"$sub_params\\\"\\n\\n33d.\\n+# Specifying only_local 'on' which indicates that the publisher should only\\n\\n\\\"Specifying\\\" => \\\"Specify\\\"\\n\\n------\\n[1] https://www.postgresql.org/docs/15/sql-createsubscription.html\\n[2] https://www.postgresql.org/docs/current/bgworker.html\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 May 2022 18:36:41 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, May 20, 2022 at 3:31 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Wed, May 18, 2022 at 4:22 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > 5.\\n> > * It is quite possible that subscriber has not yet pulled data to\\n> > + * the tables, but in ideal cases the table data will be subscribed.\\n> > + * To keep the code simple it is not checked if the subscriber table\\n> > + * has pulled the data or not.\\n> > + */\\n> > + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\\n> >\\n> > Sorry, but I don't understand what you intend to say by the above\\n> > comment. Can you please explain?\\n>\\n> When the user specifies copy_data as on, we should check if the\\n> publisher has the publication tables being subscribed from a remote\\n> publisher. If so throw an error as it remote origin data present.\\n> Ex:\\n> Node1 - pub1 for table t1 -- no data\\n> Node2 - Sub1 subscribing to data from pub1\\n> Node2 - pub2 for table t1 -- no data\\n> Node3 - create subscription to Node2 with copy_data = ON\\n>\\n> In this case even though the table does not have any remote origin\\n> data, as Node2 is subscribing to data from Node1, throw an error.\\n> We throw an error irrespective of data present in Node1 or not to keep\\n> the code simple.\\n>\\n\\nI think we can change the contents of comments something like: \\\"Throw\\nan error if the publisher has subscribed to the same table from some\\nother publisher. We cannot differentiate between the local and\\nnon-local data that is present in the HEAP during the initial sync.\\nIdentification of local data can be done only from the WAL by using\\nthe origin id. XXX: For simplicity, we don't check whether the table\\nhas any data or not. If the table doesn't have any data then we don't\\nneed to distinguish between local and non-local data so we can avoid\\nthrowing error in that case.\\\"\\n\\nFew more comments:\\n==================\\nPatch 0002\\n======\\n1.\\n+ if (copydata == COPY_DATA_ON && only_local && !slot_attisnull(slot, 3))\\n+ ereport(ERROR,\\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n+ errmsg(\\\"CREATE/ALTER SUBSCRIPTION with only_local and copy_data as\\ntrue is not allowed when the publisher might have replicated data,\\ntable:%s.%s might have replicated data in the publisher\\\",\\n+ nspname, relname),\\n+ errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\\"));\\n\\nCan we split the error message? errmsg: table:%s.%s has replicated\\ndata in the publisher; errdetail:CREATE/ALTER SUBSCRIPTION with\\nonly_local and copy_data as true is not allowed when the publisher has\\nreplicated data\\n\\n2.\\n+ <para>\\n+ Lock the required tables in the new node until the setup is complete.\\n+ </para>\\n+ <para>\\n+ Create subscriptions on existing nodes pointing to publication on\\n+ the new node with <literal>only_local</literal> option specified as\\n+ <literal>on</literal> and <literal>copy_data</literal> specified as\\n+ <literal>on</literal>.\\n+ </para>\\n+ <para>\\n+ Wait for data to be copied from the new node to existing nodes.\\n+ </para>\\n+ <para>\\n+ Alter the publication in new node so that the truncate operation is not\\n+ replicated to the subscribers.\\n+ </para>\\n\\nHere and at other places, we should specify that the lock mode should\\nto acquire the lock on table should be EXCLUSIVE so that no concurrent\\nDML is allowed on it. Also, it is better if somewhere we explain why\\nand which nodes need locks?\\n\\nPatch 0001:\\n==========\\n1.\\n+$node_A->append_conf(\\n+ 'postgresql.conf', qq(\\n+max_prepared_transactions = 10\\n+logical_decoding_work_mem = 64kB\\n+));\\n\\nI don't see why you need to set these parameters. There is no test\\ncase that needs these parameters. Please remove these from here and\\nall other similar places in 032_onlylocal.pl.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 May 2022 16:40:22 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 6, 2022 at 9:36 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Apr 5, 2022 at 7:06 PM Ashutosh Bapat\\n> <ashutosh.bapat.oss@gmail.com> wrote:\\n> >\\n> > On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > >\\n> > > Below are some other name ideas. Maybe they are not improvements, but\\n> > > it might help other people to come up with something better.\\n> > >\\n> > > subscribe_publocal_only = true/false\\n> > > origin = pub_only/any\\n> > > adjacent_data_only = true/false\\n> > > direct_subscriptions_only = true/false\\n> > > ...\\n> >\\n> > FWIW, The subscriber wants \\\"changes originated on publisher\\\". From\\n> > that angle origin = publisher/any looks attractive. It also leaves\\n> > open the possibility that the subscriber may ask changes from a set of\\n> > origins or even non-publisher origins.\\n> >\\n>\\n> So, how are you imagining extending it for multiple origins? I think\\n> we can have multiple origins data on a node when there are multiple\\n> subscriptions pointing to the different or same node. The origin names\\n> are internally generated names but are present in\\n> pg_replication_origin. So, are we going to ask the user to check that\\n> table and specify it as an option? But, note, that the user may not be\\n> able to immediately recognize which origin data is useful for her.\\n>\\n\\nI still don't have a very clear answer for the usability aspect but it\\nseems this was discussed in PGCon-Unconference [1] (Section: Origin\\nfilter) and there also it was suggested to allow users to specify\\nmultiple origin names. So, probably Ashutosh's idea makes sense and we\\nshould use \\\"origin = publisher/any or origin=local/any\\\". Among these,\\nI would prefer later.\\n\\nNote: I don't have email-id of Julian Markwort who has raised this\\ntopic in Unconference. Euler, if you know, it might be better to add\\nhim here, so that he is aware of what is going on and share his\\nopinion if he wishes to.\\n\\n[1] - https://wiki.postgresql.org/wiki/PgCon_2022_Developer_Unconference\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 27 May 2022 07:54:02 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are some more review comments for v15-0002 I found while working\\nthrough the documented examples. These are all for\\ndoc/src/sgml/logical-replication.sgml. The section numbers (e.g.\\n31.11.1) are the assigned numbers after HTML rendering.\\n\\nThe general comment is that the sections appeared at first to be\\nindependent of each other, but they are not really - sometimes\\nexisting data and existing pub/sub are assumed to remain. Therefore, I\\nthink some more text needs to be added to clarify the initial state\\nfor each of these sections.\\n\\n======\\n\\n1. Section id.\\n\\n<sect1 id=\\\"bidirectional-logical-replication\\\">\\n\\nIMO this section ID should be renamed\\n\\\"logical-replication-bidirectional\\\" then it will be more consistent\\nfilenames with all the other logical replication section names.\\n\\n~~~\\n\\n2. Section 31.11.2. Adding a new node when there is no data in any of the nodes\\n\\n2a.\\nThis seems like a continuation of 31.11.1 because pub/sub of\\nnode1,node2 is assumed to exist still.\\n\\n2b.\\nThe bottom of this section should say after these steps any subsequent\\nincremental data changes on any node will be replicated to all nodes\\n(node1, node2, node3)\\n\\n~~~\\n\\n3. Section 31.11.3. Adding a new node when data is present in the existing nodes\\n\\n3a.\\nThis seems to be a continuation of 31.11.1 because pub/sub (and\\ninitial data) of node1/node2 is assumed to exist.\\n\\n3b.\\n\\nThe bottom of this section should say after these steps that any\\ninitial data of node1/node2 will be seen in node3, and any subsequent\\nincremental data changes on any node will be replicated to all nodes\\n(node1, node2, node3)\\n\\n~~~\\n\\n4. Section 31.11.4. Adding a new node when data is present in the new node\\n\\n4a.\\nThis seems to be a continuation of 31.11.1 because pub/sub (and\\ninitial data) of node1/node2 is assumed to exist.\\n\\n4b.\\nIt is not made very clear up-front if the tables on node1 and node2\\nare empty or not. Apparently, they are considered NOT empty because\\nlater in the example you are using \\\"force\\\" when you create the\\nsubscription on node3.\\n\\n4c.\\nThe SQL wrapping here is strangely different from others (put the\\nsubscription name on 1st line)\\nnode3=# CREATE SUBSCRIPTION\\nnode3-# sub_node3_node1 CONNECTION 'dbname=foo host=node1 user=repuser'\\nnode3-# PUBLICATION pub_node1\\nnode3-# WITH (copy_data = force, only_local = on);\\nCREATE SUBSCRIPTION\\n\\n4d.\\nThe SQL wrapping here is strangely different from others (put the\\nsubscription name on 1st line)\\nnode3=# CREATE SUBSCRIPTION\\nnode3-# sub_node3_node2 CONNECTION 'dbname=foo host=node2 user=repuser'\\nnode3-# PUBLICATION pub_node2\\nnode3-# WITH (copy_data = off, only_local = on);\\nCREATE SUBSCRIPTION\\n\\n4e.\\nThe bottom of this section should say after this that any initial data\\nof node1/node2 will be seen in node3, and any subsequent incremental\\ndata changes on any node will be replicated to all nodes (node1,\\nnode2, node3)\\n\\n~~~\\n\\n5. Section 31.11.5. Generic steps to add a new node to the existing set of nodes\\n\\n5a\\nCreate subscriptions on the new node pointing to publication on the\\nfirst node with only_local option specified as on and copy_data option\\nspecified as \\\"force\\\".\\n\\n-> that should say \\\"Create a subscription\\\" (singular)\\n\\n5b.\\nCreate subscriptions on the new node pointing to publications on the\\nremaining node with only_local option specified as on and copy_data\\noption specified as off.\\n\\n-> that should say \\\"on the remaining node\\\" (plural)\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 27 May 2022 15:25:36 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, May 25, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> The attached v16 patch has the changes for the same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch.\\r\\n\\r\\nSome comments for the document in 0002 patch.\\r\\n\\r\\n1.\\r\\n+ <para>\\r\\n+ Lock the required tables in <literal>node1</literal> and\\r\\n+ <literal>node2</literal> till the setup is completed.\\r\\n+ </para>\\r\\n+\\r\\n+ <para>\\r\\n+ Create a publication in <literal>node1</literal>:\\r\\n+<programlisting>\\r\\n+node1=# CREATE PUBLICATION pub_node1 FOR TABLE t1;\\r\\n+CREATE PUBLICATION\\r\\n+</programlisting></para>\\r\\n\\r\\nIf the table is locked in the very beginning, we will not be able to create the\\r\\npublication (because the locks have conflict). Maybe we should switch the order\\r\\nof creating publication and locking tables here.\\r\\n\\r\\n2.\\r\\nIn the case of \\\"Adding a new node when data is present in the new node\\\", we need\\r\\nto truncate table t1 in node3, but the truncate operation would be blocked\\r\\nbecause the table has be locked before. Maybe we need some changes for it.\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\\r\\n\",\n \"msg_date\": \"Fri, 27 May 2022 07:04:18 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, May 27, 2022 at 5:04 PM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Wed, May 25, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > The attached v16 patch has the changes for the same.\\n> >\\n>\\n> Thanks for updating the patch.\\n>\\n> Some comments for the document in 0002 patch.\\n>\\n> 1.\\n> + <para>\\n> + Lock the required tables in <literal>node1</literal> and\\n> + <literal>node2</literal> till the setup is completed.\\n> + </para>\\n> +\\n> + <para>\\n> + Create a publication in <literal>node1</literal>:\\n> +<programlisting>\\n> +node1=# CREATE PUBLICATION pub_node1 FOR TABLE t1;\\n> +CREATE PUBLICATION\\n> +</programlisting></para>\\n>\\n> If the table is locked in the very beginning, we will not be able to create the\\n> publication (because the locks have conflict). Maybe we should switch the order\\n> of creating publication and locking tables here.\\n>\\n\\nI agree. It seems some of the locking instructions in the earlier\\nsections 31.11.1 - 31.11.4 are contradictory to the later \\\"generic\\\"\\nsteps given in \\\"31.11.5. Generic steps to add a new node to the\\nexisting set of nodes\\\". I'm assuming the generic steps are the\\n\\\"correct\\\" steps\\n\\ne.g. generic steps say get the lock on new node tables AFTER the\\npublication of new node.\\ne.g. generic steps say do NOT get a table lock on the node one you are\\n(first) joining to.\\n\\n~~~\\n\\nFurthermore, the generic steps are describing attaching a new N+1th\\nnode to some (1 ... N) other nodes.\\n\\nSo I think in the TWO-node case (section 31.11.1) node2 should be\\ntreated as the \\\"new\\\" node that you are attaching to the \\\"first\\\" node1.\\nIMO the section 31.11.1 example should be reversed so that it obeys\\nthe \\\"generic\\\" pattern.\\ne.g. It should be doing the CREATE PUBLICATION pub_node2 first (since\\nthat is the \\\"new\\\" node)\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 27 May 2022 18:38:25 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, May 26, 2022 at 7:06 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for patch v15-0001.\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> Should this also say new test cases were added for the test_decoding plugin?\\n\\nSeparated the test as a 0001 patch.\\n\\n> ~~~\\n>\\n> 2. contrib/test_decoding/sql/replorigin.sql\\n>\\n> @@ -119,3 +119,18 @@ SELECT data FROM\\n> pg_logical_slot_get_changes('regression_slot_no_lsn', NULL, NUL\\n> SELECT pg_replication_origin_session_reset();\\n> SELECT pg_drop_replication_slot('regression_slot_no_lsn');\\n> SELECT pg_replication_origin_drop('regress_test_decoding:\\n> regression_slot_no_lsn');\\n> +\\n> +-- Verify that remote origin data is not returned with only-local option\\n> +SELECT 'init' FROM\\n> pg_create_logical_replication_slot('regression_slot_only_local',\\n> 'test_decoding');\\n> +SELECT pg_replication_origin_create('regress_test_decoding:\\n> regression_slot_only_local');\\n> +SELECT pg_replication_origin_session_setup('regress_test_decoding:\\n> regression_slot_only_local');\\n> +INSERT INTO origin_tbl(data) VALUES ('only_local, commit1');\\n> +-- remote origin data returned when only-local option is not set\\n> +SELECT data FROM\\n> pg_logical_slot_get_changes('regression_slot_only_local', NULL, NULL,\\n> 'skip-empty-xacts', '1', 'include-xids', '0', 'only-local', '0');\\n> +INSERT INTO origin_tbl(data) VALUES ('only_local, commit2');\\n> +-- remote origin data not returned when only-local option is set\\n> +SELECT data FROM\\n> pg_logical_slot_get_changes('regression_slot_only_local', NULL, NULL,\\n> 'skip-empty-xacts', '1', 'include-xids', '0', 'only-local', '1');\\n> +-- Clean up\\n> +SELECT pg_replication_origin_session_reset();\\n> +SELECT pg_drop_replication_slot('regression_slot_only_local');\\n> +SELECT pg_replication_origin_drop('regress_test_decoding:\\n> regression_slot_only_local');\\n>\\n> All the new comments should consistently start with upper case.\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. doc/src/sgml/catalogs.sgml\\n>\\n> + <para>\\n> + If true, the subscription will request that the publisher send locally\\n> + originated changes. False indicates that the publisher sends any changes\\n> + regardless of their origin.\\n> + </para></entry>\\n>\\n> SUGGESTION\\n> If true, the subscription will request the publisher to only send\\n> changes that originated locally.\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> + <para>\\n> + Specifies whether the subscription will request the publisher to send\\n> + locally originated changes at the publisher node, or send any\\n> + publisher node changes regardless of their origin. The default is\\n> + <literal>false</literal>.\\n> + </para>\\n>\\n> This wording should be more similar to the same information in catalogs.sgml\\n>\\n> SUGGESTION\\n> Specifies whether the subscription will request the publisher to only\\n> send changes that originated locally, or to send any changes\\n> regardless of origin.\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. src/include/replication/walreceiver.h\\n>\\n> @@ -183,6 +183,7 @@ typedef struct\\n> bool streaming; /* Streaming of large transactions */\\n> bool twophase; /* Streaming of two-phase transactions at\\n> * prepare time */\\n> + bool only_local; /* publish only locally originated data */\\n> } logical;\\n> } proto;\\n> } WalRcvStreamOptions;\\n>\\n> SUGGESTION\\n> /* Publish only local origin data */\\n\\nModified\\n\\n> ~~~\\n>\\n> 6. src/test/subscription/t/032_onlylocal.pl - cosmetic changes\\n>\\n> 6a.\\n> +# Setup a bidirectional logical replication between Node_A & Node_B\\n>\\n> SUGGESTION\\n> \\\"... node_A and node_B\\\"\\n\\nModified\\n\\n> 6b.\\n> +is(1, 1, \\\"Circular replication setup is complete\\\");\\n>\\n> SUGGESTION (or maybe saying \\\"circular\\\" is also OK - I wasn't sure)\\n> \\\"Bidirectional replication setup is complete\\\"\\n\\nModified\\n\\n> 6c.\\n> +# check that bidirectional logical replication setup...\\n>\\n> Start comment sentence with upper case.\\n\\nModified\\n\\n> 6d.\\n> +###############################################################################\\n> +# check that remote data that is originated from node_C to node_B is not\\n> +# published to node_A\\n> +###############################################################################\\n>\\n> SUGGESTION\\n> Check that remote data of node_B (that originated from node_C) is not\\n> published to node_A\\n\\nModified\\n\\n> 6e.\\n> +is($result, qq(11\\n> +12\\n> +13), 'Node_C data replicated to Node_B'\\n> +);\\n>\\n> SUGGESTION for message\\n> 'node_C data replicated to node_B'\\n\\nModified\\n\\n> 6f.\\n> +is($result, qq(11\\n> +12), 'Remote data originated from other node is not replicated when\\n> only_local option is ON'\\n> +);\\n>\\n> SUGGESTION for message\\n> 'Remote data originating from another node (not the publisher) is not\\n> replicated when only_local option is ON'\\n\\nModified\\n\\n> 6g.\\n> \\\"Circular replication setup is complete\\\"\\n> 'Inserted successfully without leading to infinite recursion in\\n> bidirectional replication setup'\\n> 'Inserted successfully without leading to infinite recursion in\\n> bidirectional replication setup'\\n> 'Node_C data replicated to Node_B'\\n> 'Remote data originated from other node is not replicated when\\n> only_local option is ON'\\n>\\n> Why do some of the \\\"is\\\" messages have single quotes and others have\\n> double quotes? Should be consistent.\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. src/test/subscription/t/032_onlylocal.pl\\n>\\n> +my $appname_B2 = 'tap_sub_B2';\\n> +$node_B->safe_psql(\\n> + 'postgres', \\\"\\n> + CREATE SUBSCRIPTION tap_sub_B2\\n> + CONNECTION '$node_C_connstr application_name=$appname_B2'\\n> + PUBLICATION tap_pub_C\\n> + WITH (only_local = on)\\\");\\n> +\\n>\\n> AFAIK the \\\"WITH (only_local = on)\\\" is unnecessary here. We don't care\\n> where the node_C data came from for this test case.\\n\\nThis test was added to handle the comment as in [1]\\n\\nThanks for the comments, the attached v17 patch has the fixes for the same.\\nI have also separated the bidirectional logical replication steps,\\nsince the steps are slightly complex and keeping it separate will help\\nin easier review and modifying.\\n[1] - https://www.postgresql.org/message-id/CAHut%2BPvwqwzuoQio-hJniarDUOTyxFrwrS9hucd47gEW-9wu-g%40mail.gmail.com\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 1 Jun 2022 23:08:40 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, May 26, 2022 at 2:06 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> This post completes my review of patch v15*.\\n>\\n> ======\\n>\\n> 1. A general comment affecting both patches\\n>\\n> I realised the things defined in the subscription WITH are referred to\\n> in the PG DOCS as \\\"parameters\\\" (not \\\"options\\\"). See CREATE\\n> SUBSCRIPTION [1]. This is already reflected in some of my review\\n> comments for the v15-0002 below., but I'd already posted my v15-0001\\n> review comments before noticing this.\\n>\\n> Please also go through the v15-0001 patch to make all the necessary\\n> terminology changes (including in the v15-0001 commit message).\\n\\nModified\\n\\n> =====\\n>\\n> Please find below review comments for patch v15-0002.\\n>\\n> 2. Commit message\\n>\\n> This patch does couple of things:\\n> change 1) Added force option for copy_data.\\n> change 2) Check and throw an error if the publication tables were also\\n> subscribing data in the publisher from other publishers.\\n>\\n> --\\n>\\n> 2a.\\n> \\\"couple\\\" -> \\\"a couple\\\"\\n> \\\"Added force\\\" -> \\\"Adds force ...\\\"\\n> \\\"Check and throw\\\" -> \\\"Checks and throws ...\\\"\\n\\nModified\\n\\n> 2b.\\n> Isn't \\\"Change 2)\\\" only applicable when copy_data and only_local are\\n> on? It should say so here.\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. Commit message\\n>\\n> node1: Table t1 (c1 int) has data 1, 2, 3, 4\\n> node2: Table t1 (c1 int) has data 5, 6, 7, 8\\n>\\n> --\\n>\\n> I feel these kinds of replication examples are always much easier to\\n> visualize when the data somehow indicates where it came from. E.g.\\n> consider using different data in your example like:\\n>\\n> node1: Table t1 (c1 int) has data 11, 12, 13, 14\\n> node2: Table t1 (c1 int) has data 21, 22, 23, 24\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. Commit message\\n>\\n> Examples in this message exceed 80 chars. Please wrap the SQL appropriately\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. Commit message\\n>\\n> After this the data will be something like this:\\n> node1:\\n> 1, 2, 3, 4, 5, 6, 7, 8\\n>\\n> node2:\\n> 1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\\n>\\n> So, you can see that data on node2 (5, 6, 7, 8) is duplicated.\\n>\\n> --\\n>\\n> Yeah, but is that entirely correct though? AFAIK the scenario\\n> described by this example is going to recurse forever.\\n\\nModified\\n\\n> ~~~\\n>\\n> 6. Commit message\\n>\\n> This problem can be solved by using only_local and copy_data option as\\n> given below\\n>\\n> --\\n>\\n> \\\"option\\\" -> \\\"parameters\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. Commit message\\n>\\n> There are other minor review comments but it's too messy to report in\\n> detail here. I suggest checking all the \\\":\\\" (some were missing) and\\n> the letter case of/within the sentences. I also suggest\\n> cutting/pasting all this text into MSWord or some grammar checker to\\n> correct anything else reported.\\n\\nModified\\n\\n> ~~~\\n>\\n> 8. Commit message\\n>\\n> CREATE SUBSCRIPTION sub_node2_node1 CONNECTION '<node1 details>'\\n> PUBLICATION pub_node1 WITH (copy_data = force, only_local = on);\\n> ERROR: CREATE/ALTER SUBSCRIPTION with only_local and copy_data as\\n> true is not allowed when the publisher might have replicated data\\n>\\n> --\\n>\\n> Looks like a typo: e.g. If copy_data is \\\"force\\\" then how is the error happening?\\n\\nModified\\n\\n> ~~~\\n>\\n> 9. Commit message\\n>\\n> The examples seem confused/muddled in some way IMO.\\n>\\n> IIUC the new ERROR is like a new integrity check for when the\\n> subscription is using the only_local = on, along with copy_data = on.\\n>\\n> And so then the copy_data = force was added as a way to override that\\n> previous error being reported.\\n>\\n> So unless you know about the possible error then the \\\"force\\\" option\\n> makes no sense. So I think you need to change the order of the\\n> explanations in the commit message to describe the new error first.\\n\\nModified\\n\\n> ======\\n>\\n> 10. doc/src/sgml/logical-replication.sgml\\n>\\n> +\\n> + <sect1 id=\\\"bidirectional-logical-replication\\\">\\n> + <title>Bidirectional logical replication</title>\\n> +\\n> + <sect2 id=\\\"setting-bidirectional-replication-two-nodes\\\">\\n> + <title>Setting bidirectional replication between two nodes</title>\\n> + <para>\\n> + Bidirectional replication is useful in creating a multi-master database\\n> + which helps in performing read/write operations from any of the nodes.\\n> + Setting up bidirectional logical replication between two nodes requires\\n> + creation of a publication in all the nodes, creating subscriptions in\\n> + each of the nodes that subscribes to data from all the nodes. The steps\\n> + to create a two-node bidirectional replication when there is no data in\\n> + both the nodes are given below:\\n> + </para>\\n>\\n> This description is confusing. It seems to be trying to say something\\n> about a multi-master system (e.g. you say \\\"all the nodes\\\") yet this is\\n> all written inside the section about \\\"... two nodes\\\", so saying \\\"all\\n> the nodes\\\" makes no sense here.\\n>\\n> I think you need to split this information and put some kind of blurb\\n> text at the top level (section 31.11), and then this section (for \\\"two\\n> nodes\\\") should ONLY be talking about the case when there are just TWO\\n> nodes.\\n\\nModified\\n\\n> ~~~\\n>\\n> 11. doc/src/sgml/logical-replication.sgml\\n>\\n> + <para>\\n> + Lock the required tables in <literal>node1</literal> and\\n> + <literal>node2</literal> till the setup is completed.\\n> + </para>\\n>\\n> \\\"till\\\" -> \\\"until\\\" (make this same change in multiple places)\\n\\nModified\\n\\n> ~~~\\n>\\n> 12. doc/src/sgml/logical-replication.sgml\\n>\\n> + <sect2 id=\\\"add-new-node-data-in-existing-node\\\">\\n> + <title>Adding a new node when data is present in the existing nodes</title>\\n> + <para>\\n> + Adding a new node <literal>node3</literal> to the existing\\n> + <literal>node1</literal> and <literal>node2</literal> when data is present\\n> + in existing nodes <literal>node1</literal> and <literal>node2</literal>\\n> + needs similar steps. The only change required here is that\\n> + <literal>node3</literal> should create a subscription with\\n> + <literal>copy_data = force</literal> to one of the existing nodes to\\n> + receive the existing data during initial data synchronization.\\n> + </para>\\n>\\n> I think perhaps this should clarify that there is NO data in node3 for\\n> this example.\\n\\nModified\\n\\n> ~~~\\n>\\n> 13. doc/src/sgml/logical-replication.sgml - section 31.11.3\\n>\\n> + <para>\\n> + Lock the required tables in <literal>node2</literal> and\\n> + <literal>node3</literal> till the setup is completed.\\n> + </para>\\n>\\n> Should you describe why you say no lock is required for tables on\\n> node1? (Similar information is in section 31.11.4 also)\\n\\nModified\\n\\n> ~~~\\n>\\n> 14. doc/src/sgml/logical-replication.sgml - section 31.11.3\\n>\\n> + <para>\\n> + Create a subscription in <literal>node3</literal> to subscribe to\\n> + <literal>node2</literal>:\\n> +<programlisting>\\n> +node3=# CREATE SUBSCRIPTION sub_node3_node2\\n> +node3-# CONNECTION 'dbname=foo host=node2 user=repuser'\\n> +node3-# PUBLICATION pub_node2\\n> +node3-# WITH (copy_data = off, only_local = on);\\n> +CREATE SUBSCRIPTION\\n> +</programlisting></para>\\n>\\n> Using a similar description as in section 31.11.4 this should probably\\n> also say something like:\\n> \\\"Use copy_data specified as off because the initial table data would\\n> have been already copied in the previous step\\\".\\n\\nModified\\n\\n> ~~~\\n>\\n> 15. doc/src/sgml/logical-replication.sgml - section 31.11.4\\n>\\n> + <sect2 id=\\\"add-node-data-present-in-new-node\\\">\\n> + <title>Adding a new node when data is present in the new node</title>\\n> + <para>\\n> + Adding a new node <literal>node3</literal> to the existing\\n> + <literal>node1</literal> and <literal>node2</literal> when data is present\\n> + in the new node <literal>node3</literal> needs similar steps. A\\n> few changes\\n> + are required here to get the existing data from <literal>node3</literal>\\n> + to <literal>node1</literal> and <literal>node2</literal> and later\\n> + cleaning up of data in <literal>node3</literal> before synchronization of\\n> + all the data from the existing nodes.\\n> + </para>\\n>\\n> I think perhaps this should clarify that there is NO data in node1 or\\n> node2 for this example.\\n\\nThis section contents have been removed because this is not feasible currently.\\n\\n> ~~~\\n>\\n> 16. doc/src/sgml/logical-replication.sgml\\n>\\n> + <sect2 id=\\\"generic-steps-add-new-node\\\">\\n> + <title>Generic steps to add a new node to the existing set of nodes</title>\\n>\\n> I think all the steps in this section should be numbered because the\\n> order is important.\\n\\nModified\\n\\n> ~~~\\n>\\n> 17. doc/src/sgml/logical-replication.sgml\\n>\\n> + <sect2 id=\\\"generic-steps-add-new-node\\\">\\n> + <title>Generic steps to add a new node to the existing set of nodes</title>\\n>\\n> IIUC the word \\\"parameter\\\" should replace the word \\\"option\\\" in all the\\n> text of this section.\\n\\nModified\\n\\n> ~~~\\n>\\n> 18. doc/src/sgml/logical-replication.sgml - Notes\\n>\\n> + <sect2>\\n> + <title>Notes</title>\\n> + <para>\\n> + Setting up bidirectional logical replication across nodes requires multiple\\n> + steps to be performed on various nodes, as all operations are not\\n> + transactional, user is advised to take backup of existing data to avoid any\\n> + inconsistency.\\n> + </para>\\n> + </sect2>\\n>\\n> 18a.\\n> I thought this \\\"Notes\\\" section should be rendered more like an sgml\\n> note or a warning. E.g. like the warning on this page [2].\\n\\nModified\\n\\n> 18b.\\n> SUGGESTION (split the para into multiple sentences). e.g.\\n> Setting up bidirectional logical replication across nodes requires\\n> multiple steps to be performed on various nodes. Because not all\\n> operations are transactional, the user is advised to take backups.\\n\\nModified\\n\\n> ======\\n>\\n> 19. doc/src/sgml/ref/alter_subscription.sgml\\n>\\n> + <para>\\n> + There is some interaction between the <literal>only_local</literal>\\n> + option and <literal>copy_data</literal> option. Refer to the\\n> + <command>CREATE SUBSCRIPTION</command>\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n> + details and usage of <literal>force</literal> for\\n> + <literal>copy_data</literal> option.\\n> </para>\\n>\\n> \\\"option\\\" -> \\\"parameter\\\" (3x)\\n\\nModified\\n\\n> ======\\n>\\n> 20. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> + <para>\\n> + There is some interaction between the <literal>only_local</literal>\\n> + option and <literal>copy_data</literal> option. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n> + details and usage of <literal>force</literal> for\\n> + <literal>copy_data</literal> option.\\n> + </para>\\n>\\n> \\\"option\\\" -> \\\"parameter\\\" (3x)\\n\\nModified\\n\\n> ======\\n>\\n> 21. src/backend/commands/subscriptioncmds.c\\n>\\n> @@ -69,6 +69,18 @@\\n> /* check if the 'val' has 'bits' set */\\n> #define IsSet(val, bits) (((val) & (bits)) == (bits))\\n>\\n> +#define IS_COPY_DATA_ON_OR_FORCE(copy_data) ((copy_data) != COPY_DATA_OFF)\\n> +\\n> +/*\\n> + * Represents whether copy_data option is specified with off, on or force.\\n> + */\\n> +typedef enum CopyData\\n> +{\\n> + COPY_DATA_OFF,\\n> + COPY_DATA_ON,\\n> + COPY_DATA_FORCE\\n> +} CopyData;\\n> +\\n>\\n> 21a.\\n> \\\"option\\\" -> \\\"parameter\\\"\\n\\nModified\\n\\n> 21b.\\n> I think the new #define is redundant and can be removed (see the next\\n> review comment)\\n\\nModified\\n\\n> 21c.\\n> I think you set the OFF enum value like:\\n> typedef enum CopyData\\n> {\\n> COPY_DATA_OFF = 0,\\n> COPY_DATA_ON,\\n> COPY_DATA_FORCE\\n> } CopyData;\\n>\\n> then it will greatly simplify things; many changes in this file will\\n> be unnecessary (see subsequent review comments)\\n\\nModified\\n\\n> ~~~\\n>\\n> 22. src/backend/commands/subscriptioncmds.c\\n>\\n> +/*\\n> + * Validate the value specified for copy_data option.\\n> + */\\n> +static CopyData\\n> +DefGetCopyData(DefElem *def)\\n>\\n> \\\"option\\\" -> \\\"parameter\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 23. src/backend/commands/subscriptioncmds.c\\n>\\n> + /*\\n> + * Allow 0, 1, \\\"true\\\", \\\"false\\\", \\\"on\\\", \\\"off\\\" or \\\"force\\\".\\n> + */\\n> + switch (nodeTag(def->arg))\\n> + {\\n> + case T_Integer:\\n> + switch (intVal(def->arg))\\n> + {\\n> + case 0:\\n> + return COPY_DATA_OFF;\\n> + case 1:\\n> + return COPY_DATA_ON;\\n> + default:\\n> + /* otherwise, error out below */\\n> + break;\\n> + }\\n> + break;\\n>\\n> What about also allowing copy_data = 2, and making it equivalent to \\\"force\\\"?\\n\\nI felt the existing looks ok, no need to support 2. It might confuse the user.\\n\\n> ~~~\\n>\\n> 24. src/backend/commands/subscriptioncmds.c\\n>\\n> @@ -333,17 +400,17 @@ parse_subscription_options(ParseState *pstate,\\n> List *stmt_options,\\n> errmsg(\\\"%s and %s are mutually exclusive options\\\",\\n> \\\"connect = false\\\", \\\"create_slot = true\\\")));\\n>\\n> - if (opts->copy_data &&\\n> + if (IS_COPY_DATA_ON_OR_FORCE(opts->copy_data) &&\\n> IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\\n>\\n> I think this change would not be needed if you set the OFF enum = 0.\\n\\nModified\\n\\n> ~~~\\n> 25. src/backend/commands/subscriptioncmds.c\\n>\\n> @@ -671,13 +738,14 @@ CreateSubscription(ParseState *pstate,\\n> CreateSubscriptionStmt *stmt,\\n> * Set sync state based on if we were asked to do data copy or\\n> * not.\\n> */\\n> - table_state = opts.copy_data ? SUBREL_STATE_INIT : SUBREL_STATE_READY;\\n> + table_state = IS_COPY_DATA_ON_OR_FORCE(opts.copy_data) ?\\n> SUBREL_STATE_INIT : SUBREL_STATE_READY;\\n>\\n> I think this change would not be needed if you set the OFF enum = 0.\\n\\nModified\\n\\n> ~~~\\n>\\n> 26. src/backend/commands/subscriptioncmds.c\\n>\\n> @@ -720,7 +788,8 @@ CreateSubscription(ParseState *pstate,\\n> CreateSubscriptionStmt *stmt,\\n> * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\\n> * PUBLICATION to work.\\n> */\\n> - if (opts.twophase && !opts.copy_data && tables != NIL)\\n> + if (opts.twophase && opts.copy_data == COPY_DATA_OFF &&\\n> + tables != NIL)\\n> twophase_enabled = true;\\n>\\n> I think this change would not be needed if you set the OFF enum = 0.\\n\\nModified\\n\\n> ~~~\\n>\\n> 27. src/backend/commands/subscriptioncmds.c\\n>\\n> @@ -851,7 +921,7 @@ AlterSubscription_refresh(Subscription *sub, bool copy_data,\\n> list_length(subrel_states), sizeof(Oid), oid_cmp))\\n> {\\n> AddSubscriptionRelState(sub->oid, relid,\\n> - copy_data ? SUBREL_STATE_INIT : SUBREL_STATE_READY,\\n> + IS_COPY_DATA_ON_OR_FORCE(copy_data) ? SUBREL_STATE_INIT : SUBREL_STATE_READY,\\n> InvalidXLogRecPtr);\\n>\\n> I think this change would not be needed if you set the OFF enum = 0.\\n>\\n\\nModified\\n\\n> ~~~\\n> 28. src/backend/commands/subscriptioncmds.c\\n>\\n> @@ -1157,7 +1227,7 @@ AlterSubscription(ParseState *pstate,\\n> AlterSubscriptionStmt *stmt,\\n> * See ALTER_SUBSCRIPTION_REFRESH for details why this is\\n> * not allowed.\\n> */\\n> - if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\n> + if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\\n> IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\\n> ereport(ERROR,\\n> (errcode(ERRCODE_SYNTAX_ERROR),\\n> errmsg(\\\"ALTER SUBSCRIPTION with refresh and copy_data is not allowed\\n> when two_phase is enabled\\\"),\\n>\\n> I think this change would not be needed if you set the OFF enum = 0.\\n\\nModified\\n\\n> ~~~\\n>\\n> 29. src/backend/commands/subscriptioncmds.c\\n>\\n> @@ -1209,7 +1279,7 @@ AlterSubscription(ParseState *pstate,\\n> AlterSubscriptionStmt *stmt,\\n> * See ALTER_SUBSCRIPTION_REFRESH for details why this is\\n> * not allowed.\\n> */\\n> - if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\n> + if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\\n> IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\\n> ereport(ERROR,\\n> (errcode(ERRCODE_SYNTAX_ERROR),\\n> errmsg(\\\"ALTER SUBSCRIPTION with refresh and copy_data is not allowed\\n> when two_phase is enabled\\\"),\\n>\\n> I think this change would not be needed if you set the OFF enum = 0.\\n\\nModified\\n\\n> ~~~\\n>\\n> 30. src/backend/commands/subscriptioncmds.c\\n>\\n> @@ -1255,7 +1325,8 @@ AlterSubscription(ParseState *pstate,\\n> AlterSubscriptionStmt *stmt,\\n> *\\n> * For more details see comments atop worker.c.\\n> */\\n> - if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\n> + if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\\n> + IS_COPY_DATA_ON_OR_FORCE(opts.copy_data))\\n> ereport(ERROR,\\n> (errcode(ERRCODE_SYNTAX_ERROR),\\n> errmsg(\\\"ALTER SUBSCRIPTION ... REFRESH with copy_data is not allowed\\n> when two_phase is enabled\\\"),\\n>\\n> I think this change would not be needed if you set the OFF enum = 0.\\n\\nModified\\n\\n> ~~~\\n>\\n> 31. src/backend/commands/subscriptioncmds.c\\n>\\n> + appendStringInfoString(&cmd,\\n> + \\\"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\\n> PS.srrelid as replicated\\\\n\\\"\\n> + \\\"FROM pg_publication P,\\\\n\\\"\\n> + \\\"LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n> + \\\"LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n> + \\\"pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n> + \\\"WHERE C.oid = GPT.relid AND P.pubname in (\\\");\\n>\\n> use upper case \\\"in\\\" -> \\\"IN\\\"\\n\\nModified\\n\\n> ======\\n>\\n> 32. src/test/regress/sql/subscription.sql\\n>\\n> @@ -40,6 +40,9 @@ SET SESSION AUTHORIZATION 'regress_subscription_user';\\n>\\n> -- fail - invalid option combinations\\n> CREATE SUBSCRIPTION regress_testsub2 CONNECTION\\n> 'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\\n> false, copy_data = true);\\n> +CREATE SUBSCRIPTION regress_testsub2 CONNECTION\\n> 'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\\n> false, copy_data = on);\\n> +CREATE SUBSCRIPTION regress_testsub2 CONNECTION\\n> 'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\\n> false, copy_data = 1);\\n> +CREATE SUBSCRIPTION regress_testsub2 CONNECTION\\n> 'dbname=regress_doesnotexist' PUBLICATION testpub WITH (connect =\\n> false, copy_data = force);\\n>\\n> 32a.\\n> The test for \\\"copy_data = 1\\\" should confirm it is the same as \\\"on\\\".\\n\\nWe have the test for copy_data = on and copy_data = 1, both the\\nresults are the same. Isn't that sufficient?\\n\\n> 32b.\\n> Should also test copy_data = 2 (and confirm it is the same as \\\"force\\\").\\n\\nCurrently only on, off, true, false, 0 & 1 is supported. 2 is not\\nsupported. Added a test for copy_data as 2\\n\\n> ======\\n>\\n> 33. src/test/subscription/t/032_localonly.pl - cosmetic changes\\n>\\n> 33a.\\n> +# Detach node C from the node-group of (A, B, C) and clean the table contents\\n> +# from all nodes.\\n>\\n> SUGGESTION\\n> Detach node_C from the node-group of (node_A, node_B, node_C) and ...\\n\\nModified\\n\\n> 33b.\\n> +# Subroutine to create subscription and wait till the initial sync is\\n> completed.\\n>\\n> \\\"till\\\" -> \\\"until\\\"\\n\\nModified\\n\\n> 33c.\\n> +# Subroutine to create subscription and wait till the initial sync is\\n> completed.\\n> +# Subroutine expects subscriber node, publisher node, subscription name,\\n> +# destination connection string, publication name and the subscription with\\n> +# options to be passed as input parameters.\\n> +sub create_subscription\\n> +{\\n> + my ($node_subscriber, $node_publisher, $sub_name, $node_connstr,\\n> + $pub_name, $with_options)\\n> + = @_;\\n>\\n> \\\"subscription with options\\\" => \\\"subscription parameters\\\"\\n> \\\"$with_options\\\" -> \\\"$sub_params\\\"\\n\\nModified\\n\\n> 33d.\\n> +# Specifying only_local 'on' which indicates that the publisher should only\\n>\\n> \\\"Specifying\\\" => \\\"Specify\\\"\\n\\nModified\\n\\nThanks for the comments, the v17 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Jun 2022 23:16:16 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, May 26, 2022 at 4:40 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, May 20, 2022 at 3:31 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Wed, May 18, 2022 at 4:22 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > >\\n> > > 5.\\n> > > * It is quite possible that subscriber has not yet pulled data to\\n> > > + * the tables, but in ideal cases the table data will be subscribed.\\n> > > + * To keep the code simple it is not checked if the subscriber table\\n> > > + * has pulled the data or not.\\n> > > + */\\n> > > + if (copydata == COPY_DATA_ON && local_only && !slot_attisnull(slot, 3))\\n> > >\\n> > > Sorry, but I don't understand what you intend to say by the above\\n> > > comment. Can you please explain?\\n> >\\n> > When the user specifies copy_data as on, we should check if the\\n> > publisher has the publication tables being subscribed from a remote\\n> > publisher. If so throw an error as it remote origin data present.\\n> > Ex:\\n> > Node1 - pub1 for table t1 -- no data\\n> > Node2 - Sub1 subscribing to data from pub1\\n> > Node2 - pub2 for table t1 -- no data\\n> > Node3 - create subscription to Node2 with copy_data = ON\\n> >\\n> > In this case even though the table does not have any remote origin\\n> > data, as Node2 is subscribing to data from Node1, throw an error.\\n> > We throw an error irrespective of data present in Node1 or not to keep\\n> > the code simple.\\n> >\\n>\\n> I think we can change the contents of comments something like: \\\"Throw\\n> an error if the publisher has subscribed to the same table from some\\n> other publisher. We cannot differentiate between the local and\\n> non-local data that is present in the HEAP during the initial sync.\\n> Identification of local data can be done only from the WAL by using\\n> the origin id. XXX: For simplicity, we don't check whether the table\\n> has any data or not. If the table doesn't have any data then we don't\\n> need to distinguish between local and non-local data so we can avoid\\n> throwing error in that case.\\\"\\n\\nModified\\n\\n> Few more comments:\\n> ==================\\n> Patch 0002\\n> ======\\n> 1.\\n> + if (copydata == COPY_DATA_ON && only_local && !slot_attisnull(slot, 3))\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"CREATE/ALTER SUBSCRIPTION with only_local and copy_data as\\n> true is not allowed when the publisher might have replicated data,\\n> table:%s.%s might have replicated data in the publisher\\\",\\n> + nspname, relname),\\n> + errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force\\\"));\\n>\\n> Can we split the error message? errmsg: table:%s.%s has replicated\\n> data in the publisher; errdetail:CREATE/ALTER SUBSCRIPTION with\\n> only_local and copy_data as true is not allowed when the publisher has\\n> replicated data\\n\\nModified\\n\\n> 2.\\n> + <para>\\n> + Lock the required tables in the new node until the setup is complete.\\n> + </para>\\n> + <para>\\n> + Create subscriptions on existing nodes pointing to publication on\\n> + the new node with <literal>only_local</literal> option specified as\\n> + <literal>on</literal> and <literal>copy_data</literal> specified as\\n> + <literal>on</literal>.\\n> + </para>\\n> + <para>\\n> + Wait for data to be copied from the new node to existing nodes.\\n> + </para>\\n> + <para>\\n> + Alter the publication in new node so that the truncate operation is not\\n> + replicated to the subscribers.\\n> + </para>\\n>\\n> Here and at other places, we should specify that the lock mode should\\n> to acquire the lock on table should be EXCLUSIVE so that no concurrent\\n> DML is allowed on it. Also, it is better if somewhere we explain why\\n> and which nodes need locks?\\n\\nModified\\n\\n> Patch 0001:\\n> ==========\\n> 1.\\n> +$node_A->append_conf(\\n> + 'postgresql.conf', qq(\\n> +max_prepared_transactions = 10\\n> +logical_decoding_work_mem = 64kB\\n> +));\\n>\\n> I don't see why you need to set these parameters. There is no test\\n> case that needs these parameters. Please remove these from here and\\n> all other similar places in 032_onlylocal.pl.\\n\\nRemoved it.\\n\\nThanks for the comments, the v17 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Jun 2022 23:18:07 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, May 27, 2022 at 7:54 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Apr 6, 2022 at 9:36 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Tue, Apr 5, 2022 at 7:06 PM Ashutosh Bapat\\n> > <ashutosh.bapat.oss@gmail.com> wrote:\\n> > >\\n> > > On Tue, Apr 5, 2022 at 6:21 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > > >\\n> > > > Below are some other name ideas. Maybe they are not improvements, but\\n> > > > it might help other people to come up with something better.\\n> > > >\\n> > > > subscribe_publocal_only = true/false\\n> > > > origin = pub_only/any\\n> > > > adjacent_data_only = true/false\\n> > > > direct_subscriptions_only = true/false\\n> > > > ...\\n> > >\\n> > > FWIW, The subscriber wants \\\"changes originated on publisher\\\". From\\n> > > that angle origin = publisher/any looks attractive. It also leaves\\n> > > open the possibility that the subscriber may ask changes from a set of\\n> > > origins or even non-publisher origins.\\n> > >\\n> >\\n> > So, how are you imagining extending it for multiple origins? I think\\n> > we can have multiple origins data on a node when there are multiple\\n> > subscriptions pointing to the different or same node. The origin names\\n> > are internally generated names but are present in\\n> > pg_replication_origin. So, are we going to ask the user to check that\\n> > table and specify it as an option? But, note, that the user may not be\\n> > able to immediately recognize which origin data is useful for her.\\n> >\\n>\\n> I still don't have a very clear answer for the usability aspect but it\\n> seems this was discussed in PGCon-Unconference [1] (Section: Origin\\n> filter) and there also it was suggested to allow users to specify\\n> multiple origin names. So, probably Ashutosh's idea makes sense and we\\n> should use \\\"origin = publisher/any or origin=local/any\\\". Among these,\\n> I would prefer later.\\n\\nI have changed it to origin = local/any.\\n\\nThe v17 patch attached at [1] has the changes for the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Jun 2022 23:19:03 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, May 27, 2022 at 10:56 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some more review comments for v15-0002 I found while working\\n> through the documented examples. These are all for\\n> doc/src/sgml/logical-replication.sgml. The section numbers (e.g.\\n> 31.11.1) are the assigned numbers after HTML rendering.\\n>\\n> The general comment is that the sections appeared at first to be\\n> independent of each other, but they are not really - sometimes\\n> existing data and existing pub/sub are assumed to remain. Therefore, I\\n> think some more text needs to be added to clarify the initial state\\n> for each of these sections.\\n>\\n> ======\\n>\\n> 1. Section id.\\n>\\n> <sect1 id=\\\"bidirectional-logical-replication\\\">\\n>\\n> IMO this section ID should be renamed\\n> \\\"logical-replication-bidirectional\\\" then it will be more consistent\\n> filenames with all the other logical replication section names.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2. Section 31.11.2. Adding a new node when there is no data in any of the nodes\\n>\\n> 2a.\\n> This seems like a continuation of 31.11.1 because pub/sub of\\n> node1,node2 is assumed to exist still.\\n\\nAdded a note\\n\\n> 2b.\\n> The bottom of this section should say after these steps any subsequent\\n> incremental data changes on any node will be replicated to all nodes\\n> (node1, node2, node3)\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. Section 31.11.3. Adding a new node when data is present in the existing nodes\\n>\\n> 3a.\\n> This seems to be a continuation of 31.11.1 because pub/sub (and\\n> initial data) of node1/node2 is assumed to exist.\\n\\nAdded a note\\n\\n> 3b.\\n>\\n> The bottom of this section should say after these steps that any\\n> initial data of node1/node2 will be seen in node3, and any subsequent\\n> incremental data changes on any node will be replicated to all nodes\\n> (node1, node2, node3)\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. Section 31.11.4. Adding a new node when data is present in the new node\\n>\\n> 4a.\\n> This seems to be a continuation of 31.11.1 because pub/sub (and\\n> initial data) of node1/node2 is assumed to exist.\\n\\nThis section contents have been removed because this is not feasible currently\\n\\n> 4b.\\n> It is not made very clear up-front if the tables on node1 and node2\\n> are empty or not. Apparently, they are considered NOT empty because\\n> later in the example you are using \\\"force\\\" when you create the\\n> subscription on node3.\\n\\nThis section contents have been removed because this is not feasible currently\\n\\n> 4c.\\n> The SQL wrapping here is strangely different from others (put the\\n> subscription name on 1st line)\\n> node3=# CREATE SUBSCRIPTION\\n> node3-# sub_node3_node1 CONNECTION 'dbname=foo host=node1 user=repuser'\\n> node3-# PUBLICATION pub_node1\\n> node3-# WITH (copy_data = force, only_local = on);\\n> CREATE SUBSCRIPTION\\n\\nThis section contents have been removed because this is not feasible currently\\n\\n> 4d.\\n> The SQL wrapping here is strangely different from others (put the\\n> subscription name on 1st line)\\n> node3=# CREATE SUBSCRIPTION\\n> node3-# sub_node3_node2 CONNECTION 'dbname=foo host=node2 user=repuser'\\n> node3-# PUBLICATION pub_node2\\n> node3-# WITH (copy_data = off, only_local = on);\\n> CREATE SUBSCRIPTION\\n\\nThis section contents have been removed because this is not feasible currently\\n\\n> 4e.\\n> The bottom of this section should say after this that any initial data\\n> of node1/node2 will be seen in node3, and any subsequent incremental\\n> data changes on any node will be replicated to all nodes (node1,\\n> node2, node3)\\n\\nThis section contents have been removed because this is not feasible currently\\n\\n> ~~~\\n>\\n> 5. Section 31.11.5. Generic steps to add a new node to the existing set of nodes\\n>\\n> 5a\\n> Create subscriptions on the new node pointing to publication on the\\n> first node with only_local option specified as on and copy_data option\\n> specified as \\\"force\\\".\\n>\\n> -> that should say \\\"Create a subscription\\\" (singular)\\n\\nModified\\n\\n> 5b.\\n> Create subscriptions on the new node pointing to publications on the\\n> remaining node with only_local option specified as on and copy_data\\n> option specified as off.\\n>\\n> -> that should say \\\"on the remaining node\\\" (plural)\\n\\nModified\\n\\nThanks for the comments, the v17 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Jun 2022 23:23:15 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, May 27, 2022 at 12:34 PM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Wed, May 25, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > The attached v16 patch has the changes for the same.\\n> >\\n>\\n> Thanks for updating the patch.\\n>\\n> Some comments for the document in 0002 patch.\\n>\\n> 1.\\n> + <para>\\n> + Lock the required tables in <literal>node1</literal> and\\n> + <literal>node2</literal> till the setup is completed.\\n> + </para>\\n> +\\n> + <para>\\n> + Create a publication in <literal>node1</literal>:\\n> +<programlisting>\\n> +node1=# CREATE PUBLICATION pub_node1 FOR TABLE t1;\\n> +CREATE PUBLICATION\\n> +</programlisting></para>\\n>\\n> If the table is locked in the very beginning, we will not be able to create the\\n> publication (because the locks have conflict). Maybe we should switch the order\\n> of creating publication and locking tables here.\\n\\nModified\\n\\n> 2.\\n> In the case of \\\"Adding a new node when data is present in the new node\\\", we need\\n> to truncate table t1 in node3, but the truncate operation would be blocked\\n> because the table has be locked before. Maybe we need some changes for it.\\n\\nThis section contents have been removed because this is not feasible currently\\n\\nThanks for the comments, the v17 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Jun 2022 23:25:02 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, May 27, 2022 at 2:08 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Fri, May 27, 2022 at 5:04 PM shiy.fnst@fujitsu.com\\n> <shiy.fnst@fujitsu.com> wrote:\\n> >\\n> > On Wed, May 25, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > The attached v16 patch has the changes for the same.\\n> > >\\n> >\\n> > Thanks for updating the patch.\\n> >\\n> > Some comments for the document in 0002 patch.\\n> >\\n> > 1.\\n> > + <para>\\n> > + Lock the required tables in <literal>node1</literal> and\\n> > + <literal>node2</literal> till the setup is completed.\\n> > + </para>\\n> > +\\n> > + <para>\\n> > + Create a publication in <literal>node1</literal>:\\n> > +<programlisting>\\n> > +node1=# CREATE PUBLICATION pub_node1 FOR TABLE t1;\\n> > +CREATE PUBLICATION\\n> > +</programlisting></para>\\n> >\\n> > If the table is locked in the very beginning, we will not be able to create the\\n> > publication (because the locks have conflict). Maybe we should switch the order\\n> > of creating publication and locking tables here.\\n> >\\n>\\n> I agree. It seems some of the locking instructions in the earlier\\n> sections 31.11.1 - 31.11.4 are contradictory to the later \\\"generic\\\"\\n> steps given in \\\"31.11.5. Generic steps to add a new node to the\\n> existing set of nodes\\\". I'm assuming the generic steps are the\\n> \\\"correct\\\" steps\\n>\\n> e.g. generic steps say get the lock on new node tables AFTER the\\n> publication of new node.\\n> e.g. generic steps say do NOT get a table lock on the node one you are\\n> (first) joining to.\\n\\nYes, the generic steps are correct. modified\\n\\n> ~~~\\n>\\n> Furthermore, the generic steps are describing attaching a new N+1th\\n> node to some (1 ... N) other nodes.\\n>\\n> So I think in the TWO-node case (section 31.11.1) node2 should be\\n> treated as the \\\"new\\\" node that you are attaching to the \\\"first\\\" node1.\\n> IMO the section 31.11.1 example should be reversed so that it obeys\\n> the \\\"generic\\\" pattern.\\n> e.g. It should be doing the CREATE PUBLICATION pub_node2 first (since\\n> that is the \\\"new\\\" node)\\n\\nIn generic steps we mention the publication must be created in a new\\nnode and we don't say about the existing nodes, because the\\npublication would be existing already. I feel the existing order of\\npublication creation in the TWO-node case (section 31.11.1) is ok.\\n\\nThanks for the comments, the v17 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm1rMihO7daiFyLdxkqArrC%2BdtM61oPXc-XrTYBYnJg3nw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Jun 2022 23:27:36 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Please find below my review comments for patch v17-0001:\\n\\n======\\n\\n1. Commit message\\nAdd a missing test to verify only-local option in test_decoding plugin.\\n\\n\\\"option\\\" -> \\\"parameter\\\"\\n\\n======\\n\\n2. contrib/test_decoding/sql/replorigin.sql\\n\\n2a.\\n+-- Verify that remote origin data is not returned with only-local option\\n\\nSUGGESTION\\n-- Verify the behaviour of the only-local parameter\\n\\n2b.\\n+-- Remote origin data returned when only-local option is not set\\n\\n\\\"option\\\" -> \\\"parameter\\\"\\n\\n2c.\\n+-- Remote origin data not returned when only-local option is set\\n\\n\\\"option\\\" -> \\\"parameter\\\"\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 3 Jun 2022 15:26:29 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Please find below my review comments for patch v17-0002:\\n\\n======\\n\\n1. Commit message\\n\\nThis patch adds a new SUBSCRIPTION parameter \\\"origin\\\". It Specifies whether\\nthe subscription will request the publisher to only send changes that\\noriginated locally, or to send any changes regardless of origin.\\n\\n\\\"It Specifies\\\" -> \\\"It specifies\\\"\\n\\n~~~\\n\\n2. Commit message\\n\\nUsage:\\nCREATE SUBSCRIPTION sub1 CONNECTION 'dbname=postgres port=9999'\\nPUBLICATION pub1 with (origin = local);\\n\\n\\\"with\\\" -> \\\"WITH\\\"\\n\\n======\\n\\n3. doc/src/sgml/catalogs.sgml\\n\\n+ <para>\\n+ Possible origin values are <literal>local</literal>,\\n+ <literal>any</literal>, or <literal>NULL</literal> if none is specified.\\n+ If <literal>local</literal>, the subscription will request the\\n+ publisher to only send changes that originated locally. If\\n+ <literal>any</literal>, the publisher sends any changes regardless of\\n+ their origin.\\n+ </para></entry>\\n\\nShould this also mention that NULL (default) is equivalent to 'any'?\\n\\nSUGGESTION\\nIf <literal>any</literal> (or <literal>NULL</literal>), the publisher\\nsends any changes regardless of their origin.\\n\\n======\\n\\n4. src/backend/catalog/pg_subscription.c\\n\\n@@ -72,6 +72,16 @@ GetSubscription(Oid subid, bool missing_ok)\\n sub->twophasestate = subform->subtwophasestate;\\n sub->disableonerr = subform->subdisableonerr;\\n\\n+ datum = SysCacheGetAttr(SUBSCRIPTIONOID,\\n+ tup,\\n+ Anum_pg_subscription_suborigin,\\n+ &isnull);\\n+\\n+ if (!isnull)\\n+ sub->origin = TextDatumGetCString(datum);\\n+ else\\n+ sub->origin = NULL;\\n+\\n /* Get conninfo */\\n datum = SysCacheGetAttr(SUBSCRIPTIONOID,\\n tup,\\n\\nMissing comment like the nearby code has:\\n/* Get origin */\\n\\n======\\n\\n5. src/backend/replication/logical/worker.c\\n\\n+/* Macro for comparing string fields that might be NULL */\\n+#define equalstr(a, b) \\\\\\n+ (((a) != NULL && (b) != NULL) ? (strcmp(a, b) == 0) : (a) == (b))\\n+\\n\\nShould that have some extra parens for the macro args?\\ne.g. \\\"strcmp((a), (b))\\\"\\n\\n~~~\\n\\n6. src/backend/replication/logical/worker.c - maybe_reread_subscription\\n\\n@@ -3059,6 +3063,7 @@ maybe_reread_subscription(void)\\n strcmp(newsub->slotname, MySubscription->slotname) != 0 ||\\n newsub->binary != MySubscription->binary ||\\n newsub->stream != MySubscription->stream ||\\n+ equalstr(newsub->origin, MySubscription->origin) ||\\n newsub->owner != MySubscription->owner ||\\n !equal(newsub->publications, MySubscription->publications))\\n {\\n\\nIs that right? Shouldn't it say !equalstr(...)?\\n\\n======\\n\\n7. src/backend/replication/pgoutput/pgoutput.c - parse_output_parameters\\n\\n@@ -380,6 +382,16 @@ parse_output_parameters(List *options, PGOutputData *data)\\n\\n data->two_phase = defGetBoolean(defel);\\n }\\n+ else if (strcmp(defel->defname, \\\"origin\\\") == 0)\\n+ {\\n+ if (origin_option_given)\\n+ ereport(ERROR,\\n+ (errcode(ERRCODE_SYNTAX_ERROR),\\n+ errmsg(\\\"conflicting or redundant options\\\")));\\n+ origin_option_given = true;\\n+\\n+ data->origin = defGetString(defel);\\n+ }\\n\\nShould this function also be validating that the origin parameter\\nvalue is only permitted to be one of \\\"local\\\" or \\\"any\\\"?\\n\\n~~~\\n\\n8. src/backend/replication/pgoutput/pgoutput.c - pgoutput_origin_filter\\n\\n@@ -1698,12 +1710,20 @@ pgoutput_message(LogicalDecodingContext *ctx,\\nReorderBufferTXN *txn,\\n }\\n\\n /*\\n- * Currently we always forward.\\n+ * Return true if the data source (origin) is remote and user has requested\\n+ * only local data, false otherwise.\\n */\\n static bool\\n pgoutput_origin_filter(LogicalDecodingContext *ctx,\\n RepOriginId origin_id)\\n {\\n+ PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\\n+\\n+ if (data->origin &&\\n+ (strcmp(data->origin, \\\"local\\\") == 0) &&\\n+ origin_id != InvalidRepOriginId)\\n+ return true;\\n+\\n return false;\\n }\\n\\n8a.\\nCould rewrite the condition by putting the strcmp last so you can\\navoid doing unnecessary strcmp.\\n\\ne.g\\n+ if (data->origin &&\\n+ origin_id != InvalidRepOriginId &&\\n+ strcmp(data->origin, \\\"local\\\" == 0)\\n\\n\\n8b.\\nI also wondered if it might be worth considering caching the origin\\nparameter value when it was parsed so that you can avoid doing any\\nstrcmp at all during this function. Because otherwise this might be\\ncalled millions of times, right?\\n\\n======\\n\\n9. src/include/catalog/pg_subscription.h\\n\\n@@ -87,6 +87,9 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\\nBKI_SHARED_RELATION BKI_ROW\\n\\n /* List of publications subscribed to */\\n text subpublications[1] BKI_FORCE_NOT_NULL;\\n+\\n+ /* Publish the data originated from the specified origin */\\n+ text suborigin;\\n #endif\\n } FormData_pg_subscription;\\n\\nSUGGESTION (for the comment text)\\n/* Only publish data originating from the specified origin */\\n\\n~~~\\n\\n10 src/include/catalog/pg_subscription.h - Subscription\\n\\n@@ -118,6 +121,9 @@ typedef struct Subscription\\n char *slotname; /* Name of the replication slot */\\n char *synccommit; /* Synchronous commit setting for worker */\\n List *publications; /* List of publication names to subscribe to */\\n+ char *origin; /* Publish the data originated from the\\n+ * specified origin */\\n+\\n } Subscription;\\n\\n10a.\\nReword comment same as suggested in review comment #9\\n\\n10b.\\nRemove spurious blank line\\n\\n======\\n\\n11. src/include/replication/walreceiver.h\\n\\n@@ -183,6 +183,8 @@ typedef struct\\n bool streaming; /* Streaming of large transactions */\\n bool twophase; /* Streaming of two-phase transactions at\\n * prepare time */\\n+ char *origin; /* Publish the data originated from the\\n+ * specified origin */\\n } logical;\\n\\nReword comment same as suggested in review comment #9\\n\\n======\\n\\n12. src/test/subscription/t/032_origin.pl\\n\\n+# Test logical replication using origin option.\\n\\n# Test the CREATE SUBSCRIPTION 'origin' parameter\\n\\n~~~\\n\\n13. src/test/subscription/t/032_origin.pl\\n\\n+# check that the data published from node_C to node_B is not sent to node_A\\n+$result = $node_A->safe_psql('postgres', \\\"SELECT * FROM tab_full ORDER BY 1;\\\");\\n+is($result, qq(11\\n+12), 'Remote data originating from another node (not the publisher)\\nis not replicated when origin option is local'\\n+);\\n\\n\\\"option\\\" -> \\\"parameter\\\"\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 3 Jun 2022 15:30:49 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jun 3, 2022 at 10:56 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Please find below my review comments for patch v17-0001:\\n>\\n> ======\\n>\\n> 1. Commit message\\n> Add a missing test to verify only-local option in test_decoding plugin.\\n>\\n> \\\"option\\\" -> \\\"parameter\\\"\\n\\nModified\\n\\n> ======\\n>\\n> 2. contrib/test_decoding/sql/replorigin.sql\\n>\\n> 2a.\\n> +-- Verify that remote origin data is not returned with only-local option\\n>\\n> SUGGESTION\\n> -- Verify the behaviour of the only-local parameter\\n\\nModified\\n\\n> 2b.\\n> +-- Remote origin data returned when only-local option is not set\\n>\\n> \\\"option\\\" -> \\\"parameter\\\"\\n\\nModified\\n\\n> 2c.\\n> +-- Remote origin data not returned when only-local option is set\\n>\\n> \\\"option\\\" -> \\\"parameter\\\"\\n\\nModified\\n\\nThe attached v18 patch has the fixes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Mon, 6 Jun 2022 10:44:08 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jun 3, 2022 at 11:01 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Please find below my review comments for patch v17-0002:\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> This patch adds a new SUBSCRIPTION parameter \\\"origin\\\". It Specifies whether\\n> the subscription will request the publisher to only send changes that\\n> originated locally, or to send any changes regardless of origin.\\n>\\n> \\\"It Specifies\\\" -> \\\"It specifies\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 2. Commit message\\n>\\n> Usage:\\n> CREATE SUBSCRIPTION sub1 CONNECTION 'dbname=postgres port=9999'\\n> PUBLICATION pub1 with (origin = local);\\n>\\n> \\\"with\\\" -> \\\"WITH\\\"\\n\\nModified\\n\\n> ======\\n>\\n> 3. doc/src/sgml/catalogs.sgml\\n>\\n> + <para>\\n> + Possible origin values are <literal>local</literal>,\\n> + <literal>any</literal>, or <literal>NULL</literal> if none is specified.\\n> + If <literal>local</literal>, the subscription will request the\\n> + publisher to only send changes that originated locally. If\\n> + <literal>any</literal>, the publisher sends any changes regardless of\\n> + their origin.\\n> + </para></entry>\\n>\\n> Should this also mention that NULL (default) is equivalent to 'any'?\\n>\\n> SUGGESTION\\n> If <literal>any</literal> (or <literal>NULL</literal>), the publisher\\n> sends any changes regardless of their origin.\\n\\nModified\\n\\n> ======\\n>\\n> 4. src/backend/catalog/pg_subscription.c\\n>\\n> @@ -72,6 +72,16 @@ GetSubscription(Oid subid, bool missing_ok)\\n> sub->twophasestate = subform->subtwophasestate;\\n> sub->disableonerr = subform->subdisableonerr;\\n>\\n> + datum = SysCacheGetAttr(SUBSCRIPTIONOID,\\n> + tup,\\n> + Anum_pg_subscription_suborigin,\\n> + &isnull);\\n> +\\n> + if (!isnull)\\n> + sub->origin = TextDatumGetCString(datum);\\n> + else\\n> + sub->origin = NULL;\\n> +\\n> /* Get conninfo */\\n> datum = SysCacheGetAttr(SUBSCRIPTIONOID,\\n> tup,\\n>\\n> Missing comment like the nearby code has:\\n> /* Get origin */\\n\\nModified\\n\\n> ======\\n>\\n> 5. src/backend/replication/logical/worker.c\\n>\\n> +/* Macro for comparing string fields that might be NULL */\\n> +#define equalstr(a, b) \\\\\\n> + (((a) != NULL && (b) != NULL) ? (strcmp(a, b) == 0) : (a) == (b))\\n> +\\n>\\n> Should that have some extra parens for the macro args?\\n> e.g. \\\"strcmp((a), (b))\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 6. src/backend/replication/logical/worker.c - maybe_reread_subscription\\n>\\n> @@ -3059,6 +3063,7 @@ maybe_reread_subscription(void)\\n> strcmp(newsub->slotname, MySubscription->slotname) != 0 ||\\n> newsub->binary != MySubscription->binary ||\\n> newsub->stream != MySubscription->stream ||\\n> + equalstr(newsub->origin, MySubscription->origin) ||\\n> newsub->owner != MySubscription->owner ||\\n> !equal(newsub->publications, MySubscription->publications))\\n> {\\n>\\n> Is that right? Shouldn't it say !equalstr(...)?\\n\\nModified\\n\\n> ======\\n>\\n> 7. src/backend/replication/pgoutput/pgoutput.c - parse_output_parameters\\n>\\n> @@ -380,6 +382,16 @@ parse_output_parameters(List *options, PGOutputData *data)\\n>\\n> data->two_phase = defGetBoolean(defel);\\n> }\\n> + else if (strcmp(defel->defname, \\\"origin\\\") == 0)\\n> + {\\n> + if (origin_option_given)\\n> + ereport(ERROR,\\n> + (errcode(ERRCODE_SYNTAX_ERROR),\\n> + errmsg(\\\"conflicting or redundant options\\\")));\\n> + origin_option_given = true;\\n> +\\n> + data->origin = defGetString(defel);\\n> + }\\n>\\n> Should this function also be validating that the origin parameter\\n> value is only permitted to be one of \\\"local\\\" or \\\"any\\\"?\\n\\nModified\\n\\n> ~~~\\n>\\n> 8. src/backend/replication/pgoutput/pgoutput.c - pgoutput_origin_filter\\n>\\n> @@ -1698,12 +1710,20 @@ pgoutput_message(LogicalDecodingContext *ctx,\\n> ReorderBufferTXN *txn,\\n> }\\n>\\n> /*\\n> - * Currently we always forward.\\n> + * Return true if the data source (origin) is remote and user has requested\\n> + * only local data, false otherwise.\\n> */\\n> static bool\\n> pgoutput_origin_filter(LogicalDecodingContext *ctx,\\n> RepOriginId origin_id)\\n> {\\n> + PGOutputData *data = (PGOutputData *) ctx->output_plugin_private;\\n> +\\n> + if (data->origin &&\\n> + (strcmp(data->origin, \\\"local\\\") == 0) &&\\n> + origin_id != InvalidRepOriginId)\\n> + return true;\\n> +\\n> return false;\\n> }\\n>\\n> 8a.\\n> Could rewrite the condition by putting the strcmp last so you can\\n> avoid doing unnecessary strcmp.\\n>\\n> e.g\\n> + if (data->origin &&\\n> + origin_id != InvalidRepOriginId &&\\n> + strcmp(data->origin, \\\"local\\\" == 0)\\n\\nI have avoided the strcmp by caching locally as suggested in 8b. I\\nhave not made any change for this.\\n\\n>\\n> 8b.\\n> I also wondered if it might be worth considering caching the origin\\n> parameter value when it was parsed so that you can avoid doing any\\n> strcmp at all during this function. Because otherwise this might be\\n> called millions of times, right?\\n\\nModified\\n\\n> ======\\n>\\n> 9. src/include/catalog/pg_subscription.h\\n>\\n> @@ -87,6 +87,9 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\\n> BKI_SHARED_RELATION BKI_ROW\\n>\\n> /* List of publications subscribed to */\\n> text subpublications[1] BKI_FORCE_NOT_NULL;\\n> +\\n> + /* Publish the data originated from the specified origin */\\n> + text suborigin;\\n> #endif\\n> } FormData_pg_subscription;\\n>\\n> SUGGESTION (for the comment text)\\n> /* Only publish data originating from the specified origin */\\n\\nModified\\n\\n> ~~~\\n>\\n> 10 src/include/catalog/pg_subscription.h - Subscription\\n>\\n> @@ -118,6 +121,9 @@ typedef struct Subscription\\n> char *slotname; /* Name of the replication slot */\\n> char *synccommit; /* Synchronous commit setting for worker */\\n> List *publications; /* List of publication names to subscribe to */\\n> + char *origin; /* Publish the data originated from the\\n> + * specified origin */\\n> +\\n> } Subscription;\\n>\\n> 10a.\\n> Reword comment same as suggested in review comment #9\\n\\nModified\\n\\n> 10b.\\n> Remove spurious blank line\\n\\nModified\\n\\n> ======\\n>\\n> 11. src/include/replication/walreceiver.h\\n>\\n> @@ -183,6 +183,8 @@ typedef struct\\n> bool streaming; /* Streaming of large transactions */\\n> bool twophase; /* Streaming of two-phase transactions at\\n> * prepare time */\\n> + char *origin; /* Publish the data originated from the\\n> + * specified origin */\\n> } logical;\\n>\\n> Reword comment same as suggested in review comment #9\\n\\nModified\\n\\n> ======\\n>\\n> 12. src/test/subscription/t/032_origin.pl\\n>\\n> +# Test logical replication using origin option.\\n>\\n> # Test the CREATE SUBSCRIPTION 'origin' parameter\\n\\nModified\\n\\n> ~~~\\n>\\n> 13. src/test/subscription/t/032_origin.pl\\n>\\n> +# check that the data published from node_C to node_B is not sent to node_A\\n>+$result = $node_A->safe_psql('postgres', \\\"SELECT * FROM tab_full ORDER BY 1;\\\");\\n> +is($result, qq(11\\n> +12), 'Remote data originating from another node (not the publisher)\\n> is not replicated when origin option is local'\\n> +);\\n>\\n> \\\"option\\\" -> \\\"parameter\\\"\\n\\nModified\\n\\nThanks for the comments, the v18 patch attached at [1] has the fixes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm0Haovukx2q7Yd987Sm8fbQ0nsh8F0EWaO_qsw0uObGBQ%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 Jun 2022 10:47:09 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 6, 2022 1:14 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> The attached v18 patch has the fixes for the same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch, here are some comments.\\r\\n\\r\\n0002 patch\\r\\n==============\\r\\n1. \\r\\n+ <varlistentry>\\r\\n+ <term><literal>origin</literal> (<type>string</type>)</term>\\r\\n+ <listitem>\\r\\n+ <para>\\r\\n\\r\\nIt maybe better if the type of \\\"origin\\\" parameter is enum, as it cannot be any\\r\\nstring and only has two valid values.\\r\\n\\r\\n2.\\r\\n@@ -607,6 +626,11 @@ CreateSubscription(ParseState *pstate, CreateSubscriptionStmt *stmt,\\r\\n \\t\\t\\t\\t\\t LOGICALREP_TWOPHASE_STATE_PENDING :\\r\\n \\t\\t\\t\\t\\t LOGICALREP_TWOPHASE_STATE_DISABLED);\\r\\n \\tvalues[Anum_pg_subscription_subdisableonerr - 1] = BoolGetDatum(opts.disableonerr);\\r\\n+\\tif (opts.origin)\\r\\n+\\t\\tvalues[Anum_pg_subscription_suborigin - 1] =\\r\\n+\\t\\t\\tCStringGetTextDatum(opts.origin);\\r\\n+\\telse\\r\\n+\\t\\tnulls[Anum_pg_subscription_suborigin - 1] = true;\\r\\n \\tvalues[Anum_pg_subscription_subconninfo - 1] =\\r\\n \\t\\tCStringGetTextDatum(conninfo);\\r\\n \\tif (opts.slot_name)\\r\\n\\r\\nDocument of \\\"CREATE SUBSCRIPTION\\\" says, the default value of \\\"origin\\\" is \\\"any\\\", so why not set\\r\\nsuborigin to \\\"any\\\" when user doesn't specify this parameter?\\r\\n\\r\\n\\r\\n0003 patch\\r\\n==============\\r\\n1.\\r\\n@@ -300,6 +310,11 @@ CREATE SUBSCRIPTION <replaceable class=\\\"parameter\\\">subscription_name</replaceabl\\r\\n replication from the publisher. The default is\\r\\n <literal>false</literal>.\\r\\n </para>\\r\\n+ <para>\\r\\n+ There is some interaction between the <literal>origin</literal>\\r\\n+ parameter and <literal>copy_data</literal> parameter. Refer to the\\r\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\r\\n+ </para>\\r\\n </listitem>\\r\\n </varlistentry>\\r\\n\\r\\nI think this change should be put together with \\\"origin\\\" parameter, instead of\\r\\n\\\"disable_on_error\\\".\\r\\n\\r\\n\\r\\n0004 patch\\r\\n==============\\r\\n1.\\r\\n+ <para>\\r\\n+ Now the bidirectional logical replication setup is complete between\\r\\n+ <literal>node1</literal>, <literal>node2</literal> and\\r\\n+ <literal>node2</literal>. Any subsequent changes in one node will\\r\\n+ replicate the changes to the other nodes.\\r\\n+ </para>\\r\\n\\r\\nI think \\\"node1, node2 and node2\\\" should be \\\"node1, node2 and node3\\\".\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Mon, 6 Jun 2022 08:58:57 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 6, 2022 at 2:29 PM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Mon, Jun 6, 2022 1:14 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > The attached v18 patch has the fixes for the same.\\n> >\\n>\\n> Thanks for updating the patch, here are some comments.\\n>\\n> 0002 patch\\n> ==============\\n> 1.\\n> + <varlistentry>\\n> + <term><literal>origin</literal> (<type>string</type>)</term>\\n> + <listitem>\\n> + <para>\\n>\\n> It maybe better if the type of \\\"origin\\\" parameter is enum, as it cannot be any\\n> string and only has two valid values.\\n\\nCurrently we only support local and any. But this was designed so that\\nit can be extended to support origin names. Users can provide a\\nparticular origin name to be filtered.\\nThe same was also discussed in pg unconference as mentioned in [1]\\n\\n> 2.\\n> @@ -607,6 +626,11 @@ CreateSubscription(ParseState *pstate, CreateSubscriptionStmt *stmt,\\n> LOGICALREP_TWOPHASE_STATE_PENDING :\\n> LOGICALREP_TWOPHASE_STATE_DISABLED);\\n> values[Anum_pg_subscription_subdisableonerr - 1] = BoolGetDatum(opts.disableonerr);\\n> + if (opts.origin)\\n> + values[Anum_pg_subscription_suborigin - 1] =\\n> + CStringGetTextDatum(opts.origin);\\n> + else\\n> + nulls[Anum_pg_subscription_suborigin - 1] = true;\\n> values[Anum_pg_subscription_subconninfo - 1] =\\n> CStringGetTextDatum(conninfo);\\n> if (opts.slot_name)\\n>\\n> Document of \\\"CREATE SUBSCRIPTION\\\" says, the default value of \\\"origin\\\" is \\\"any\\\", so why not set\\n> suborigin to \\\"any\\\" when user doesn't specify this parameter?\\n\\nModified\\n\\n>\\n> 0003 patch\\n> ==============\\n> 1.\\n> @@ -300,6 +310,11 @@ CREATE SUBSCRIPTION <replaceable class=\\\"parameter\\\">subscription_name</replaceabl\\n> replication from the publisher. The default is\\n> <literal>false</literal>.\\n> </para>\\n> + <para>\\n> + There is some interaction between the <literal>origin</literal>\\n> + parameter and <literal>copy_data</literal> parameter. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n> + </para>\\n> </listitem>\\n> </varlistentry>\\n>\\n> I think this change should be put together with \\\"origin\\\" parameter, instead of\\n> \\\"disable_on_error\\\".\\n\\nModified\\n\\n>\\n> 0004 patch\\n> ==============\\n> 1.\\n> + <para>\\n> + Now the bidirectional logical replication setup is complete between\\n> + <literal>node1</literal>, <literal>node2</literal> and\\n> + <literal>node2</literal>. Any subsequent changes in one node will\\n> + replicate the changes to the other nodes.\\n> + </para>\\n>\\n> I think \\\"node1, node2 and node2\\\" should be \\\"node1, node2 and node3\\\".\\n\\nModified\\n\\nThanks for the comments, the attached v19 patch has the changes for the same.\\n[1] - https://wiki.postgresql.org/wiki/PgCon_2022_Developer_Unconference#Logical_Replication_Origin_Filtering_and_Consistency\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 8 Jun 2022 16:21:52 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Below are some review comments for the patch v18-0003\\n\\n1. Commit message\\n\\nThis patch does a couple of things:\\nchange 1) Checks and throws an error if the publication tables were also\\nsubscribing data in the publisher from other publishers when copy_data\\nparameter is specified as 'ON' and origin parameter is specified as\\n'local'.\\nchange 2) Adds force value for copy_data parameter.\\n\\nSUGGESTION\\nThis patch does a couple of things:\\nchange 1) Checks and throws an error if 'copy_data = on' and 'origin =\\nlocal' but the publication tables were also subscribing data in the\\npublisher from other publishers.\\nchange 2) Adds 'force' value for copy_data parameter.\\n\\n~~~\\n\\n2. Commit message - about the example\\n\\nAll my following review comments for the commit message are assuming\\nthat the example steps are as they are written in the patch, but\\nactually I felt that the example might be more complex than it needed\\nto be: e.g\\n- You don’t really need the node2 to have data\\n- Therefore you don’t need all the added TRUNCATE complications\\n\\nE.g. I think you only need node1 (with data) and node2 (no data).\\n\\nThen node1 subscribes node2 with (origin=local, copy_data=off).\\nThen node2 subscribes node1 with (origin=local, copy_data=on).\\n- Demonstrates exception happens because node1 already had a subscription\\n- Demonstrates need for the copy_data=force to override that exception\\n\\nSo please consider using a simpler example for this commit message\\n\\n~~~\\n\\n3. Commit message\\n\\nThe below help us understand how the first change will be useful:\\n\\nIf copy_data parameter was used with 'on' in step 5, then an error\\nwill be thrown\\nto alert the user to prevent inconsistent data being populated:\\nCREATE SUBSCRIPTION sub_node2_node1 CONNECTION '<node1 details>'\\nPUBLICATION pub_node1 WITH (copy_data = on, origin = local);\\nERROR: CREATE/ALTER SUBSCRIPTION with origin and copy_data as true is not\\nallowed when the publisher might have replicated data\\n\\nSUGGESTION\\nThe steps below help to demonstrate how the new exception is useful:\\n\\nThe initial copy phase has no way to know the origin of the row data,\\nso if 'copy_data = on' in the step 5 below, then an error will be\\nthrown to prevent any potentially non-local data from being copied:\\n<blank line>\\ne.g\\nCREATE SUBSCRIPTION ...\\n\\n~~~\\n\\n4. Commit message\\n\\nThe following will help us understand how the second change will be useful:\\nLet's take a simple case where user is trying to setup bidirectional logical\\nreplication between node1 and node2 where the two nodes have some pre-existing\\ndata like below:\\nnode1: Table t1 (c1 int) has data 11, 12, 13, 14\\nnode2: Table t1 (c1 int) has data 21, 22, 23, 24\\n\\nSUGGESTION\\nThe following steps help to demonstrate how the 'copy_data = force'\\nchange will be useful:\\n<blank line>\\nLet's take a scenario where the user wants to set up bidirectional\\nlogical replication between node1 and node2 where the same table on\\neach node has pre-existing data. e.g.\\nnode1: Table t1 (c1 int) has data 11, 12, 13, 14\\nnode2: Table t1 (c1 int) has data 21, 22, 23, 24\\n\\n~~~\\n\\n5. Commit message\\n\\nstep 4:\\nnode2=# CREATE SUBSCRIPTION sub_node2_node1 Connection '<node1 details>'\\nnode2-# PUBLICATION pub_node1;\\nCREATE SUBSCRIPTION\\n\\n\\\"Connection\\\" => \\\"CONNECTION\\\"\\n\\n~~~\\n\\n6. Commit message\\n\\nIf table t1 has a unique key, it will cause a unique key\\nviolation and replication won't proceed.\\n\\nSUGGESTION\\nIn case, table t1 has a unique key, it will lead to a unique key\\nviolation and replication won't proceed.\\n\\n~~~\\n\\n7. Commit message\\n\\nstep 3: Create a subscription in node1 to subscribe to node2. Use\\ncopy_data specified as on so that the existing table data is copied during\\ninitial sync:\\n\\nSUGGESTION\\nstep 3: Create a subscription in node1 to subscribe to node2. Use\\n'copy_data = on' so that the existing table data is copied during\\ninitial sync:\\n\\n~~~\\n\\n8. Commit message\\n\\nstep 4: Adjust the publication publish settings so that truncate is not\\npublished to the subscribers and truncate the table data in node2:\\n\\nSUGGESTION (only added a comma)\\nstep 4: Adjust the publication publish settings so that truncate is\\nnot published to the subscribers, and truncate the table data in\\nnode2:\\n\\n~~~\\n\\n9. Commit message\\n\\nstep 5: Create a subscription in node2 to subscribe to node1. Use copy_data\\nspecified as force when creating a subscription to node1 so that the existing\\ntable data is copied during initial sync:\\n\\nSUGGESTION\\nstep 5: Create a subscription in node2 to subscribe to node1. Use\\n'copy_data = force' when creating a subscription to node1 so that the\\nexisting table data is copied during initial sync:\\n\\n======\\n\\n10. doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n can have non-existent publications.\\n </para>\\n\\n+ <para>\\n+ If subscription is created with <literal>origin = local</literal> and\\n+ <literal>copy_data = on</literal>, it will check if the publisher tables are\\n+ being subscribed to any other publisher and throw an error to prevent\\n+ inconsistent data in the subscription. The user can continue with the copy\\n+ operation without throwing any error in this case by specifying\\n+ <literal>copy_data = force</literal>.\\n+ </para>\\n\\nSUGGESTION (minor rewording)\\nIf the subscription is created with <literal>origin = local</literal>\\nand <literal>copy_data = on</literal>, it will check if the publisher\\ntables are being subscribed to any other publisher and, if so, then\\nthrow an error to prevent possible non-local data from being copied.\\nThe user can override this check and continue with the copy operation\\nby specifying <literal>copy_data = force</literal>.\\n\\n======\\n\\n11. src/backend/commands/subscriptioncmds.c - parse_subscription_options\\n\\n From [1]:\\n>> What about also allowing copy_data = 2, and making it equivalent to \\\"force\\\"?\\n> Vignesh: I felt the existing looks ok, no need to support 2. It might confuse the user.\\n\\nI don't think it would be confusing, but I also don't feel strongly\\nenough to debate it. Anyway, I could not find a similar precedent, so\\nyour decision is fine.\\n\\n~~~\\n\\n12. src/backend/commands/subscriptioncmds.c - parse_subscription_options\\n\\n@@ -339,17 +406,16 @@ parse_subscription_options(ParseState *pstate,\\nList *stmt_options,\\n errmsg(\\\"%s and %s are mutually exclusive options\\\",\\n \\\"connect = false\\\", \\\"create_slot = true\\\")));\\n\\n- if (opts->copy_data &&\\n- IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\\n+ if (opts->copy_data && IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\\n ereport(ERROR,\\n\\nThis is just a formatting change. Is it needed for this patch? patch.\\n\\n~~~\\n\\n13. src/backend/commands/subscriptioncmds.c - AlterSubscription_refresh\\n\\n@@ -730,7 +798,8 @@ CreateSubscription(ParseState *pstate,\\nCreateSubscriptionStmt *stmt,\\n * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\\n * PUBLICATION to work.\\n */\\n- if (opts.twophase && !opts.copy_data && tables != NIL)\\n+ if (opts.twophase && opts.copy_data == COPY_DATA_OFF &&\\n+ tables != NIL)\\n twophase_enabled = true;\\n\\nWhy is this change needed? I thought the original code is OK now since\\nCOPY_DATA_OFF = 0\\n\\n~~~\\n\\n14. src/backend/commands/subscriptioncmds.c - AlterSubscription\\n\\n@@ -1265,7 +1337,8 @@ AlterSubscription(ParseState *pstate,\\nAlterSubscriptionStmt *stmt,\\n *\\n * For more details see comments atop worker.c.\\n */\\n- if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\n+ if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\\n+ opts.copy_data)\\n\\nThis is just a formatting change. Is it needed for this patch?\\n\\n~~~\\n\\n15. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n\\n+\\n+ if (!origin || (strcmp(origin, \\\"local\\\") != 0) || copydata != COPY_DATA_ON)\\n+ return;\\n+\\n\\nThis condition could be rearranged to put the strcmp last so it is not\\ncalled unless absolutely necessary.\\n\\n~~~\\n\\n16. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n\\n+ appendStringInfoString(&cmd,\\n+ \\\"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\\nPS.srrelid as replicated\\\\n\\\"\\n+ \\\"FROM pg_publication P,\\\\n\\\"\\n\\nThe line is too long; needs wrapping.\\n\\n~~~\\n\\n17. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n\\n+ if (!slot_attisnull(slot, 3))\\n+ ereport(ERROR,\\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n+ errmsg(\\\"table:%s.%s might have replicated data in the publisher\\\",\\n+ nspname, relname),\\n+ errdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin as local and\\ncopy_data as true is not allowed when the publisher might have\\nreplicated data.\\\"),\\n+ errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force.\\\"));\\n\\nI felt the errmsg may be easier to read using \\\"=\\\" instead of \\\"as\\\".\\nAnyway, it would be more consistent with the errhint. Also, change the\\n\\\"true\\\" to \\\"on\\\" to be consistent with the errhint.\\n\\nSUGGESTION\\nerrdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data\\n= on is not allowed when the publisher might have replicated data.\\\"),\\n\\n\\n------\\n[1] https://www.postgresql.org/message-id/CALDaNm3iLLxP4OV%2ByQHs-c71P6zQ9W8D30DGsve1SQs_1pFsSQ%40mail.gmail.com\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 10 Jun 2022 14:53:27 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Below are some review comments for the patch v18-0004\\n\\n======\\n\\n1. Commit message\\n\\nDocument the steps for the following:\\na) Create a two-node bidirectional replication when there is no data in both the\\nnodes.\\nb) Adding a new node when there is no data in any of the nodes.\\nc) Adding a new node when data is present in the existing nodes.\\nd) Generic steps to add a new node to the existing set of nodes.\\n\\nSUGGESTION (minor changes to make the tense consistent)\\nDocuments the steps for the following:\\na) Creating a two-node bidirectional replication when there is no data\\nin both nodes.\\nb) Adding a new node when there is no data in any of the nodes.\\nc) Adding a new node when data is present in the existing nodes.\\nd) Generic steps for adding a new node to an existing set of nodes.\\n\\n======\\n\\n2. doc/src/sgml/logical-replication.sgml - blurb\\n\\n+ <para>\\n+ Bidirectional replication is useful in creating a multi-master database\\n+ which helps in performing read/write operations from any of the nodes.\\n\\nSUGGESTION\\nBidirectional replication is useful for creating a multi-master\\ndatabase environment for replicating read/write operations performed\\nby any of the member nodes.\\n\\n~~~\\n\\n3. doc/src/sgml/logical-replication.sgml - warning\\n\\n+ <warning>\\n+ <para>\\n+ Setting up bidirectional logical replication across nodes\\nrequires multiple\\n+ steps to be performed on various nodes. Because not all operations are\\n+ transactional, the user is advised to take backups.\\n+ </para>\\n+ </warning>\\n\\nSUGGESTION (but keep your formatting)\\nSetting up bidirectional logical replication requires multiple steps\\nto be performed on various nodes. Because...\\n\\n~~~\\n\\n4. doc/src/sgml/logical-replication.sgml - Setting bidirectional\\nreplication between two nodes\\n\\n+ <para>\\n+ The steps to create a two-node bidirectional replication when there is no\\n+ data in both the nodes <literal>node1</literal> and\\n+ <literal>node2</literal> are given below:\\n+ </para>\\n\\nSUGGESTION (but keep your formatting)\\nThe following steps demonstrate how to create a two-node bidirectional\\nreplication when there is no table data present in both nodes node1\\nand node2:\\n\\n~~~\\n\\n5. doc/src/sgml/logical-replication.sgml - Setting bidirectional\\nreplication between two nodes\\n\\n+ <para>\\n+ Lock the required tables of <literal>node1</literal> and\\n+ <literal>node2</literal> in <literal>EXCLUSIVE</literal> mode until the\\n+ setup is completed.\\n+ </para>\\n\\nInstead of \\\"the required tables\\\" shouldn't this just say \\\"table t1\\\"?\\n\\n~~~\\n\\n6. doc/src/sgml/logical-replication.sgml - Setting bidirectional\\nreplication between two nodes\\n\\n+ <para>\\n+ Now the bidirectional logical replication setup is complete between\\n+ <literal>node1</literal>, <literal>node2</literal> and\\n+ <literal>node2</literal>. Any subsequent changes in one node will\\n+ replicate the changes to the other nodes.\\n+ </para>\\n\\nSUGGESTION (for 2nd sentence, and keep your formatting)\\nAny incremental changes from node1 will be replicated to node2, and\\nany incremental changes from node2 will be replicated to node1.\\n\\n~~~\\n\\n7. doc/src/sgml/logical-replication.sgml - Adding a new node when\\nthere is no data in any of the nodes\\n\\n+ <para>\\n+ Adding a new node <literal>node3</literal> to the existing\\n+ <literal>node1</literal> and <literal>node2</literal> when there is no data\\n+ in any of the nodes requires setting up subscription in\\n+ <literal>node1</literal> and <literal>node2</literal> to replicate the data\\n+ from <literal>node3</literal> and setting up subscription in\\n+ <literal>node3</literal> to replicate data from <literal>node1</literal>\\n+ and <literal>node2</literal>.\\n+ </para>\\n\\nSUGGESTION (but keep your formatting)\\nThe following steps demonstrate adding a new node node3 to the\\nexisting node1 and node2 when there is no t1 data in any of the nodes.\\nThis requires creating subscriptions in node1 and node2 to replicate\\nthe data from node3 and creating subscriptions in node3 to replicate\\ndata from node1 and node2.\\n\\n~~~\\n\\n8. doc/src/sgml/logical-replication.sgml - Adding a new node when\\nthere is no data in any of the nodes\\n\\n+ <note>\\n+ <para>\\n+ It is assumed that the bidirectional logical replication between\\n+ <literal>node1</literal> and <literal>node2</literal> is already\\ncompleted.\\n+ </para>\\n+ </note>\\n\\nIMO this note should just be a text note in the previous paragraph\\ninstead of an SGML <note>. e.g. \\\"Note: These steps assume that...\\\"\\n\\n~~~\\n\\n9. doc/src/sgml/logical-replication.sgml - Adding a new node when\\nthere is no data in any of the nodes\\n\\n+ <para>\\n+ Lock the required tables of all the nodes <literal>node1</literal>,\\n+ <literal>node2</literal> and <literal>node3</literal> in\\n+ <literal>EXCLUSIVE</literal> mode until the setup is completed.\\n+ </para>\\n\\nInstead of \\\"the required tables\\\" shouldn't this just say \\\"table t1\\\"?\\n\\n~~~\\n\\n10. doc/src/sgml/logical-replication.sgml - Adding a new node when\\nthere is no data in any of the nodes\\n\\n+ <para>\\n+ Now the bidirectional logical replication setup is complete between\\n+ <literal>node1</literal>, <literal>node2</literal> and\\n+ <literal>node2</literal>. Any subsequent changes in one node will\\n+ replicate the changes to the other nodes.\\n+ </para>\\n\\nSUGGESTION (2nd sentence)\\nIncremental changes made in any node will be replicated to the other two nodes.\\n\\n~~~\\n\\n11. doc/src/sgml/logical-replication.sgml - Adding a new node when\\ndata is present in the existing nodes\\n\\n+ <para>\\n+ Adding a new node <literal>node3</literal> which has no data to the\\n+ existing <literal>node1</literal> and <literal>node2</literal> when data\\n+ is present in existing nodes <literal>node1</literal> and\\n+ <literal>node2</literal> needs similar steps. The only change required\\n+ here is that <literal>node3</literal> should create a subscription with\\n+ <literal>copy_data = force</literal> to one of the existing nodes to\\n+ receive the existing data during initial data synchronization.\\n+ </para>\\n\\nSUGGESTION (but keep your formatting)\\nThe following steps demonstrate adding a new node node3 which has no\\nt1 data to the existing node1 and node2 where t1 data is present. This\\nneeds similar steps; the only change required here is that node3\\nshould create a subscription with copy_data = force to one of the\\nexisting nodes so it can receive the existing t1 data during initial\\ndata synchronization.\\n\\n~~~\\n\\n12 doc/src/sgml/logical-replication.sgml - Adding a new node when data\\nis present in the existing nodes\\n\\n+ <note>\\n+ <para>\\n+ It is assumed that the bidirectional logical replication between\\n+ <literal>node1</literal> and <literal>node2</literal> is already\\ncompleted.\\n+ The nodes <literal>node1</literal> and <literal>node2</literal> has some\\n+ pre-existing data in table t1 that is synchronized in both the nodes.\\n+ </para>\\n+ </note>\\n\\n12a.\\nIMO this note should just be text in the previous paragraph instead of\\nan SGML <note>. e.g. \\\"Note: These steps assume that...\\\"\\n\\n12b.\\nSUGGESTION (minor rewording; keep your formatting)\\nNote: These steps assume that the bidirectional logical replication\\nbetween node1 and node2 is already completed, and the pre-existing\\ndata in table t1 is already synchronized in both those nodes.\\n\\n~~~\\n\\n13. doc/src/sgml/logical-replication.sgml - Adding a new node when\\ndata is present in the existing nodes\\n\\n+ <para>\\n+ Lock the required tables of <literal>node2</literal> and\\n+ <literal>node3</literal> in <literal>EXCLUSIVE</literal> mode until the\\n+ setup is completed. No need to lock the tables in <literal>node1</literal>\\n+ as any data changes made will be synchronized while creating the\\n+ subscription with <literal>copy_data</literal> specified as\\n+ <literal>force</literal>.\\n+ </para>\\n\\n13a.\\nInstead of \\\"the required tables\\\" shouldn't this just say \\\"table t1\\\"?\\n\\n13b.\\nSUGGESTION (2nd sentence; keep your formatting)\\nThere is no need to lock table t1 in node1 because any data changes\\nmade will be synchronized while creating the subscription with\\ncopy_data = force.\\n\\n~~~\\n\\n14. doc/src/sgml/logical-replication.sgml - Adding a new node when\\ndata is present in the existing nodes\\n\\n+ <para>\\n+ Create a subscription in <literal>node3</literal> to subscribe to\\n+ <literal>node1</literal>. Use <literal>copy_data</literal> specified as\\n+ <literal>force</literal> so that the existing table data is\\n+ copied during initial sync:\\n\\nSUGGESTION (2nd sentence; keep your formatting)\\nUse copy_data = force so that the existing table data is copied during\\nthe initial sync:\\n\\n~~~\\n\\n15. doc/src/sgml/logical-replication.sgml - Adding a new node when\\ndata is present in the existing nodes\\n\\n+ <para>\\n+ Create a subscription in <literal>node3</literal> to subscribe to\\n+ <literal>node2</literal>. Use <literal>copy_data</literal> specified as\\n+ <literal>off</literal> because the initial table data would have been\\n+ already copied in the previous step:\\n\\nSUGGESTION (2nd sentence; keep your formatting)\\nUse copy_data = off because the initial table data would have been\\nalready copied in the previous step:\\n\\n~~~\\n\\n16. doc/src/sgml/logical-replication.sgml - Adding a new node when\\ndata is present in the existing nodes\\n\\n+ <para>\\n+ Now the bidirectional logical replication setup is complete between\\n+ <literal>node1</literal>, <literal>node2</literal> and\\n+ <literal>node2</literal>. Any subsequent changes in one node will\\n+ replicate the changes to the other nodes.\\n+ </para>\\n\\n16a.\\nShould say node1, node2 and node3\\n\\n16b.\\nSUGGESTION (2nd sentence – same as the previous comment)\\nIncremental changes made in any node will be replicated to the other two nodes.\\n\\n~~~\\n\\n17. doc/src/sgml/logical-replication.sgml - Adding a new node when\\ndata is present in the new node\\n\\n+ <warning>\\n+ <para>\\n+ Adding a new node <literal>node3</literal> to the existing\\n+ <literal>node1</literal> and <literal>node2</literal> when data is present\\n+ in the new node <literal>node3</literal> is not possible.\\n+ </para>\\n+ </warning>\\n\\n17a.\\nIMO\\n- Not really necessary to name the nodes, because this is a generic statement\\n- Maybe say \\\"not supported\\\" instead of \\\"not possible\\\". e.g. there is\\nno ERROR check for this case is there?\\n\\nSUGGESTION\\nAdding a new node when data is present in the new node tables is not supported.\\n\\n17b.\\nI am not sure but I felt this advice seemed more like an SGML <note>;\\nnote a <warning>\\n\\n~~~\\n\\n18. doc/src/sgml/logical-replication.sgml - Generic steps to add a new\\nnode to the existing set of nodes\\n\\n+ <title>Generic steps to add a new node to the existing set of nodes</title>\\n\\nSUGGESTION\\nGeneric steps for adding a new node to an existing set of nodes\\n\\n~~~\\n\\n19. doc/src/sgml/logical-replication.sgml - Generic steps to add a new\\nnode to the existing set of nodes\\n\\n+ <para>\\n+ 1. Create the required publication on the new node.\\n+ </para>\\n+ <para>\\n+ 2. Lock the required tables of the new node in <literal>EXCLUSIVE</literal>\\n+ mode until the setup is complete. This is required to prevent any\\n+ modifications from happening in the new node. If data modifications occur\\n+ after step-3, there is a chance that the modifications will be published to\\n+ the first node and then synchronized back to the new node while creating\\n+ subscription in step-5 resulting in inconsistent data.\\n+ </para>\\n+ <para>\\n+ 3. Create subscriptions on existing nodes pointing to publication on\\n+ the new node with <literal>origin</literal> parameter specified as\\n+ <literal>local</literal> and <literal>copy_data</literal> specified as\\n+ <literal>off</literal>.\\n+ </para>\\n+ <para>\\n+ 4. Lock the required tables of the existing nodes except the first node in\\n+ <literal>EXCLUSIVE</literal> mode until the setup is complete. This is\\n+ required to prevent any modifications from happening. If data modifications\\n+ occur, there is a chance that the modifications done in between step-5 and\\n+ step-6 will not be synchronized to the new node resulting in inconsistent\\n+ data. No need to lock the tables in the first node as any data changes\\n+ made will be synchronized while creating the subscription with\\n+ <literal>copy_data</literal> specified as <literal>force</literal>.\\n+ </para>\\n+ <para>\\n+ 5. Create a subscription on the new node pointing to publication on the\\n+ first node with <literal>origin</literal> parameter specified as\\n+ <literal>local</literal> and <literal>copy_data</literal> parameter\\n+ specified as <literal>force</literal>.\\n+ </para>\\n+ <para>\\n+ 6. Create subscriptions on the new node pointing to publications on the\\n+ remaining nodes with <literal>origin</literal> parameter specified as\\n+ <literal>local</literal> and <literal>copy_data</literal> parameter\\n+ specifiedas <literal>off</literal>.\\n+ </para>\\n\\nFollowing suggestions make the following changes:\\n- Some minor rewording\\n- Change the step names\\n- Use \\\"=\\\" instead of \\\"specified as\\\" for the parameters\\n- I felt it is more readable if the explanatory notes are separate\\nparagraphs from the steps. Maybe if you could indent them or something\\nit would be even better.\\n- Added a couple more explanatory notes\\n\\nSUGGESTION (using those above changes; please keep your formatting)\\n\\nStep-1: Create a publication on the new node.\\n\\nStep-2: Lock the required tables of the new node in EXCLUSIVE mode\\nuntil the setup is complete.\\n\\n(This lock is necessary to prevent any modifications from happening in\\nthe new node because if data modifications occurred after Step-3,\\nthere is a chance that the modifications will be published to the\\nfirst node and then synchronized back to the new node while creating\\nthe subscription in Step-5. This would result in inconsistent data).\\n\\nStep-3. Create subscriptions on existing nodes to the publication on\\nthe new node with origin = local and copy_data = off.\\n\\n(The copy_data = off is OK here because it is asserted that the\\npublished tables of the new node will have no pre-existing data).\\n\\nStep-4. Lock the required tables of the existing nodes except the\\nfirst node in EXCLUSIVE mode until the setup is complete.\\n\\n(This lock is necessary to prevent any modifications from happening.\\nIf data modifications occur, there is a chance that modifications done\\nbetween Step-5 and Step-6 will not be synchronized to the new node.\\nThis would result in inconsistent data. There is no need to lock table\\nt1 in node1 because any data changes made will be synchronized while\\ncreating the subscription with copy_data = force).\\n\\nStep-5. Create a subscription on the new node to the publication on\\nthe first node with origin = local and copy_data = force.\\n\\n(This will copy the same table data from the existing nodes to the new node)\\n\\nStep-6. Create subscriptions on the new node to publications on the\\nremaining nodes with origin = local and copy_data = off.\\n\\n(The copy_data = off is OK here because the existing node data was\\nalready copied to the new node in Step-5)\\n\\n======\\n\\n20. doc/src/sgml/ref/create_subscription.sgml\\n\\n- <literal>copy_data = force</literal>.\\n+ <literal>copy_data = force</literal>. Refer to the\\n+ <xref linkend=\\\"logical-replication-bidirectional\\\"/> on how\\n+ <literal>copy_data</literal> and <literal>origin</literal> can be used\\n+ in bidirectional replication.\\n </para>\\n\\nSUGGESTION (keep your formatting)\\nRefer to <xref linkend=\\\"logical-replication-bidirectional\\\"/> for how\\n<literal>copy_data</literal> and <literal>origin</literal> can be used\\nin bidirectional replication.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 10 Jun 2022 18:39:04 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jun 10, 2022 at 10:23 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Below are some review comments for the patch v18-0003\\n>\\n> 1. Commit message\\n>\\n> This patch does a couple of things:\\n> change 1) Checks and throws an error if the publication tables were also\\n> subscribing data in the publisher from other publishers when copy_data\\n> parameter is specified as 'ON' and origin parameter is specified as\\n> 'local'.\\n> change 2) Adds force value for copy_data parameter.\\n>\\n> SUGGESTION\\n> This patch does a couple of things:\\n> change 1) Checks and throws an error if 'copy_data = on' and 'origin =\\n> local' but the publication tables were also subscribing data in the\\n> publisher from other publishers.\\n> change 2) Adds 'force' value for copy_data parameter.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2. Commit message - about the example\\n>\\n> All my following review comments for the commit message are assuming\\n> that the example steps are as they are written in the patch, but\\n> actually I felt that the example might be more complex than it needed\\n> to be: e.g\\n> - You don’t really need the node2 to have data\\n> - Therefore you don’t need all the added TRUNCATE complications\\n>\\n> E.g. I think you only need node1 (with data) and node2 (no data).\\n>\\n> Then node1 subscribes node2 with (origin=local, copy_data=off).\\n> Then node2 subscribes node1 with (origin=local, copy_data=on).\\n> - Demonstrates exception happens because node1 already had a subscription\\n> - Demonstrates need for the copy_data=force to override that exception\\n>\\n> So please consider using a simpler example for this commit message\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. Commit message\\n>\\n> The below help us understand how the first change will be useful:\\n>\\n> If copy_data parameter was used with 'on' in step 5, then an error\\n> will be thrown\\n> to alert the user to prevent inconsistent data being populated:\\n> CREATE SUBSCRIPTION sub_node2_node1 CONNECTION '<node1 details>'\\n> PUBLICATION pub_node1 WITH (copy_data = on, origin = local);\\n> ERROR: CREATE/ALTER SUBSCRIPTION with origin and copy_data as true is not\\n> allowed when the publisher might have replicated data\\n>\\n> SUGGESTION\\n> The steps below help to demonstrate how the new exception is useful:\\n>\\n> The initial copy phase has no way to know the origin of the row data,\\n> so if 'copy_data = on' in the step 5 below, then an error will be\\n> thrown to prevent any potentially non-local data from being copied:\\n> <blank line>\\n> e.g\\n> CREATE SUBSCRIPTION ...\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. Commit message\\n>\\n> The following will help us understand how the second change will be useful:\\n> Let's take a simple case where user is trying to setup bidirectional logical\\n> replication between node1 and node2 where the two nodes have some pre-existing\\n> data like below:\\n> node1: Table t1 (c1 int) has data 11, 12, 13, 14\\n> node2: Table t1 (c1 int) has data 21, 22, 23, 24\\n>\\n> SUGGESTION\\n> The following steps help to demonstrate how the 'copy_data = force'\\n> change will be useful:\\n> <blank line>\\n> Let's take a scenario where the user wants to set up bidirectional\\n> logical replication between node1 and node2 where the same table on\\n> each node has pre-existing data. e.g.\\n> node1: Table t1 (c1 int) has data 11, 12, 13, 14\\n> node2: Table t1 (c1 int) has data 21, 22, 23, 24\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. Commit message\\n>\\n> step 4:\\n> node2=# CREATE SUBSCRIPTION sub_node2_node1 Connection '<node1 details>'\\n> node2-# PUBLICATION pub_node1;\\n> CREATE SUBSCRIPTION\\n>\\n> \\\"Connection\\\" => \\\"CONNECTION\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 6. Commit message\\n>\\n> If table t1 has a unique key, it will cause a unique key\\n> violation and replication won't proceed.\\n>\\n> SUGGESTION\\n> In case, table t1 has a unique key, it will lead to a unique key\\n> violation and replication won't proceed.\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. Commit message\\n>\\n> step 3: Create a subscription in node1 to subscribe to node2. Use\\n> copy_data specified as on so that the existing table data is copied during\\n> initial sync:\\n>\\n> SUGGESTION\\n> step 3: Create a subscription in node1 to subscribe to node2. Use\\n> 'copy_data = on' so that the existing table data is copied during\\n> initial sync:\\n\\nModified\\n\\n> ~~~\\n>\\n> 8. Commit message\\n>\\n> step 4: Adjust the publication publish settings so that truncate is not\\n> published to the subscribers and truncate the table data in node2:\\n>\\n> SUGGESTION (only added a comma)\\n> step 4: Adjust the publication publish settings so that truncate is\\n> not published to the subscribers, and truncate the table data in\\n> node2:\\n\\nThis content is not required any more, I have removed it.\\n\\n> ~~~\\n>\\n> 9. Commit message\\n>\\n> step 5: Create a subscription in node2 to subscribe to node1. Use copy_data\\n> specified as force when creating a subscription to node1 so that the existing\\n> table data is copied during initial sync:\\n>\\n> SUGGESTION\\n> step 5: Create a subscription in node2 to subscribe to node1. Use\\n> 'copy_data = force' when creating a subscription to node1 so that the\\n> existing table data is copied during initial sync:\\n\\nModified\\n\\n> ======\\n>\\n> 10. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> can have non-existent publications.\\n> </para>\\n>\\n> + <para>\\n> + If subscription is created with <literal>origin = local</literal> and\\n> + <literal>copy_data = on</literal>, it will check if the publisher tables are\\n> + being subscribed to any other publisher and throw an error to prevent\\n> + inconsistent data in the subscription. The user can continue with the copy\\n> + operation without throwing any error in this case by specifying\\n> + <literal>copy_data = force</literal>.\\n> + </para>\\n>\\n> SUGGESTION (minor rewording)\\n> If the subscription is created with <literal>origin = local</literal>\\n> and <literal>copy_data = on</literal>, it will check if the publisher\\n> tables are being subscribed to any other publisher and, if so, then\\n> throw an error to prevent possible non-local data from being copied.\\n> The user can override this check and continue with the copy operation\\n> by specifying <literal>copy_data = force</literal>.\\n\\nModified\\n\\n> ======\\n>\\n> 11. src/backend/commands/subscriptioncmds.c - parse_subscription_options\\n>\\n> From [1]:\\n> >> What about also allowing copy_data = 2, and making it equivalent to \\\"force\\\"?\\n> > Vignesh: I felt the existing looks ok, no need to support 2. It might confuse the user.\\n>\\n> I don't think it would be confusing, but I also don't feel strongly\\n> enough to debate it. Anyway, I could not find a similar precedent, so\\n> your decision is fine.\\n\\nOk\\n\\n> ~~~\\n>\\n> 12. src/backend/commands/subscriptioncmds.c - parse_subscription_options\\n>\\n> @@ -339,17 +406,16 @@ parse_subscription_options(ParseState *pstate,\\n> List *stmt_options,\\n> errmsg(\\\"%s and %s are mutually exclusive options\\\",\\n> \\\"connect = false\\\", \\\"create_slot = true\\\")));\\n>\\n> - if (opts->copy_data &&\\n> - IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\\n> + if (opts->copy_data && IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\\n> ereport(ERROR,\\n>\\n> This is just a formatting change. Is it needed for this patch? patch.\\n\\nModified\\n\\n> ~~~\\n>\\n> 13. src/backend/commands/subscriptioncmds.c - AlterSubscription_refresh\\n>\\n> @@ -730,7 +798,8 @@ CreateSubscription(ParseState *pstate,\\n> CreateSubscriptionStmt *stmt,\\n> * PENDING, to allow ALTER SUBSCRIPTION ... REFRESH\\n> * PUBLICATION to work.\\n> */\\n> - if (opts.twophase && !opts.copy_data && tables != NIL)\\n> + if (opts.twophase && opts.copy_data == COPY_DATA_OFF &&\\n> + tables != NIL)\\n> twophase_enabled = true;\\n>\\n> Why is this change needed? I thought the original code is OK now since\\n> COPY_DATA_OFF = 0\\n\\nModified\\n\\n> ~~~\\n>\\n> 14. src/backend/commands/subscriptioncmds.c - AlterSubscription\\n>\\n> @@ -1265,7 +1337,8 @@ AlterSubscription(ParseState *pstate,\\n> AlterSubscriptionStmt *stmt,\\n> *\\n> * For more details see comments atop worker.c.\\n> */\\n> - if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED && opts.copy_data)\\n> + if (sub->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED &&\\n> + opts.copy_data)\\n>\\n> This is just a formatting change. Is it needed for this patch?\\n\\nModified\\n\\n> ~~~\\n>\\n> 15. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> +\\n> + if (!origin || (strcmp(origin, \\\"local\\\") != 0) || copydata != COPY_DATA_ON)\\n> + return;\\n> +\\n>\\n> This condition could be rearranged to put the strcmp last so it is not\\n> called unless absolutely necessary.\\n\\nModified\\n\\n> ~~~\\n>\\n> 16. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> + appendStringInfoString(&cmd,\\n> + \\\"SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename,\\n> PS.srrelid as replicated\\\\n\\\"\\n> + \\\"FROM pg_publication P,\\\\n\\\"\\n>\\n> The line is too long; needs wrapping.\\n\\nModified\\n\\n> ~~~\\n>\\n> 17. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> + if (!slot_attisnull(slot, 3))\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"table:%s.%s might have replicated data in the publisher\\\",\\n> + nspname, relname),\\n> + errdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin as local and\\n> copy_data as true is not allowed when the publisher might have\\n> replicated data.\\\"),\\n> + errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force.\\\"));\\n>\\n> I felt the errmsg may be easier to read using \\\"=\\\" instead of \\\"as\\\".\\n> Anyway, it would be more consistent with the errhint. Also, change the\\n> \\\"true\\\" to \\\"on\\\" to be consistent with the errhint.\\n>\\n> SUGGESTION\\n> errdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data\\n> = on is not allowed when the publisher might have replicated data.\\\"),\\n\\nModified\\n\\nThanks for the comments, the attached v20 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 14 Jun 2022 17:34:52 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jun 10, 2022 at 2:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Below are some review comments for the patch v18-0004\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> Document the steps for the following:\\n> a) Create a two-node bidirectional replication when there is no data in both the\\n> nodes.\\n> b) Adding a new node when there is no data in any of the nodes.\\n> c) Adding a new node when data is present in the existing nodes.\\n> d) Generic steps to add a new node to the existing set of nodes.\\n>\\n> SUGGESTION (minor changes to make the tense consistent)\\n> Documents the steps for the following:\\n> a) Creating a two-node bidirectional replication when there is no data\\n> in both nodes.\\n> b) Adding a new node when there is no data in any of the nodes.\\n> c) Adding a new node when data is present in the existing nodes.\\n> d) Generic steps for adding a new node to an existing set of nodes.\\n\\nModified\\n\\n> ======\\n>\\n> 2. doc/src/sgml/logical-replication.sgml - blurb\\n>\\n> + <para>\\n> + Bidirectional replication is useful in creating a multi-master database\\n> + which helps in performing read/write operations from any of the nodes.\\n>\\n> SUGGESTION\\n> Bidirectional replication is useful for creating a multi-master\\n> database environment for replicating read/write operations performed\\n> by any of the member nodes.\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. doc/src/sgml/logical-replication.sgml - warning\\n>\\n> + <warning>\\n> + <para>\\n> + Setting up bidirectional logical replication across nodes\\n> requires multiple\\n> + steps to be performed on various nodes. Because not all operations are\\n> + transactional, the user is advised to take backups.\\n> + </para>\\n> + </warning>\\n>\\n> SUGGESTION (but keep your formatting)\\n> Setting up bidirectional logical replication requires multiple steps\\n> to be performed on various nodes. Because...\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. doc/src/sgml/logical-replication.sgml - Setting bidirectional\\n> replication between two nodes\\n>\\n> + <para>\\n> + The steps to create a two-node bidirectional replication when there is no\\n> + data in both the nodes <literal>node1</literal> and\\n> + <literal>node2</literal> are given below:\\n> + </para>\\n>\\n> SUGGESTION (but keep your formatting)\\n> The following steps demonstrate how to create a two-node bidirectional\\n> replication when there is no table data present in both nodes node1\\n> and node2:\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. doc/src/sgml/logical-replication.sgml - Setting bidirectional\\n> replication between two nodes\\n>\\n> + <para>\\n> + Lock the required tables of <literal>node1</literal> and\\n> + <literal>node2</literal> in <literal>EXCLUSIVE</literal> mode until the\\n> + setup is completed.\\n> + </para>\\n>\\n> Instead of \\\"the required tables\\\" shouldn't this just say \\\"table t1\\\"?\\n\\nModified\\n\\n> ~~~\\n>\\n> 6. doc/src/sgml/logical-replication.sgml - Setting bidirectional\\n> replication between two nodes\\n>\\n> + <para>\\n> + Now the bidirectional logical replication setup is complete between\\n> + <literal>node1</literal>, <literal>node2</literal> and\\n> + <literal>node2</literal>. Any subsequent changes in one node will\\n> + replicate the changes to the other nodes.\\n> + </para>\\n>\\n> SUGGESTION (for 2nd sentence, and keep your formatting)\\n> Any incremental changes from node1 will be replicated to node2, and\\n> any incremental changes from node2 will be replicated to node1.\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> there is no data in any of the nodes\\n>\\n> + <para>\\n> + Adding a new node <literal>node3</literal> to the existing\\n> + <literal>node1</literal> and <literal>node2</literal> when there is no data\\n> + in any of the nodes requires setting up subscription in\\n> + <literal>node1</literal> and <literal>node2</literal> to replicate the data\\n> + from <literal>node3</literal> and setting up subscription in\\n> + <literal>node3</literal> to replicate data from <literal>node1</literal>\\n> + and <literal>node2</literal>.\\n> + </para>\\n>\\n> SUGGESTION (but keep your formatting)\\n> The following steps demonstrate adding a new node node3 to the\\n> existing node1 and node2 when there is no t1 data in any of the nodes.\\n> This requires creating subscriptions in node1 and node2 to replicate\\n> the data from node3 and creating subscriptions in node3 to replicate\\n> data from node1 and node2.\\n\\nModified\\n\\n> ~~~\\n>\\n> 8. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> there is no data in any of the nodes\\n>\\n> + <note>\\n> + <para>\\n> + It is assumed that the bidirectional logical replication between\\n> + <literal>node1</literal> and <literal>node2</literal> is already\\n> completed.\\n> + </para>\\n> + </note>\\n>\\n> IMO this note should just be a text note in the previous paragraph\\n> instead of an SGML <note>. e.g. \\\"Note: These steps assume that...\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 9. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> there is no data in any of the nodes\\n>\\n> + <para>\\n> + Lock the required tables of all the nodes <literal>node1</literal>,\\n> + <literal>node2</literal> and <literal>node3</literal> in\\n> + <literal>EXCLUSIVE</literal> mode until the setup is completed.\\n> + </para>\\n>\\n> Instead of \\\"the required tables\\\" shouldn't this just say \\\"table t1\\\"?\\n\\nModified\\n\\n> ~~~\\n>\\n> 10. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> there is no data in any of the nodes\\n>\\n> + <para>\\n> + Now the bidirectional logical replication setup is complete between\\n> + <literal>node1</literal>, <literal>node2</literal> and\\n> + <literal>node2</literal>. Any subsequent changes in one node will\\n> + replicate the changes to the other nodes.\\n> + </para>\\n>\\n> SUGGESTION (2nd sentence)\\n> Incremental changes made in any node will be replicated to the other two nodes.\\n\\nModified\\n\\n> ~~~\\n>\\n> 11. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> data is present in the existing nodes\\n>\\n> + <para>\\n> + Adding a new node <literal>node3</literal> which has no data to the\\n> + existing <literal>node1</literal> and <literal>node2</literal> when data\\n> + is present in existing nodes <literal>node1</literal> and\\n> + <literal>node2</literal> needs similar steps. The only change required\\n> + here is that <literal>node3</literal> should create a subscription with\\n> + <literal>copy_data = force</literal> to one of the existing nodes to\\n> + receive the existing data during initial data synchronization.\\n> + </para>\\n>\\n> SUGGESTION (but keep your formatting)\\n> The following steps demonstrate adding a new node node3 which has no\\n> t1 data to the existing node1 and node2 where t1 data is present. This\\n> needs similar steps; the only change required here is that node3\\n> should create a subscription with copy_data = force to one of the\\n> existing nodes so it can receive the existing t1 data during initial\\n> data synchronization.\\n\\nModified\\n\\n> ~~~\\n>\\n> 12 doc/src/sgml/logical-replication.sgml - Adding a new node when data\\n> is present in the existing nodes\\n>\\n> + <note>\\n> + <para>\\n> + It is assumed that the bidirectional logical replication between\\n> + <literal>node1</literal> and <literal>node2</literal> is already\\n> completed.\\n> + The nodes <literal>node1</literal> and <literal>node2</literal> has some\\n> + pre-existing data in table t1 that is synchronized in both the nodes.\\n> + </para>\\n> + </note>\\n>\\n> 12a.\\n> IMO this note should just be text in the previous paragraph instead of\\n> an SGML <note>. e.g. \\\"Note: These steps assume that...\\\"\\n\\nModified\\n\\n> 12b.\\n> SUGGESTION (minor rewording; keep your formatting)\\n> Note: These steps assume that the bidirectional logical replication\\n> between node1 and node2 is already completed, and the pre-existing\\n> data in table t1 is already synchronized in both those nodes.\\n\\nModified\\n\\n> ~~~\\n>\\n> 13. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> data is present in the existing nodes\\n>\\n> + <para>\\n> + Lock the required tables of <literal>node2</literal> and\\n> + <literal>node3</literal> in <literal>EXCLUSIVE</literal> mode until the\\n> + setup is completed. No need to lock the tables in <literal>node1</literal>\\n> + as any data changes made will be synchronized while creating the\\n> + subscription with <literal>copy_data</literal> specified as\\n> + <literal>force</literal>.\\n> + </para>\\n>\\n> 13a.\\n> Instead of \\\"the required tables\\\" shouldn't this just say \\\"table t1\\\"?\\n\\nModified\\n\\n> 13b.\\n> SUGGESTION (2nd sentence; keep your formatting)\\n> There is no need to lock table t1 in node1 because any data changes\\n> made will be synchronized while creating the subscription with\\n> copy_data = force.\\n\\nModified\\n\\n> ~~~\\n>\\n> 14. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> data is present in the existing nodes\\n>\\n> + <para>\\n> + Create a subscription in <literal>node3</literal> to subscribe to\\n> + <literal>node1</literal>. Use <literal>copy_data</literal> specified as\\n> + <literal>force</literal> so that the existing table data is\\n> + copied during initial sync:\\n>\\n> SUGGESTION (2nd sentence; keep your formatting)\\n> Use copy_data = force so that the existing table data is copied during\\n> the initial sync:\\n\\nModified\\n\\n> ~~~\\n>\\n> 15. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> data is present in the existing nodes\\n>\\n> + <para>\\n> + Create a subscription in <literal>node3</literal> to subscribe to\\n> + <literal>node2</literal>. Use <literal>copy_data</literal> specified as\\n> + <literal>off</literal> because the initial table data would have been\\n> + already copied in the previous step:\\n>\\n> SUGGESTION (2nd sentence; keep your formatting)\\n> Use copy_data = off because the initial table data would have been\\n> already copied in the previous step:\\n\\nModified\\n\\n> ~~~\\n>\\n> 16. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> data is present in the existing nodes\\n>\\n> + <para>\\n> + Now the bidirectional logical replication setup is complete between\\n> + <literal>node1</literal>, <literal>node2</literal> and\\n> + <literal>node2</literal>. Any subsequent changes in one node will\\n> + replicate the changes to the other nodes.\\n> + </para>\\n>\\n> 16a.\\n> Should say node1, node2 and node3\\n\\nModified it in v19\\n\\n> 16b.\\n> SUGGESTION (2nd sentence – same as the previous comment)\\n> Incremental changes made in any node will be replicated to the other two nodes.\\n\\nModified\\n\\n> ~~~\\n>\\n> 17. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> data is present in the new node\\n>\\n> + <warning>\\n> + <para>\\n> + Adding a new node <literal>node3</literal> to the existing\\n> + <literal>node1</literal> and <literal>node2</literal> when data is present\\n> + in the new node <literal>node3</literal> is not possible.\\n> + </para>\\n> + </warning>\\n>\\n> 17a.\\n> IMO\\n> - Not really necessary to name the nodes, because this is a generic statement\\n> - Maybe say \\\"not supported\\\" instead of \\\"not possible\\\". e.g. there is\\n> no ERROR check for this case is there?\\n>\\n> SUGGESTION\\n> Adding a new node when data is present in the new node tables is not supported.\\n\\nModified\\n\\n> 17b.\\n> I am not sure but I felt this advice seemed more like an SGML <note>;\\n> note a <warning>\\n\\nModified\\n\\n> ~~~\\n>\\n> 18. doc/src/sgml/logical-replication.sgml - Generic steps to add a new\\n> node to the existing set of nodes\\n>\\n> + <title>Generic steps to add a new node to the existing set of nodes</title>\\n>\\n> SUGGESTION\\n> Generic steps for adding a new node to an existing set of nodes\\n\\nModified\\n\\n> ~~~\\n>\\n> 19. doc/src/sgml/logical-replication.sgml - Generic steps to add a new\\n> node to the existing set of nodes\\n>\\n> + <para>\\n> + 1. Create the required publication on the new node.\\n> + </para>\\n> + <para>\\n> + 2. Lock the required tables of the new node in <literal>EXCLUSIVE</literal>\\n> + mode until the setup is complete. This is required to prevent any\\n> + modifications from happening in the new node. If data modifications occur\\n> + after step-3, there is a chance that the modifications will be published to\\n> + the first node and then synchronized back to the new node while creating\\n> + subscription in step-5 resulting in inconsistent data.\\n> + </para>\\n> + <para>\\n> + 3. Create subscriptions on existing nodes pointing to publication on\\n> + the new node with <literal>origin</literal> parameter specified as\\n> + <literal>local</literal> and <literal>copy_data</literal> specified as\\n> + <literal>off</literal>.\\n> + </para>\\n> + <para>\\n> + 4. Lock the required tables of the existing nodes except the first node in\\n> + <literal>EXCLUSIVE</literal> mode until the setup is complete. This is\\n> + required to prevent any modifications from happening. If data modifications\\n> + occur, there is a chance that the modifications done in between step-5 and\\n> + step-6 will not be synchronized to the new node resulting in inconsistent\\n> + data. No need to lock the tables in the first node as any data changes\\n> + made will be synchronized while creating the subscription with\\n> + <literal>copy_data</literal> specified as <literal>force</literal>.\\n> + </para>\\n> + <para>\\n> + 5. Create a subscription on the new node pointing to publication on the\\n> + first node with <literal>origin</literal> parameter specified as\\n> + <literal>local</literal> and <literal>copy_data</literal> parameter\\n> + specified as <literal>force</literal>.\\n> + </para>\\n> + <para>\\n> + 6. Create subscriptions on the new node pointing to publications on the\\n> + remaining nodes with <literal>origin</literal> parameter specified as\\n> + <literal>local</literal> and <literal>copy_data</literal> parameter\\n> + specifiedas <literal>off</literal>.\\n> + </para>\\n>\\n> Following suggestions make the following changes:\\n> - Some minor rewording\\n> - Change the step names\\n> - Use \\\"=\\\" instead of \\\"specified as\\\" for the parameters\\n> - I felt it is more readable if the explanatory notes are separate\\n> paragraphs from the steps. Maybe if you could indent them or something\\n> it would be even better.\\n> - Added a couple more explanatory notes\\n>\\n> SUGGESTION (using those above changes; please keep your formatting)\\n>\\n> Step-1: Create a publication on the new node.\\n>\\n> Step-2: Lock the required tables of the new node in EXCLUSIVE mode\\n> until the setup is complete.\\n>\\n> (This lock is necessary to prevent any modifications from happening in\\n> the new node because if data modifications occurred after Step-3,\\n> there is a chance that the modifications will be published to the\\n> first node and then synchronized back to the new node while creating\\n> the subscription in Step-5. This would result in inconsistent data).\\n>\\n> Step-3. Create subscriptions on existing nodes to the publication on\\n> the new node with origin = local and copy_data = off.\\n>\\n> (The copy_data = off is OK here because it is asserted that the\\n> published tables of the new node will have no pre-existing data).\\n>\\n> Step-4. Lock the required tables of the existing nodes except the\\n> first node in EXCLUSIVE mode until the setup is complete.\\n>\\n> (This lock is necessary to prevent any modifications from happening.\\n> If data modifications occur, there is a chance that modifications done\\n> between Step-5 and Step-6 will not be synchronized to the new node.\\n> This would result in inconsistent data. There is no need to lock table\\n> t1 in node1 because any data changes made will be synchronized while\\n> creating the subscription with copy_data = force).\\n>\\n> Step-5. Create a subscription on the new node to the publication on\\n> the first node with origin = local and copy_data = force.\\n>\\n> (This will copy the same table data from the existing nodes to the new node)\\n>\\n> Step-6. Create subscriptions on the new node to publications on the\\n> remaining nodes with origin = local and copy_data = off.\\n>\\n> (The copy_data = off is OK here because the existing node data was\\n> already copied to the new node in Step-5)\\n\\nModified\\n\\n> ======\\n>\\n> 20. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> - <literal>copy_data = force</literal>.\\n> + <literal>copy_data = force</literal>. Refer to the\\n> + <xref linkend=\\\"logical-replication-bidirectional\\\"/> on how\\n> + <literal>copy_data</literal> and <literal>origin</literal> can be used\\n> + in bidirectional replication.\\n> </para>\\n>\\n> SUGGESTION (keep your formatting)\\n> Refer to <xref linkend=\\\"logical-replication-bidirectional\\\"/> for how\\n> <literal>copy_data</literal> and <literal>origin</literal> can be used\\n> in bidirectional replication.\\n\\nModified\\n\\nThanks for the comments. The comments are handled as part of the v20\\npatch attached at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm0XtQVX3taeLKWE-gPQyppqs34ipXawAPOyO%3Dhe37MQSg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 14 Jun 2022 17:38:50 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are some review comments for patch v20-0002\\n\\n======\\n\\n1. General comment\\n\\nSomething subtle but significant changed since I last reviewed v18*.\\nNow the describe.c is changed so that the catalog will never display a\\nNULL origin column; it would always be \\\"any\\\". But now I am not sure if\\nit is a good idea to still allow the NULL in this catalog column while\\nat the same time you are pretending it is not there. I felt it might\\nbe less confusing, and would simplify the code (e.g. remove quite a\\nfew null checks) to have just used a single concept of the default -\\ne.g. Just assign the default as \\\"any\\\" everywhere. The column would be\\ndefined as NOT NULL. Most of the following review comments are related\\nto this point.\\n\\n======\\n\\n2. doc/src/sgml/catalogs.sgml\\n\\n+ <para>\\n+ Possible origin values are <literal>local</literal>,\\n+ <literal>any</literal>, or <literal>NULL</literal> if none is specified.\\n+ If <literal>local</literal>, the subscription will request the\\n+ publisher to only send changes that originated locally. If\\n+ <literal>any</literal> (or <literal>NULL</literal>), the publisher sends\\n+ any changes regardless of their origin.\\n+ </para></entry>\\n\\nIs NULL still possible? Perhaps it would be better if it was not and\\nthe default \\\"any\\\" was always written instead.\\n\\n======\\n\\n3. src/backend/catalog/pg_subscription.c\\n\\n+ if (!isnull)\\n+ sub->origin = TextDatumGetCString(datum);\\n+ else\\n+ sub->origin = NULL;\\n+\\n\\nMaybe better to either disallow NULL in the first place or assign the\\n\\\"any\\\" here instead of NULL.\\n\\n======\\n\\n4. src/backend/commands/subscriptioncmds.c - parse_subscription_options\\n\\n@@ -137,6 +139,8 @@ parse_subscription_options(ParseState *pstate,\\nList *stmt_options,\\n opts->twophase = false;\\n if (IsSet(supported_opts, SUBOPT_DISABLE_ON_ERR))\\n opts->disableonerr = false;\\n+ if (IsSet(supported_opts, SUBOPT_ORIGIN))\\n+ opts->origin = NULL;\\n\\nIf opt->origin was assigned to \\\"any\\\" then other code would be simplified.\\n\\n~~~\\n\\n5. src/backend/commands/subscriptioncmds.c - CreateSubscription\\n\\n@@ -607,6 +626,11 @@ CreateSubscription(ParseState *pstate,\\nCreateSubscriptionStmt *stmt,\\n LOGICALREP_TWOPHASE_STATE_PENDING :\\n LOGICALREP_TWOPHASE_STATE_DISABLED);\\n values[Anum_pg_subscription_subdisableonerr - 1] =\\nBoolGetDatum(opts.disableonerr);\\n+ if (opts.origin)\\n+ values[Anum_pg_subscription_suborigin - 1] =\\n+ CStringGetTextDatum(opts.origin);\\n+ else\\n+ values[Anum_pg_subscription_suborigin - 1] =\\nCStringGetTextDatum(LOGICALREP_ORIGIN_ANY);\\n\\nIf NULL was not possible then this would just be one line:\\nvalues[Anum_pg_subscription_suborigin - 1] = CStringGetTextDatum(opts.origin);\\n\\n======\\n\\n6. src/backend/replication/logical/worker.c\\n\\n@@ -276,6 +276,10 @@ static TransactionId stream_xid = InvalidTransactionId;\\n static XLogRecPtr skip_xact_finish_lsn = InvalidXLogRecPtr;\\n #define is_skipping_changes()\\n(unlikely(!XLogRecPtrIsInvalid(skip_xact_finish_lsn)))\\n\\n+/* Macro for comparing string fields that might be NULL */\\n+#define equalstr(a, b) \\\\\\n+ (((a) != NULL && (b) != NULL) ? (strcmp((a), (b)) == 0) : (a) == (b))\\n+\\n\\nIf the NULL was not allowed in the first place then I think this macro\\nwould just become redundant.\\n\\n7. src/backend/replication/logical/worker.c - ApplyWorkerMain\\n\\n@@ -3741,6 +3746,11 @@ ApplyWorkerMain(Datum main_arg)\\n options.proto.logical.streaming = MySubscription->stream;\\n options.proto.logical.twophase = false;\\n\\n+ if (MySubscription->origin)\\n+ options.proto.logical.origin = pstrdup(MySubscription->origin);\\n+ else\\n+ options.proto.logical.origin = NULL;\\n+\\n\\nCan't the if/else be avoided if you always assigned the \\\"any\\\" default\\nin the first place?\\n\\n======\\n\\n8. src/backend/replication/pgoutput/pgoutput.c - parse_output_parameters\\n\\n@@ -287,11 +289,13 @@ parse_output_parameters(List *options, PGOutputData *data)\\n bool messages_option_given = false;\\n bool streaming_given = false;\\n bool two_phase_option_given = false;\\n+ bool origin_option_given = false;\\n\\n data->binary = false;\\n data->streaming = false;\\n data->messages = false;\\n data->two_phase = false;\\n+ data->origin = NULL;\\n\\nConsider assigning default \\\"any\\\" here instead of NULL.\\n\\n======\\n\\n9. src/bin/pg_dump/pg_dump.c - getSubscriptions\\n\\n+ /* FIXME: 150000 should be changed to 160000 later for PG16. */\\n+ if (fout->remoteVersion >= 150000)\\n+ appendPQExpBufferStr(query, \\\" s.suborigin\\\\n\\\");\\n+ else\\n+ appendPQExpBufferStr(query, \\\" NULL AS suborigin\\\\n\\\");\\n+\\n\\nMaybe say: 'any' AS suborigin?\\n\\n~~~\\n\\n10. src/bin/pg_dump/pg_dump.c - getSubscriptions\\n\\n@@ -4517,6 +4525,11 @@ getSubscriptions(Archive *fout)\\n subinfo[i].subdisableonerr =\\n pg_strdup(PQgetvalue(res, i, i_subdisableonerr));\\n\\n+ if (PQgetisnull(res, i, i_suborigin))\\n+ subinfo[i].suborigin = NULL;\\n+ else\\n+ subinfo[i].suborigin = pg_strdup(PQgetvalue(res, i, i_suborigin));\\n+\\n\\nIf you disallow the NULL in the first place this condition maybe is no\\nlonger needed.\\n\\n~~~\\n\\n11. src/bin/pg_dump/pg_dump.c - dumpSubscription\\n\\n@@ -4589,6 +4602,9 @@ dumpSubscription(Archive *fout, const\\nSubscriptionInfo *subinfo)\\n if (strcmp(subinfo->subdisableonerr, \\\"t\\\") == 0)\\n appendPQExpBufferStr(query, \\\", disable_on_error = true\\\");\\n\\n+ if (subinfo->suborigin)\\n+ appendPQExpBuffer(query, \\\", origin = %s\\\", subinfo->suborigin);\\n+\\n\\nIf NULL cannot happen then maybe this test is also redundant.\\n\\n======\\n\\n12. src/bin/pg_dump/t/002_pg_dump.pl\\n\\nAFAICT there is a test for no origin (default), and a test for\\nexplicit origin = local, but there is no test case for the explicit\\norigin = any.\\n\\n======\\n\\n13. src/include/catalog/pg_subscription.h\\n\\n@@ -31,6 +31,9 @@\\n #define LOGICALREP_TWOPHASE_STATE_PENDING 'p'\\n #define LOGICALREP_TWOPHASE_STATE_ENABLED 'e'\\n\\n+#define LOGICALREP_ORIGIN_LOCAL \\\"local\\\"\\n+#define LOGICALREP_ORIGIN_ANY \\\"any\\\"\\n+\\n\\nI thought there should be a comment above these new constants.\\n\\n~~~\\n\\n14. src/include/catalog/pg_subscription.h\\n\\n@@ -87,6 +90,9 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\\nBKI_SHARED_RELATION BKI_ROW\\n\\n /* List of publications subscribed to */\\n text subpublications[1] BKI_FORCE_NOT_NULL;\\n+\\n+ /* Only publish data originating from the specified origin */\\n+ text suborigin;\\n #endif\\n } FormData_pg_subscription;\\n\\nPerhaps it would be better if this new column was also forced to be NOT NULL.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 15 Jun 2022 16:39:29 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are some review comments for patch v20-0003.\\n\\n======\\n\\n1. Commit message\\n\\nIn case, table t1 has a unique key, it will lead to a unique key\\nviolation and replication won't proceed.\\n\\nSUGGESTION\\nIf table t1 has a unique key, this will lead to a unique key\\nviolation, and replication won't proceed.\\n\\n~~~\\n\\n2. Commit message\\n\\nThis problem can be solved by using...\\n\\nSUGGESTION\\nThis problem can be avoided by using...\\n\\n~~~\\n\\n3. Commit message\\n\\nstep 3: Create a subscription in node1 to subscribe to node2. Use\\n'copy_data = on' so that the existing table data is copied during\\ninitial sync:\\nnode1=# CREATE SUBSCRIPTION sub_node1_node2 CONNECTION '<node2 details>'\\nnode1-# PUBLICATION pub_node2 WITH (copy_data = off, origin = local);\\nCREATE SUBSCRIPTION\\n\\nThis is wrong information. The table on node2 has no data, so talking\\nabout 'copy_data = on' is inappropriate here.\\n\\n======\\n\\n4. Commit message\\n\\nIMO it might be better to refer to subscription/publication/table \\\"on\\\"\\nnodeXXX, instead of saying \\\"in\\\" nodeXXX.\\n\\n4a.\\n\\\"the publication tables were also subscribing data in the publisher\\nfrom other publishers.\\\" -> \\\"the publication tables were also\\nsubscribing from other publishers.\\n\\n4b.\\n\\\"After the subscription is created in node2\\\" -> \\\"After the\\nsubscription is created on node2\\\"\\n\\n4c.\\n\\\"step 3: Create a subscription in node1 to subscribe to node2.\\\" ->\\n\\\"step 3: Create a subscription on node1 to subscribe to node2.\\\"\\n\\n4d.\\n\\\"step 4: Create a subscription in node2 to subscribe to node1.\\\" ->\\n\\\"step 4: Create a subscription on node2 to subscribe to node1.\\\"\\n\\n======\\n\\n5. doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -383,6 +397,15 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n can have non-existent publications.\\n </para>\\n\\n+ <para>\\n+ If the subscription is created with <literal>origin = local</literal> and\\n+ <literal>copy_data = on</literal>, it will check if the publisher tables are\\n+ being subscribed to any other publisher and, if so, then throw an error to\\n+ prevent possible non-local data from being copied. The user can override\\n+ this check and continue with the copy operation by specifying\\n+ <literal>copy_data = force</literal>.\\n+ </para>\\n\\nPerhaps it is better here to say 'copy_data = true' instead of\\n'copy_data = on', simply because the value 'true' was mentioned\\nearlier on this page (but this would be the first mention of 'on').\\n\\n======\\n\\n6. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n\\n+ errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force.\\\"));\\n\\nSaying \\\"off or force\\\" is not consistent with the other message wording\\nin this patch, which used \\\"/\\\" for multiple enums.\\n(e.g. \\\"connect = false\\\", \\\"copy_data = true/force\\\").\\n\\nSo perhaps this errhint should be worded similarly:\\n\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\\"\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 15 Jun 2022 16:41:21 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are some review comments for patch v20-0004.\\n\\n======\\n\\n1. General\\n\\nI thought that it is better to refer to the\\nsubscription/publications/table \\\"on\\\" the node, rather than \\\"in\\\" the\\nnode. Most of the review comments below are related to this point.\\n\\n======\\n\\n2. Commit message\\n\\na) Creating a two-node bidirectional replication when there is no data\\nin both nodes.\\nb) Adding a new node when there is no data in any of the nodes.\\nc) Adding a new node when data is present in the existing nodes.\\n\\n\\\"in both nodes\\\" -> \\\"on both nodes\\\"\\n\\\"in any of the nodes\\\" -> \\\"on any of the nodes\\\"\\n\\\"in the existing nodes\\\" -> \\\"on the existing nodes\\\"\\n\\n======\\n\\n3. doc/src/sgml/logical-replication.sgml - Setting bidirectional\\nreplication between two nodes\\n\\n3a.\\n+ <para>\\n+ The following steps demonstrate how to create a two-node bidirectional\\n+ replication when there is no table data present in both nodes\\n+ <literal>node1</literal> and <literal>node2</literal>:\\n+ </para>\\n-> \\\"on both nodes\\\"\\n\\n3b.\\n+ Create a publication in <literal>node1</literal>:\\n-> \\\"on\\\"\\n\\n3c.\\n+ Create a publication in <literal>node2</literal>:\\n-> \\\"on\\\"\\n\\n3d.\\n+ <para>\\n+ Lock the table <literal>t1</literal> in <literal>node1</literal> and\\n+ <literal>node2</literal> in <literal>EXCLUSIVE</literal> mode until the\\n+ setup is completed.\\n+ </para>\\n-> \\\"on <literal>node1</literal>\\\"\\n\\n3e.\\n+ Create a subscription in <literal>node2</literal> to subscribe to\\n-> \\\"on\\\"\\n\\n3f.\\n+ Create a subscription in <literal>node1</literal> to subscribe to\\n+ <literal>node2</literal>:\\n-> \\\"on\\\"\\n\\n~~~\\n\\n4. doc/src/sgml/logical-replication.sgml - Adding a new node when\\nthere is no data in any of the nodes\\n\\n4a.\\n+ <title>Adding a new node when there is no data in any of the nodes</title>\\nSUGGESTION\\nAdding a new node when there is no table data on any of the nodes\\n\\n4b.\\n+ <para>\\n+ The following steps demonstrate adding a new node <literal>node3</literal>\\n+ to the existing <literal>node1</literal> and <literal>node2</literal> when\\n+ there is no <literal>t1</literal> data in any of the nodes. This requires\\n+ creating subscriptions in <literal>node1</literal> and\\n+ <literal>node2</literal> to replicate the data from\\n+ <literal>node3</literal> and creating subscriptions in\\n+ <literal>node3</literal> to replicate data from <literal>node1</literal>\\n+ and <literal>node2</literal>. Note: These steps assume that the\\n+ bidirectional logical replication between <literal>node1</literal> and\\n+ <literal>node2</literal> is already completed.\\n+ </para>\\n\\n\\\"data in any of the nodes\\\" -> \\\"data on any of the nodes\\\"\\n\\\"creating subscriptions in <literal>node1</literal>\\\" -> \\\"creating\\nsubscriptions on <literal>node1</literal>\\\"\\n\\\"creating subscriptions in <literal>node3</literal>\\\" -> \\\"creating\\nsubscriptions on <literal>node3</literal>\\\"\\n\\n4c.\\n+ Create a publication in <literal>node3</literal>:\\n-> \\\"on\\\"\\n\\n4d.\\n+ Lock table <literal>t1</literal> in all the nodes\\n-> \\\"on\\\"\\n\\n4e.\\n+ Create a subscription in <literal>node1</literal> to subscribe to\\n+ <literal>node3</literal>:\\n-> \\\"on\\\"\\n\\n4f.\\n+ Create a subscription in <literal>node2</literal> to subscribe to\\n+ <literal>node3</literal>:\\n-> \\\"on\\\"\\n\\n4g.\\n+ Create a subscription in <literal>node3</literal> to subscribe to\\n+ <literal>node1</literal>:\\n-> \\\"on\\\"\\n\\n4h.\\n+ Create a subscription in <literal>node3</literal> to subscribe to\\n+ <literal>node2</literal>:\\n\\n4i.\\n+ <literal>node3</literal>. Incremental changes made in any node will be\\n+ replicated to the other two nodes.\\n\\\"in any node\\\" -> \\\"on any node\\\"\\n\\n~~~\\n\\n5. doc/src/sgml/logical-replication.sgml - Adding a new node when data\\nis present in the existing nodes\\n\\n5a.\\n+ <title>Adding a new node when data is present in the existing nodes</title>\\nSUGGESTION\\nAdding a new node when table data is present on the existing nodes\\n\\n5b.\\n+ during initial data synchronization. Note: These steps assume that the\\n+ bidirectional logical replication between <literal>node1</literal> and\\n+ <literal>node2</literal> is already completed, and the pre-existing data\\n+ in table <literal>t1</literal> is already synchronized in both those\\n+ nodes.\\n+ </para>\\n\\\"in both those nodes\\\" -> \\\"on both those nodes\\\"\\n\\n5c.\\n+ Create a publication in <literal>node3</literal>\\n-> \\\"on\\\"\\n\\n5d.\\n+ Lock table <literal>t1</literal> in <literal>node2</literal> and\\n-> \\\"on\\\"\\n\\n5e.\\n+ Create a subscription in <literal>node1</literal> to subscribe to\\n+ <literal>node3</literal>:\\n-> \\\"on\\\"\\n\\n5f.\\n+ Create a subscription in <literal>node2</literal> to subscribe to\\n+ <literal>node3</literal>:\\n-> \\\"on\\\"\\n\\n5g.\\n+ Create a subscription in <literal>node3</literal> to subscribe to\\n+ <literal>node1</literal>. Use <literal>copy_data = force </literal> so that\\n+ the existing table data is copied during initial sync:\\n-> \\\"on\\\"\\n\\n\\n5h.\\n+ Create a subscription in <literal>node3</literal> to subscribe to\\n+ <literal>node2</literal>. Use <literal>copy_data = off</literal>\\n-> \\\"on\\\"\\n\\n5i.\\n+ <literal>node3</literal>. Incremental changes made in any node will be\\n+ replicated to the other two nodes.\\n\\\"in any node\\\" -> \\\"on any node\\\"\\n\\n~~~\\n\\n6. doc/src/sgml/logical-replication.sgml - Adding a new node when data\\nis present in the new node\\n\\n+ <title>Adding a new node when data is present in the new node</title>\\nSUGGESTION\\nAdding a new node when table data is present on the new node\\n\\n~~~\\n\\n7. doc/src/sgml/logical-replication.sgml - Generic steps for adding a\\nnew node to an existing set of nodes\\n\\n7a.\\n+ <para>\\n+ Step-2: Lock the required tables of the new node in EXCLUSIVE mode until\\n+ the setup is complete. (This lock is necessary to prevent any modifications\\n+ from happening in the new node because if data modifications occurred after\\n+ Step-3, there is a chance that the modifications will be published to the\\n+ first node and then synchronized back to the new node while creating the\\n+ subscription in Step-5. This would result in inconsistent data).\\n+ </para>\\n\\\"happening in the new node\\\" -> \\\"happening on the new node\\\"\\n\\n7b.\\n+ not be synchronized to the new node. This would result in inconsistent\\n+ data. There is no need to lock the required tables in\\n+ <literal>node1</literal> because any data changes made will be synchronized\\n+ while creating the subscription with <literal>copy_data = force</literal>).\\n+ </para>\\n\\\"no need to lock the required tables in\\\" -> \\\"no need to lock the\\nrequired tables on\\\"\\n\\n======\\n\\n8. doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -403,7 +403,10 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n being subscribed to any other publisher and, if so, then throw an error to\\n prevent possible non-local data from being copied. The user can override\\n this check and continue with the copy operation by specifying\\n- <literal>copy_data = force</literal>.\\n+ <literal>copy_data = force</literal>. Refer to\\n+ <xref linkend=\\\"logical-replication-bidirectional\\\"/> for how\\n+ <literal>copy_data</literal> and <literal>origin</literal> can be used\\n+ in bidirectional replication.\\n </para>\\n\\n\\\"can be used in bidirectional replication\\\" -> \\\"can be used to set up\\nbidirectional replication\\\"\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 15 Jun 2022 16:43:35 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"PSA a test script that demonstrates all the documented steps for\\nsetting up n-way bidirectional replication. These steps are the same\\nas those documented [1] on the new page \\\"Bidirectional logical\\nreplication\\\".\\n\\nThis script works using the current latest v20* patch set. Each of the\\nsections of 31.11.1 - 31.11.5 (see below) can be run independently\\n(just edit the script and at the bottom uncomment the part you want to\\ntest):\\n31.11.1. Setting bidirectional replication between two nodes\\n31.11.2. Adding a new node when there is no data in any of the nodes\\n31.11.3. Adding a new node when data is present in the existing nodes\\n31.11.4. Adding a new node when data is present in the new node\\n31.11.5. Generic steps for adding a new node to an existing set of nodes\\n\\n~~\\n\\nSome sample output is also attached.\\n\\n------\\n[1] https://www.postgresql.org/message-id/attachment/134464/v20-0004-Document-bidirectional-logical-replication-steps.patch\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\",\n \"msg_date\": \"Wed, 15 Jun 2022 18:56:24 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some review comments for patch v20-0002\\n>\\n> ======\\n>\\n> 1. General comment\\n>\\n> Something subtle but significant changed since I last reviewed v18*.\\n> Now the describe.c is changed so that the catalog will never display a\\n> NULL origin column; it would always be \\\"any\\\". But now I am not sure if\\n> it is a good idea to still allow the NULL in this catalog column while\\n> at the same time you are pretending it is not there. I felt it might\\n> be less confusing, and would simplify the code (e.g. remove quite a\\n> few null checks) to have just used a single concept of the default -\\n> e.g. Just assign the default as \\\"any\\\" everywhere. The column would be\\n> defined as NOT NULL. Most of the following review comments are related\\n> to this point.\\n\\nOk, I was initially feeling having NULL value will help in the upgrade\\ncases where we upgrade from a lower version which does not have origin\\nto current. But setting the default value handles the upgrade scenario\\ntoo.\\n\\n> ======\\n>\\n> 2. doc/src/sgml/catalogs.sgml\\n>\\n> + <para>\\n> + Possible origin values are <literal>local</literal>,\\n> + <literal>any</literal>, or <literal>NULL</literal> if none is specified.\\n> + If <literal>local</literal>, the subscription will request the\\n> + publisher to only send changes that originated locally. If\\n> + <literal>any</literal> (or <literal>NULL</literal>), the publisher sends\\n> + any changes regardless of their origin.\\n> + </para></entry>\\n>\\n> Is NULL still possible? Perhaps it would be better if it was not and\\n> the default \\\"any\\\" was always written instead.\\n\\nModified\\n\\n> ======\\n>\\n> 3. src/backend/catalog/pg_subscription.c\\n>\\n> + if (!isnull)\\n> + sub->origin = TextDatumGetCString(datum);\\n> + else\\n> + sub->origin = NULL;\\n> +\\n>\\n> Maybe better to either disallow NULL in the first place or assign the\\n> \\\"any\\\" here instead of NULL.\\n\\nModified\\n\\n> ======\\n>\\n> 4. src/backend/commands/subscriptioncmds.c - parse_subscription_options\\n>\\n> @@ -137,6 +139,8 @@ parse_subscription_options(ParseState *pstate,\\n> List *stmt_options,\\n> opts->twophase = false;\\n> if (IsSet(supported_opts, SUBOPT_DISABLE_ON_ERR))\\n> opts->disableonerr = false;\\n> + if (IsSet(supported_opts, SUBOPT_ORIGIN))\\n> + opts->origin = NULL;\\n>\\n> If opt->origin was assigned to \\\"any\\\" then other code would be simplified.\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. src/backend/commands/subscriptioncmds.c - CreateSubscription\\n>\\n> @@ -607,6 +626,11 @@ CreateSubscription(ParseState *pstate,\\n> CreateSubscriptionStmt *stmt,\\n> LOGICALREP_TWOPHASE_STATE_PENDING :\\n> LOGICALREP_TWOPHASE_STATE_DISABLED);\\n> values[Anum_pg_subscription_subdisableonerr - 1] =\\n> BoolGetDatum(opts.disableonerr);\\n> + if (opts.origin)\\n> + values[Anum_pg_subscription_suborigin - 1] =\\n> + CStringGetTextDatum(opts.origin);\\n> + else\\n> + values[Anum_pg_subscription_suborigin - 1] =\\n> CStringGetTextDatum(LOGICALREP_ORIGIN_ANY);\\n>\\n> If NULL was not possible then this would just be one line:\\n> values[Anum_pg_subscription_suborigin - 1] = CStringGetTextDatum(opts.origin);\\n\\nModified\\n\\n> ======\\n>\\n> 6. src/backend/replication/logical/worker.c\\n>\\n> @@ -276,6 +276,10 @@ static TransactionId stream_xid = InvalidTransactionId;\\n> static XLogRecPtr skip_xact_finish_lsn = InvalidXLogRecPtr;\\n> #define is_skipping_changes()\\n> (unlikely(!XLogRecPtrIsInvalid(skip_xact_finish_lsn)))\\n>\\n> +/* Macro for comparing string fields that might be NULL */\\n> +#define equalstr(a, b) \\\\\\n> + (((a) != NULL && (b) != NULL) ? (strcmp((a), (b)) == 0) : (a) == (b))\\n> +\\n>\\n> If the NULL was not allowed in the first place then I think this macro\\n> would just become redundant.\\n\\nRemoved it\\n\\n> 7. src/backend/replication/logical/worker.c - ApplyWorkerMain\\n>\\n> @@ -3741,6 +3746,11 @@ ApplyWorkerMain(Datum main_arg)\\n> options.proto.logical.streaming = MySubscription->stream;\\n> options.proto.logical.twophase = false;\\n>\\n> + if (MySubscription->origin)\\n> + options.proto.logical.origin = pstrdup(MySubscription->origin);\\n> + else\\n> + options.proto.logical.origin = NULL;\\n> +\\n>\\n> Can't the if/else be avoided if you always assigned the \\\"any\\\" default\\n> in the first place?\\n\\nModified\\n\\n> ======\\n>\\n> 8. src/backend/replication/pgoutput/pgoutput.c - parse_output_parameters\\n>\\n> @@ -287,11 +289,13 @@ parse_output_parameters(List *options, PGOutputData *data)\\n> bool messages_option_given = false;\\n> bool streaming_given = false;\\n> bool two_phase_option_given = false;\\n> + bool origin_option_given = false;\\n>\\n> data->binary = false;\\n> data->streaming = false;\\n> data->messages = false;\\n> data->two_phase = false;\\n> + data->origin = NULL;\\n>\\n> Consider assigning default \\\"any\\\" here instead of NULL.\\n\\nThis is no more required because of setting default to any and the\\nsame value will be passed as an option\\n\\n> ======\\n>\\n> 9. src/bin/pg_dump/pg_dump.c - getSubscriptions\\n>\\n> + /* FIXME: 150000 should be changed to 160000 later for PG16. */\\n> + if (fout->remoteVersion >= 150000)\\n> + appendPQExpBufferStr(query, \\\" s.suborigin\\\\n\\\");\\n> + else\\n> + appendPQExpBufferStr(query, \\\" NULL AS suborigin\\\\n\\\");\\n> +\\n>\\n> Maybe say: 'any' AS suborigin?\\n\\nModified\\n\\n> ~~~\\n>\\n> 10. src/bin/pg_dump/pg_dump.c - getSubscriptions\\n>\\n> @@ -4517,6 +4525,11 @@ getSubscriptions(Archive *fout)\\n> subinfo[i].subdisableonerr =\\n> pg_strdup(PQgetvalue(res, i, i_subdisableonerr));\\n>\\n> + if (PQgetisnull(res, i, i_suborigin))\\n> + subinfo[i].suborigin = NULL;\\n> + else\\n> + subinfo[i].suborigin = pg_strdup(PQgetvalue(res, i, i_suborigin));\\n> +\\n>\\n> If you disallow the NULL in the first place this condition maybe is no\\n> longer needed.\\n\\nModified\\n\\n> ~~~\\n>\\n> 11. src/bin/pg_dump/pg_dump.c - dumpSubscription\\n>\\n> @@ -4589,6 +4602,9 @@ dumpSubscription(Archive *fout, const\\n> SubscriptionInfo *subinfo)\\n> if (strcmp(subinfo->subdisableonerr, \\\"t\\\") == 0)\\n> appendPQExpBufferStr(query, \\\", disable_on_error = true\\\");\\n>\\n> + if (subinfo->suborigin)\\n> + appendPQExpBuffer(query, \\\", origin = %s\\\", subinfo->suborigin);\\n> +\\n>\\n> If NULL cannot happen then maybe this test is also redundant.\\n\\nModified\\n\\n> ======\\n>\\n> 12. src/bin/pg_dump/t/002_pg_dump.pl\\n>\\n> AFAICT there is a test for no origin (default), and a test for\\n> explicit origin = local, but there is no test case for the explicit\\n> origin = any.\\n\\nAdded a test for the same\\n\\n> ======\\n>\\n> 13. src/include/catalog/pg_subscription.h\\n>\\n> @@ -31,6 +31,9 @@\\n> #define LOGICALREP_TWOPHASE_STATE_PENDING 'p'\\n> #define LOGICALREP_TWOPHASE_STATE_ENABLED 'e'\\n>\\n> +#define LOGICALREP_ORIGIN_LOCAL \\\"local\\\"\\n> +#define LOGICALREP_ORIGIN_ANY \\\"any\\\"\\n> +\\n>\\n> I thought there should be a comment above these new constants.\\n\\nAdded comments\\n\\n> ~~~\\n>\\n> 14. src/include/catalog/pg_subscription.h\\n>\\n> @@ -87,6 +90,9 @@ CATALOG(pg_subscription,6100,SubscriptionRelationId)\\n> BKI_SHARED_RELATION BKI_ROW\\n>\\n> /* List of publications subscribed to */\\n> text subpublications[1] BKI_FORCE_NOT_NULL;\\n> +\\n> + /* Only publish data originating from the specified origin */\\n> + text suborigin;\\n> #endif\\n> } FormData_pg_subscription;\\n>\\n> Perhaps it would be better if this new column was also forced to be NOT NULL.\\n\\nI have set a default value so need to set NOT NULL.\\n\\nThanks for the comments, the attached v21 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Thu, 16 Jun 2022 15:47:50 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 15, 2022 at 12:11 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some review comments for patch v20-0003.\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> In case, table t1 has a unique key, it will lead to a unique key\\n> violation and replication won't proceed.\\n>\\n> SUGGESTION\\n> If table t1 has a unique key, this will lead to a unique key\\n> violation, and replication won't proceed.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2. Commit message\\n>\\n> This problem can be solved by using...\\n>\\n> SUGGESTION\\n> This problem can be avoided by using...\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. Commit message\\n>\\n> step 3: Create a subscription in node1 to subscribe to node2. Use\\n> 'copy_data = on' so that the existing table data is copied during\\n> initial sync:\\n> node1=# CREATE SUBSCRIPTION sub_node1_node2 CONNECTION '<node2 details>'\\n> node1-# PUBLICATION pub_node2 WITH (copy_data = off, origin = local);\\n> CREATE SUBSCRIPTION\\n>\\n> This is wrong information. The table on node2 has no data, so talking\\n> about 'copy_data = on' is inappropriate here.\\n\\nModified\\n\\n> ======\\n>\\n> 4. Commit message\\n>\\n> IMO it might be better to refer to subscription/publication/table \\\"on\\\"\\n> nodeXXX, instead of saying \\\"in\\\" nodeXXX.\\n>\\n> 4a.\\n> \\\"the publication tables were also subscribing data in the publisher\\n> from other publishers.\\\" -> \\\"the publication tables were also\\n> subscribing from other publishers.\\n\\nModified\\n\\n> 4b.\\n> \\\"After the subscription is created in node2\\\" -> \\\"After the\\n> subscription is created on node2\\\"\\n\\nModified\\n\\n> 4c.\\n> \\\"step 3: Create a subscription in node1 to subscribe to node2.\\\" ->\\n> \\\"step 3: Create a subscription on node1 to subscribe to node2.\\\"\\n\\nModified\\n\\n> 4d.\\n> \\\"step 4: Create a subscription in node2 to subscribe to node1.\\\" ->\\n> \\\"step 4: Create a subscription on node2 to subscribe to node1.\\\"\\n\\nModified\\n\\n> ======\\n>\\n> 5. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -383,6 +397,15 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> can have non-existent publications.\\n> </para>\\n>\\n> + <para>\\n> + If the subscription is created with <literal>origin = local</literal> and\\n> + <literal>copy_data = on</literal>, it will check if the publisher tables are\\n> + being subscribed to any other publisher and, if so, then throw an error to\\n> + prevent possible non-local data from being copied. The user can override\\n> + this check and continue with the copy operation by specifying\\n> + <literal>copy_data = force</literal>.\\n> + </para>\\n>\\n> Perhaps it is better here to say 'copy_data = true' instead of\\n> 'copy_data = on', simply because the value 'true' was mentioned\\n> earlier on this page (but this would be the first mention of 'on').\\n\\nModified\\n\\n> ======\\n>\\n> 6. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> + errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off or force.\\\"));\\n>\\n> Saying \\\"off or force\\\" is not consistent with the other message wording\\n> in this patch, which used \\\"/\\\" for multiple enums.\\n> (e.g. \\\"connect = false\\\", \\\"copy_data = true/force\\\").\\n>\\n> So perhaps this errhint should be worded similarly:\\n> \\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\\"\\n\\nModified\\n\\nThanks for the comments, the v21 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm3%2B6cey0rcDft1ZUCjSUtLDM0xmU_Q%2BYhcsBrqe1RH8%3Dw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Jun 2022 15:50:02 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 15, 2022 at 12:13 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some review comments for patch v20-0004.\\n>\\n> ======\\n>\\n> 1. General\\n>\\n> I thought that it is better to refer to the\\n> subscription/publications/table \\\"on\\\" the node, rather than \\\"in\\\" the\\n> node. Most of the review comments below are related to this point.\\n\\nModified\\n\\n> ======\\n>\\n> 2. Commit message\\n>\\n> a) Creating a two-node bidirectional replication when there is no data\\n> in both nodes.\\n> b) Adding a new node when there is no data in any of the nodes.\\n> c) Adding a new node when data is present in the existing nodes.\\n>\\n> \\\"in both nodes\\\" -> \\\"on both nodes\\\"\\n> \\\"in any of the nodes\\\" -> \\\"on any of the nodes\\\"\\n> \\\"in the existing nodes\\\" -> \\\"on the existing nodes\\\"\\n\\nModified\\n\\n> ======\\n>\\n> 3. doc/src/sgml/logical-replication.sgml - Setting bidirectional\\n> replication between two nodes\\n>\\n> 3a.\\n> + <para>\\n> + The following steps demonstrate how to create a two-node bidirectional\\n> + replication when there is no table data present in both nodes\\n> + <literal>node1</literal> and <literal>node2</literal>:\\n> + </para>\\n> -> \\\"on both nodes\\\"\\n\\nModified\\n\\n> 3b.\\n> + Create a publication in <literal>node1</literal>:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 3c.\\n> + Create a publication in <literal>node2</literal>:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 3d.\\n> + <para>\\n> + Lock the table <literal>t1</literal> in <literal>node1</literal> and\\n> + <literal>node2</literal> in <literal>EXCLUSIVE</literal> mode until the\\n> + setup is completed.\\n> + </para>\\n> -> \\\"on <literal>node1</literal>\\\"\\n\\nModified\\n\\n> 3e.\\n> + Create a subscription in <literal>node2</literal> to subscribe to\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 3f.\\n> + Create a subscription in <literal>node1</literal> to subscribe to\\n> + <literal>node2</literal>:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. doc/src/sgml/logical-replication.sgml - Adding a new node when\\n> there is no data in any of the nodes\\n>\\n> 4a.\\n> + <title>Adding a new node when there is no data in any of the nodes</title>\\n> SUGGESTION\\n> Adding a new node when there is no table data on any of the nodes\\n\\nModified\\n\\n> 4b.\\n> + <para>\\n> + The following steps demonstrate adding a new node <literal>node3</literal>\\n> + to the existing <literal>node1</literal> and <literal>node2</literal> when\\n> + there is no <literal>t1</literal> data in any of the nodes. This requires\\n> + creating subscriptions in <literal>node1</literal> and\\n> + <literal>node2</literal> to replicate the data from\\n> + <literal>node3</literal> and creating subscriptions in\\n> + <literal>node3</literal> to replicate data from <literal>node1</literal>\\n> + and <literal>node2</literal>. Note: These steps assume that the\\n> + bidirectional logical replication between <literal>node1</literal> and\\n> + <literal>node2</literal> is already completed.\\n> + </para>\\n>\\n> \\\"data in any of the nodes\\\" -> \\\"data on any of the nodes\\\"\\n> \\\"creating subscriptions in <literal>node1</literal>\\\" -> \\\"creating\\n> subscriptions on <literal>node1</literal>\\\"\\n> \\\"creating subscriptions in <literal>node3</literal>\\\" -> \\\"creating\\n> subscriptions on <literal>node3</literal>\\\"\\n\\nModified\\n\\n> 4c.\\n> + Create a publication in <literal>node3</literal>:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 4d.\\n> + Lock table <literal>t1</literal> in all the nodes\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 4e.\\n> + Create a subscription in <literal>node1</literal> to subscribe to\\n> + <literal>node3</literal>:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 4f.\\n> + Create a subscription in <literal>node2</literal> to subscribe to\\n> + <literal>node3</literal>:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 4g.\\n> + Create a subscription in <literal>node3</literal> to subscribe to\\n> + <literal>node1</literal>:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 4h.\\n> + Create a subscription in <literal>node3</literal> to subscribe to\\n> + <literal>node2</literal>:\\n\\nModified\\n\\n> 4i.\\n> + <literal>node3</literal>. Incremental changes made in any node will be\\n> + replicated to the other two nodes.\\n> \\\"in any node\\\" -> \\\"on any node\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. doc/src/sgml/logical-replication.sgml - Adding a new node when data\\n> is present in the existing nodes\\n>\\n> 5a.\\n> + <title>Adding a new node when data is present in the existing nodes</title>\\n> SUGGESTION\\n> Adding a new node when table data is present on the existing nodes\\n\\nModified\\n\\n> 5b.\\n> + during initial data synchronization. Note: These steps assume that the\\n> + bidirectional logical replication between <literal>node1</literal> and\\n> + <literal>node2</literal> is already completed, and the pre-existing data\\n> + in table <literal>t1</literal> is already synchronized in both those\\n> + nodes.\\n> + </para>\\n> \\\"in both those nodes\\\" -> \\\"on both those nodes\\\"\\n\\nModified\\n\\n> 5c.\\n> + Create a publication in <literal>node3</literal>\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 5d.\\n> + Lock table <literal>t1</literal> in <literal>node2</literal> and\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 5e.\\n> + Create a subscription in <literal>node1</literal> to subscribe to\\n> + <literal>node3</literal>:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 5f.\\n> + Create a subscription in <literal>node2</literal> to subscribe to\\n> + <literal>node3</literal>:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 5g.\\n> + Create a subscription in <literal>node3</literal> to subscribe to\\n> + <literal>node1</literal>. Use <literal>copy_data = force </literal> so that\\n> + the existing table data is copied during initial sync:\\n> -> \\\"on\\\"\\n\\nModified\\n\\n>\\n> 5h.\\n> + Create a subscription in <literal>node3</literal> to subscribe to\\n> + <literal>node2</literal>. Use <literal>copy_data = off</literal>\\n> -> \\\"on\\\"\\n\\nModified\\n\\n> 5i.\\n> + <literal>node3</literal>. Incremental changes made in any node will be\\n> + replicated to the other two nodes.\\n> \\\"in any node\\\" -> \\\"on any node\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 6. doc/src/sgml/logical-replication.sgml - Adding a new node when data\\n> is present in the new node\\n>\\n> + <title>Adding a new node when data is present in the new node</title>\\n> SUGGESTION\\n> Adding a new node when table data is present on the new node\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. doc/src/sgml/logical-replication.sgml - Generic steps for adding a\\n> new node to an existing set of nodes\\n>\\n> 7a.\\n> + <para>\\n> + Step-2: Lock the required tables of the new node in EXCLUSIVE mode until\\n> + the setup is complete. (This lock is necessary to prevent any modifications\\n> + from happening in the new node because if data modifications occurred after\\n> + Step-3, there is a chance that the modifications will be published to the\\n> + first node and then synchronized back to the new node while creating the\\n> + subscription in Step-5. This would result in inconsistent data).\\n> + </para>\\n> \\\"happening in the new node\\\" -> \\\"happening on the new node\\\"\\n\\nModified\\n\\n> 7b.\\n> + not be synchronized to the new node. This would result in inconsistent\\n> + data. There is no need to lock the required tables in\\n> + <literal>node1</literal> because any data changes made will be synchronized\\n> + while creating the subscription with <literal>copy_data = force</literal>).\\n> + </para>\\n> \\\"no need to lock the required tables in\\\" -> \\\"no need to lock the\\n> required tables on\\\"\\n\\nModified\\n\\n> ======\\n>\\n> 8. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -403,7 +403,10 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> being subscribed to any other publisher and, if so, then throw an error to\\n> prevent possible non-local data from being copied. The user can override\\n> this check and continue with the copy operation by specifying\\n> - <literal>copy_data = force</literal>.\\n> + <literal>copy_data = force</literal>. Refer to\\n> + <xref linkend=\\\"logical-replication-bidirectional\\\"/> for how\\n> + <literal>copy_data</literal> and <literal>origin</literal> can be used\\n> + in bidirectional replication.\\n> </para>\\n>\\n> \\\"can be used in bidirectional replication\\\" -> \\\"can be used to set up\\n> bidirectional replication\\\"\\n\\nModified\\n\\nThanks for the comments, the v21 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm3%2B6cey0rcDft1ZUCjSUtLDM0xmU_Q%2BYhcsBrqe1RH8%3Dw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Jun 2022 15:53:18 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 16, 2022 6:18 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> Thanks for the comments, the attached v21 patch has the changes for the\\r\\n> same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch. Here are some comments.\\r\\n\\r\\n0002 patch\\r\\n==============\\r\\n1.\\r\\n+ publisher to only send changes that originated locally. Setting\\r\\n+ <literal>origin</literal> to <literal>any</literal> means that that\\r\\n+ the publisher sends any changes regardless of their origin. The\\r\\n+ default is <literal>any</literal>.\\r\\n\\r\\nIt seems there's a redundant \\\"that\\\" at the end of second line.\\r\\n\\r\\n2.\\r\\n+ <row>\\r\\n+ <entry role=\\\"catalog_table_entry\\\"><para role=\\\"column_definition\\\">\\r\\n+ <structfield>suborigin</structfield> <type>text</type>\\r\\n+ </para>\\r\\n+ <para>\\r\\n+ Possible origin values are <literal>local</literal> or\\r\\n+ <literal>any</literal>. The default is <literal>any</literal>.\\r\\n+ If <literal>local</literal>, the subscription will request the publisher\\r\\n+ to only send changes that originated locally. If <literal>any</literal>\\r\\n+ the publisher sends any changes regardless of their origin.\\r\\n+ </para></entry>\\r\\n+ </row>.\\r\\n\\r\\nA comma can be added after \\\"If any\\\".\\r\\n\\r\\n3.\\r\\n@@ -4589,6 +4598,8 @@ dumpSubscription(Archive *fout, const SubscriptionInfo *subinfo)\\r\\n \\tif (strcmp(subinfo->subdisableonerr, \\\"t\\\") == 0)\\r\\n \\t\\tappendPQExpBufferStr(query, \\\", disable_on_error = true\\\");\\r\\n \\r\\n+\\tappendPQExpBuffer(query, \\\", origin = %s\\\", subinfo->suborigin);\\r\\n+\\r\\n \\tif (strcmp(subinfo->subsynccommit, \\\"off\\\") != 0)\\r\\n \\t\\tappendPQExpBuffer(query, \\\", synchronous_commit = %s\\\", fmtId(subinfo->subsynccommit));\\r\\n\\r\\nDo we need to append anything if it's the default value (\\\"any\\\")? I saw that some\\r\\nother parameters, they will be appended only if they are not the default value.\\r\\n\\r\\n0003 patch\\r\\n==============\\r\\n1. \\r\\nin create_subscription.sgml:\\r\\n (You cannot combine setting <literal>connect</literal>\\r\\n to <literal>false</literal> with\\r\\n setting <literal>create_slot</literal>, <literal>enabled</literal>,\\r\\n or <literal>copy_data</literal> to <literal>true</literal>.)\\r\\n\\r\\nIn this description about \\\"connect\\\" parameter in CREATE SUBSCIPTION document,\\r\\nmaybe it would be better to change \\\"copy_data to true\\\" to \\\"copy_data to\\r\\ntrue/force\\\".\\r\\n\\r\\n2.\\r\\n+\\tappendStringInfoString(&cmd,\\r\\n+\\t\\t\\t\\t\\t\\t \\\"SELECT DISTINCT N.nspname AS schemaname,\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"\\t\\t\\t\\tC.relname AS tablename,\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"\\t\\t\\t\\tPS.srrelid as replicated\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"FROM pg_publication P,\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"\\t LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"\\t LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"\\t pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"WHERE C.oid = GPT.relid AND P.pubname IN (\\\");\\r\\n\\r\\n\\\"PS.srrelid as replicated\\\" can be modified to \\\"PS.srrelid AS replicated\\\".\\r\\n\\r\\nBesides, I think we can filter out the tables which are not subscribing data in\\r\\nthis SQL statement, then later processing can be simplified.\\r\\n\\r\\nSomething like:\\r\\nSELECT DISTINCT N.nspname AS schemaname,\\r\\n C.relname AS tablename\\r\\nFROM pg_publication P,\\r\\n LATERAL pg_get_publication_tables(P.pubname) GPT\\r\\n LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\r\\n pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\r\\nWHERE C.oid = GPT.relid AND P.pubname IN ('pa') AND PS.srrelid IS NOT NULL;\\r\\n\\r\\n0004 patch\\r\\n==============\\r\\n1. Generic steps for adding a new node to an existing set of nodes\\r\\n\\r\\n+ Step-2: Lock the required tables of the new node in EXCLUSIVE mode until\\r\\n+ the setup is complete. (This lock is necessary to prevent any modifications\\r\\n\\r\\n+ Step-4. Lock the required tables of the existing nodes except the first node\\r\\n+ in EXCLUSIVE mode until the setup is complete. (This lock is necessary to\\r\\n\\r\\nShould \\\"in EXCLUSIVE mode\\\" be modified to \\\"in <literal>EXCLUSIVE</literal>\\r\\nmode\\\"?\\r\\n\\r\\n2. Generic steps for adding a new node to an existing set of nodes\\r\\n\\r\\n+ data. There is no need to lock the required tables on\\r\\n+ <literal>node1</literal> because any data changes made will be synchronized\\r\\n+ while creating the subscription with <literal>copy_data = force</literal>).\\r\\n\\r\\nI think it would be better to say \\\"on the first node\\\" here, instead of \\\"node1\\\",\\r\\nbecause in this section, node1 is not mentioned before.\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\\r\\n\\r\\n\",\n \"msg_date\": \"Mon, 20 Jun 2022 03:52:01 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 16, 2022 at 3:48 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n\\n> Thanks for the comments, the attached v21 patch has the changes for the same.\\n\\nI have done some basic review of v21 and I have a few comments,\\n\\n1.\\n+/*\\n+ * The subscription will request the publisher to only send changes that\\n+ * originated locally.\\n+ */\\n+#define LOGICALREP_ORIGIN_LOCAL \\\"local\\\"\\n+\\n+/*\\n+ * The subscription will request the publisher to send any changes regardless\\n+ * of their origin\\n+ */\\n+#define LOGICALREP_ORIGIN_ANY \\\"any\\\"\\n\\nAre we planning to extend this to more options or are we planning to\\nsupport the actual origin name here? If not then why isn't it just\\nbool? I think the comments and the patch commit message should\\nexplain the details behind it if it has been already discussed and\\nconcluded.\\n\\n2.\\n+/*\\n+ * Check and throw an error if the publisher has subscribed to the same table\\n+ * from some other publisher. This check is required only if copydata is ON and\\n+ * the origin is local.\\n+ */\\n\\nI think it should also explain why this combination is not allowed and\\nif it is already explained in code\\nthen this code can add comments to refer to that part of the code.\\n\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 20 Jun 2022 14:37:32 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 20, 2022 at 2:37 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>\\n> On Thu, Jun 16, 2022 at 3:48 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > >\\n>\\n> > Thanks for the comments, the attached v21 patch has the changes for the same.\\n>\\n> I have done some basic review of v21 and I have a few comments,\\n>\\n> 1.\\n> +/*\\n> + * The subscription will request the publisher to only send changes that\\n> + * originated locally.\\n> + */\\n> +#define LOGICALREP_ORIGIN_LOCAL \\\"local\\\"\\n> +\\n> +/*\\n> + * The subscription will request the publisher to send any changes regardless\\n> + * of their origin\\n> + */\\n> +#define LOGICALREP_ORIGIN_ANY \\\"any\\\"\\n>\\n> Are we planning to extend this to more options or are we planning to\\n> support the actual origin name here? If not then why isn't it just\\n> bool? I think the comments and the patch commit message should\\n> explain the details behind it if it has been already discussed and\\n> concluded.\\n\\nCurrently we only support local and any. But this was designed to\\naccept string instead of boolean type, so that it can be extended\\nlater to support filtering of origin names specified by the user in\\nthe later versions. The same was also discussed in pg unconference as\\nmentioned in [1]. I will add it to the commit message and a comment\\nfor the same in the next version.\\n\\n> 2.\\n> +/*\\n> + * Check and throw an error if the publisher has subscribed to the same table\\n> + * from some other publisher. This check is required only if copydata is ON and\\n> + * the origin is local.\\n> + */\\n>\\n> I think it should also explain why this combination is not allowed and\\n> if it is already explained in code\\n> then this code can add comments to refer to that part of the code.\\n\\nIn the same function, the reason for this is mentioned detailly just\\nabove the place where error is thrown. I think that should be enough.\\nHave a look and let me know if that is not sufficient:\\n+ /*\\n+ * Throw an error if the publisher has subscribed to\\nthe same table\\n+ * from some other publisher. We cannot differentiate\\nbetween the\\n+ * local and non-local data that is present in the\\nHEAP during the\\n+ * initial sync. Identification of local data can be\\ndone only from\\n+ * the WAL by using the origin id. XXX: For\\nsimplicity, we don't check\\n+ * whether the table has any data or not. If the table\\ndoesn't have\\n+ * any data then we don't need to distinguish between local and\\n+ * non-local data so we can avoid throwing error in that case.\\n+ */\\n+ if (!slot_attisnull(slot, 3))\\n+ ereport(ERROR,\\n+\\nerrcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n+ errmsg(\\\"table:%s.%s might have\\nreplicated data in the publisher\\\",\\n+ nspname, relname),\\n+ errdetail(\\\"CREATE/ALTER\\nSUBSCRIPTION with origin = local and copy_data = on is not allowed\\nwhen the publisher might have replicated data.\\\"),\\n+ errhint(\\\"Use CREATE/ALTER\\nSUBSCRIPTION with copy_data = off/force.\\\"));\\n\\n\\n[1] - https://wiki.postgresql.org/wiki/PgCon_2022_Developer_Unconference#Logical_Replication_Origin_Filtering_and_Consistency\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 20 Jun 2022 15:16:43 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 20, 2022 at 3:16 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Jun 20, 2022 at 2:37 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >\\n> > On Thu, Jun 16, 2022 at 3:48 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > > >\\n> >\\n> > > Thanks for the comments, the attached v21 patch has the changes for the same.\\n> >\\n> > I have done some basic review of v21 and I have a few comments,\\n> >\\n> > 1.\\n> > +/*\\n> > + * The subscription will request the publisher to only send changes that\\n> > + * originated locally.\\n> > + */\\n> > +#define LOGICALREP_ORIGIN_LOCAL \\\"local\\\"\\n> > +\\n> > +/*\\n> > + * The subscription will request the publisher to send any changes regardless\\n> > + * of their origin\\n> > + */\\n> > +#define LOGICALREP_ORIGIN_ANY \\\"any\\\"\\n> >\\n> > Are we planning to extend this to more options or are we planning to\\n> > support the actual origin name here? If not then why isn't it just\\n> > bool? I think the comments and the patch commit message should\\n> > explain the details behind it if it has been already discussed and\\n> > concluded.\\n>\\n> Currently we only support local and any. But this was designed to\\n> accept string instead of boolean type, so that it can be extended\\n> later to support filtering of origin names specified by the user in\\n> the later versions. The same was also discussed in pg unconference as\\n> mentioned in [1]. I will add it to the commit message and a comment\\n> for the same in the next version.\\n>\\n> > 2.\\n> > +/*\\n> > + * Check and throw an error if the publisher has subscribed to the same table\\n> > + * from some other publisher. This check is required only if copydata is ON and\\n> > + * the origin is local.\\n> > + */\\n> >\\n> > I think it should also explain why this combination is not allowed and\\n> > if it is already explained in code\\n> > then this code can add comments to refer to that part of the code.\\n>\\n> In the same function, the reason for this is mentioned detailly just\\n> above the place where error is thrown. I think that should be enough.\\n> Have a look and let me know if that is not sufficient:\\n\\nI think that should be sufficient, thanks.\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 20 Jun 2022 16:02:13 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 20, 2022 at 9:22 AM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Thu, Jun 16, 2022 6:18 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached v21 patch has the changes for the\\n> > same.\\n> >\\n>\\n> Thanks for updating the patch. Here are some comments.\\n>\\n> 0002 patch\\n> ==============\\n> 1.\\n> + publisher to only send changes that originated locally. Setting\\n> + <literal>origin</literal> to <literal>any</literal> means that that\\n> + the publisher sends any changes regardless of their origin. The\\n> + default is <literal>any</literal>.\\n>\\n> It seems there's a redundant \\\"that\\\" at the end of second line.\\n\\nModified\\n\\n> 2.\\n> + <row>\\n> + <entry role=\\\"catalog_table_entry\\\"><para role=\\\"column_definition\\\">\\n> + <structfield>suborigin</structfield> <type>text</type>\\n> + </para>\\n> + <para>\\n> + Possible origin values are <literal>local</literal> or\\n> + <literal>any</literal>. The default is <literal>any</literal>.\\n> + If <literal>local</literal>, the subscription will request the publisher\\n> + to only send changes that originated locally. If <literal>any</literal>\\n> + the publisher sends any changes regardless of their origin.\\n> + </para></entry>\\n> + </row>.\\n>\\n> A comma can be added after \\\"If any\\\".\\n\\nModified\\n\\n> 3.\\n> @@ -4589,6 +4598,8 @@ dumpSubscription(Archive *fout, const SubscriptionInfo *subinfo)\\n> if (strcmp(subinfo->subdisableonerr, \\\"t\\\") == 0)\\n> appendPQExpBufferStr(query, \\\", disable_on_error = true\\\");\\n>\\n> + appendPQExpBuffer(query, \\\", origin = %s\\\", subinfo->suborigin);\\n> +\\n> if (strcmp(subinfo->subsynccommit, \\\"off\\\") != 0)\\n> appendPQExpBuffer(query, \\\", synchronous_commit = %s\\\", fmtId(subinfo->subsynccommit));\\n>\\n> Do we need to append anything if it's the default value (\\\"any\\\")? I saw that some\\n> other parameters, they will be appended only if they are not the default value.\\n\\nModified\\n\\n> 0003 patch\\n> ==============\\n> 1.\\n> in create_subscription.sgml:\\n> (You cannot combine setting <literal>connect</literal>\\n> to <literal>false</literal> with\\n> setting <literal>create_slot</literal>, <literal>enabled</literal>,\\n> or <literal>copy_data</literal> to <literal>true</literal>.)\\n>\\n> In this description about \\\"connect\\\" parameter in CREATE SUBSCIPTION document,\\n> maybe it would be better to change \\\"copy_data to true\\\" to \\\"copy_data to\\n> true/force\\\".\\n\\nModified\\n\\n> 2.\\n> + appendStringInfoString(&cmd,\\n> + \\\"SELECT DISTINCT N.nspname AS schemaname,\\\\n\\\"\\n> + \\\" C.relname AS tablename,\\\\n\\\"\\n> + \\\" PS.srrelid as replicated\\\\n\\\"\\n> + \\\"FROM pg_publication P,\\\\n\\\"\\n> + \\\" LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n> + \\\" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n> + \\\" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n> + \\\"WHERE C.oid = GPT.relid AND P.pubname IN (\\\");\\n>\\n> \\\"PS.srrelid as replicated\\\" can be modified to \\\"PS.srrelid AS replicated\\\".\\n>\\n> Besides, I think we can filter out the tables which are not subscribing data in\\n> this SQL statement, then later processing can be simplified.\\n>\\n> Something like:\\n> SELECT DISTINCT N.nspname AS schemaname,\\n> C.relname AS tablename\\n> FROM pg_publication P,\\n> LATERAL pg_get_publication_tables(P.pubname) GPT\\n> LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n> pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n> WHERE C.oid = GPT.relid AND P.pubname IN ('pa') AND PS.srrelid IS NOT NULL;\\n\\nModified\\n\\n> 0004 patch\\n> ==============\\n> 1. Generic steps for adding a new node to an existing set of nodes\\n>\\n> + Step-2: Lock the required tables of the new node in EXCLUSIVE mode until\\n> + the setup is complete. (This lock is necessary to prevent any modifications\\n>\\n> + Step-4. Lock the required tables of the existing nodes except the first node\\n> + in EXCLUSIVE mode until the setup is complete. (This lock is necessary to\\n>\\n> Should \\\"in EXCLUSIVE mode\\\" be modified to \\\"in <literal>EXCLUSIVE</literal>\\n> mode\\\"?\\n\\nModified\\n\\n> 2. Generic steps for adding a new node to an existing set of nodes\\n>\\n> + data. There is no need to lock the required tables on\\n> + <literal>node1</literal> because any data changes made will be synchronized\\n> + while creating the subscription with <literal>copy_data = force</literal>).\\n>\\n> I think it would be better to say \\\"on the first node\\\" here, instead of \\\"node1\\\",\\n> because in this section, node1 is not mentioned before.\\n\\nModified\\n\\nThanks for the comment, the v22 patch attached has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Mon, 20 Jun 2022 17:25:24 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 20, 2022 at 3:16 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Jun 20, 2022 at 2:37 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >\\n> > On Thu, Jun 16, 2022 at 3:48 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > On Wed, Jun 15, 2022 at 12:09 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > > >\\n> >\\n> > > Thanks for the comments, the attached v21 patch has the changes for the same.\\n> >\\n> > I have done some basic review of v21 and I have a few comments,\\n> >\\n> > 1.\\n> > +/*\\n> > + * The subscription will request the publisher to only send changes that\\n> > + * originated locally.\\n> > + */\\n> > +#define LOGICALREP_ORIGIN_LOCAL \\\"local\\\"\\n> > +\\n> > +/*\\n> > + * The subscription will request the publisher to send any changes regardless\\n> > + * of their origin\\n> > + */\\n> > +#define LOGICALREP_ORIGIN_ANY \\\"any\\\"\\n> >\\n> > Are we planning to extend this to more options or are we planning to\\n> > support the actual origin name here? If not then why isn't it just\\n> > bool? I think the comments and the patch commit message should\\n> > explain the details behind it if it has been already discussed and\\n> > concluded.\\n>\\n> Currently we only support local and any. But this was designed to\\n> accept string instead of boolean type, so that it can be extended\\n> later to support filtering of origin names specified by the user in\\n> the later versions. The same was also discussed in pg unconference as\\n> mentioned in [1]. I will add it to the commit message and a comment\\n> for the same in the next version.\\n\\nThanks for the comment, the v22 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm1h-9UNi_Jo_K%2BPK34tXBmV7fhj5C_nB8YzGA9rmUwHEA%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 20 Jun 2022 17:27:29 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for the v22* patch set.\\n\\n========\\nv22-0001\\n========\\n\\nNo comments. LGTM\\n\\n========\\nV22-0002\\n========\\n\\n2.1 doc/src/sgml/catalogs.sgml\\n\\n+ Possible origin values are <literal>local</literal> or\\n+ <literal>any</literal>. The default is <literal>any</literal>.\\n\\nIMO the word \\\"Possible\\\" here is giving a sense of vagueness.\\n\\nSUGGESTION\\nThe origin value must be either <literal>local</literal> or\\n<literal>any</literal>.\\n\\n~~~\\n\\n2.2 src/backend/commands/subscriptioncmds.c\\n\\n@@ -265,6 +271,29 @@ parse_subscription_options(ParseState *pstate,\\nList *stmt_options,\\n opts->specified_opts |= SUBOPT_DISABLE_ON_ERR;\\n opts->disableonerr = defGetBoolean(defel);\\n }\\n+ else if (IsSet(supported_opts, SUBOPT_ORIGIN) &&\\n+ strcmp(defel->defname, \\\"origin\\\") == 0)\\n+ {\\n+ if (IsSet(opts->specified_opts, SUBOPT_ORIGIN))\\n+ errorConflictingDefElem(defel, pstate);\\n+\\n+ opts->specified_opts |= SUBOPT_ORIGIN;\\n+\\n+ /*\\n+ * Even though \\\"origin\\\" parameter allows only \\\"local\\\" and \\\"any\\\"\\n+ * values, the \\\"origin\\\" parameter type is implemented as string\\n+ * type instead of boolean to extend the \\\"origin\\\" parameter to\\n+ * support filtering of origin name specified by the user in the\\n+ * later versions.\\n+ */\\n+ opts->origin = defGetString(defel);\\n+\\n+ if ((strcmp(opts->origin, LOGICALREP_ORIGIN_LOCAL) != 0) &&\\n+ (strcmp(opts->origin, LOGICALREP_ORIGIN_ANY) != 0))\\n+ ereport(ERROR,\\n+ errcode(ERRCODE_SYNTAX_ERROR),\\n+ errmsg(\\\"unrecognized origin value: \\\\\\\"%s\\\\\\\"\\\", opts->origin));\\n+ }\\n\\nI was wondering if it might be wise now to do a pfree(opts->origin)\\nhere before setting the new option value which overwrites the\\nstrdup-ed default \\\"any\\\". OTOH maybe it is overkill to worry about the\\ntiny leak? I am not sure what is the convention for this.\\n\\n~~~\\n\\n2.3 src/backend/replication/pgoutput/pgoutput.c\\n\\n@@ -1698,12 +1719,16 @@ pgoutput_message(LogicalDecodingContext *ctx,\\nReorderBufferTXN *txn,\\n }\\n\\n /*\\n- * Currently we always forward.\\n+ * Return true if the data source (origin) is remote and user has requested\\n+ * only local data, false otherwise.\\n */\\n static bool\\n pgoutput_origin_filter(LogicalDecodingContext *ctx,\\n RepOriginId origin_id)\\n\\n\\\"user\\\" -> \\\"the user\\\"\\n\\n~~~\\n\\n2.4 src/include/catalog/pg_subscription.h\\n\\n+/*\\n+ * The subscription will request the publisher to send any changes regardless\\n+ * of their origin\\n+ */\\n+#define LOGICALREP_ORIGIN_ANY \\\"any\\\"\\n\\nSUGGESTION (remove 'any'; add period)\\nThe subscription will request the publisher to send changes regardless\\nof their origin.\\n\\n========\\nv22-0003\\n========\\n\\n3.1 Commit message\\n\\nchange 1) Checks and throws an error if 'copy_data = on' and 'origin =\\nlocal' but the publication tables were also subscribing from other publishers.\\n\\n\\\"also subscribing from other publishers.\\\" -> \\\"also replicating from\\nother publishers.\\\"\\n\\n~~~\\n\\n3.2 doc/src/sgml/ref/alter_subscription.sgml\\n\\n+ <para>\\n+ There is some interaction between the <literal>origin</literal>\\n+ parameter and <literal>copy_data</literal> parameter.\\n\\n\\\"and copy_data parameter.\\\" -> \\\"and the copy_data parameter.\\\"\\n\\n~~~\\n\\n3.3 doc/src/sgml/ref/create_subscription.sgml\\n\\n+ <para>\\n+ There is some interaction between the <literal>origin</literal>\\n+ parameter and <literal>copy_data</literal> parameter.\\n\\n\\\"and copy_data parameter.\\\" -> \\\"and the copy_data parameter.\\\"\\n\\n~~~\\n\\n3.4 doc/src/sgml/ref/create_subscription.sgml\\n\\n+ <para>\\n+ There is some interaction between the <literal>origin</literal>\\n+ parameter and <literal>copy_data</literal> parameter. Refer to the\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n+ </para>\\n\\n\\\"and copy_data parameter.\\\" -> \\\"and the copy_data parameter.\\\"\\n\\n~~~\\n\\n3.5 doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n can have non-existent publications.\\n </para>\\n\\n+ <para>\\n+ If the subscription is created with <literal>origin = local</literal> and\\n+ <literal>copy_data = true</literal>, it will check if the\\npublisher tables are\\n+ being subscribed to any other publisher and, if so, then throw an error to\\n+ prevent possible non-local data from being copied. The user can override\\n+ this check and continue with the copy operation by specifying\\n+ <literal>copy_data = force</literal>.\\n+ </para>\\n+\\n\\nFor the part \\\"it will check if the publisher tables are being\\nsubscribed to any other publisher...\\\", the \\\"being subscribed to\\\"\\nsounds a bit strange. Maybe it is better to reword this like you\\nalready worded some of the code comments.\\n\\nSUGGESTION\\n-> \\\"... it will check if the publisher has subscribed to the same\\ntable from other publishers and ...\\\"\\n\\n~~~\\n\\n3.6 src/backend/commands/subscriptioncmds.c\\n\\n+ /*\\n+ * Throw an error if the publisher has subscribed to the same table\\n+ * from some other publisher. We cannot differentiate between the\\n+ * local and non-local data that is present in the HEAP during the\\n+ * initial sync. Identification of local data can be done only from\\n+ * the WAL by using the origin id. XXX: For simplicity, we don't check\\n+ * whether the table has any data or not. If the table doesn't have\\n+ * any data then we don't need to distinguish between local and\\n+ * non-local data so we can avoid throwing error in that case.\\n+ */\\n\\nI think the XXX part should be after a blank like so it is more\\nprominent, instead of being buried in the other text.\\n\\ne.g.\\n\\n+ /*\\n+ * Throw an error if the publisher has subscribed to the same table\\n+ * from some other publisher. We cannot differentiate between the\\n+ * local and non-local data that is present in the HEAP during the\\n+ * initial sync. Identification of local data can be done only from\\n+ * the WAL by using the origin id.\\n+ *\\n+ * XXX: For simplicity, we don't check\\n+ * whether the table has any data or not. If the table doesn't have\\n+ * any data then we don't need to distinguish between local and\\n+ * non-local data so we can avoid throwing error in that case.\\n+ */\\n\\n~~~\\n\\n3.7 src/test/regress/sql/subscription.sql\\n\\nElsewhere, there was code in this patch that apparently accepts\\n'copy_data' parameter but with no parameter value specified. But this\\ncombination has no test case.\\n\\n========\\nv22-0004\\n========\\n\\n4.1 doc/src/sgml/logical-replication.sgml\\n\\n+ <sect2 id=\\\"add-node-data-present-on-new-node\\\">\\n+ <title>Adding a new node when data is present on the new node</title>\\n\\n\\\"when data is present\\\" -> \\\"when table data is present\\\"\\n\\n~~~\\n\\n4.2 doc/src/sgml/logical-replication.sgml\\n\\n+ <note>\\n+ <para>\\n+ Adding a new node when data is present on the new node tables is not\\n+ supported.\\n+ </para>\\n+ </note>\\n\\nI think the note text should match the title text (see #4.1)\\n\\nSUGGESTION\\nAdding a new node when table data is present on the new node is not supported.\\n\\n~~~\\n\\n4.3\\n\\n+ <para>\\n+ Step-2: Lock the required tables of the new node in\\n+ <literal>EXCLUSIVE</literal> mode untilthe setup is complete. (This lock is\\n+ necessary to prevent any modifications from happening on the new node\\n+ because if data modifications occurred after Step-3, there is a chance that\\n+ the modifications will be published to the first node and then synchronized\\n+ back to the new node while creating the subscription in Step-5. This would\\n+ result in inconsistent data).\\n+ </para>\\n+ <para>\\n\\n4.3.a\\ntypo: \\\"untilthe\\\" -> \\\"until the\\\"\\n\\n4.3.b\\nSUGGESTION (just for the 2nd sentence)\\nThis lock is necessary to prevent any modifications from happening on\\nthe new node. If data modifications occurred after Step-3, there is a\\nchance they could be published to the first node and then synchronized\\nback to the new node while creating the subscription in Step-5.\\n\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 22 Jun 2022 16:46:15 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 20, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> Thanks for the comment, the v22 patch attached has the changes for the\\r\\n> same.\\r\\n\\r\\nThanks for updating the patch, here are some comments on 0003 patch.\\r\\n\\r\\n1. 032_origin.pl\\r\\n+###############################################################################\\r\\n+# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\\r\\n+# replication setup when the existing nodes (node_A & node_B) and the new node\\r\\n+# (node_C) does not have any data.\\r\\n+###############################################################################\\r\\n\\r\\n\\\"does not have any data\\\" should be \\\"do not have any data\\\" I think.\\r\\n\\r\\n2.\\r\\nThe comment for 032_origin.pl:\\r\\n\\r\\n# Test the CREATE SUBSCRIPTION 'origin' parameter.\\r\\n\\r\\nAfter applying this patch, this file tests no only 'origin' parameter, but also\\r\\n\\\"copy_data\\\" parameter, so should we modify this comment?\\r\\n\\r\\nBesides, should we change the file name in this patch? It looks more like test\\r\\ncases for bidirectional logical replication.\\r\\n\\r\\n3. subscriptioncmds.c\\r\\n\\t/* Set default values for the boolean supported options. */\\r\\n...\\r\\n\\tif (IsSet(supported_opts, SUBOPT_CREATE_SLOT))\\r\\n \\t\\topts->create_slot = true;\\r\\n \\tif (IsSet(supported_opts, SUBOPT_COPY_DATA))\\r\\n-\\t\\topts->copy_data = true;\\r\\n+\\t\\topts->copy_data = COPY_DATA_ON;\\r\\n \\tif (IsSet(supported_opts, SUBOPT_REFRESH))\\r\\n \\t\\topts->refresh = true;\\r\\n \\tif (IsSet(supported_opts, SUBOPT_BINARY))\\r\\n\\r\\n\\\"copy_data\\\" option is not Boolean now, which is inconsistent with the comment.\\r\\nSo maybe we can change the comment here? (\\\"the boolean supported options\\\" ->\\r\\n\\\"the supported options\\\")\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Thu, 23 Jun 2022 01:35:15 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 22, 2022 at 12:16 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for the v22* patch set.\\n>\\n> ========\\n> v22-0001\\n> ========\\n>\\n> No comments. LGTM\\n>\\n> ========\\n> V22-0002\\n> ========\\n>\\n> 2.1 doc/src/sgml/catalogs.sgml\\n>\\n> + Possible origin values are <literal>local</literal> or\\n> + <literal>any</literal>. The default is <literal>any</literal>.\\n>\\n> IMO the word \\\"Possible\\\" here is giving a sense of vagueness.\\n>\\n> SUGGESTION\\n> The origin value must be either <literal>local</literal> or\\n> <literal>any</literal>.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.2 src/backend/commands/subscriptioncmds.c\\n>\\n> @@ -265,6 +271,29 @@ parse_subscription_options(ParseState *pstate,\\n> List *stmt_options,\\n> opts->specified_opts |= SUBOPT_DISABLE_ON_ERR;\\n> opts->disableonerr = defGetBoolean(defel);\\n> }\\n> + else if (IsSet(supported_opts, SUBOPT_ORIGIN) &&\\n> + strcmp(defel->defname, \\\"origin\\\") == 0)\\n> + {\\n> + if (IsSet(opts->specified_opts, SUBOPT_ORIGIN))\\n> + errorConflictingDefElem(defel, pstate);\\n> +\\n> + opts->specified_opts |= SUBOPT_ORIGIN;\\n> +\\n> + /*\\n> + * Even though \\\"origin\\\" parameter allows only \\\"local\\\" and \\\"any\\\"\\n> + * values, the \\\"origin\\\" parameter type is implemented as string\\n> + * type instead of boolean to extend the \\\"origin\\\" parameter to\\n> + * support filtering of origin name specified by the user in the\\n> + * later versions.\\n> + */\\n> + opts->origin = defGetString(defel);\\n> +\\n> + if ((strcmp(opts->origin, LOGICALREP_ORIGIN_LOCAL) != 0) &&\\n> + (strcmp(opts->origin, LOGICALREP_ORIGIN_ANY) != 0))\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_SYNTAX_ERROR),\\n> + errmsg(\\\"unrecognized origin value: \\\\\\\"%s\\\\\\\"\\\", opts->origin));\\n> + }\\n>\\n> I was wondering if it might be wise now to do a pfree(opts->origin)\\n> here before setting the new option value which overwrites the\\n> strdup-ed default \\\"any\\\". OTOH maybe it is overkill to worry about the\\n> tiny leak? I am not sure what is the convention for this.\\n\\nI have freed the memory to avoid the leak, I felt it is better to\\nclean the memory as it is a positive flow.\\n\\n> ~~~\\n>\\n> 2.3 src/backend/replication/pgoutput/pgoutput.c\\n>\\n> @@ -1698,12 +1719,16 @@ pgoutput_message(LogicalDecodingContext *ctx,\\n> ReorderBufferTXN *txn,\\n> }\\n>\\n> /*\\n> - * Currently we always forward.\\n> + * Return true if the data source (origin) is remote and user has requested\\n> + * only local data, false otherwise.\\n> */\\n> static bool\\n> pgoutput_origin_filter(LogicalDecodingContext *ctx,\\n> RepOriginId origin_id)\\n>\\n> \\\"user\\\" -> \\\"the user\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.4 src/include/catalog/pg_subscription.h\\n>\\n> +/*\\n> + * The subscription will request the publisher to send any changes regardless\\n> + * of their origin\\n> + */\\n> +#define LOGICALREP_ORIGIN_ANY \\\"any\\\"\\n>\\n> SUGGESTION (remove 'any'; add period)\\n> The subscription will request the publisher to send changes regardless\\n> of their origin.\\n\\nModified\\n\\n> ========\\n> v22-0003\\n> ========\\n>\\n> 3.1 Commit message\\n>\\n> change 1) Checks and throws an error if 'copy_data = on' and 'origin =\\n> local' but the publication tables were also subscribing from other publishers.\\n>\\n> \\\"also subscribing from other publishers.\\\" -> \\\"also replicating from\\n> other publishers.\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 3.2 doc/src/sgml/ref/alter_subscription.sgml\\n>\\n> + <para>\\n> + There is some interaction between the <literal>origin</literal>\\n> + parameter and <literal>copy_data</literal> parameter.\\n>\\n> \\\"and copy_data parameter.\\\" -> \\\"and the copy_data parameter.\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 3.3 doc/src/sgml/ref/create_subscription.sgml\\n>\\n> + <para>\\n> + There is some interaction between the <literal>origin</literal>\\n> + parameter and <literal>copy_data</literal> parameter.\\n>\\n> \\\"and copy_data parameter.\\\" -> \\\"and the copy_data parameter.\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 3.4 doc/src/sgml/ref/create_subscription.sgml\\n>\\n> + <para>\\n> + There is some interaction between the <literal>origin</literal>\\n> + parameter and <literal>copy_data</literal> parameter. Refer to the\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for details.\\n> + </para>\\n>\\n> \\\"and copy_data parameter.\\\" -> \\\"and the copy_data parameter.\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 3.5 doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> can have non-existent publications.\\n> </para>\\n>\\n> + <para>\\n> + If the subscription is created with <literal>origin = local</literal> and\\n> + <literal>copy_data = true</literal>, it will check if the\\n> publisher tables are\\n> + being subscribed to any other publisher and, if so, then throw an error to\\n> + prevent possible non-local data from being copied. The user can override\\n> + this check and continue with the copy operation by specifying\\n> + <literal>copy_data = force</literal>.\\n> + </para>\\n> +\\n>\\n> For the part \\\"it will check if the publisher tables are being\\n> subscribed to any other publisher...\\\", the \\\"being subscribed to\\\"\\n> sounds a bit strange. Maybe it is better to reword this like you\\n> already worded some of the code comments.\\n>\\n> SUGGESTION\\n> -> \\\"... it will check if the publisher has subscribed to the same\\n> table from other publishers and ...\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 3.6 src/backend/commands/subscriptioncmds.c\\n>\\n> + /*\\n> + * Throw an error if the publisher has subscribed to the same table\\n> + * from some other publisher. We cannot differentiate between the\\n> + * local and non-local data that is present in the HEAP during the\\n> + * initial sync. Identification of local data can be done only from\\n> + * the WAL by using the origin id. XXX: For simplicity, we don't check\\n> + * whether the table has any data or not. If the table doesn't have\\n> + * any data then we don't need to distinguish between local and\\n> + * non-local data so we can avoid throwing error in that case.\\n> + */\\n>\\n> I think the XXX part should be after a blank like so it is more\\n> prominent, instead of being buried in the other text.\\n>\\n> e.g.\\n>\\n> + /*\\n> + * Throw an error if the publisher has subscribed to the same table\\n> + * from some other publisher. We cannot differentiate between the\\n> + * local and non-local data that is present in the HEAP during the\\n> + * initial sync. Identification of local data can be done only from\\n> + * the WAL by using the origin id.\\n> + *\\n> + * XXX: For simplicity, we don't check\\n> + * whether the table has any data or not. If the table doesn't have\\n> + * any data then we don't need to distinguish between local and\\n> + * non-local data so we can avoid throwing error in that case.\\n> + */\\n\\nModified\\n\\n> ~~~\\n>\\n> 3.7 src/test/regress/sql/subscription.sql\\n>\\n> Elsewhere, there was code in this patch that apparently accepts\\n> 'copy_data' parameter but with no parameter value specified. But this\\n> combination has no test case.\\n\\nI have added a test for the same\\n\\n> ========\\n> v22-0004\\n> ========\\n>\\n> 4.1 doc/src/sgml/logical-replication.sgml\\n>\\n> + <sect2 id=\\\"add-node-data-present-on-new-node\\\">\\n> + <title>Adding a new node when data is present on the new node</title>\\n>\\n> \\\"when data is present\\\" -> \\\"when table data is present\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 4.2 doc/src/sgml/logical-replication.sgml\\n>\\n> + <note>\\n> + <para>\\n> + Adding a new node when data is present on the new node tables is not\\n> + supported.\\n> + </para>\\n> + </note>\\n>\\n> I think the note text should match the title text (see #4.1)\\n>\\n> SUGGESTION\\n> Adding a new node when table data is present on the new node is not supported.\\n\\nModified\\n\\n> ~~~\\n>\\n> 4.3\\n>\\n> + <para>\\n> + Step-2: Lock the required tables of the new node in\\n> + <literal>EXCLUSIVE</literal> mode untilthe setup is complete. (This lock is\\n> + necessary to prevent any modifications from happening on the new node\\n> + because if data modifications occurred after Step-3, there is a chance that\\n> + the modifications will be published to the first node and then synchronized\\n> + back to the new node while creating the subscription in Step-5. This would\\n> + result in inconsistent data).\\n> + </para>\\n> + <para>\\n>\\n> 4.3.a\\n> typo: \\\"untilthe\\\" -> \\\"until the\\\"\\n\\nModified\\n\\n> 4.3.b\\n> SUGGESTION (just for the 2nd sentence)\\n> This lock is necessary to prevent any modifications from happening on\\n> the new node. If data modifications occurred after Step-3, there is a\\n> chance they could be published to the first node and then synchronized\\n> back to the new node while creating the subscription in Step-5.\\n\\nModified\\n\\nThanks for the comments, the attached v23 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Thu, 23 Jun 2022 15:36:45 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 23, 2022 at 7:05 AM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Mon, Jun 20, 2022 7:55 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comment, the v22 patch attached has the changes for the\\n> > same.\\n>\\n> Thanks for updating the patch, here are some comments on 0003 patch.\\n>\\n> 1. 032_origin.pl\\n> +###############################################################################\\n> +# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\\n> +# replication setup when the existing nodes (node_A & node_B) and the new node\\n> +# (node_C) does not have any data.\\n> +###############################################################################\\n>\\n> \\\"does not have any data\\\" should be \\\"do not have any data\\\" I think.\\n\\nI felt the existing is ok, I did not find it gramatically wrong. I did\\nnot make any changes for this.\\n\\n> 2.\\n> The comment for 032_origin.pl:\\n>\\n> # Test the CREATE SUBSCRIPTION 'origin' parameter.\\n>\\n> After applying this patch, this file tests no only 'origin' parameter, but also\\n> \\\"copy_data\\\" parameter, so should we modify this comment?\\n\\nModified\\n\\n> Besides, should we change the file name in this patch? It looks more like test\\n> cases for bidirectional logical replication.\\n\\nI felt let's keep it as origin, as most of the tests are based on\\n'origin' parameter, also later features will be able to filter a\\nparticular origin, those tests can easily fit in the same file.\\n\\n> 3. subscriptioncmds.c\\n> /* Set default values for the boolean supported options. */\\n> ...\\n> if (IsSet(supported_opts, SUBOPT_CREATE_SLOT))\\n> opts->create_slot = true;\\n> if (IsSet(supported_opts, SUBOPT_COPY_DATA))\\n> - opts->copy_data = true;\\n> + opts->copy_data = COPY_DATA_ON;\\n> if (IsSet(supported_opts, SUBOPT_REFRESH))\\n> opts->refresh = true;\\n> if (IsSet(supported_opts, SUBOPT_BINARY))\\n>\\n> \\\"copy_data\\\" option is not Boolean now, which is inconsistent with the comment.\\n> So maybe we can change the comment here? (\\\"the boolean supported options\\\" ->\\n> \\\"the supported options\\\")\\n\\nModified\\n\\nThanks for the comments, the v23 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm1-ZrG%3DhaAoiB2yFKYc%2Bckcd1NLaU8QB3SWs32wPsph4w%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Thu, 23 Jun 2022 15:38:50 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for the v23* patch set.\\n\\n========\\nv23-0001\\n========\\n\\nNo comments. LGTM\\n\\n========\\nV23-0002\\n========\\n\\n2.1 src/backend/commands/subscriptioncmds.c\\n\\n+ opts->origin = defGetString(defel);\\n+\\n+ if ((strcmp(opts->origin, LOGICALREP_ORIGIN_LOCAL) != 0) &&\\n+ (strcmp(opts->origin, LOGICALREP_ORIGIN_ANY) != 0))\\n+ ereport(ERROR,\\n+ errcode(ERRCODE_SYNTAX_ERROR),\\n+ errmsg(\\\"unrecognized origin value: \\\\\\\"%s\\\\\\\"\\\", opts->origin));\\n+ }\\n\\nI thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\\n\\n~~~\\n\\n2.2 src/backend/replication/pgoutput/pgoutput.c\\n\\n+ data->origin = defGetString(defel);\\n+ if (strcmp(data->origin, LOGICALREP_ORIGIN_LOCAL) == 0)\\n+ publish_local_origin = true;\\n+ else if (strcmp(data->origin, LOGICALREP_ORIGIN_ANY) == 0)\\n+ publish_local_origin = false;\\n+ else\\n+ ereport(ERROR,\\n+ errcode(ERRCODE_SYNTAX_ERROR),\\n+ errmsg(\\\"unrecognized origin value: \\\\\\\"%s\\\\\\\"\\\", data->origin));\\n+ }\\n\\nI thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\\n\\n========\\nv23-0003\\n========\\n\\n3.1 Commit message\\n\\nAfter the subscription is created on node2, node1 will be synced to\\nnode2 and the newly synced data will be sent to node2, this process of\\nnode1 sending data to node2 and node2 sending data to node1 will repeat\\ninfinitely. If table t1 has a unique key, this will lead to a unique key\\nviolation, and replication won't proceed.\\n\\n~\\n\\n31a.\\n\\\"node2, this process of\\\" -> \\\"node2; this process of\\\"\\nOR\\n\\\"node2, this process of\\\" -> \\\"node2. This process of\\\"\\n\\n31b.\\nAlso, my grammar checker recommends removing the comma after \\\"violation\\\"\\n\\n~~~\\n\\n3.2 doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -115,7 +115,8 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n (You cannot combine setting <literal>connect</literal>\\n to <literal>false</literal> with\\n setting <literal>create_slot</literal>, <literal>enabled</literal>,\\n- or <literal>copy_data</literal> to <literal>true</literal>.)\\n+ or <literal>copy_data</literal> to\\n+ <literal>true</literal>/<literal>force</literal>.)\\n </para>\\n\\nI am not sure why that last sentence needs to be in parentheses, but\\nOTOH it seems to be that way already in PG15.\\n\\n~~~\\n\\n3.3 doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n can have non-existent publications.\\n </para>\\n\\n+ <para>\\n+ If the subscription is created with <literal>origin = local</literal> and\\n+ <literal>copy_data = true</literal>, it will check if the publisher has\\n+ subscribed to the same table from other publishers and, if so, then throw an\\n+ error to prevent possible non-local data from being copied. The user can\\n+ override this check and continue with the copy operation by specifying\\n+ <literal>copy_data = force</literal>.\\n+ </para>\\n\\n\\\"and, if so, then throw...\\\" -> \\\"and, if so, throw...\\\"\\n\\n~~~\\n\\n3.4 src/backend/commands/subscriptioncmds.c\\n\\n+ ereport(ERROR,\\n+ errcode(ERRCODE_SYNTAX_ERROR),\\n+ errmsg(\\\"%s requires a boolean or \\\\\\\"force\\\\\\\"\\\", def->defname));\\n+ return COPY_DATA_OFF; /* keep compiler quiet */\\n+}\\n\\nI thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\\n\\n========\\nv23-0004\\n========\\n\\n4.1 Commit message\\n\\nDocument the steps for the following:\\na) Creating a two-node bidirectional replication when there is no data\\non both nodes.\\nb) Adding a new node when there is no data on any of the nodes.\\nc) Adding a new node when data is present on the existing nodes.\\nd) Generic steps for adding a new node to an existing set of nodes.\\n\\n~\\n\\nThese pgdocs titles have changed slightly. I think this commit message\\ntext should use match the current pgdocs titles.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 24 Jun 2022 14:50:19 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jun 24, 2022 at 10:20 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for the v23* patch set.\\n>\\n> ========\\n> v23-0001\\n> ========\\n>\\n> No comments. LGTM\\n>\\n> ========\\n> V23-0002\\n> ========\\n>\\n> 2.1 src/backend/commands/subscriptioncmds.c\\n>\\n> + opts->origin = defGetString(defel);\\n> +\\n> + if ((strcmp(opts->origin, LOGICALREP_ORIGIN_LOCAL) != 0) &&\\n> + (strcmp(opts->origin, LOGICALREP_ORIGIN_ANY) != 0))\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_SYNTAX_ERROR),\\n> + errmsg(\\\"unrecognized origin value: \\\\\\\"%s\\\\\\\"\\\", opts->origin));\\n> + }\\n>\\n> I thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.2 src/backend/replication/pgoutput/pgoutput.c\\n>\\n> + data->origin = defGetString(defel);\\n> + if (strcmp(data->origin, LOGICALREP_ORIGIN_LOCAL) == 0)\\n> + publish_local_origin = true;\\n> + else if (strcmp(data->origin, LOGICALREP_ORIGIN_ANY) == 0)\\n> + publish_local_origin = false;\\n> + else\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_SYNTAX_ERROR),\\n> + errmsg(\\\"unrecognized origin value: \\\\\\\"%s\\\\\\\"\\\", data->origin));\\n> + }\\n>\\n> I thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\\n\\nModified\\n\\n> ========\\n> v23-0003\\n> ========\\n>\\n> 3.1 Commit message\\n>\\n> After the subscription is created on node2, node1 will be synced to\\n> node2 and the newly synced data will be sent to node2, this process of\\n> node1 sending data to node2 and node2 sending data to node1 will repeat\\n> infinitely. If table t1 has a unique key, this will lead to a unique key\\n> violation, and replication won't proceed.\\n>\\n> ~\\n>\\n> 31a.\\n> \\\"node2, this process of\\\" -> \\\"node2; this process of\\\"\\n> OR\\n> \\\"node2, this process of\\\" -> \\\"node2. This process of\\\"\\n\\nModified\\n\\n> 31b.\\n> Also, my grammar checker recommends removing the comma after \\\"violation\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 3.2 doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -115,7 +115,8 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> (You cannot combine setting <literal>connect</literal>\\n> to <literal>false</literal> with\\n> setting <literal>create_slot</literal>, <literal>enabled</literal>,\\n> - or <literal>copy_data</literal> to <literal>true</literal>.)\\n> + or <literal>copy_data</literal> to\\n> + <literal>true</literal>/<literal>force</literal>.)\\n> </para>\\n>\\n> I am not sure why that last sentence needs to be in parentheses, but\\n> OTOH it seems to be that way already in PG15.\\n\\nI feel since it is like that since PG15, let's keep it in parenthesis\\nlike earlier. I have not made any changes for this.\\n\\n> ~~~\\n>\\n> 3.3 doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -383,6 +398,15 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> can have non-existent publications.\\n> </para>\\n>\\n> + <para>\\n> + If the subscription is created with <literal>origin = local</literal> and\\n> + <literal>copy_data = true</literal>, it will check if the publisher has\\n> + subscribed to the same table from other publishers and, if so, then throw an\\n> + error to prevent possible non-local data from being copied. The user can\\n> + override this check and continue with the copy operation by specifying\\n> + <literal>copy_data = force</literal>.\\n> + </para>\\n>\\n> \\\"and, if so, then throw...\\\" -> \\\"and, if so, throw...\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 3.4 src/backend/commands/subscriptioncmds.c\\n>\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_SYNTAX_ERROR),\\n> + errmsg(\\\"%s requires a boolean or \\\\\\\"force\\\\\\\"\\\", def->defname));\\n> + return COPY_DATA_OFF; /* keep compiler quiet */\\n> +}\\n>\\n> I thought maybe this should be ERRCODE_INVALID_PARAMETER_VALUE.\\n\\nModified\\n\\n> ========\\n> v23-0004\\n> ========\\n>\\n> 4.1 Commit message\\n>\\n> Document the steps for the following:\\n> a) Creating a two-node bidirectional replication when there is no data\\n> on both nodes.\\n> b) Adding a new node when there is no data on any of the nodes.\\n> c) Adding a new node when data is present on the existing nodes.\\n> d) Generic steps for adding a new node to an existing set of nodes.\\n>\\n> ~\\n>\\n> These pgdocs titles have changed slightly. I think this commit message\\n> text should use match the current pgdocs titles.\\n\\nModified\\n\\nThanks for the comments, the attached v24 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Mon, 27 Jun 2022 19:08:44 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for the v24* patch set.\\n\\nNow, these few comments are all trivial and non-functional. Apart from\\nthese, everything looks good to me.\\n\\n========\\nv24-0001\\n========\\n\\nNo comments. LGTM\\n\\n========\\nV24-0002\\n========\\n\\n2.1 doc/src/sgml/ref/create_subscription.sgml\\n\\n+ <para>\\n+ Specifies whether the subscription will request the publisher to only\\n+ send changes that originated locally, or to send any changes\\n+ regardless of origin. Setting <literal>origin</literal> to\\n+ <literal>local</literal> means that the subscription will request the\\n+ publisher to only send changes that originated locally. Setting\\n+ <literal>origin</literal> to <literal>any</literal> means that the\\n+ publisher sends any changes regardless of their origin. The default\\n+ is <literal>any</literal>.\\n+ </para>\\n\\n2.1a.\\nIMO remove the word \\\"any\\\" from \\\"any changes\\\". Then the text will match\\nwhat is written in catalogs.sgml.\\n\\\"send any changes regardless of origin\\\" -> \\\"send changes regardless of\\norigin\\\" (occurs 2x)\\n\\n2.1b.\\nThis same text is cut/paste to the commit message so that can also be updated.\\n\\n\\n~~~\\n\\n2.2 src/backend/commands/subscriptioncmds.c\\n\\n+ /*\\n+ * Even though \\\"origin\\\" parameter allows only \\\"local\\\" and \\\"any\\\"\\n+ * values, the \\\"origin\\\" parameter type is implemented as string\\n+ * type instead of boolean to extend the \\\"origin\\\" parameter to\\n+ * support filtering of origin name specified by the user in the\\n+ * later versions.\\n+ */\\n\\n2.2a.\\nSUGGESTION\\nEven though \\\"origin\\\" parameter allows only \\\"local\\\" and \\\"any\\\" values,\\nit is implemented as a string type so that the parameter can be\\nextended in future versions to support filtering using origin names\\nspecified by the user.\\n\\n2.2b.\\nThis same text is cut/paste to the commit message so that can also be updated.\\n\\n\\n========\\nv24-0003\\n========\\n\\n3.1 Commit message\\n\\nThis patch does a couple of things:\\nchange 1) Checks and throws an error if 'copy_data = on' and 'origin =\\nlocal' but the publication tables were also replicating from other publishers.\\nchange 2) Adds 'force' value for copy_data parameter.\\n\\n~\\n\\n\\\"replicating\\\" -> \\\"replicated\\\"\\n\\\"change 1)\\\" -> \\\"1)\\\"\\n\\\"change 2)\\\" -> \\\"2)\\\"\\n\\n========\\nv24-0004\\n========\\n\\nNo comments. LGTM\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 28 Jun 2022 11:58:49 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jun 28, 2022 at 7:29 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for the v24* patch set.\\n>\\n> Now, these few comments are all trivial and non-functional. Apart from\\n> these, everything looks good to me.\\n>\\n> ========\\n> v24-0001\\n> ========\\n>\\n> No comments. LGTM\\n>\\n> ========\\n> V24-0002\\n> ========\\n>\\n> 2.1 doc/src/sgml/ref/create_subscription.sgml\\n>\\n> + <para>\\n> + Specifies whether the subscription will request the publisher to only\\n> + send changes that originated locally, or to send any changes\\n> + regardless of origin. Setting <literal>origin</literal> to\\n> + <literal>local</literal> means that the subscription will request the\\n> + publisher to only send changes that originated locally. Setting\\n> + <literal>origin</literal> to <literal>any</literal> means that the\\n> + publisher sends any changes regardless of their origin. The default\\n> + is <literal>any</literal>.\\n> + </para>\\n>\\n> 2.1a.\\n> IMO remove the word \\\"any\\\" from \\\"any changes\\\". Then the text will match\\n> what is written in catalogs.sgml.\\n> \\\"send any changes regardless of origin\\\" -> \\\"send changes regardless of\\n> origin\\\" (occurs 2x)\\n\\nModified\\n\\n> 2.1b.\\n> This same text is cut/paste to the commit message so that can also be updated.\\n\\nModified\\n\\n>\\n> ~~~\\n>\\n> 2.2 src/backend/commands/subscriptioncmds.c\\n>\\n> + /*\\n> + * Even though \\\"origin\\\" parameter allows only \\\"local\\\" and \\\"any\\\"\\n> + * values, the \\\"origin\\\" parameter type is implemented as string\\n> + * type instead of boolean to extend the \\\"origin\\\" parameter to\\n> + * support filtering of origin name specified by the user in the\\n> + * later versions.\\n> + */\\n>\\n> 2.2a.\\n> SUGGESTION\\n> Even though \\\"origin\\\" parameter allows only \\\"local\\\" and \\\"any\\\" values,\\n> it is implemented as a string type so that the parameter can be\\n> extended in future versions to support filtering using origin names\\n> specified by the user.\\n\\nModified\\n\\n> 2.2b.\\n> This same text is cut/paste to the commit message so that can also be updated.\\n\\nModified\\n\\n>\\n> ========\\n> v24-0003\\n> ========\\n>\\n> 3.1 Commit message\\n>\\n\\n> v24-0004\\n> ========\\n>\\n> No comments. LGTM\\n>\\n> ------\\n> Kind Regards,\\n> Peter Smith.\\n> This patch does a couple of things:\\n> change 1) Checks and throws an error if 'copy_data = on' and 'origin =\\n> local' but the publication tables were also replicating from other publishers.\\n> change 2) Adds 'force' value for copy_data parameter.\\n>\\n> ~\\n>\\n> \\\"replicating\\\" -> \\\"replicated\\\"\\n> \\\"change 1)\\\" -> \\\"1)\\\"\\n> \\\"change 2)\\\" -> \\\"2)\\\"\\n\\nModified\\n\\nThanks for the comments, the attached v25 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 28 Jun 2022 11:47:53 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jun 28, 2022 2:18 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> Thanks for the comments, the attached v25 patch has the changes for the\\r\\n> same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch. Here are some comments.\\r\\n\\r\\n0002 patch:\\r\\n==============\\r\\n1.\\r\\n+# Test the CREATE SUBSCRIPTION 'origin' parameter and its interaction with\\r\\n+# 'copy_data' parameter.\\r\\n\\r\\nIt seems we should move \\\"and its interaction with 'copy_data' parameter\\\" to\\r\\n0003 patch.\\r\\n\\r\\n0003 patch\\r\\n==============\\r\\n1.\\r\\nWhen using ALTER SUBSCRIPTION ... REFRESH, subscription will throw an error if\\r\\nany table is subscribed in publisher, even if the table has been subscribed\\r\\nbefore refresh (which won't do the initial copy when refreshing). It looks the\\r\\npreviously subscribed tables don't need this check. Would it be better that we\\r\\nonly check the tables which need to do the initial copy?\\r\\n\\r\\n2.\\r\\n+\\t\\t\\t\\terrmsg(\\\"table:%s.%s might have replicated data in the publisher\\\",\\r\\n+\\t\\t\\t\\t\\t nspname, relname),\\r\\n\\r\\nI think the table name needs to be enclosed in double quotes, which is\\r\\nconsistent with other messages.\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Thu, 30 Jun 2022 03:47:24 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 30, 2022 at 9:17 AM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Tue, Jun 28, 2022 2:18 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached v25 patch has the changes for the\\n> > same.\\n> >\\n>\\n> Thanks for updating the patch. Here are some comments.\\n>\\n> 0002 patch:\\n> ==============\\n> 1.\\n> +# Test the CREATE SUBSCRIPTION 'origin' parameter and its interaction with\\n> +# 'copy_data' parameter.\\n>\\n> It seems we should move \\\"and its interaction with 'copy_data' parameter\\\" to\\n> 0003 patch.\\n\\nModified\\n\\n> 0003 patch\\n> ==============\\n> 1.\\n> When using ALTER SUBSCRIPTION ... REFRESH, subscription will throw an error if\\n> any table is subscribed in publisher, even if the table has been subscribed\\n> before refresh (which won't do the initial copy when refreshing). It looks the\\n> previously subscribed tables don't need this check. Would it be better that we\\n> only check the tables which need to do the initial copy?\\n\\nModified\\n\\n> 2.\\n> + errmsg(\\\"table:%s.%s might have replicated data in the publisher\\\",\\n> + nspname, relname),\\n>\\n> I think the table name needs to be enclosed in double quotes, which is\\n> consistent with other messages.\\n\\nModified\\n\\nThanks for the comments, the attached v26 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Thu, 30 Jun 2022 21:40:14 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 30, 2022 at 9:40 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Thu, Jun 30, 2022 at 9:17 AM shiy.fnst@fujitsu.com\\n> <shiy.fnst@fujitsu.com> wrote:\\n> >\\n\\nThe first patch that adds a test case for existing functionality looks\\ngood to me and I'll push that early next week (by Tuesday) unless\\nthere are more comments on it.\\n\\nFew minor comments on 0002\\n========================\\n1.\\n+ /* FIXME: 150000 should be changed to 160000 later for PG16. */\\n+ if (options->proto.logical.origin &&\\n+ PQserverVersion(conn->streamConn) >= 150000)\\n+ appendStringInfo(&cmd, \\\", origin '%s'\\\",\\n+ options->proto.logical.origin);\\n\\n...\\n...\\n+ /* FIXME: 150000 should be changed to 160000 later for PG16. */\\n+ if (fout->remoteVersion >= 150000)\\n+ appendPQExpBufferStr(query, \\\" s.suborigin\\\\n\\\");\\n+ else\\n+ appendPQExpBuffer(query, \\\" '%s' AS suborigin\\\\n\\\", LOGICALREP_ORIGIN_ANY);\\n...\\n...\\n+ /* FIXME: 150000 should be changed to 160000 later for PG16 */\\n+ if (pset.sversion >= 150000)\\n+ appendPQExpBuffer(&buf,\\n+ \\\", suborigin AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n+ gettext_noop(\\\"Origin\\\"));\\n\\nAll these should now change to 16.\\n\\n2.\\n/* ALTER SUBSCRIPTION <name> SET ( */\\n else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\nTailMatches(\\\"SET\\\", \\\"(\\\"))\\n- COMPLETE_WITH(\\\"binary\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n\\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n+ COMPLETE_WITH(\\\"binary\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n\\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n /* ALTER SUBSCRIPTION <name> SKIP ( */\\n else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\nTailMatches(\\\"SKIP\\\", \\\"(\\\"))\\n COMPLETE_WITH(\\\"lsn\\\");\\n@@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\\n /* Complete \\\"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\\\" */\\n else if (HeadMatches(\\\"CREATE\\\", \\\"SUBSCRIPTION\\\") && TailMatches(\\\"WITH\\\", \\\"(\\\"))\\n COMPLETE_WITH(\\\"binary\\\", \\\"connect\\\", \\\"copy_data\\\", \\\"create_slot\\\",\\n- \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n+ \\\"enabled\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n \\\"synchronous_commit\\\", \\\"two_phase\\\", \\\"disable_on_error\\\");\\n\\nWhy do you choose to add a new option in-between other parameters\\ninstead of at the end which we normally do? The one possible reason I\\ncan think of is that all the parameters at the end are boolean so you\\nwant to add this before those but then why before slot_name, and again\\nI don't see such a rule being followed for other parameters.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Sat, 2 Jul 2022 12:12:35 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sat, Jul 2, 2022 at 12:12 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Thu, Jun 30, 2022 at 9:40 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Thu, Jun 30, 2022 at 9:17 AM shiy.fnst@fujitsu.com\\n> > <shiy.fnst@fujitsu.com> wrote:\\n> > >\\n>\\n> The first patch that adds a test case for existing functionality looks\\n> good to me and I'll push that early next week (by Tuesday) unless\\n> there are more comments on it.\\n>\\n> Few minor comments on 0002\\n> ========================\\n> 1.\\n> + /* FIXME: 150000 should be changed to 160000 later for PG16. */\\n> + if (options->proto.logical.origin &&\\n> + PQserverVersion(conn->streamConn) >= 150000)\\n> + appendStringInfo(&cmd, \\\", origin '%s'\\\",\\n> + options->proto.logical.origin);\\n>\\n> ...\\n> ...\\n> + /* FIXME: 150000 should be changed to 160000 later for PG16. */\\n> + if (fout->remoteVersion >= 150000)\\n> + appendPQExpBufferStr(query, \\\" s.suborigin\\\\n\\\");\\n> + else\\n> + appendPQExpBuffer(query, \\\" '%s' AS suborigin\\\\n\\\", LOGICALREP_ORIGIN_ANY);\\n> ...\\n> ...\\n> + /* FIXME: 150000 should be changed to 160000 later for PG16 */\\n> + if (pset.sversion >= 150000)\\n> + appendPQExpBuffer(&buf,\\n> + \\\", suborigin AS \\\\\\\"%s\\\\\\\"\\\\n\\\",\\n> + gettext_noop(\\\"Origin\\\"));\\n>\\n> All these should now change to 16.\\n\\nModified\\n\\n> 2.\\n> /* ALTER SUBSCRIPTION <name> SET ( */\\n> else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\n> TailMatches(\\\"SET\\\", \\\"(\\\"))\\n> - COMPLETE_WITH(\\\"binary\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n> + COMPLETE_WITH(\\\"binary\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n> /* ALTER SUBSCRIPTION <name> SKIP ( */\\n> else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\n> TailMatches(\\\"SKIP\\\", \\\"(\\\"))\\n> COMPLETE_WITH(\\\"lsn\\\");\\n> @@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\\n> /* Complete \\\"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\\\" */\\n> else if (HeadMatches(\\\"CREATE\\\", \\\"SUBSCRIPTION\\\") && TailMatches(\\\"WITH\\\", \\\"(\\\"))\\n> COMPLETE_WITH(\\\"binary\\\", \\\"connect\\\", \\\"copy_data\\\", \\\"create_slot\\\",\\n> - \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> + \\\"enabled\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> \\\"synchronous_commit\\\", \\\"two_phase\\\", \\\"disable_on_error\\\");\\n>\\n> Why do you choose to add a new option in-between other parameters\\n> instead of at the end which we normally do? The one possible reason I\\n> can think of is that all the parameters at the end are boolean so you\\n> want to add this before those but then why before slot_name, and again\\n> I don't see such a rule being followed for other parameters.\\n\\nI was not sure if it should be maintained in alphabetical order,\\nanyway since the last option \\\"disable_on_error\\\" is at the end, I have\\nchanged it to the end.\\n\\nThanks for the comments, the attached v27 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Sun, 3 Jul 2022 20:29:37 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 4, 2022 at 12:59 AM vignesh C <vignesh21@gmail.com> wrote:\\n...\\n> > 2.\\n> > /* ALTER SUBSCRIPTION <name> SET ( */\\n> > else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\n> > TailMatches(\\\"SET\\\", \\\"(\\\"))\\n> > - COMPLETE_WITH(\\\"binary\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n> > + COMPLETE_WITH(\\\"binary\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n> > /* ALTER SUBSCRIPTION <name> SKIP ( */\\n> > else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\n> > TailMatches(\\\"SKIP\\\", \\\"(\\\"))\\n> > COMPLETE_WITH(\\\"lsn\\\");\\n> > @@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\\n> > /* Complete \\\"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\\\" */\\n> > else if (HeadMatches(\\\"CREATE\\\", \\\"SUBSCRIPTION\\\") && TailMatches(\\\"WITH\\\", \\\"(\\\"))\\n> > COMPLETE_WITH(\\\"binary\\\", \\\"connect\\\", \\\"copy_data\\\", \\\"create_slot\\\",\\n> > - \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > + \\\"enabled\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > \\\"synchronous_commit\\\", \\\"two_phase\\\", \\\"disable_on_error\\\");\\n> >\\n> > Why do you choose to add a new option in-between other parameters\\n> > instead of at the end which we normally do? The one possible reason I\\n> > can think of is that all the parameters at the end are boolean so you\\n> > want to add this before those but then why before slot_name, and again\\n> > I don't see such a rule being followed for other parameters.\\n>\\n> I was not sure if it should be maintained in alphabetical order,\\n> anyway since the last option \\\"disable_on_error\\\" is at the end, I have\\n> changed it to the end.\\n>\\n\\nAlthough it seems it is not a hard rule, mostly the COMPLETE_WITH are\\ncoded using alphabetical order. Anyway, I think that was a clear\\nintention here too since 13 of 14 parameters were already in\\nalphabetical order; it is actually only that \\\"disable_on_error\\\"\\nparameter that was misplaced; not the new \\\"origin\\\" parameter.\\n\\nAlso, in practice, on <tab> those completions will get output in\\nalphabetical order, so IMO it makes more sense for the code to be\\nconsistent with the output:\\n\\ne.g.\\ntest_sub=# create subscription sub xxx connection '' publication pub WITH (\\nBINARY DISABLE_ON_ERROR STREAMING\\nCONNECT ENABLED SYNCHRONOUS_COMMIT\\nCOPY_DATA ORIGIN TWO_PHASE\\nCREATE_SLOT SLOT_NAME\\ntest_sub=#\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Jul 2022 08:29:03 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 4, 2022 at 3:59 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Mon, Jul 4, 2022 at 12:59 AM vignesh C <vignesh21@gmail.com> wrote:\\n> ...\\n> > > 2.\\n> > > /* ALTER SUBSCRIPTION <name> SET ( */\\n> > > else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\n> > > TailMatches(\\\"SET\\\", \\\"(\\\"))\\n> > > - COMPLETE_WITH(\\\"binary\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > > \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n> > > + COMPLETE_WITH(\\\"binary\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > > \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n> > > /* ALTER SUBSCRIPTION <name> SKIP ( */\\n> > > else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\n> > > TailMatches(\\\"SKIP\\\", \\\"(\\\"))\\n> > > COMPLETE_WITH(\\\"lsn\\\");\\n> > > @@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\\n> > > /* Complete \\\"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\\\" */\\n> > > else if (HeadMatches(\\\"CREATE\\\", \\\"SUBSCRIPTION\\\") && TailMatches(\\\"WITH\\\", \\\"(\\\"))\\n> > > COMPLETE_WITH(\\\"binary\\\", \\\"connect\\\", \\\"copy_data\\\", \\\"create_slot\\\",\\n> > > - \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > > + \\\"enabled\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > > \\\"synchronous_commit\\\", \\\"two_phase\\\", \\\"disable_on_error\\\");\\n> > >\\n> > > Why do you choose to add a new option in-between other parameters\\n> > > instead of at the end which we normally do? The one possible reason I\\n> > > can think of is that all the parameters at the end are boolean so you\\n> > > want to add this before those but then why before slot_name, and again\\n> > > I don't see such a rule being followed for other parameters.\\n> >\\n> > I was not sure if it should be maintained in alphabetical order,\\n> > anyway since the last option \\\"disable_on_error\\\" is at the end, I have\\n> > changed it to the end.\\n> >\\n>\\n> Although it seems it is not a hard rule, mostly the COMPLETE_WITH are\\n> coded using alphabetical order. Anyway, I think that was a clear\\n> intention here too since 13 of 14 parameters were already in\\n> alphabetical order; it is actually only that \\\"disable_on_error\\\"\\n> parameter that was misplaced; not the new \\\"origin\\\" parameter.\\n>\\n\\nAgreed, but let's not change disable_on_error as part of this patch.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Jul 2022 09:22:58 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sun, Jul 3, 2022 11:00 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> Thanks for the comments, the attached v27 patch has the changes for the\\r\\n> same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch.\\r\\n\\r\\nA comment on 0003 patch:\\r\\n+\\t\\t/*\\r\\n+\\t\\t * No need to throw an error for the tables that are in ready state,\\r\\n+\\t\\t * as the walsender will send the changes from WAL in case of tables\\r\\n+\\t\\t * in ready state.\\r\\n+\\t\\t */\\r\\n+\\t\\tif (isreadytable)\\r\\n+\\t\\t\\tcontinue;\\r\\n+\\r\\n...\\r\\n+\\t\\tereport(ERROR,\\r\\n+\\t\\t\\t\\terrcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\r\\n+\\t\\t\\t\\terrmsg(\\\"table: \\\\\\\"%s.%s\\\\\\\" might have replicated data in the publisher\\\",\\r\\n+\\t\\t\\t\\t\\t nspname, relname),\\r\\n+\\t\\t\\t\\terrdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data = on is not allowed when the publisher might have replicated data.\\\"),\\r\\n+\\t\\t\\t\\terrhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\\"));\\r\\n+\\r\\n+\\t\\tExecClearTuple(slot);\\r\\n\\r\\nI think we should call ExecClearTuple() before getting next tuple, so it should\\r\\nbe called if the table is in ready state. How about modifying it to:\\r\\n\\t\\tif (!isreadytable)\\r\\n\\t\\t\\tereport(ERROR,\\r\\n\\t\\t\\t\\t\\terrcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\r\\n\\t\\t\\t\\t\\terrmsg(\\\"table: \\\\\\\"%s.%s\\\\\\\" might have replicated data in the publisher\\\",\\r\\n\\t\\t\\t\\t\\t\\t nspname, relname),\\r\\n\\t\\t\\t\\t\\terrdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data = on is not allowed when the publisher might have replicated data.\\\"),\\r\\n\\t\\t\\t\\t\\terrhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\\"));\\r\\n\\r\\n\\t\\tExecClearTuple(slot);\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\\r\\n\",\n \"msg_date\": \"Mon, 4 Jul 2022 08:16:45 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 4, 2022 4:17 PM shiy.fnst@fujitsu.com <shiy.fnst@fujitsu.com> wrote:\\r\\n> \\r\\n> On Sun, Jul 3, 2022 11:00 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> >\\r\\n> > Thanks for the comments, the attached v27 patch has the changes for the\\r\\n> > same.\\r\\n> >\\r\\n> \\r\\n> Thanks for updating the patch.\\r\\n> \\r\\n> A comment on 0003 patch:\\r\\n> \\r\\n> I think we should call ExecClearTuple() before getting next tuple, so it should\\r\\n> be called if the table is in ready state. How about modifying it to:\\r\\n> \\r\\n\\r\\nBy the way, I have tested pg_dump/psql changes in this patch with older version\\r\\n(server: pg10 ~ pg15, pg_dump/psql: pg16), and it worked ok.\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Mon, 4 Jul 2022 08:41:17 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sun, Jul 3, 2022 at 8:29 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n\\nReview comments\\n===============\\n1.\\n@@ -530,7 +557,7 @@ CreateSubscription(ParseState *pstate,\\nCreateSubscriptionStmt *stmt,\\n SUBOPT_SLOT_NAME | SUBOPT_COPY_DATA |\\n SUBOPT_SYNCHRONOUS_COMMIT | SUBOPT_BINARY |\\n SUBOPT_STREAMING | SUBOPT_TWOPHASE_COMMIT |\\n- SUBOPT_DISABLE_ON_ERR);\\n+ SUBOPT_DISABLE_ON_ERR | SUBOPT_ORIGIN);\\n parse_subscription_options(pstate, stmt->options, supported_opts, &opts);\\n\\n /*\\n@@ -606,6 +633,8 @@ CreateSubscription(ParseState *pstate,\\nCreateSubscriptionStmt *stmt,\\n LOGICALREP_TWOPHASE_STATE_PENDING :\\n LOGICALREP_TWOPHASE_STATE_DISABLED);\\n values[Anum_pg_subscription_subdisableonerr - 1] =\\nBoolGetDatum(opts.disableonerr);\\n+ values[Anum_pg_subscription_suborigin - 1] =\\n+ CStringGetTextDatum(opts.origin);\\n values[Anum_pg_subscription_subconninfo - 1] =\\n CStringGetTextDatum(conninfo);\\n if (opts.slot_name)\\n...\\n...\\n\\n /* List of publications subscribed to */\\n text subpublications[1] BKI_FORCE_NOT_NULL;\\n+\\n+ /* Only publish data originating from the specified origin */\\n+ text suborigin BKI_DEFAULT(LOGICALREP_ORIGIN_ANY);\\n #endif\\n } FormData_pg_subscription;\\n\\nThe order of declaration and assignment for 'suborigin' should match\\nin above usage.\\n\\n2. Similarly the changes in GetSubscription() should also match the\\ndeclaration of the origin column.\\n\\n3.\\n GRANT SELECT (oid, subdbid, subskiplsn, subname, subowner, subenabled,\\n- subbinary, substream, subtwophasestate,\\nsubdisableonerr, subslotname,\\n- subsynccommit, subpublications)\\n+ subbinary, substream, subtwophasestate, subdisableonerr,\\n+ suborigin, subslotname, subsynccommit, subpublications)\\n ON pg_subscription TO public;\\n\\nThis should also match the order of columns as in pg_subscription.h\\nunless there is a reason for not doing so.\\n\\n4.\\n+ /*\\n+ * Even though \\\"origin\\\" parameter allows only \\\"local\\\" and \\\"any\\\"\\n+ * values, it is implemented as a string type so that the parameter\\n+ * can be extended in future versions to support filtering using\\n+ * origin names specified by the user.\\n\\n/Even though \\\"origin\\\" .../Even though the \\\"origin\\\" parameter ...\\n\\n5.\\n+\\n+# Create tables on node_A\\n+$node_A->safe_psql('postgres', \\\"CREATE TABLE tab_full (a int PRIMARY KEY)\\\");\\n+\\n+# Create the same tables on node_B\\n+$node_B->safe_psql('postgres', \\\"CREATE TABLE tab_full (a int PRIMARY KEY)\\\");\\n\\nIn both the above comments, you should use table instead of tables as\\nthe test creates only one table.\\n\\n6.\\n+$node_A->safe_psql('postgres', \\\"DELETE FROM tab_full;\\\");\\n\\nAfter this, the test didn't ensure that this operation is replicated.\\nCan't that lead to unpredictable results for the other tests after\\nthis test?\\n\\n7.\\n+# Setup logical replication\\n+# node_C (pub) -> node_B (sub)\\n+my $node_C_connstr = $node_C->connstr . ' dbname=postgres';\\n+$node_C->safe_psql('postgres',\\n+ \\\"CREATE PUBLICATION tap_pub_C FOR TABLE tab_full\\\");\\n+\\n+my $appname_B2 = 'tap_sub_B2';\\n+$node_B->safe_psql(\\n+ 'postgres', \\\"\\n+ CREATE SUBSCRIPTION tap_sub_B2\\n+ CONNECTION '$node_C_connstr application_name=$appname_B2'\\n+ PUBLICATION tap_pub_C\\n+ WITH (origin = local)\\\");\\n+\\n+$node_C->wait_for_catchup($appname_B2);\\n+\\n+$node_C->poll_query_until('postgres', $synced_query)\\n+ or die \\\"Timed out while waiting for subscriber to synchronize data\\\";\\n\\nThis test allows the publisher (node_C) to poll for sync but it should\\nbe the subscriber (node_B) that needs to poll to allow the initial\\nsync to finish.\\n\\n8. Do you think it makes sense to see if this new option can also be\\nsupported by pg_recvlogical? I see that previously we have not\\nextended pg_recvlogical for all the newly added options but I feel we\\nshould keep pg_recvlogical up to date w.r.t new options. We can do\\nthis as a separate patch if we agree?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Jul 2022 15:48:33 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"> From d8f8844f877806527b6f3f45320b6ba55a8e3154 Mon Sep 17 00:00:00 2001\\n> From: Vigneshwaran C <vignesh21@gmail.com>\\n> Date: Thu, 26 May 2022 19:29:33 +0530\\n> Subject: [PATCH v27 1/4] Add a missing test to verify only-local parameter in\\n> test_decoding plugin.\\n> \\n> Add a missing test to verify only-local parameter in test_decoding plugin.\\n\\nI don't get it. replorigin.sql already has some lines to test\\nlocal-only. What is your patch adding that is new? Maybe instead of\\nadding some more lines at the end of the script, you should add lines\\nwhere this stuff is already being tested. But that assumes that there\\nis something new that is being tested; if so what is it?\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\\"Investigación es lo que hago cuando no sé lo que estoy haciendo\\\"\\n(Wernher von Braun)\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Jul 2022 16:14:02 +0200\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 4, 2022 at 3:48 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Sun, Jul 3, 2022 at 8:29 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n>\\n> Review comments\\n> ===============\\n> 1.\\n> @@ -530,7 +557,7 @@ CreateSubscription(ParseState *pstate,\\n> CreateSubscriptionStmt *stmt,\\n> SUBOPT_SLOT_NAME | SUBOPT_COPY_DATA |\\n> SUBOPT_SYNCHRONOUS_COMMIT | SUBOPT_BINARY |\\n> SUBOPT_STREAMING | SUBOPT_TWOPHASE_COMMIT |\\n> - SUBOPT_DISABLE_ON_ERR);\\n> + SUBOPT_DISABLE_ON_ERR | SUBOPT_ORIGIN);\\n> parse_subscription_options(pstate, stmt->options, supported_opts, &opts);\\n>\\n> /*\\n> @@ -606,6 +633,8 @@ CreateSubscription(ParseState *pstate,\\n> CreateSubscriptionStmt *stmt,\\n> LOGICALREP_TWOPHASE_STATE_PENDING :\\n> LOGICALREP_TWOPHASE_STATE_DISABLED);\\n> values[Anum_pg_subscription_subdisableonerr - 1] =\\n> BoolGetDatum(opts.disableonerr);\\n> + values[Anum_pg_subscription_suborigin - 1] =\\n> + CStringGetTextDatum(opts.origin);\\n> values[Anum_pg_subscription_subconninfo - 1] =\\n> CStringGetTextDatum(conninfo);\\n> if (opts.slot_name)\\n> ...\\n> ...\\n>\\n> /* List of publications subscribed to */\\n> text subpublications[1] BKI_FORCE_NOT_NULL;\\n> +\\n> + /* Only publish data originating from the specified origin */\\n> + text suborigin BKI_DEFAULT(LOGICALREP_ORIGIN_ANY);\\n> #endif\\n> } FormData_pg_subscription;\\n>\\n> The order of declaration and assignment for 'suborigin' should match\\n> in above usage.\\n\\nModified\\n\\n> 2. Similarly the changes in GetSubscription() should also match the\\n> declaration of the origin column.\\n\\nModified\\n\\n> 3.\\n> GRANT SELECT (oid, subdbid, subskiplsn, subname, subowner, subenabled,\\n> - subbinary, substream, subtwophasestate,\\n> subdisableonerr, subslotname,\\n> - subsynccommit, subpublications)\\n> + subbinary, substream, subtwophasestate, subdisableonerr,\\n> + suborigin, subslotname, subsynccommit, subpublications)\\n> ON pg_subscription TO public;\\n>\\n> This should also match the order of columns as in pg_subscription.h\\n> unless there is a reason for not doing so.\\n\\nModified\\n\\n> 4.\\n> + /*\\n> + * Even though \\\"origin\\\" parameter allows only \\\"local\\\" and \\\"any\\\"\\n> + * values, it is implemented as a string type so that the parameter\\n> + * can be extended in future versions to support filtering using\\n> + * origin names specified by the user.\\n>\\n> /Even though \\\"origin\\\" .../Even though the \\\"origin\\\" parameter ...\\n\\nModified\\n\\n> 5.\\n> +\\n> +# Create tables on node_A\\n> +$node_A->safe_psql('postgres', \\\"CREATE TABLE tab_full (a int PRIMARY KEY)\\\");\\n> +\\n> +# Create the same tables on node_B\\n> +$node_B->safe_psql('postgres', \\\"CREATE TABLE tab_full (a int PRIMARY KEY)\\\");\\n>\\n> In both the above comments, you should use table instead of tables as\\n> the test creates only one table.\\n\\nModified\\n\\n> 6.\\n> +$node_A->safe_psql('postgres', \\\"DELETE FROM tab_full;\\\");\\n>\\n> After this, the test didn't ensure that this operation is replicated.\\n> Can't that lead to unpredictable results for the other tests after\\n> this test?\\n\\nModified to include wait_for_catchup and verified the data at the\\nbeginning of the next test\\n\\n> 7.\\n> +# Setup logical replication\\n> +# node_C (pub) -> node_B (sub)\\n> +my $node_C_connstr = $node_C->connstr . ' dbname=postgres';\\n> +$node_C->safe_psql('postgres',\\n> + \\\"CREATE PUBLICATION tap_pub_C FOR TABLE tab_full\\\");\\n> +\\n> +my $appname_B2 = 'tap_sub_B2';\\n> +$node_B->safe_psql(\\n> + 'postgres', \\\"\\n> + CREATE SUBSCRIPTION tap_sub_B2\\n> + CONNECTION '$node_C_connstr application_name=$appname_B2'\\n> + PUBLICATION tap_pub_C\\n> + WITH (origin = local)\\\");\\n> +\\n> +$node_C->wait_for_catchup($appname_B2);\\n> +\\n> +$node_C->poll_query_until('postgres', $synced_query)\\n> + or die \\\"Timed out while waiting for subscriber to synchronize data\\\";\\n>\\n> This test allows the publisher (node_C) to poll for sync but it should\\n> be the subscriber (node_B) that needs to poll to allow the initial\\n> sync to finish.\\n\\nModified\\n\\n> 8. Do you think it makes sense to see if this new option can also be\\n> supported by pg_recvlogical? I see that previously we have not\\n> extended pg_recvlogical for all the newly added options but I feel we\\n> should keep pg_recvlogical up to date w.r.t new options. We can do\\n> this as a separate patch if we agree?\\n\\nI will analyze this and post my analysis soon.\\n\\nThanks for the comments, the v28 patch attached has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Mon, 4 Jul 2022 23:29:27 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 4, 2022 at 1:46 PM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Sun, Jul 3, 2022 11:00 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached v27 patch has the changes for the\\n> > same.\\n> >\\n>\\n> Thanks for updating the patch.\\n>\\n> A comment on 0003 patch:\\n> + /*\\n> + * No need to throw an error for the tables that are in ready state,\\n> + * as the walsender will send the changes from WAL in case of tables\\n> + * in ready state.\\n> + */\\n> + if (isreadytable)\\n> + continue;\\n> +\\n> ...\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"table: \\\\\\\"%s.%s\\\\\\\" might have replicated data in the publisher\\\",\\n> + nspname, relname),\\n> + errdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data = on is not allowed when the publisher might have replicated data.\\\"),\\n> + errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\\"));\\n> +\\n> + ExecClearTuple(slot);\\n>\\n> I think we should call ExecClearTuple() before getting next tuple, so it should\\n> be called if the table is in ready state. How about modifying it to:\\n> if (!isreadytable)\\n> ereport(ERROR,\\n> errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> errmsg(\\\"table: \\\\\\\"%s.%s\\\\\\\" might have replicated data in the publisher\\\",\\n> nspname, relname),\\n> errdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin = local and copy_data = on is not allowed when the publisher might have replicated data.\\\"),\\n> errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\\"));\\n>\\n> ExecClearTuple(slot);\\n\\nThanks for the comment, I have modified this in the v28 patch attached at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm3MNK2hMYroTiHGS9HkSxiA-az1QC1mpa0YwDZ8nGmmZg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Jul 2022 23:30:53 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 4, 2022 at 9:23 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Jul 4, 2022 at 3:59 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > On Mon, Jul 4, 2022 at 12:59 AM vignesh C <vignesh21@gmail.com> wrote:\\n> > ...\\n> > > > 2.\\n> > > > /* ALTER SUBSCRIPTION <name> SET ( */\\n> > > > else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\n> > > > TailMatches(\\\"SET\\\", \\\"(\\\"))\\n> > > > - COMPLETE_WITH(\\\"binary\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > > > \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n> > > > + COMPLETE_WITH(\\\"binary\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > > > \\\"synchronous_commit\\\", \\\"disable_on_error\\\");\\n> > > > /* ALTER SUBSCRIPTION <name> SKIP ( */\\n> > > > else if (HeadMatches(\\\"ALTER\\\", \\\"SUBSCRIPTION\\\", MatchAny) &&\\n> > > > TailMatches(\\\"SKIP\\\", \\\"(\\\"))\\n> > > > COMPLETE_WITH(\\\"lsn\\\");\\n> > > > @@ -3152,7 +3152,7 @@ psql_completion(const char *text, int start, int end)\\n> > > > /* Complete \\\"CREATE SUBSCRIPTION <name> ... WITH ( <opt>\\\" */\\n> > > > else if (HeadMatches(\\\"CREATE\\\", \\\"SUBSCRIPTION\\\") && TailMatches(\\\"WITH\\\", \\\"(\\\"))\\n> > > > COMPLETE_WITH(\\\"binary\\\", \\\"connect\\\", \\\"copy_data\\\", \\\"create_slot\\\",\\n> > > > - \\\"enabled\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > > > + \\\"enabled\\\", \\\"origin\\\", \\\"slot_name\\\", \\\"streaming\\\",\\n> > > > \\\"synchronous_commit\\\", \\\"two_phase\\\", \\\"disable_on_error\\\");\\n> > > >\\n> > > > Why do you choose to add a new option in-between other parameters\\n> > > > instead of at the end which we normally do? The one possible reason I\\n> > > > can think of is that all the parameters at the end are boolean so you\\n> > > > want to add this before those but then why before slot_name, and again\\n> > > > I don't see such a rule being followed for other parameters.\\n> > >\\n> > > I was not sure if it should be maintained in alphabetical order,\\n> > > anyway since the last option \\\"disable_on_error\\\" is at the end, I have\\n> > > changed it to the end.\\n> > >\\n> >\\n> > Although it seems it is not a hard rule, mostly the COMPLETE_WITH are\\n> > coded using alphabetical order. Anyway, I think that was a clear\\n> > intention here too since 13 of 14 parameters were already in\\n> > alphabetical order; it is actually only that \\\"disable_on_error\\\"\\n> > parameter that was misplaced; not the new \\\"origin\\\" parameter.\\n> >\\n>\\n> Agreed, but let's not change disable_on_error as part of this patch.\\n\\nI have changed this in the v28 patch attached at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm3MNK2hMYroTiHGS9HkSxiA-az1QC1mpa0YwDZ8nGmmZg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Jul 2022 23:31:55 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 4, 2022 at 7:44 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n>\\n> > From d8f8844f877806527b6f3f45320b6ba55a8e3154 Mon Sep 17 00:00:00 2001\\n> > From: Vigneshwaran C <vignesh21@gmail.com>\\n> > Date: Thu, 26 May 2022 19:29:33 +0530\\n> > Subject: [PATCH v27 1/4] Add a missing test to verify only-local parameter in\\n> > test_decoding plugin.\\n> >\\n> > Add a missing test to verify only-local parameter in test_decoding plugin.\\n>\\n> I don't get it. replorigin.sql already has some lines to test\\n> local-only. What is your patch adding that is new? Maybe instead of\\n> adding some more lines at the end of the script, you should add lines\\n> where this stuff is already being tested. But that assumes that there\\n> is something new that is being tested; if so what is it?\\n\\nThe test is to check that remote origin data (i.e. replication origin\\nbeing set) will be filtered when only-local parameter is set. I felt\\nthat this scenario is not covered, unless I'm missing something. I\\nhave moved the test as suggested.\\nI have changed this in the v28 patch attached at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm3MNK2hMYroTiHGS9HkSxiA-az1QC1mpa0YwDZ8nGmmZg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Jul 2022 23:33:13 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"I checked again the v26* patch set.\\n\\nI had no more comments for v26-0001, v26-0002, or v26-0004, but below\\nare some comments for v26-0003.\\n\\n========\\nv26-0003\\n========\\n\\n3.1 src/backend/catalog/pg_subscription.c\\n\\n3.1.a\\n+/*\\n+ * Get all relations for subscription that are in a ready state.\\n+ *\\n+ * Returned list is palloc'ed in current memory context.\\n+ */\\n+List *\\n+GetSubscriptionReadyRelations(Oid subid)\\n\\n\\\"subscription\\\" -> \\\"the subscription\\\"\\n\\nAlso, It might be better to say in the function header what kind of\\nstructures are in the returned List. E.g. Firstly, I'd assumed it was\\nthe same return type as the other function\\nGetSubscriptionReadyRelations, but it isn’t.\\n\\n3.1.b\\nI think there might be a case to be made for *combining* those\\nSubscriptionRelState and SubscripotionRel structs into a single common\\nstruct. Then all those GetSubscriptionNotReadyRelations and\\nGetSubscriptionNotReadyRelations (also GetSubscriptionRelations?) can\\nbe merged to be just one common function that takes a parameter to say\\ndo you want to return the relation List of ALL / READY / NOT_READY?\\nWhat do you think?\\n\\n======\\n\\n3.2 src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n\\n+/*\\n+ * Check and throw an error if the publisher has subscribed to the same table\\n+ * from some other publisher. This check is required only if copydata is ON and\\n+ * the origin is local.\\n+ */\\n+static void\\n+check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\\n+ CopyData copydata, char *origin, Oid subid)\\n\\nThe function comment probably should say something about why the new\\nREADY logic was added in v26.\\n\\n~~~\\n\\n3.3\\n\\n3.3.a\\n+ /*\\n+ * The subid will be valid only for ALTER SUBSCRIPTION ... REFRESH\\n+ * PUBLICATION. Get the ready relations for the subscription only in case\\n+ * of ALTER SUBSCRIPTION case as there will be no relations in ready state\\n+ * while the subscription is created.\\n+ */\\n+ if (subid != InvalidOid)\\n+ subreadyrels = GetSubscriptionReadyRelations(subid);\\n\\nThe word \\\"case\\\" is 2x in the same sentence. I also paraphrased my\\nunderstanding of this comment below. Maybe it is simpler?\\n\\nSUGGESTION\\nGet the ready relations for the subscription. The subid will be valid\\nonly for ALTER SUBSCRIPTION ... REFRESH because there will be no\\nrelations in ready state while the subscription is created.\\n\\n3.3.b\\n+ if (subid != InvalidOid)\\n+ subreadyrels = GetSubscriptionReadyRelations(subid);\\n\\nSUGGESTION\\nif (OidIsValid(subid))\\n\\n======\\n\\n3.4 src/test/subscription/t/032_localonly.pl\\n\\nNow all the test cases are re-using the same data (e.g. 11,12,13) and\\nyou are deleting the data between the tests. I guess it is OK, but IMO\\nthe tests are easier to read when each test part was using unique data\\n(e.g. 11,12,13, then 12,22,32, then 13,23,33 etc)\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Jul 2022 10:44:32 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 4, 2022 at 11:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Jul 4, 2022 at 7:44 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> >\\n> > > From d8f8844f877806527b6f3f45320b6ba55a8e3154 Mon Sep 17 00:00:00 2001\\n> > > From: Vigneshwaran C <vignesh21@gmail.com>\\n> > > Date: Thu, 26 May 2022 19:29:33 +0530\\n> > > Subject: [PATCH v27 1/4] Add a missing test to verify only-local parameter in\\n> > > test_decoding plugin.\\n> > >\\n> > > Add a missing test to verify only-local parameter in test_decoding plugin.\\n> >\\n> > I don't get it. replorigin.sql already has some lines to test\\n> > local-only. What is your patch adding that is new? Maybe instead of\\n> > adding some more lines at the end of the script, you should add lines\\n> > where this stuff is already being tested. But that assumes that there\\n> > is something new that is being tested; if so what is it?\\n>\\n> The test is to check that remote origin data (i.e. replication origin\\n> being set) will be filtered when only-local parameter is set. I felt\\n> that this scenario is not covered, unless I'm missing something.\\n>\\n\\nIf we change only-local option to '0' in the below part of the test,\\nwe can see a different result:\\n-- and magically the replayed xact will be filtered!\\nSELECT data FROM pg_logical_slot_get_changes('regression_slot', NULL,\\nNULL, 'include-xids', '0', 'skip-empty-xacts', '1', 'only-local',\\n'1');\\n\\nSo, I think only-local functionality is being tested but I feel your\\ntest is clear in the sense that it clearly shows that remote data can\\nonly be fetched with 'only-local' as '0' when the origin is set. It\\nseems to me that in existing tests, we are not testing the combination\\nwhere the origin is set and 'only-local' is '0'. So, there is some\\nvalue to having the test you are proposing but if Alvaro, you, or\\nothers don't think it is worth adding a new test for it then we can\\nprobably drop this one.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Jul 2022 13:59:28 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 5, 2022 at 1:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Jul 4, 2022 at 11:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Mon, Jul 4, 2022 at 7:44 PM Alvaro Herrera <alvherre@alvh.no-ip.org> wrote:\\n> > >\\n> > > > From d8f8844f877806527b6f3f45320b6ba55a8e3154 Mon Sep 17 00:00:00 2001\\n> > > > From: Vigneshwaran C <vignesh21@gmail.com>\\n> > > > Date: Thu, 26 May 2022 19:29:33 +0530\\n> > > > Subject: [PATCH v27 1/4] Add a missing test to verify only-local parameter in\\n> > > > test_decoding plugin.\\n> > > >\\n> > > > Add a missing test to verify only-local parameter in test_decoding plugin.\\n> > >\\n> > > I don't get it. replorigin.sql already has some lines to test\\n> > > local-only. What is your patch adding that is new? Maybe instead of\\n> > > adding some more lines at the end of the script, you should add lines\\n> > > where this stuff is already being tested. But that assumes that there\\n> > > is something new that is being tested; if so what is it?\\n> >\\n> > The test is to check that remote origin data (i.e. replication origin\\n> > being set) will be filtered when only-local parameter is set. I felt\\n> > that this scenario is not covered, unless I'm missing something.\\n> >\\n>\\n> If we change only-local option to '0' in the below part of the test,\\n> we can see a different result:\\n> -- and magically the replayed xact will be filtered!\\n> SELECT data FROM pg_logical_slot_get_changes('regression_slot', NULL,\\n> NULL, 'include-xids', '0', 'skip-empty-xacts', '1', 'only-local',\\n> '1');\\n>\\n> So, I think only-local functionality is being tested but I feel your\\n> test is clear in the sense that it clearly shows that remote data can\\n> only be fetched with 'only-local' as '0' when the origin is set. It\\n> seems to me that in existing tests, we are not testing the combination\\n> where the origin is set and 'only-local' is '0'. So, there is some\\n> value to having the test you are proposing but if Alvaro, you, or\\n> others don't think it is worth adding a new test for it then we can\\n> probably drop this one.\\n\\nSince the existing test is already handling the verification of this\\nscenario, I felt no need to add the test. Updated v29 patch removes\\nthe 0001 patch which had the test case.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 5 Jul 2022 21:33:18 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 5, 2022 at 6:14 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> I checked again the v26* patch set.\\n>\\n> I had no more comments for v26-0001, v26-0002, or v26-0004, but below\\n> are some comments for v26-0003.\\n>\\n> ========\\n> v26-0003\\n> ========\\n>\\n> 3.1 src/backend/catalog/pg_subscription.c\\n>\\n> 3.1.a\\n> +/*\\n> + * Get all relations for subscription that are in a ready state.\\n> + *\\n> + * Returned list is palloc'ed in current memory context.\\n> + */\\n> +List *\\n> +GetSubscriptionReadyRelations(Oid subid)\\n>\\n> \\\"subscription\\\" -> \\\"the subscription\\\"\\n>\\n> Also, It might be better to say in the function header what kind of\\n> structures are in the returned List. E.g. Firstly, I'd assumed it was\\n> the same return type as the other function\\n> GetSubscriptionReadyRelations, but it isn’t.\\n\\nThis function has been removed and GetSubscriptionRelations is used\\nwith slight changes.\\n\\n> 3.1.b\\n> I think there might be a case to be made for *combining* those\\n> SubscriptionRelState and SubscripotionRel structs into a single common\\n> struct. Then all those GetSubscriptionNotReadyRelations and\\n> GetSubscriptionNotReadyRelations (also GetSubscriptionRelations?) can\\n> be merged to be just one common function that takes a parameter to say\\n> do you want to return the relation List of ALL / READY / NOT_READY?\\n> What do you think?\\n\\nI have used a common function that can get all relations, ready\\nrelations and not ready relations instead of the 3 functions.\\n\\n> ======\\n>\\n> 3.2 src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> +/*\\n> + * Check and throw an error if the publisher has subscribed to the same table\\n> + * from some other publisher. This check is required only if copydata is ON and\\n> + * the origin is local.\\n> + */\\n> +static void\\n> +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\\n> + CopyData copydata, char *origin, Oid subid)\\n>\\n> The function comment probably should say something about why the new\\n> READY logic was added in v26.\\n\\nAdded a comment for the same.\\n\\n> ~~~\\n>\\n> 3.3\\n>\\n> 3.3.a\\n> + /*\\n> + * The subid will be valid only for ALTER SUBSCRIPTION ... REFRESH\\n> + * PUBLICATION. Get the ready relations for the subscription only in case\\n> + * of ALTER SUBSCRIPTION case as there will be no relations in ready state\\n> + * while the subscription is created.\\n> + */\\n> + if (subid != InvalidOid)\\n> + subreadyrels = GetSubscriptionReadyRelations(subid);\\n>\\n> The word \\\"case\\\" is 2x in the same sentence. I also paraphrased my\\n> understanding of this comment below. Maybe it is simpler?\\n>\\n> SUGGESTION\\n> Get the ready relations for the subscription. The subid will be valid\\n> only for ALTER SUBSCRIPTION ... REFRESH because there will be no\\n> relations in ready state while the subscription is created.\\n\\nModified\\n\\n> 3.3.b\\n> + if (subid != InvalidOid)\\n> + subreadyrels = GetSubscriptionReadyRelations(subid);\\n>\\n> SUGGESTION\\n> if (OidIsValid(subid))\\n\\nModified\\n\\n> ======\\n>\\n> 3.4 src/test/subscription/t/032_localonly.pl\\n>\\n> Now all the test cases are re-using the same data (e.g. 11,12,13) and\\n> you are deleting the data between the tests. I guess it is OK, but IMO\\n> the tests are easier to read when each test part was using unique data\\n> (e.g. 11,12,13, then 12,22,32, then 13,23,33 etc)\\n\\nModified\\n\\nThanks for the comments, the v29 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm1T5utq60qVx%3DRN60rHcg7wt2psM7PpCQ2fDiB-R8oLGg%40mail.gmail.com\\n\\nRegards,\\n Vignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Jul 2022 21:36:17 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 5, 2022 at 9:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> Since the existing test is already handling the verification of this\\n> scenario, I felt no need to add the test. Updated v29 patch removes\\n> the 0001 patch which had the test case.\\n>\\n\\nI have again looked at the first and it looks good to me. I would like\\nto push it after some more review but before that, I would like to\\nknow if someone else has any suggestions/objections to this patch, so\\nlet me summarize the idea of the first patch. It will allow users to\\nskip the replication of remote data. Here remote data is the data that\\nthe publisher node has received from some other node. The primary use\\ncase is to avoid loops (infinite replication of the same data) among\\nnodes as shown in the initial email of this thread.\\n\\nTo achieve that this patch adds a new SUBSCRIPTION parameter \\\"origin\\\".\\nIt specifies whether the subscription will request the publisher to\\nonly send changes that originated locally or to send changes\\nregardless of origin. Setting it to \\\"local\\\" means that the\\nsubscription will request the publisher to only send changes that\\noriginated locally. Setting it to \\\"any\\\" means that the publisher sends\\nchanges regardless of their origin. The default is \\\"any\\\".\\n\\nUsage:\\nCREATE SUBSCRIPTION sub1 CONNECTION 'dbname=postgres port=9999'\\nPUBLICATION pub1 WITH (origin = local);\\n\\nFor now, even though the \\\"origin\\\" parameter allows only \\\"local\\\" and\\n\\\"any\\\" values, it is implemented as a string type so that the parameter\\ncan be extended in future versions to support filtering using origin\\nnames specified by the user.\\n\\nThis feature allows filtering only the replication data originated\\nfrom WAL but for initial sync (initial copy of table data) we don't\\nhave such a facility as we can only distinguish the data based on\\norigin from WAL. As a separate patch (v29-0002*), we are planning to\\nforbid the initial sync if we notice that the publication tables were\\nalso replicated from other publishers to avoid duplicate data or\\nloops. They will be allowed to copy with the 'force' option in such\\ncases.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Jul 2022 17:09:30 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 5, 2022 at 9:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> Since the existing test is already handling the verification of this\\n> scenario, I felt no need to add the test. Updated v29 patch removes\\n> the 0001 patch which had the test case.\\n>\\n\\nI am not able to apply 0001.\\npatching file src/bin/psql/tab-complete.c\\nHunk #1 FAILED at 1873.\\nHunk #2 FAILED at 3152.\\n\\nFew comments on 0002\\n=====================\\n1.\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n\\nIs there a need for space before / in above? If not, please remove it\\nwith similar extra space from other similar usages.\\n\\n2.\\n+ <para>\\n+ There is some interaction between the <literal>origin</literal>\\n+ parameter and the <literal>copy_data</literal> parameter. Refer to\\n+ the <command>CREATE SUBSCRIPTION</command>\\n+ <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n+ details and usage of <literal>force</literal> for\\n+ <literal>copy_data</literal> parameter.\\n </para>\\n\\nI think this is bit long. We can try to reduce this by something like:\\nRefer <xref linkend=\\\"sql-createsubscription-notes\\\"/> for the usage of\\n<literal>force</literal> option and its interaction with the\\n<literal>origin</literal> parameter.\\n\\nAlso, adopt the same other places if you agree with the above change.\\n\\n4.\\n@@ -601,16 +549,28 @@ GetSubscriptionNotReadyRelations(Oid subid)\\n SysScanDesc scan;\\n\\n rel = table_open(SubscriptionRelRelationId, AccessShareLock);\\n-\\n ScanKeyInit(&skey[nkeys++],\\n Anum_pg_subscription_rel_srsubid,\\n\\nSpurious line removal.\\n\\n5.\\n+static void\\n+check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\\n+ CopyData copydata, char *origin, Oid subid)\\n{\\n...\\n...\\n+ /*\\n+ * Get the ready relations for the subscription. The subid will be valid\\n+ * only for ALTER SUBSCRIPTION ... REFRESH because there will be no\\n+ * relations in ready state while the subscription is created.\\n+ */\\n+ if (OidIsValid(subid))\\n+ subreadyrels = GetSubscriptionRelations(subid, READY_STATE);\\n\\nWhy do we want to consider only READY_STATE relations here? If you see\\nits caller AlterSubscription_refresh(), we don't consider copying the\\nrelation if it exists on subscribers in any state. If my observation\\nis correct then you probably don't need to introduce SubRelStateType.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Jul 2022 16:38:57 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jul 8, 2022 at 4:39 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Jul 5, 2022 at 9:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Since the existing test is already handling the verification of this\\n> > scenario, I felt no need to add the test. Updated v29 patch removes\\n> > the 0001 patch which had the test case.\\n> >\\n>\\n> I am not able to apply 0001.\\n> patching file src/bin/psql/tab-complete.c\\n> Hunk #1 FAILED at 1873.\\n> Hunk #2 FAILED at 3152.\\n\\nI have rebased the patch on top of HEAD.\\n\\n> Few comments on 0002\\n> =====================\\n> 1.\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n>\\n> Is there a need for space before / in above? If not, please remove it\\n> with similar extra space from other similar usages.\\n\\nModified\\n\\n> 2.\\n> + <para>\\n> + There is some interaction between the <literal>origin</literal>\\n> + parameter and the <literal>copy_data</literal> parameter. Refer to\\n> + the <command>CREATE SUBSCRIPTION</command>\\n> + <xref linkend=\\\"sql-createsubscription-notes\\\" /> for interaction\\n> + details and usage of <literal>force</literal> for\\n> + <literal>copy_data</literal> parameter.\\n> </para>\\n>\\n> I think this is bit long. We can try to reduce this by something like:\\n> Refer <xref linkend=\\\"sql-createsubscription-notes\\\"/> for the usage of\\n> <literal>force</literal> option and its interaction with the\\n> <literal>origin</literal> parameter.\\n>\\n> Also, adopt the same other places if you agree with the above change.\\n\\nModified with slight rewording.\\n\\n> 4.\\n> @@ -601,16 +549,28 @@ GetSubscriptionNotReadyRelations(Oid subid)\\n> SysScanDesc scan;\\n>\\n> rel = table_open(SubscriptionRelRelationId, AccessShareLock);\\n> -\\n> ScanKeyInit(&skey[nkeys++],\\n> Anum_pg_subscription_rel_srsubid,\\n>\\n> Spurious line removal.\\n\\nI have removed this\\n\\n> 5.\\n> +static void\\n> +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\\n> + CopyData copydata, char *origin, Oid subid)\\n> {\\n> ...\\n> ...\\n> + /*\\n> + * Get the ready relations for the subscription. The subid will be valid\\n> + * only for ALTER SUBSCRIPTION ... REFRESH because there will be no\\n> + * relations in ready state while the subscription is created.\\n> + */\\n> + if (OidIsValid(subid))\\n> + subreadyrels = GetSubscriptionRelations(subid, READY_STATE);\\n>\\n> Why do we want to consider only READY_STATE relations here? If you see\\n> its caller AlterSubscription_refresh(), we don't consider copying the\\n> relation if it exists on subscribers in any state. If my observation\\n> is correct then you probably don't need to introduce SubRelStateType.\\n\\nI have changed it similar to the caller function and removed SubRelStateType\\n\\nThanks for the comments, the v30 patch attached has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Sat, 9 Jul 2022 20:10:56 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for the v30* patches:\\n\\n========\\nv30-0001\\n========\\n\\n1.1 <general>\\n\\nI was wondering if it is better to implement a new defGetOrigin method\\nnow instead of just using the defGetString to process the 'origin',\\nsince you may need to do that in future anyway if the 'origin' name is\\nplanned to become specifiable by the user. OTOH maybe you prefer to\\nchange this code later when the time comes. I am not sure what way is\\nbest.\\n\\n~~~\\n\\n1.2. src/include/replication/walreceiver.h\\n\\n@@ -183,6 +183,8 @@ typedef struct\\n bool streaming; /* Streaming of large transactions */\\n bool twophase; /* Streaming of two-phase transactions at\\n * prepare time */\\n+ char *origin; /* Only publish data originating from the\\n+ * specified origin */\\n } logical;\\n } proto;\\n } WalRcvStreamOptions;\\n\\nShould the new comments be aligned with the other ones?\\n\\n========\\nv30-0002\\n========\\n\\n2.1 doc/src/sgml/ref/alter_subscription.sgml\\n\\n+ <para>\\n+ Refer <xref linkend=\\\"sql-createsubscription-notes\\\"/> for the usage of\\n+ <literal>force</literal> for <literal>copy_data</literal> parameter\\n+ and its interaction with the <literal>origin</literal> parameter.\\n </para>\\n\\nIMO it's better to say \\\"Refer to the Notes\\\" or \\\"Refer to CREATE\\nSUBSCRIPTION Notes\\\" instead of just \\\"Refer Notes\\\"\\n\\n~~~\\n\\n2.2 doc/src/sgml/ref/create_subscription.sgml\\n\\n2.2.a\\n+ <para>\\n+ Refer <xref linkend=\\\"sql-createsubscription-notes\\\"/> for the usage of\\n+ <literal>force</literal> for <literal>copy_data</literal> parameter\\n+ and its interaction with the <literal>origin</literal> parameter.\\n+ </para>\\n\\nIMO it's better to say \\\"Refer to the Notes\\\" (same as other xref on\\nthis page) instead of \\\"Refer Notes\\\"\\n\\n2.2.b\\n@@ -316,6 +324,11 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n publisher sends changes regardless of their origin. The default is\\n <literal>any</literal>.\\n </para>\\n+ <para>\\n+ Refer <xref linkend=\\\"sql-createsubscription-notes\\\"/> for the usage of\\n+ <literal>force</literal> for <literal>copy_data</literal> parameter\\n+ and its interaction with the <literal>origin</literal> parameter.\\n+ </para>\\n\\nDitto\\n\\n~~~\\n\\n2.3 src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n\\n+/*\\n+ * Validate the value specified for copy_data parameter.\\n+ */\\n+static CopyData\\n+DefGetCopyData(DefElem *def)\\n\\n~~~\\n\\n2.4\\n\\n+ /*\\n+ * If no parameter given, assume \\\"true\\\" is meant.\\n+ */\\n\\nPlease modify the comment to match the recent push [1].\\n\\n~~~\\n\\n2.5 src/test/subscription/t/032_localonly.pl\\n\\n2.5.a\\n+# Check Alter subscription ... refresh publication when there is a new\\n+# table that is subscribing data from a different publication\\n+$node_A->safe_psql('postgres', \\\"CREATE TABLE tab_full1 (a int PRIMARY KEY)\\\");\\n+$node_B->safe_psql('postgres', \\\"CREATE TABLE tab_full1 (a int PRIMARY KEY)\\\");\\n\\nUnfortunately, I think tab_full1 is a terrible table name because in\\nmy screen font the 'l' and the '1' look exactly the same so it just\\nlooks like a typo. Maybe change it to \\\"tab_new\\\" or something?\\n\\n2.5b\\nWhat exactly is the purpose of \\\"full\\\" in all these test table names?\\nAFAIK \\\"full\\\" is just some name baggage inherited from completely\\ndifferent tests which were doing full versus partial table\\nreplication. I'm not sure it is relevant here.\\n\\n========\\nv30-0002\\n========\\n\\nNo comments.\\n\\n------\\n[1] https://github.com/postgres/postgres/commit/8445f5a21d40b969673ca03918c74b4fbc882bf4\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 11 Jul 2022 13:41:26 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sat, Jul 9, 2022 at 8:11 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n\\nThanks, a few more comments on v30_0002*\\n1.\\n+/*\\n+ * Represents whether copy_data parameter is specified with off, on or force.\\n\\nA comma is required after on.\\n\\n2.\\n qsort(subrel_local_oids, list_length(subrel_states),\\n sizeof(Oid), oid_cmp);\\n\\n+ check_pub_table_subscribed(wrconn, sub->publications, copy_data,\\n+ sub->origin, subrel_local_oids,\\n+ list_length(subrel_states));\\n\\nWe can avoid using list_length by using an additional variable in this case.\\n\\n3.\\nerrmsg(\\\"table: \\\\\\\"%s.%s\\\\\\\" might have replicated data in the publisher\\\",\\n+ nspname, relname),\\n\\nWhy ':' is used after the table in the above message? I don't see such\\na convention at other places in the code. Also, having might in the\\nerror messages makes it less clear, so, can we slightly change the\\nmessage as in the attached patch?\\n\\n4. I have made some additional changes in the comments, kindly check\\nthe attached and merge those if you are okay.\\n\\n5.\\n+$node_C->safe_psql(\\n+ 'postgres', \\\"\\n+ DELETE FROM tab_full\\\");\\n+$node_B->safe_psql(\\n+ 'postgres', \\\"\\n+ DELETE FROM tab_full where a = 13\\\");\\n\\nDon't we need to wait for these operations to replicate?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\",\n \"msg_date\": \"Mon, 11 Jul 2022 17:03:34 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 11, 2022 at 9:11 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for the v30* patches:\\n>\\n> ========\\n> v30-0001\\n> ========\\n>\\n> 1.1 <general>\\n>\\n> I was wondering if it is better to implement a new defGetOrigin method\\n> now instead of just using the defGetString to process the 'origin',\\n> since you may need to do that in future anyway if the 'origin' name is\\n> planned to become specifiable by the user. OTOH maybe you prefer to\\n> change this code later when the time comes. I am not sure what way is\\n> best.\\n\\nI preferred to do that change when the feature is getting extended.\\n\\n> ~~~\\n>\\n> 1.2. src/include/replication/walreceiver.h\\n>\\n> @@ -183,6 +183,8 @@ typedef struct\\n> bool streaming; /* Streaming of large transactions */\\n> bool twophase; /* Streaming of two-phase transactions at\\n> * prepare time */\\n> + char *origin; /* Only publish data originating from the\\n> + * specified origin */\\n> } logical;\\n> } proto;\\n> } WalRcvStreamOptions;\\n>\\n> Should the new comments be aligned with the other ones?\\n\\nI kept it like this as pgindent also is aligning it as the current code.\\n\\n> ========\\n> v30-0002\\n> ========\\n>\\n> 2.1 doc/src/sgml/ref/alter_subscription.sgml\\n>\\n> + <para>\\n> + Refer <xref linkend=\\\"sql-createsubscription-notes\\\"/> for the usage of\\n> + <literal>force</literal> for <literal>copy_data</literal> parameter\\n> + and its interaction with the <literal>origin</literal> parameter.\\n> </para>\\n>\\n> IMO it's better to say \\\"Refer to the Notes\\\" or \\\"Refer to CREATE\\n> SUBSCRIPTION Notes\\\" instead of just \\\"Refer Notes\\\"\\n\\nModified to \\\"Refer to the Notes\\\"\\n\\n> ~~~\\n>\\n> 2.2 doc/src/sgml/ref/create_subscription.sgml\\n>\\n> 2.2.a\\n> + <para>\\n> + Refer <xref linkend=\\\"sql-createsubscription-notes\\\"/> for the usage of\\n> + <literal>force</literal> for <literal>copy_data</literal> parameter\\n> + and its interaction with the <literal>origin</literal> parameter.\\n> + </para>\\n>\\n> IMO it's better to say \\\"Refer to the Notes\\\" (same as other xref on\\n> this page) instead of \\\"Refer Notes\\\"\\n\\nModified to \\\"Refer to the Notes\\\"\\n\\n> 2.2.b\\n> @@ -316,6 +324,11 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> publisher sends changes regardless of their origin. The default is\\n> <literal>any</literal>.\\n> </para>\\n> + <para>\\n> + Refer <xref linkend=\\\"sql-createsubscription-notes\\\"/> for the usage of\\n> + <literal>force</literal> for <literal>copy_data</literal> parameter\\n> + and its interaction with the <literal>origin</literal> parameter.\\n> + </para>\\n>\\n> Ditto\\n\\nModified to \\\"Refer to the Notes\\\"\\n\\n> ~~~\\n>\\n> 2.3 src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n>\\n> +/*\\n> + * Validate the value specified for copy_data parameter.\\n> + */\\n> +static CopyData\\n> +DefGetCopyData(DefElem *def)\\n\\nChanged it to defGetCopyData to keep the naming similar to others\\n\\n> ~~~\\n>\\n> 2.4\\n>\\n> + /*\\n> + * If no parameter given, assume \\\"true\\\" is meant.\\n> + */\\n>\\n> Please modify the comment to match the recent push [1].\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.5 src/test/subscription/t/032_localonly.pl\\n>\\n> 2.5.a\\n> +# Check Alter subscription ... refresh publication when there is a new\\n> +# table that is subscribing data from a different publication\\n> +$node_A->safe_psql('postgres', \\\"CREATE TABLE tab_full1 (a int PRIMARY KEY)\\\");\\n> +$node_B->safe_psql('postgres', \\\"CREATE TABLE tab_full1 (a int PRIMARY KEY)\\\");\\n>\\n> Unfortunately, I think tab_full1 is a terrible table name because in\\n> my screen font the 'l' and the '1' look exactly the same so it just\\n> looks like a typo. Maybe change it to \\\"tab_new\\\" or something?\\n\\nModified to tab_new\\n\\n> 2.5b\\n> What exactly is the purpose of \\\"full\\\" in all these test table names?\\n> AFAIK \\\"full\\\" is just some name baggage inherited from completely\\n> different tests which were doing full versus partial table\\n> replication. I'm not sure it is relevant here.\\n\\nModified to tab\\n\\nThanks for the comments, the v31 patch attached has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 12 Jul 2022 08:42:55 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 11, 2022 at 5:03 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Sat, Jul 9, 2022 at 8:11 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n>\\n> Thanks, a few more comments on v30_0002*\\n> 1.\\n> +/*\\n> + * Represents whether copy_data parameter is specified with off, on or force.\\n>\\n> A comma is required after on.\\n\\nModified\\n\\n> 2.\\n> qsort(subrel_local_oids, list_length(subrel_states),\\n> sizeof(Oid), oid_cmp);\\n>\\n> + check_pub_table_subscribed(wrconn, sub->publications, copy_data,\\n> + sub->origin, subrel_local_oids,\\n> + list_length(subrel_states));\\n>\\n> We can avoid using list_length by using an additional variable in this case.\\n\\nModified\\n\\n> 3.\\n> errmsg(\\\"table: \\\\\\\"%s.%s\\\\\\\" might have replicated data in the publisher\\\",\\n> + nspname, relname),\\n>\\n> Why ':' is used after the table in the above message? I don't see such\\n> a convention at other places in the code. Also, having might in the\\n> error messages makes it less clear, so, can we slightly change the\\n> message as in the attached patch?\\n\\nModified as suggested\\n\\n> 4. I have made some additional changes in the comments, kindly check\\n> the attached and merge those if you are okay.\\n\\nI have merged the changes\\n\\n> 5.\\n> +$node_C->safe_psql(\\n> + 'postgres', \\\"\\n> + DELETE FROM tab_full\\\");\\n> +$node_B->safe_psql(\\n> + 'postgres', \\\"\\n> + DELETE FROM tab_full where a = 13\\\");\\n>\\n> Don't we need to wait for these operations to replicate?\\n\\nModified to include wait\\n\\nThanks for the comments, the v31 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm2-D860yULtcmZAzDbdiof-Dg6Y_YaY4owbO6Rj%3DXEHMw%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Jul 2022 08:47:06 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 12, 2022 at 8:43 AM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Jul 11, 2022 at 9:11 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Here are my review comments for the v30* patches:\\n> >\\n> > ========\\n> > v30-0001\\n> > ========\\n> >\\n> > 1.1 <general>\\n> >\\n> > I was wondering if it is better to implement a new defGetOrigin method\\n> > now instead of just using the defGetString to process the 'origin',\\n> > since you may need to do that in future anyway if the 'origin' name is\\n> > planned to become specifiable by the user. OTOH maybe you prefer to\\n> > change this code later when the time comes. I am not sure what way is\\n> > best.\\n>\\n> I preferred to do that change when the feature is getting extended.\\n>\\n\\n+1.\\n\\n*\\n+$node_C->safe_psql(\\n+ 'postgres', \\\"\\n+ DELETE FROM tab\\\");\\n+$node_B->safe_psql(\\n+ 'postgres', \\\"\\n+ DELETE FROM tab where a = 32\\\");\\n+\\n+$node_A->wait_for_catchup($subname_BA);\\n+$node_B->wait_for_catchup($subname_AB);\\n\\nHere, don't we need to use node_C instead of node_A for waiting as we\\nhave performed an operation on node_C?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 12 Jul 2022 09:51:31 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Jul 12, 2022 at 8:43 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Mon, Jul 11, 2022 at 9:11 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > >\\n> > > Here are my review comments for the v30* patches:\\n> > >\\n> > > ========\\n> > > v30-0001\\n> > > ========\\n> > >\\n> > > 1.1 <general>\\n> > >\\n> > > I was wondering if it is better to implement a new defGetOrigin method\\n> > > now instead of just using the defGetString to process the 'origin',\\n> > > since you may need to do that in future anyway if the 'origin' name is\\n> > > planned to become specifiable by the user. OTOH maybe you prefer to\\n> > > change this code later when the time comes. I am not sure what way is\\n> > > best.\\n> >\\n> > I preferred to do that change when the feature is getting extended.\\n> >\\n>\\n> +1.\\n>\\n> *\\n> +$node_C->safe_psql(\\n> + 'postgres', \\\"\\n> + DELETE FROM tab\\\");\\n> +$node_B->safe_psql(\\n> + 'postgres', \\\"\\n> + DELETE FROM tab where a = 32\\\");\\n> +\\n> +$node_A->wait_for_catchup($subname_BA);\\n> +$node_B->wait_for_catchup($subname_AB);\\n>\\n> Here, don't we need to use node_C instead of node_A for waiting as we\\n> have performed an operation on node_C?\\n\\nNo need to wait for node_C as we have dropped the subscription before\\nthe delete operation. I have added the comment to make it clear.\\nThe attached v32 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 12 Jul 2022 14:58:04 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n\\nI find one thing confusing about this patch. Basically, this has two\\noption 'local' and 'any', so I would assume that all the local server\\nchanges should be covered under the 'local' but now if we set some\\norigin using 'select pg_replication_origin_session_setup('aa');' then\\nchanges from that session will be ignored because it has an origin id.\\nI think actually the name is creating confusion, because by local it\\nseems like a change which originated locally and the document is also\\nspecifying the same.\\n\\n+ If <literal>local</literal>, the subscription will request the publisher\\n+ to only send changes that originated locally. If <literal>any</literal>,\\n\\nI think if we want to keep the option local then we should look up all\\nthe origin in the replication origin catalog and identify whether it\\nis a local origin id or remote origin id and based on that filter out\\nthe changes.\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Jul 2022 16:49:56 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>\\n> On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> I find one thing confusing about this patch. Basically, this has two\\n> option 'local' and 'any', so I would assume that all the local server\\n> changes should be covered under the 'local' but now if we set some\\n> origin using 'select pg_replication_origin_session_setup('aa');' then\\n> changes from that session will be ignored because it has an origin id.\\n> I think actually the name is creating confusion, because by local it\\n> seems like a change which originated locally and the document is also\\n> specifying the same.\\n>\\n> + If <literal>local</literal>, the subscription will request the publisher\\n> + to only send changes that originated locally. If <literal>any</literal>,\\n>\\n> I think if we want to keep the option local then we should look up all\\n> the origin in the replication origin catalog and identify whether it\\n> is a local origin id or remote origin id and based on that filter out\\n> the changes.\\n\\nOn the other hand if we are interested in receiving the changes which\\nare generated without any origin then I think we should change 'local'\\nto 'none' and then in future we can provide a new option which can\\nsend the changes generated by all the local origin? I think other\\nthan this the patch LGTM.\\n\\n\\n--\\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Jul 2022 11:26:32 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>\\n> On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >\\n> > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > I find one thing confusing about this patch. Basically, this has two\\n> > option 'local' and 'any', so I would assume that all the local server\\n> > changes should be covered under the 'local' but now if we set some\\n> > origin using 'select pg_replication_origin_session_setup('aa');' then\\n> > changes from that session will be ignored because it has an origin id.\\n> > I think actually the name is creating confusion, because by local it\\n> > seems like a change which originated locally and the document is also\\n> > specifying the same.\\n> >\\n> > + If <literal>local</literal>, the subscription will request the publisher\\n> > + to only send changes that originated locally. If <literal>any</literal>,\\n> >\\n> > I think if we want to keep the option local then we should look up all\\n> > the origin in the replication origin catalog and identify whether it\\n> > is a local origin id or remote origin id and based on that filter out\\n> > the changes.\\n>\\n> On the other hand if we are interested in receiving the changes which\\n> are generated without any origin then I think we should change 'local'\\n> to 'none' and then in future we can provide a new option which can\\n> send the changes generated by all the local origin? I think other\\n> than this the patch LGTM.\\n\\nThanks for the comment. The attached v33 patch has the changes to\\nspecify origin as 'none' instead of 'local' which will not publish the\\ndata having any origin.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Thu, 14 Jul 2022 18:34:02 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, 14 Jul 2022 at 6:34 PM, vignesh C <vignesh21@gmail.com> wrote:\\n\\n> On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com>\\n> wrote:\\n> >\\n> > On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com>\\n> wrote:\\n> > >\\n> > > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > > >\\n> > > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com>\\n> wrote:\\n> > >\\n> > > I find one thing confusing about this patch. Basically, this has two\\n> > > option 'local' and 'any', so I would assume that all the local server\\n> > > changes should be covered under the 'local' but now if we set some\\n> > > origin using 'select pg_replication_origin_session_setup('aa');' then\\n> > > changes from that session will be ignored because it has an origin id.\\n> > > I think actually the name is creating confusion, because by local it\\n> > > seems like a change which originated locally and the document is also\\n> > > specifying the same.\\n> > >\\n> > > + If <literal>local</literal>, the subscription will request the\\n> publisher\\n> > > + to only send changes that originated locally. If\\n> <literal>any</literal>,\\n> > >\\n> > > I think if we want to keep the option local then we should look up all\\n> > > the origin in the replication origin catalog and identify whether it\\n> > > is a local origin id or remote origin id and based on that filter out\\n> > > the changes.\\n> >\\n> > On the other hand if we are interested in receiving the changes which\\n> > are generated without any origin then I think we should change 'local'\\n> > to 'none' and then in future we can provide a new option which can\\n> > send the changes generated by all the local origin? I think other\\n> > than this the patch LGTM.\\n>\\n> Thanks for the comment. The attached v33 patch has the changes to\\n> specify origin as 'none' instead of 'local' which will not publish the\\n> data having any origin.\\n\\n\\nI think the ‘none’ might have problem from expand ability pov? what if in\\nfuture we support the actual origin name and than what none mean? no origin\\nor origin name none? Should we just give origin name empty name ‘’? Or is\\nthere some other issue?\\n\\n>\\n—\\nDilip\\n\\n> --\\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\nOn Thu, 14 Jul 2022 at 6:34 PM, vignesh C <vignesh21@gmail.com> wrote:On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>\\n> On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >\\n> > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > I find one thing confusing about this patch. Basically, this has two\\n> > option 'local' and 'any', so I would assume that all the local server\\n> > changes should be covered under the 'local' but now if we set some\\n> > origin using 'select pg_replication_origin_session_setup('aa');' then\\n> > changes from that session will be ignored because it has an origin id.\\n> > I think actually the name is creating confusion, because by local it\\n> > seems like a change which originated locally and the document is also\\n> > specifying the same.\\n> >\\n> > + If <literal>local</literal>, the subscription will request the publisher\\n> > + to only send changes that originated locally. If <literal>any</literal>,\\n> >\\n> > I think if we want to keep the option local then we should look up all\\n> > the origin in the replication origin catalog and identify whether it\\n> > is a local origin id or remote origin id and based on that filter out\\n> > the changes.\\n>\\n> On the other hand if we are interested in receiving the changes which\\n> are generated without any origin then I think we should change 'local'\\n> to 'none' and then in future we can provide a new option which can\\n> send the changes generated by all the local origin? I think other\\n> than this the patch LGTM.\\n\\nThanks for the comment. The attached v33 patch has the changes to\\nspecify origin as 'none' instead of 'local' which will not publish the\\ndata having any origin.I think the ‘none’ might have problem from expand ability pov? what if in future we support the actual origin name and than what none mean? no origin or origin name none? Should we just give origin name empty name ‘’? Or is there some other issue?—Dilip-- Regards,Dilip KumarEnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 14 Jul 2022 18:42:11 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jul 14, 2022 at 6:42 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>\\n> On Thu, 14 Jul 2022 at 6:34 PM, vignesh C <vignesh21@gmail.com> wrote:\\n>>\\n>> On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>> >\\n>> > On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>> > >\\n>> > > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\\n>> > > >\\n>> > > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>> > >\\n>> > > I find one thing confusing about this patch. Basically, this has two\\n>> > > option 'local' and 'any', so I would assume that all the local server\\n>> > > changes should be covered under the 'local' but now if we set some\\n>> > > origin using 'select pg_replication_origin_session_setup('aa');' then\\n>> > > changes from that session will be ignored because it has an origin id.\\n>> > > I think actually the name is creating confusion, because by local it\\n>> > > seems like a change which originated locally and the document is also\\n>> > > specifying the same.\\n>> > >\\n>> > > + If <literal>local</literal>, the subscription will request the publisher\\n>> > > + to only send changes that originated locally. If <literal>any</literal>,\\n>> > >\\n>> > > I think if we want to keep the option local then we should look up all\\n>> > > the origin in the replication origin catalog and identify whether it\\n>> > > is a local origin id or remote origin id and based on that filter out\\n>> > > the changes.\\n>> >\\n>> > On the other hand if we are interested in receiving the changes which\\n>> > are generated without any origin then I think we should change 'local'\\n>> > to 'none' and then in future we can provide a new option which can\\n>> > send the changes generated by all the local origin? I think other\\n>> > than this the patch LGTM.\\n>>\\n>> Thanks for the comment. The attached v33 patch has the changes to\\n>> specify origin as 'none' instead of 'local' which will not publish the\\n>> data having any origin.\\n>\\n>\\n> I think the ‘none’ might have problem from expand ability pov? what if in future we support the actual origin name and than what none mean? no origin or origin name none? Should we just give origin name empty name ‘’? Or is there some other issue?\\n\\nCurrently there is no restriction in the name we can specify for\\norigin, ex any, none, local, etc all are allowed as origin name. How\\nabout extending it with another parameter \\\"origin_name\\\" when we\\nsupport filtering of a particular origin like:\\n1) origin = name, origin_name = 'orig1' -- Means that the publisher\\nwill filter the changes having origin name as 'orig1' and send the\\nother changes.\\n2) origin = any -- Means that the publisher sends all changes\\nregardless of their origin.\\n3) origin = none -- Means that the subscription will request the\\npublisher to only send changes that have no origin associated.\\nWhen we need to specify filtering of a particular origin name we will\\nhave to use both origin and origin_name option, like origin = name,\\norigin_name = 'orig1' as in the first example.\\n\\nI'm not sure if there is a simpler way to do this with only a single\\noption as both 'none' and 'any' can be specified as the origin name.\\nThoughts?\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Sat, 16 Jul 2022 09:04:54 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sat, Jul 16, 2022 at 9:05 AM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Thu, Jul 14, 2022 at 6:42 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >\\n> > On Thu, 14 Jul 2022 at 6:34 PM, vignesh C <vignesh21@gmail.com> wrote:\\n> >>\\n> >> On Thu, Jul 14, 2022 at 11:26 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >> >\\n> >> > On Wed, Jul 13, 2022 at 4:49 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >> > >\\n> >> > > On Tue, Jul 12, 2022 at 2:58 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >> > > >\\n> >> > > > On Tue, Jul 12, 2022 at 9:51 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >> > >\\n> >> > > I find one thing confusing about this patch. Basically, this has two\\n> >> > > option 'local' and 'any', so I would assume that all the local server\\n> >> > > changes should be covered under the 'local' but now if we set some\\n> >> > > origin using 'select pg_replication_origin_session_setup('aa');' then\\n> >> > > changes from that session will be ignored because it has an origin id.\\n> >> > > I think actually the name is creating confusion, because by local it\\n> >> > > seems like a change which originated locally and the document is also\\n> >> > > specifying the same.\\n> >> > >\\n> >> > > + If <literal>local</literal>, the subscription will request the publisher\\n> >> > > + to only send changes that originated locally. If <literal>any</literal>,\\n> >> > >\\n> >> > > I think if we want to keep the option local then we should look up all\\n> >> > > the origin in the replication origin catalog and identify whether it\\n> >> > > is a local origin id or remote origin id and based on that filter out\\n> >> > > the changes.\\n> >> >\\n> >> > On the other hand if we are interested in receiving the changes which\\n> >> > are generated without any origin then I think we should change 'local'\\n> >> > to 'none' and then in future we can provide a new option which can\\n> >> > send the changes generated by all the local origin? I think other\\n> >> > than this the patch LGTM.\\n> >>\\n> >> Thanks for the comment. The attached v33 patch has the changes to\\n> >> specify origin as 'none' instead of 'local' which will not publish the\\n> >> data having any origin.\\n> >\\n> >\\n> > I think the ‘none’ might have problem from expand ability pov? what if in future we support the actual origin name and than what none mean? no origin or origin name none? Should we just give origin name empty name ‘’? Or is there some other issue?\\n>\\n> Currently there is no restriction in the name we can specify for\\n> origin, ex any, none, local, etc all are allowed as origin name. How\\n> about extending it with another parameter \\\"origin_name\\\" when we\\n> support filtering of a particular origin like:\\n> 1) origin = name, origin_name = 'orig1' -- Means that the publisher\\n> will filter the changes having origin name as 'orig1' and send the\\n> other changes.\\n> 2) origin = any -- Means that the publisher sends all changes\\n> regardless of their origin.\\n> 3) origin = none -- Means that the subscription will request the\\n> publisher to only send changes that have no origin associated.\\n> When we need to specify filtering of a particular origin name we will\\n> have to use both origin and origin_name option, like origin = name,\\n> origin_name = 'orig1' as in the first example.\\n>\\n> I'm not sure if there is a simpler way to do this with only a single\\n> option as both 'none' and 'any' can be specified as the origin name.\\n> Thoughts?\\n\\nI think giving two options would be really confusing from the\\nusability perspective. I think what we should be doing here is to\\nkeep these three names 'none', 'any' and 'local' as reserved names for\\nthe origin name so that those are not allowed to be set by the user\\nand they have some internal meaning. And I don't think this is going\\nto create too much trouble for anyone because those are not really the\\nnames someone wants to use for their replication origin. So I think\\npg_replication_origin_create() we can also check for the reserved\\nnames just for the replication origin.\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Sat, 16 Jul 2022 10:29:11 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sat, Jul 16, 2022 at 10:29 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>\\n> I think giving two options would be really confusing from the\\n> usability perspective. I think what we should be doing here is to\\n> keep these three names 'none', 'any' and 'local' as reserved names for\\n> the origin name so that those are not allowed to be set by the user\\n> and they have some internal meaning.\\n>\\n\\nThis makes sense to me. I think we can avoid reserving 'local' for now\\ntill we agree on its use case and implementation. One similar point\\nabout slots is that we treat 'none' slot_name in subscription commands\\nas a special value indicating no slot name whereas we do allow\\ncreating a slot with the name 'none' with\\npg_create_logical_replication_slot(). So, if we want to follow a\\nsimilar convention here, we may not need to add any restriction for\\norigin names but personally, I think it is better to add such a\\nrestriction to avoid confusion and in fact, as a separate patch we\\nshould even disallow creating slot name as 'none'.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 18 Jul 2022 10:22:50 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 18, 2022 at 10:23 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Sat, Jul 16, 2022 at 10:29 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >\\n> > I think giving two options would be really confusing from the\\n> > usability perspective. I think what we should be doing here is to\\n> > keep these three names 'none', 'any' and 'local' as reserved names for\\n> > the origin name so that those are not allowed to be set by the user\\n> > and they have some internal meaning.\\n> >\\n>\\n> This makes sense to me. I think we can avoid reserving 'local' for now\\n> till we agree on its use case and implementation. One similar point\\n> about slots is that we treat 'none' slot_name in subscription commands\\n> as a special value indicating no slot name whereas we do allow\\n> creating a slot with the name 'none' with\\n> pg_create_logical_replication_slot(). So, if we want to follow a\\n> similar convention here, we may not need to add any restriction for\\n> origin names but personally, I think it is better to add such a\\n> restriction to avoid confusion and in fact, as a separate patch we\\n> should even disallow creating slot name as 'none'.\\n\\nI have made changes to disallow the name \\\"any\\\" and \\\"none\\\" in\\npg_replication_origin_create. The attached v34 patch has the changes\\nfor the same. I will post a separate patch to disallow creation of\\nslots with names as 'none' separately later.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Mon, 18 Jul 2022 16:58:32 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 18, 2022 at 4:58 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Jul 18, 2022 at 10:23 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Sat, Jul 16, 2022 at 10:29 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> > >\\n> > > I think giving two options would be really confusing from the\\n> > > usability perspective. I think what we should be doing here is to\\n> > > keep these three names 'none', 'any' and 'local' as reserved names for\\n> > > the origin name so that those are not allowed to be set by the user\\n> > > and they have some internal meaning.\\n> > >\\n> >\\n> > This makes sense to me. I think we can avoid reserving 'local' for now\\n> > till we agree on its use case and implementation. One similar point\\n> > about slots is that we treat 'none' slot_name in subscription commands\\n> > as a special value indicating no slot name whereas we do allow\\n> > creating a slot with the name 'none' with\\n> > pg_create_logical_replication_slot(). So, if we want to follow a\\n> > similar convention here, we may not need to add any restriction for\\n> > origin names but personally, I think it is better to add such a\\n> > restriction to avoid confusion and in fact, as a separate patch we\\n> > should even disallow creating slot name as 'none'.\\n>\\n> I have made changes to disallow the name \\\"any\\\" and \\\"none\\\" in\\n> pg_replication_origin_create. The attached v34 patch has the changes\\n> for the same. I will post a separate patch to disallow creation of\\n> slots with names as 'none' separately later.\\n\\nI have updated the patch to handle the origin value case\\ninsensitively. The attached patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Mon, 18 Jul 2022 21:46:37 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 18, 2022 at 9:46 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> I have updated the patch to handle the origin value case\\n> insensitively. The attached patch has the changes for the same.\\n>\\n\\nThanks, the patch looks mostly good to me. I have made a few changes\\nin 0001 patch which are as follows: (a) make a comparison of origin\\nnames in maybe_reread_subscription similar to slot names as in future\\nwe may support origin names other than 'any' and 'none', (b) made\\ncomment changes at few places and minor change in one of the error\\nmessage, (c) ran pgindent and slightly changed the commit message.\\n\\nI am planning to push this day after tomorrow unless there are any\\ncomments/suggestions.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\",\n \"msg_date\": \"Tue, 19 Jul 2022 17:04:37 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 19, 2022 at 11:34 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Jul 18, 2022 at 9:46 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > I have updated the patch to handle the origin value case\\n> > insensitively. The attached patch has the changes for the same.\\n> >\\n>\\n> Thanks, the patch looks mostly good to me. I have made a few changes\\n> in 0001 patch which are as follows: (a) make a comparison of origin\\n> names in maybe_reread_subscription similar to slot names as in future\\n> we may support origin names other than 'any' and 'none', (b) made\\n> comment changes at few places and minor change in one of the error\\n> message, (c) ran pgindent and slightly changed the commit message.\\n>\\n> I am planning to push this day after tomorrow unless there are any\\n> comments/suggestions.\\n\\nFYI, the function name in the comment is not same as the function name here:\\n\\n+/*\\n+ * IsReservedName\\n+ * True iff name is either \\\"none\\\" or \\\"any\\\".\\n+ */\\n+static bool\\n+IsReservedOriginName(const char *name)\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 20 Jul 2022 17:08:36 +1200\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Jul 20, 2022 at 10:38 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Tue, Jul 19, 2022 at 11:34 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Mon, Jul 18, 2022 at 9:46 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > I have updated the patch to handle the origin value case\\n> > > insensitively. The attached patch has the changes for the same.\\n> > >\\n> >\\n> > Thanks, the patch looks mostly good to me. I have made a few changes\\n> > in 0001 patch which are as follows: (a) make a comparison of origin\\n> > names in maybe_reread_subscription similar to slot names as in future\\n> > we may support origin names other than 'any' and 'none', (b) made\\n> > comment changes at few places and minor change in one of the error\\n> > message, (c) ran pgindent and slightly changed the commit message.\\n> >\\n> > I am planning to push this day after tomorrow unless there are any\\n> > comments/suggestions.\\n>\\n> FYI, the function name in the comment is not same as the function name here:\\n>\\n> +/*\\n> + * IsReservedName\\n> + * True iff name is either \\\"none\\\" or \\\"any\\\".\\n> + */\\n> +static bool\\n> +IsReservedOriginName(const char *name)\\n\\nModified. Apart from this I have run pgperltidy on the perl file and\\nrenamed 032_origin.pl to 030_origin.pl as currently there is\\n029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\\nThanks for the comment, the attached patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 20 Jul 2022 14:33:16 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Jul 20, 2022 at 2:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> Modified. Apart from this I have run pgperltidy on the perl file and\\n> renamed 032_origin.pl to 030_origin.pl as currently there is\\n> 029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\\n> Thanks for the comment, the attached patch has the changes for the same.\\n>\\n\\nPushed. Kindly rebase the remaining patches.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 21 Jul 2022 14:06:31 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jul 21, 2022 at 2:06 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Jul 20, 2022 at 2:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Modified. Apart from this I have run pgperltidy on the perl file and\\n> > renamed 032_origin.pl to 030_origin.pl as currently there is\\n> > 029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\\n> > Thanks for the comment, the attached patch has the changes for the same.\\n> >\\n>\\n> Pushed. Kindly rebase the remaining patches.\\n\\nThanks for pushing the patch.\\nThe attached v37 version contains the rebased patch for the remaining patches.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Thu, 21 Jul 2022 16:04:22 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi,\\r\\n\\r\\nOn 7/21/22 6:34 AM, vignesh C wrote:\\r\\n> On Thu, Jul 21, 2022 at 2:06 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n>>\\r\\n>> On Wed, Jul 20, 2022 at 2:33 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n>>>\\r\\n>>> Modified. Apart from this I have run pgperltidy on the perl file and\\r\\n>>> renamed 032_origin.pl to 030_origin.pl as currently there is\\r\\n>>> 029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\\r\\n>>> Thanks for the comment, the attached patch has the changes for the same.\\r\\n>>>\\r\\n>>\\r\\n>> Pushed. Kindly rebase the remaining patches.\\r\\n> \\r\\n> Thanks for pushing the patch.\\r\\n> The attached v37 version contains the rebased patch for the remaining patches.\\r\\n\\r\\nThanks for the work on this feature -- this is definitely very helpful \\r\\ntowards supporting more types of use cases with logical replication!\\r\\n\\r\\nI've read through the proposed documentation and did some light testing \\r\\nof the patch. I have two general comments about the docs as they \\r\\ncurrently read:\\r\\n\\r\\n1. I'm concerned by calling this \\\"Bidirectional replication\\\" in the docs \\r\\nthat we are overstating the current capabilities. I think this is \\r\\naccentuated int he opening paragraph:\\r\\n\\r\\n==snip==\\r\\n Bidirectional replication is useful for creating a multi-master database\\r\\n environment for replicating read/write operations performed by any of the\\r\\n member nodes.\\r\\n==snip==\\r\\n\\r\\nFor one, we're not replicating reads, we're replicating writes. Amongst \\r\\nthe writes, at this point we're only replicating DML. A reader could \\r\\nthink that deploying can work for a full bidirectional solution.\\r\\n\\r\\n(Even if we're aspirationally calling this section \\\"Bidirectional \\r\\nreplication\\\", that does make it sound like we're limited to two nodes, \\r\\nwhen we can support more than two).\\r\\n\\r\\nPerhaps \\\"Logical replication between writers\\\" or \\\"Logical replication \\r\\nbetween primaries\\\" or \\\"Replicating changes between primaries\\\", or \\r\\nsomething better.\\r\\n\\r\\n2. There is no mention of conflicts in the documentation, e.g. \\r\\nreferencing the \\\"Conflicts\\\" section of the documentation. It's very easy \\r\\nto create a conflicting transaction that causes a subscriber to be \\r\\nunable to continue to apply transactions:\\r\\n\\r\\n -- DB 1\\r\\n CREATE TABLE abc (id int);\\r\\n CREATE PUBLICATION node1 FOR ALL TABLES ;\\r\\n\\r\\n -- DB2\\r\\n CREATE TABLE abc (id int);\\r\\n CREATE PUBLICATION node2 FOR ALL TABLES ;\\r\\n CREATE SUBSCRIPTION node2_node1\\r\\n CONNECTION 'dbname=logi port=5433'\\r\\n PUBLICATION node1\\r\\n WITH (copy_data = off, origin = none);\\r\\n\\r\\n -- DB1\\r\\n CREATE SUBSCRIPTION node1_node2\\r\\n CONNECTION 'dbname=logi port=5434'\\r\\n PUBLICATION node2\\r\\n WITH (copy_data = off, origin = none);\\r\\n INSERT INTO abc VALUES (1);\\r\\n\\r\\n -- DB2\\r\\n INSERT INTO abc VALUES (2);\\r\\n\\r\\n -- DB1\\r\\n ALTER TABLE abc ADD PRIMARY KEY id;\\r\\n INSERT INTO abc VALUES (3);\\r\\n\\r\\n -- DB2\\r\\n INSERT INTO abc VALUES (3);\\r\\n\\r\\n -- DB1 cannot apply the transactions\\r\\n\\r\\nAt a minimum, I think we should reference the documentation we have in \\r\\nthe logical replication section on conflicts. We may also want to advise \\r\\nthat a user is responsible for designing their schemas in a way to \\r\\nminimize the risk of conflicts.\\r\\n\\r\\nThanks,\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Thu, 21 Jul 2022 16:08:54 -0400\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> Thanks for the work on this feature -- this is definitely very helpful\\n> towards supporting more types of use cases with logical replication!\\n>\\n> I've read through the proposed documentation and did some light testing\\n> of the patch. I have two general comments about the docs as they\\n> currently read:\\n>\\n> 1. I'm concerned by calling this \\\"Bidirectional replication\\\" in the docs\\n> that we are overstating the current capabilities. I think this is\\n> accentuated int he opening paragraph:\\n>\\n> ==snip==\\n> Bidirectional replication is useful for creating a multi-master database\\n> environment for replicating read/write operations performed by any of the\\n> member nodes.\\n> ==snip==\\n>\\n> For one, we're not replicating reads, we're replicating writes. Amongst\\n> the writes, at this point we're only replicating DML. A reader could\\n> think that deploying can work for a full bidirectional solution.\\n>\\n> (Even if we're aspirationally calling this section \\\"Bidirectional\\n> replication\\\", that does make it sound like we're limited to two nodes,\\n> when we can support more than two).\\n>\\n\\nRight, I think the system can support N-Way replication.\\n\\n> Perhaps \\\"Logical replication between writers\\\" or \\\"Logical replication\\n> between primaries\\\" or \\\"Replicating changes between primaries\\\", or\\n> something better.\\n>\\n\\nAmong the above \\\"Replicating changes between primaries\\\" sounds good to\\nme or simply \\\"Replication between primaries\\\". As this is a sub-section\\non the Logical Replication page, I feel it is okay to not use Logical\\nin the title.\\n\\n> 2. There is no mention of conflicts in the documentation, e.g.\\n> referencing the \\\"Conflicts\\\" section of the documentation. It's very easy\\n> to create a conflicting transaction that causes a subscriber to be\\n> unable to continue to apply transactions:\\n>\\n> -- DB 1\\n> CREATE TABLE abc (id int);\\n> CREATE PUBLICATION node1 FOR ALL TABLES ;\\n>\\n> -- DB2\\n> CREATE TABLE abc (id int);\\n> CREATE PUBLICATION node2 FOR ALL TABLES ;\\n> CREATE SUBSCRIPTION node2_node1\\n> CONNECTION 'dbname=logi port=5433'\\n> PUBLICATION node1\\n> WITH (copy_data = off, origin = none);\\n>\\n> -- DB1\\n> CREATE SUBSCRIPTION node1_node2\\n> CONNECTION 'dbname=logi port=5434'\\n> PUBLICATION node2\\n> WITH (copy_data = off, origin = none);\\n> INSERT INTO abc VALUES (1);\\n>\\n> -- DB2\\n> INSERT INTO abc VALUES (2);\\n>\\n> -- DB1\\n> ALTER TABLE abc ADD PRIMARY KEY id;\\n> INSERT INTO abc VALUES (3);\\n>\\n> -- DB2\\n> INSERT INTO abc VALUES (3);\\n>\\n> -- DB1 cannot apply the transactions\\n>\\n> At a minimum, I think we should reference the documentation we have in\\n> the logical replication section on conflicts. We may also want to advise\\n> that a user is responsible for designing their schemas in a way to\\n> minimize the risk of conflicts.\\n>\\n\\nThis sounds reasonable to me.\\n\\nOne more point about docs, it appears to be added as the last\\nsub-section on the Logical Replication page. Is there a reason for\\ndoing so? I feel this should be third sub-section after describing\\nPublication and Subscription.\\n\\nBTW, do you have any opinion on the idea of the first remaining patch\\nwhere we accomplish two things: a) Checks and throws an error if\\n'copy_data = on' and 'origin = none' but the publication tables were\\nalso replicated from other publishers. b) Adds 'force' value for\\ncopy_data parameter to allow copying in such a case. The primary\\nreason for this patch is to avoid loops or duplicate data in the\\ninitial phase. We can't skip copying based on origin as we can do\\nwhile replicating changes from WAL. So, we detect that the publisher\\nalready has data from some other node and doesn't allow replication\\nunless the user uses the 'force' option for copy_data.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 22 Jul 2022 10:17:01 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> Hi,\\n>\\n> On 7/21/22 6:34 AM, vignesh C wrote:\\n> > On Thu, Jul 21, 2022 at 2:06 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >>\\n> >> On Wed, Jul 20, 2022 at 2:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >>>\\n> >>> Modified. Apart from this I have run pgperltidy on the perl file and\\n> >>> renamed 032_origin.pl to 030_origin.pl as currently there is\\n> >>> 029_on_error.pl, 031_column_list.pl and there is no 030_*****.pl file.\\n> >>> Thanks for the comment, the attached patch has the changes for the same.\\n> >>>\\n> >>\\n> >> Pushed. Kindly rebase the remaining patches.\\n> >\\n> > Thanks for pushing the patch.\\n> > The attached v37 version contains the rebased patch for the remaining patches.\\n>\\n> Thanks for the work on this feature -- this is definitely very helpful\\n> towards supporting more types of use cases with logical replication!\\n>\\n> I've read through the proposed documentation and did some light testing\\n> of the patch. I have two general comments about the docs as they\\n> currently read:\\n>\\n> 1. I'm concerned by calling this \\\"Bidirectional replication\\\" in the docs\\n> that we are overstating the current capabilities. I think this is\\n> accentuated int he opening paragraph:\\n>\\n> ==snip==\\n> Bidirectional replication is useful for creating a multi-master database\\n> environment for replicating read/write operations performed by any of the\\n> member nodes.\\n> ==snip==\\n>\\n> For one, we're not replicating reads, we're replicating writes. Amongst\\n> the writes, at this point we're only replicating DML. A reader could\\n> think that deploying can work for a full bidirectional solution.\\n\\nI have changed read/write operations to write operations. I have also\\nadded a note saying \\\"The logical replication restrictions applies to\\nthe replication between primaries also.\\\", to clarify that non DML\\noperations and other restrictions apply in this case too.\\n\\n> (Even if we're aspirationally calling this section \\\"Bidirectional\\n> replication\\\", that does make it sound like we're limited to two nodes,\\n> when we can support more than two).\\n>\\n> Perhaps \\\"Logical replication between writers\\\" or \\\"Logical replication\\n> between primaries\\\" or \\\"Replicating changes between primaries\\\", or\\n> something better.\\n\\nI have changed it to \\\"Replication between primaries\\\".\\n\\n> 2. There is no mention of conflicts in the documentation, e.g.\\n> referencing the \\\"Conflicts\\\" section of the documentation. It's very easy\\n> to create a conflicting transaction that causes a subscriber to be\\n> unable to continue to apply transactions:\\n>\\n> -- DB 1\\n> CREATE TABLE abc (id int);\\n> CREATE PUBLICATION node1 FOR ALL TABLES ;\\n>\\n> -- DB2\\n> CREATE TABLE abc (id int);\\n> CREATE PUBLICATION node2 FOR ALL TABLES ;\\n> CREATE SUBSCRIPTION node2_node1\\n> CONNECTION 'dbname=logi port=5433'\\n> PUBLICATION node1\\n> WITH (copy_data = off, origin = none);\\n>\\n> -- DB1\\n> CREATE SUBSCRIPTION node1_node2\\n> CONNECTION 'dbname=logi port=5434'\\n> PUBLICATION node2\\n> WITH (copy_data = off, origin = none);\\n> INSERT INTO abc VALUES (1);\\n>\\n> -- DB2\\n> INSERT INTO abc VALUES (2);\\n>\\n> -- DB1\\n> ALTER TABLE abc ADD PRIMARY KEY id;\\n> INSERT INTO abc VALUES (3);\\n>\\n> -- DB2\\n> INSERT INTO abc VALUES (3);\\n>\\n> -- DB1 cannot apply the transactions\\n>\\n> At a minimum, I think we should reference the documentation we have in\\n> the logical replication section on conflicts. We may also want to advise\\n> that a user is responsible for designing their schemas in a way to\\n> minimize the risk of conflicts.\\n\\nAdded a note for the same and referred it to the conflicts section.\\n\\nThanks for the comments, the attached v38 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Sat, 23 Jul 2022 22:57:56 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jul 22, 2022 at 10:17 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n> >\\n> > Thanks for the work on this feature -- this is definitely very helpful\\n> > towards supporting more types of use cases with logical replication!\\n> >\\n> > I've read through the proposed documentation and did some light testing\\n> > of the patch. I have two general comments about the docs as they\\n> > currently read:\\n> >\\n> > 1. I'm concerned by calling this \\\"Bidirectional replication\\\" in the docs\\n> > that we are overstating the current capabilities. I think this is\\n> > accentuated int he opening paragraph:\\n> >\\n> > ==snip==\\n> > Bidirectional replication is useful for creating a multi-master database\\n> > environment for replicating read/write operations performed by any of the\\n> > member nodes.\\n> > ==snip==\\n> >\\n> > For one, we're not replicating reads, we're replicating writes. Amongst\\n> > the writes, at this point we're only replicating DML. A reader could\\n> > think that deploying can work for a full bidirectional solution.\\n> >\\n> > (Even if we're aspirationally calling this section \\\"Bidirectional\\n> > replication\\\", that does make it sound like we're limited to two nodes,\\n> > when we can support more than two).\\n> >\\n>\\n> Right, I think the system can support N-Way replication.\\n>\\n> > Perhaps \\\"Logical replication between writers\\\" or \\\"Logical replication\\n> > between primaries\\\" or \\\"Replicating changes between primaries\\\", or\\n> > something better.\\n> >\\n>\\n> Among the above \\\"Replicating changes between primaries\\\" sounds good to\\n> me or simply \\\"Replication between primaries\\\". As this is a sub-section\\n> on the Logical Replication page, I feel it is okay to not use Logical\\n> in the title.\\n\\nI have changed it to \\\"Replication between primaries\\\".\\n\\n> > 2. There is no mention of conflicts in the documentation, e.g.\\n> > referencing the \\\"Conflicts\\\" section of the documentation. It's very easy\\n> > to create a conflicting transaction that causes a subscriber to be\\n> > unable to continue to apply transactions:\\n> >\\n> > -- DB 1\\n> > CREATE TABLE abc (id int);\\n> > CREATE PUBLICATION node1 FOR ALL TABLES ;\\n> >\\n> > -- DB2\\n> > CREATE TABLE abc (id int);\\n> > CREATE PUBLICATION node2 FOR ALL TABLES ;\\n> > CREATE SUBSCRIPTION node2_node1\\n> > CONNECTION 'dbname=logi port=5433'\\n> > PUBLICATION node1\\n> > WITH (copy_data = off, origin = none);\\n> >\\n> > -- DB1\\n> > CREATE SUBSCRIPTION node1_node2\\n> > CONNECTION 'dbname=logi port=5434'\\n> > PUBLICATION node2\\n> > WITH (copy_data = off, origin = none);\\n> > INSERT INTO abc VALUES (1);\\n> >\\n> > -- DB2\\n> > INSERT INTO abc VALUES (2);\\n> >\\n> > -- DB1\\n> > ALTER TABLE abc ADD PRIMARY KEY id;\\n> > INSERT INTO abc VALUES (3);\\n> >\\n> > -- DB2\\n> > INSERT INTO abc VALUES (3);\\n> >\\n> > -- DB1 cannot apply the transactions\\n> >\\n> > At a minimum, I think we should reference the documentation we have in\\n> > the logical replication section on conflicts. We may also want to advise\\n> > that a user is responsible for designing their schemas in a way to\\n> > minimize the risk of conflicts.\\n> >\\n>\\n> This sounds reasonable to me.\\n>\\n> One more point about docs, it appears to be added as the last\\n> sub-section on the Logical Replication page. Is there a reason for\\n> doing so? I feel this should be third sub-section after describing\\n> Publication and Subscription.\\n\\nI had initially kept it at the end since we demonstrated the various\\nsteps to create the replication between the primaries. Since it gives\\nan introduction about the \\\"Replication between primaries\\\" and then\\nstates the steps, it looks ok to move it as suggested. I have modified\\nthis in the v38 patch attached at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm01x0sLz8YzfCSjxcMFxM4NDQxcFzZa%2B4eesUmD40DdTg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Sat, 23 Jul 2022 23:02:31 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On 7/22/22 12:47 AM, Amit Kapila wrote:\\r\\n> On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\r\\n\\r\\n>> 1. I'm concerned by calling this \\\"Bidirectional replication\\\" in the docs\\r\\n>> that we are overstating the current capabilities. I think this is\\r\\n>> accentuated int he opening paragraph:\\r\\n>>\\r\\n>> ==snip==\\r\\n>> Bidirectional replication is useful for creating a multi-master database\\r\\n>> environment for replicating read/write operations performed by any of the\\r\\n>> member nodes.\\r\\n>> ==snip==\\r\\n>>\\r\\n>> For one, we're not replicating reads, we're replicating writes. Amongst\\r\\n>> the writes, at this point we're only replicating DML. A reader could\\r\\n>> think that deploying can work for a full bidirectional solution.\\r\\n>>\\r\\n>> (Even if we're aspirationally calling this section \\\"Bidirectional\\r\\n>> replication\\\", that does make it sound like we're limited to two nodes,\\r\\n>> when we can support more than two).\\r\\n>>\\r\\n> \\r\\n> Right, I think the system can support N-Way replication.\\r\\n\\r\\nI did some more testing of the feature, i.e. doing 3-node and 4-node \\r\\ntests. While logical replication today can handle replicating between \\r\\nmultiple nodes (N-way), the \\\"origin = none\\\" does require setting up \\r\\nsubscribers between each of the nodes.\\r\\n\\r\\nFor example, if I have 4 nodes A, B, C, D and I want to replicate the \\r\\nsame table between all of them, I need to set up subscribers between all \\r\\nof them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node \\r\\ncan replicate between each other in a way that's convenient (vs. having \\r\\nto do something funky with partitions) so this is still a big step forward.\\r\\n\\r\\nThis is a long way of saying that I do think it's fair to say we support \\r\\n\\\"N-way\\\" replication so long as you are set up in a mesh (vs. a ring, \\r\\ne.g. A=>B=>C=>D=>A).\\r\\n\\r\\n> Among the above \\\"Replicating changes between primaries\\\" sounds good to\\r\\n> me or simply \\\"Replication between primaries\\\". As this is a sub-section\\r\\n> on the Logical Replication page, I feel it is okay to not use Logical\\r\\n> in the title.\\r\\n\\r\\nAgreed, I think that's fine.\\r\\n\\r\\n>> At a minimum, I think we should reference the documentation we have in\\r\\n>> the logical replication section on conflicts. We may also want to advise\\r\\n>> that a user is responsible for designing their schemas in a way to\\r\\n>> minimize the risk of conflicts.\\r\\n>>\\r\\n> \\r\\n> This sounds reasonable to me.\\r\\n> \\r\\n> One more point about docs, it appears to be added as the last\\r\\n> sub-section on the Logical Replication page. Is there a reason for\\r\\n> doing so? I feel this should be third sub-section after describing\\r\\n> Publication and Subscription.\\r\\n\\r\\nWhen I first reviewed, I had not built the docs. Did so on this pass.\\r\\n\\r\\nI agree with the positioning argument, i.e. it should go after \\r\\n\\\"Subscription\\\" in the table of contents -- but it makes me question a \\r\\ncouple of things:\\r\\n\\r\\n1. The general ordering of the docs\\r\\n2. How we describe that section (more on that in a sec)\\r\\n3. If \\\"row filters\\\" should be part of \\\"subscription\\\" instead of its own \\r\\nsection.\\r\\n\\r\\nIf you look at the current order, \\\"Quick setup\\\" is the last section; one \\r\\nwould think the \\\"quick\\\" portion goes first :) Given a lot of this is for \\r\\nthe current docs, I may start a separate discussion on -docs for this part.\\r\\n\\r\\nFor the time being, I agree it should be moved to the section after \\r\\n\\\"Subscription\\\".\\r\\n\\r\\nI think what this section describes is \\\"Configuring Replication Between \\r\\nNodes\\\" as it covers a few different scenarios.\\r\\n\\r\\nI do think we need to iterate on these docs -- the examples with the \\r\\ncommands are generally OK and easy to follow, but a few things I noticed:\\r\\n\\r\\n1. The general description of the section needs work. We may want to \\r\\nrefine the description of the use cases, and in the warning, link to \\r\\ninstructions on how to take backups.\\r\\n2. We put the \\\"case not supported\\\" in the middle, not at the end.\\r\\n3. The \\\"generic steps for adding a new node...\\\" section uses a \\r\\nconvention for steps that is not found in the docs. We also don't \\r\\nprovide an example for this section, and this is the most complicated \\r\\nscenario to set up.\\r\\n\\r\\nI may be able to propose some suggestions in a few days.\\r\\n\\r\\n> BTW, do you have any opinion on the idea of the first remaining patch\\r\\n> where we accomplish two things: a) Checks and throws an error if\\r\\n> 'copy_data = on' and 'origin = none' but the publication tables were\\r\\n> also replicated from other publishers. b) Adds 'force' value for\\r\\n> copy_data parameter to allow copying in such a case. The primary\\r\\n> reason for this patch is to avoid loops or duplicate data in the\\r\\n> initial phase. We can't skip copying based on origin as we can do\\r\\n> while replicating changes from WAL. So, we detect that the publisher\\r\\n> already has data from some other node and doesn't allow replication\\r\\n> unless the user uses the 'force' option for copy_data.\\r\\n\\r\\nIn general, I agree with the patch; but I'm not sure why we are calling \\r\\n\\\"copy_data = force\\\" in this case and how it varies from \\\"on\\\". I \\r\\nunderstand the goal is to prevent the infinite loop, but is there some \\r\\ntechnical restriction why we can't set \\\"origin = none, copy_data = on\\\" \\r\\nand have this work (and apologies if I missed that upthread)?\\r\\n\\r\\nThe other concern I'll note is that we're changing a boolean parameter \\r\\nto an enum and I want to be sensitive to folks who are already using \\r\\n\\\"copy_data\\\" to be sure we don't break them.\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Sun, 24 Jul 2022 12:51:48 -0400\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> On 7/22/22 12:47 AM, Amit Kapila wrote:\\n> > On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> >> 1. I'm concerned by calling this \\\"Bidirectional replication\\\" in the docs\\n> >> that we are overstating the current capabilities. I think this is\\n> >> accentuated int he opening paragraph:\\n> >>\\n> >> ==snip==\\n> >> Bidirectional replication is useful for creating a multi-master database\\n> >> environment for replicating read/write operations performed by any of the\\n> >> member nodes.\\n> >> ==snip==\\n> >>\\n> >> For one, we're not replicating reads, we're replicating writes. Amongst\\n> >> the writes, at this point we're only replicating DML. A reader could\\n> >> think that deploying can work for a full bidirectional solution.\\n> >>\\n> >> (Even if we're aspirationally calling this section \\\"Bidirectional\\n> >> replication\\\", that does make it sound like we're limited to two nodes,\\n> >> when we can support more than two).\\n> >>\\n> >\\n> > Right, I think the system can support N-Way replication.\\n>\\n> I did some more testing of the feature, i.e. doing 3-node and 4-node\\n> tests. While logical replication today can handle replicating between\\n> multiple nodes (N-way), the \\\"origin = none\\\" does require setting up\\n> subscribers between each of the nodes.\\n>\\n> For example, if I have 4 nodes A, B, C, D and I want to replicate the\\n> same table between all of them, I need to set up subscribers between all\\n> of them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node\\n> can replicate between each other in a way that's convenient (vs. having\\n> to do something funky with partitions) so this is still a big step forward.\\n>\\n> This is a long way of saying that I do think it's fair to say we support\\n> \\\"N-way\\\" replication so long as you are set up in a mesh (vs. a ring,\\n> e.g. A=>B=>C=>D=>A).\\n>\\n\\nSorry, but I don't get your point of mesh vs. ring? I think with some\\ncare users can set up replication even in a ring topology.\\n\\n> > Among the above \\\"Replicating changes between primaries\\\" sounds good to\\n> > me or simply \\\"Replication between primaries\\\". As this is a sub-section\\n> > on the Logical Replication page, I feel it is okay to not use Logical\\n> > in the title.\\n>\\n> Agreed, I think that's fine.\\n>\\n> >> At a minimum, I think we should reference the documentation we have in\\n> >> the logical replication section on conflicts. We may also want to advise\\n> >> that a user is responsible for designing their schemas in a way to\\n> >> minimize the risk of conflicts.\\n> >>\\n> >\\n> > This sounds reasonable to me.\\n> >\\n> > One more point about docs, it appears to be added as the last\\n> > sub-section on the Logical Replication page. Is there a reason for\\n> > doing so? I feel this should be third sub-section after describing\\n> > Publication and Subscription.\\n>\\n> When I first reviewed, I had not built the docs. Did so on this pass.\\n>\\n> I agree with the positioning argument, i.e. it should go after\\n> \\\"Subscription\\\" in the table of contents -- but it makes me question a\\n> couple of things:\\n>\\n> 1. The general ordering of the docs\\n> 2. How we describe that section (more on that in a sec)\\n> 3. If \\\"row filters\\\" should be part of \\\"subscription\\\" instead of its own\\n> section.\\n>\\n\\nI don't think it is a good idea to keep \\\"row filters\\\" as a part of\\nsubscription because we define those at publisher but there are\\ncertain things like initial sync or combining of row filters that are\\nrelated to subscriptions. So, probably having it in a separate\\nsub-section seems okay to me. I have also thought about keeping it as\\na part of Publication or Subscription but left it due to the reasons\\nmentioned.\\n\\n> If you look at the current order, \\\"Quick setup\\\" is the last section; one\\n> would think the \\\"quick\\\" portion goes first :) Given a lot of this is for\\n> the current docs, I may start a separate discussion on -docs for this part.\\n>\\n> For the time being, I agree it should be moved to the section after\\n> \\\"Subscription\\\".\\n>\\n\\nOkay, thanks!\\n\\n> I think what this section describes is \\\"Configuring Replication Between\\n> Nodes\\\" as it covers a few different scenarios.\\n>\\n> I do think we need to iterate on these docs -- the examples with the\\n> commands are generally OK and easy to follow, but a few things I noticed:\\n>\\n> 1. The general description of the section needs work. We may want to\\n> refine the description of the use cases, and in the warning, link to\\n> instructions on how to take backups.\\n> 2. We put the \\\"case not supported\\\" in the middle, not at the end.\\n> 3. The \\\"generic steps for adding a new node...\\\" section uses a\\n> convention for steps that is not found in the docs. We also don't\\n> provide an example for this section, and this is the most complicated\\n> scenario to set up.\\n>\\n\\nI agree with all these points, especially your point related to a\\ncomplicated setup. This won't be easy for users without a very clear\\ndescription and examples. However, after this work, it will at least\\nbe possible for users to set up an N-Way replication with some\\nrestrictions and care. BTW, while working on this we have noticed\\nthat it is normally a lot of work for users to set up N-way\\nreplication among 3, 4, or more nodes and that is why there is another\\nproposal to make that set up easier by providing a simpler APIs which\\nwill internally do similar to all the manual steps the 0002 patch is\\ntrying to describe. See [1] (please don't be confused with the thread\\ntitle but the real intent is to allow users to provide an easier way\\nto set up an N-Way replication by having all current restrictions of\\nLogical Replication like it doesn't support DDL replication)\\n\\n> I may be able to propose some suggestions in a few days.\\n>\\n\\nOkay, thanks! The plan is to get the 0001 patch of Vignesh. Then work\\non docs to describe how users can set up replication among primary\\nnodes. After that, if we get consensus on providing simpler APIs for\\nsetting up replication among primary nodes as is being discussed in\\nthe thread [1], then work on it.\\n\\n> > BTW, do you have any opinion on the idea of the first remaining patch\\n> > where we accomplish two things: a) Checks and throws an error if\\n> > 'copy_data = on' and 'origin = none' but the publication tables were\\n> > also replicated from other publishers. b) Adds 'force' value for\\n> > copy_data parameter to allow copying in such a case. The primary\\n> > reason for this patch is to avoid loops or duplicate data in the\\n> > initial phase. We can't skip copying based on origin as we can do\\n> > while replicating changes from WAL. So, we detect that the publisher\\n> > already has data from some other node and doesn't allow replication\\n> > unless the user uses the 'force' option for copy_data.\\n>\\n> In general, I agree with the patch; but I'm not sure why we are calling\\n> \\\"copy_data = force\\\" in this case and how it varies from \\\"on\\\". I\\n> understand the goal is to prevent the infinite loop, but is there some\\n> technical restriction why we can't set \\\"origin = none, copy_data = on\\\"\\n> and have this work (and apologies if I missed that upthread)?\\n>\\n\\nThe technical restriction is that we want to throw an error when users\\nset \\\"origin = none, copy_data = on\\\" and we find that the publication\\ntables were also replicated from other publishers. Now, we can give\\nWARNING to users in this case but users won't have any way to avoid\\nduplicate data which can lead to constraint violations. So, we decided\\nto throw ERROR and allow users to perform it with a \\\"force\\\" option.\\n\\n> The other concern I'll note is that we're changing a boolean parameter\\n> to an enum and I want to be sensitive to folks who are already using\\n> \\\"copy_data\\\" to be sure we don't break them.\\n>\\n\\nOkay, but AFAIU, the patch still allows users to specify on/off, so\\nwon't that be sufficient?\\n\\n[1] - https://www.postgresql.org/message-id/CAHut%2BPuwRAoWY9pz%3DEubps3ooQCOBFiYPU9Yi%3DVB-U%2ByORU7OA%40mail.gmail.com\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Jul 2022 09:50:58 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Firstly, I have some (case-sensitivity) questions about the previous\\npatch which was already pushed [1].\\n\\nQ1. create_subscription docs\\n\\nI did not understand why the docs refer to slot_name = NONE, yet the\\nnewly added option says origin = none/any. I think that saying origin\\n= NONE/ANY would be more consistent with the existing usage of NONE in\\nthis documentation.\\n\\n~~~\\n\\nQ2. parse_subscription_options\\n\\nSimilarly, in the code (parse_subscription_options), I did not\\nunderstand why the checks for special name values are implemented\\ndifferently:\\n\\nThe new 'origin' code is using pg_strcmpcase to check special values\\n(none/any), and the old 'slot_name' code uses case-sensitive strcmp to\\ncheck the special value (none).\\n\\nFWIW, here I thought the new origin code is the correct one.\\n\\n======\\n\\nNow, here are some review comments for the patch v38-0001:\\n\\n1. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n\\n@@ -1781,6 +1858,121 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\\n table_close(rel, RowExclusiveLock);\\n }\\n\\n+/*\\n+ * Check and throw an error if the publisher has subscribed to the same table\\n+ * from some other publisher. This check is required only if copydata is ON and\\n+ * the origin is local for CREATE SUBSCRIPTION and\\n+ * ALTER SUBSCRIPTION ... REFRESH statements to avoid replicating remote data\\n+ * from the publisher.\\n+ *\\n+ * This check need not be performed on the tables that are already added as\\n+ * incremental sync for such tables will happen through WAL and the origin of\\n+ * the data can be identified from the WAL records.\\n+ *\\n+ * subrel_local_oids contains the list of relation oids that are already\\n+ * present on the subscriber.\\n+ */\\n+static void\\n+check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\\n+ CopyData copydata, char *origin,\\n+ Oid *subrel_local_oids, int subrel_count)\\n\\n1a.\\n\\\"copydata is ON\\\" --> \\\"copy_data = on\\\" (because the comment is talking\\nabout the CREATE/ALTER statements, so it seemed a bit confusing to\\nrefer to the copydata function param instead of the copy_data\\nsubscription parameter)\\n\\n1b.\\n\\\"the origin is local\\\" ?? But, \\\"local\\\" was the old special name value.\\nNow it is \\\"none\\\", so I think this part needs minor rewording.\\n\\n~~~\\n\\n2.\\n\\n+ if (copydata != COPY_DATA_ON || !origin ||\\n+ (pg_strcasecmp(origin, \\\"none\\\") != 0))\\n+ return;\\n\\nShould this be using the constant LOGICALREP_ORIGIN_NONE?\\n\\n~~~\\n\\n3.\\n\\n+ /*\\n+ * Throw an error if the publisher has subscribed to the same table\\n+ * from some other publisher. We cannot differentiate between the\\n+ * origin and non-origin data that is present in the HEAP during the\\n+ * initial sync. Identification of non-origin data can be done only\\n+ * from the WAL by using the origin id.\\n+ *\\n+ * XXX: For simplicity, we don't check whether the table has any data\\n+ * or not. If the table doesn't have any data then we don't need to\\n+ * distinguish between local and non-local data so we can avoid\\n+ * throwing an error in that case.\\n+ */\\n\\n3a.\\nWhen the special origin value changed from \\\"local\\\" to \\\"none\\\" this\\ncomment's first part seems to have got a bit lost in translation.\\n\\nSUGGESTION:\\nThrow an error if the publisher has subscribed to the same table from\\nsome other publisher. We cannot know the origin of data during the\\ninitial sync. Data origins can be found only from the WAL by looking\\nat the origin id.\\n\\n3b.\\nI think referring to \\\"local and non-local\\\" data in the XXX part of\\nthis comment also needs some minor rewording now that \\\"local\\\" is not a\\nspecial origin name anymore.\\n\\n------\\n[1] https://github.com/postgres/postgres/commit/366283961ac0ed6d89014444c6090f3fd02fce0a\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Jul 2022 17:28:40 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> On 7/22/22 12:47 AM, Amit Kapila wrote:\\n> > On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> >> 1. I'm concerned by calling this \\\"Bidirectional replication\\\" in the docs\\n> >> that we are overstating the current capabilities. I think this is\\n> >> accentuated int he opening paragraph:\\n> >>\\n> >> ==snip==\\n> >> Bidirectional replication is useful for creating a multi-master database\\n> >> environment for replicating read/write operations performed by any of the\\n> >> member nodes.\\n> >> ==snip==\\n> >>\\n> >> For one, we're not replicating reads, we're replicating writes. Amongst\\n> >> the writes, at this point we're only replicating DML. A reader could\\n> >> think that deploying can work for a full bidirectional solution.\\n> >>\\n> >> (Even if we're aspirationally calling this section \\\"Bidirectional\\n> >> replication\\\", that does make it sound like we're limited to two nodes,\\n> >> when we can support more than two).\\n> >>\\n> >\\n> > Right, I think the system can support N-Way replication.\\n>\\n> I did some more testing of the feature, i.e. doing 3-node and 4-node\\n> tests. While logical replication today can handle replicating between\\n> multiple nodes (N-way), the \\\"origin = none\\\" does require setting up\\n> subscribers between each of the nodes.\\n>\\n> For example, if I have 4 nodes A, B, C, D and I want to replicate the\\n> same table between all of them, I need to set up subscribers between all\\n> of them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node\\n> can replicate between each other in a way that's convenient (vs. having\\n> to do something funky with partitions) so this is still a big step forward.\\n>\\n> This is a long way of saying that I do think it's fair to say we support\\n> \\\"N-way\\\" replication so long as you are set up in a mesh (vs. a ring,\\n> e.g. A=>B=>C=>D=>A).\\n>\\n> > Among the above \\\"Replicating changes between primaries\\\" sounds good to\\n> > me or simply \\\"Replication between primaries\\\". As this is a sub-section\\n> > on the Logical Replication page, I feel it is okay to not use Logical\\n> > in the title.\\n>\\n> Agreed, I think that's fine.\\n>\\n> >> At a minimum, I think we should reference the documentation we have in\\n> >> the logical replication section on conflicts. We may also want to advise\\n> >> that a user is responsible for designing their schemas in a way to\\n> >> minimize the risk of conflicts.\\n> >>\\n> >\\n> > This sounds reasonable to me.\\n> >\\n> > One more point about docs, it appears to be added as the last\\n> > sub-section on the Logical Replication page. Is there a reason for\\n> > doing so? I feel this should be third sub-section after describing\\n> > Publication and Subscription.\\n>\\n> When I first reviewed, I had not built the docs. Did so on this pass.\\n>\\n> I agree with the positioning argument, i.e. it should go after\\n> \\\"Subscription\\\" in the table of contents -- but it makes me question a\\n> couple of things:\\n>\\n> 1. The general ordering of the docs\\n> 2. How we describe that section (more on that in a sec)\\n> 3. If \\\"row filters\\\" should be part of \\\"subscription\\\" instead of its own\\n> section.\\n>\\n> If you look at the current order, \\\"Quick setup\\\" is the last section; one\\n> would think the \\\"quick\\\" portion goes first :) Given a lot of this is for\\n> the current docs, I may start a separate discussion on -docs for this part.\\n>\\n> For the time being, I agree it should be moved to the section after\\n> \\\"Subscription\\\".\\n>\\n> I think what this section describes is \\\"Configuring Replication Between\\n> Nodes\\\" as it covers a few different scenarios.\\n>\\n> I do think we need to iterate on these docs -- the examples with the\\n> commands are generally OK and easy to follow, but a few things I noticed:\\n>\\n> 1. The general description of the section needs work. We may want to\\n> refine the description of the use cases, and in the warning, link to\\n> instructions on how to take backups.\\n> 2. We put the \\\"case not supported\\\" in the middle, not at the end.\\n> 3. The \\\"generic steps for adding a new node...\\\" section uses a\\n> convention for steps that is not found in the docs. We also don't\\n> provide an example for this section, and this is the most complicated\\n> scenario to set up.\\n>\\n> I may be able to propose some suggestions in a few days.\\n>\\n> > BTW, do you have any opinion on the idea of the first remaining patch\\n> > where we accomplish two things: a) Checks and throws an error if\\n> > 'copy_data = on' and 'origin = none' but the publication tables were\\n> > also replicated from other publishers. b) Adds 'force' value for\\n> > copy_data parameter to allow copying in such a case. The primary\\n> > reason for this patch is to avoid loops or duplicate data in the\\n> > initial phase. We can't skip copying based on origin as we can do\\n> > while replicating changes from WAL. So, we detect that the publisher\\n> > already has data from some other node and doesn't allow replication\\n> > unless the user uses the 'force' option for copy_data.\\n>\\n> In general, I agree with the patch; but I'm not sure why we are calling\\n> \\\"copy_data = force\\\" in this case and how it varies from \\\"on\\\". I\\n> understand the goal is to prevent the infinite loop, but is there some\\n> technical restriction why we can't set \\\"origin = none, copy_data = on\\\"\\n> and have this work (and apologies if I missed that upthread)?\\n\\nLet's take a simple case to understand why copy_data = force is\\nrequired to replicate between two primaries for table t1 which has\\ndata as given below:\\nNode-1:\\nTable t1 (c1 int) has data\\n1, 2, 3, 4\\n\\nNode-2:\\nTable t1 (c1 int) has data\\n5, 6, 7, 8\\n\\nstep1 - Node-1\\n#Publication for t1\\nCreate Publication pub1_2 For Table t1;\\n\\nstep2 - Node-2\\n#Publication for t1,\\nCreate Publication pub2_1 For Table t1;\\n\\nstep3 - Node-1:\\nCreate Subscription sub1 Connection '<node-2 details>' publication\\npub2_1 with (origin = none);\\n\\nAfter this the data will be something like this:\\nNode-1:\\n1, 2, 3, 4, 5, 6, 7, 8\\n\\nNode-2:\\n5, 6, 7, 8\\n\\nstep4 - Node-2:\\nCreate Subscription sub2 Connection '<node-1 details>' Publication\\npub1_2 with (origin = none, copy_data=on);\\nIf we had allowed the create subscription to be successful with\\ncopy_data = on. After this the data will be something like this:\\nNode-1:\\n1, 2, 3, 4, 5, 6, 7, 8\\n\\nNode-2:\\n1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\\n\\nSo, you can see that data on Node-2 (5, 6, 7, 8) is duplicated. In\\ncase, table t1 has a unique key, it will lead to a unique key\\nviolation and replication won't proceed.\\n\\nTo avoid this we will throw an error:\\nERROR: could not replicate table \\\"public.t1\\\"\\nDETAIL: CREATE/ALTER SUBSCRIPTION with origin = none and copy_data =\\non is not allowed when the publisher has subscribed same table.\\nHINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\n\\nUsers can then overcome this problem by using the following steps:\\nstep1 to step3 is the same as above.\\n\\nstep4 - Node-2\\n# Disallow truncates to be published and then truncate the table\\nAlter Publication pub2_1 Set (publish = 'insert, update, delete');\\nTruncate t1;\\n\\nAfter this the data will be like this:\\nNode-1:\\n1, 2, 3, 4, 5, 6, 7, 8\\n\\nNode-2: no data\\n\\nstep5 - Node-2\\nCreate Subscription sub2 Connection '<node-1 details>' Publication\\npub1_2 with (origin = none, copy_data = force);\\n\\nAfter this the data will be in sync:\\nNode-1:\\n1, 2, 3, 4, 5, 6, 7, 8\\n\\nNode-2:\\n1, 2, 3, 4, 5, 6, 7, 8\\n\\nstep6 - Node-1\\n# Now include truncates to be published\\nAlter Publication pub1_2 Set (publish = 'insert, update, delete, truncate');\\n\\nNow the replication setup between two primaries node1 and node2 is\\ncomplete. Any incremental changes from node1 will be replicated to\\nnode2, and any incremental changes from node2 will be replicated to\\nnode1.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Jul 2022 14:24:00 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Firstly, I have some (case-sensitivity) questions about the previous\\n> patch which was already pushed [1].\\n>\\n> Q1. create_subscription docs\\n>\\n> I did not understand why the docs refer to slot_name = NONE, yet the\\n> newly added option says origin = none/any. I think that saying origin\\n> = NONE/ANY would be more consistent with the existing usage of NONE in\\n> this documentation.\\n>\\n> ~~~\\n>\\n> Q2. parse_subscription_options\\n>\\n> Similarly, in the code (parse_subscription_options), I did not\\n> understand why the checks for special name values are implemented\\n> differently:\\n>\\n> The new 'origin' code is using pg_strcmpcase to check special values\\n> (none/any), and the old 'slot_name' code uses case-sensitive strcmp to\\n> check the special value (none).\\n>\\n\\nWe have a restriction for slot_names for lower case letters (aka\\n\\\"Replication slot names may only contain lower case letters, numbers,\\nand the underscore character.\\\") whereas there is no such restriction\\nin origin name, that's why the check is different. So, if you try with\\nslot_name = 'NONE', you will get the error.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Jul 2022 14:24:43 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Firstly, I have some (case-sensitivity) questions about the previous\\n> patch which was already pushed [1].\\n>\\n> Q1. create_subscription docs\\n>\\n> I did not understand why the docs refer to slot_name = NONE, yet the\\n> newly added option says origin = none/any. I think that saying origin\\n> = NONE/ANY would be more consistent with the existing usage of NONE in\\n> this documentation.\\n\\nBoth NONE and none are ok in the case of origin, if you want I can\\nchange it to NONE/ANY in case of origin to keep it consistent.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Jul 2022 15:03:25 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 25, 2022 at 2:24 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n> >\\n> > On 7/22/22 12:47 AM, Amit Kapila wrote:\\n> > > On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n> >\\n> > >> 1. I'm concerned by calling this \\\"Bidirectional replication\\\" in the docs\\n> > >> that we are overstating the current capabilities. I think this is\\n> > >> accentuated int he opening paragraph:\\n> > >>\\n> > >> ==snip==\\n> > >> Bidirectional replication is useful for creating a multi-master database\\n> > >> environment for replicating read/write operations performed by any of the\\n> > >> member nodes.\\n> > >> ==snip==\\n> > >>\\n> > >> For one, we're not replicating reads, we're replicating writes. Amongst\\n> > >> the writes, at this point we're only replicating DML. A reader could\\n> > >> think that deploying can work for a full bidirectional solution.\\n> > >>\\n> > >> (Even if we're aspirationally calling this section \\\"Bidirectional\\n> > >> replication\\\", that does make it sound like we're limited to two nodes,\\n> > >> when we can support more than two).\\n> > >>\\n> > >\\n> > > Right, I think the system can support N-Way replication.\\n> >\\n> > I did some more testing of the feature, i.e. doing 3-node and 4-node\\n> > tests. While logical replication today can handle replicating between\\n> > multiple nodes (N-way), the \\\"origin = none\\\" does require setting up\\n> > subscribers between each of the nodes.\\n> >\\n> > For example, if I have 4 nodes A, B, C, D and I want to replicate the\\n> > same table between all of them, I need to set up subscribers between all\\n> > of them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node\\n> > can replicate between each other in a way that's convenient (vs. having\\n> > to do something funky with partitions) so this is still a big step forward.\\n> >\\n> > This is a long way of saying that I do think it's fair to say we support\\n> > \\\"N-way\\\" replication so long as you are set up in a mesh (vs. a ring,\\n> > e.g. A=>B=>C=>D=>A).\\n> >\\n> > > Among the above \\\"Replicating changes between primaries\\\" sounds good to\\n> > > me or simply \\\"Replication between primaries\\\". As this is a sub-section\\n> > > on the Logical Replication page, I feel it is okay to not use Logical\\n> > > in the title.\\n> >\\n> > Agreed, I think that's fine.\\n> >\\n> > >> At a minimum, I think we should reference the documentation we have in\\n> > >> the logical replication section on conflicts. We may also want to advise\\n> > >> that a user is responsible for designing their schemas in a way to\\n> > >> minimize the risk of conflicts.\\n> > >>\\n> > >\\n> > > This sounds reasonable to me.\\n> > >\\n> > > One more point about docs, it appears to be added as the last\\n> > > sub-section on the Logical Replication page. Is there a reason for\\n> > > doing so? I feel this should be third sub-section after describing\\n> > > Publication and Subscription.\\n> >\\n> > When I first reviewed, I had not built the docs. Did so on this pass.\\n> >\\n> > I agree with the positioning argument, i.e. it should go after\\n> > \\\"Subscription\\\" in the table of contents -- but it makes me question a\\n> > couple of things:\\n> >\\n> > 1. The general ordering of the docs\\n> > 2. How we describe that section (more on that in a sec)\\n> > 3. If \\\"row filters\\\" should be part of \\\"subscription\\\" instead of its own\\n> > section.\\n> >\\n> > If you look at the current order, \\\"Quick setup\\\" is the last section; one\\n> > would think the \\\"quick\\\" portion goes first :) Given a lot of this is for\\n> > the current docs, I may start a separate discussion on -docs for this part.\\n> >\\n> > For the time being, I agree it should be moved to the section after\\n> > \\\"Subscription\\\".\\n> >\\n> > I think what this section describes is \\\"Configuring Replication Between\\n> > Nodes\\\" as it covers a few different scenarios.\\n> >\\n> > I do think we need to iterate on these docs -- the examples with the\\n> > commands are generally OK and easy to follow, but a few things I noticed:\\n> >\\n> > 1. The general description of the section needs work. We may want to\\n> > refine the description of the use cases, and in the warning, link to\\n> > instructions on how to take backups.\\n> > 2. We put the \\\"case not supported\\\" in the middle, not at the end.\\n> > 3. The \\\"generic steps for adding a new node...\\\" section uses a\\n> > convention for steps that is not found in the docs. We also don't\\n> > provide an example for this section, and this is the most complicated\\n> > scenario to set up.\\n> >\\n> > I may be able to propose some suggestions in a few days.\\n> >\\n> > > BTW, do you have any opinion on the idea of the first remaining patch\\n> > > where we accomplish two things: a) Checks and throws an error if\\n> > > 'copy_data = on' and 'origin = none' but the publication tables were\\n> > > also replicated from other publishers. b) Adds 'force' value for\\n> > > copy_data parameter to allow copying in such a case. The primary\\n> > > reason for this patch is to avoid loops or duplicate data in the\\n> > > initial phase. We can't skip copying based on origin as we can do\\n> > > while replicating changes from WAL. So, we detect that the publisher\\n> > > already has data from some other node and doesn't allow replication\\n> > > unless the user uses the 'force' option for copy_data.\\n> >\\n> > In general, I agree with the patch; but I'm not sure why we are calling\\n> > \\\"copy_data = force\\\" in this case and how it varies from \\\"on\\\". I\\n> > understand the goal is to prevent the infinite loop, but is there some\\n> > technical restriction why we can't set \\\"origin = none, copy_data = on\\\"\\n> > and have this work (and apologies if I missed that upthread)?\\n>\\n> Let's take a simple case to understand why copy_data = force is\\n> required to replicate between two primaries for table t1 which has\\n> data as given below:\\n> Node-1:\\n> Table t1 (c1 int) has data\\n> 1, 2, 3, 4\\n>\\n> Node-2:\\n> Table t1 (c1 int) has data\\n> 5, 6, 7, 8\\n>\\n> step1 - Node-1\\n> #Publication for t1\\n> Create Publication pub1_2 For Table t1;\\n>\\n> step2 - Node-2\\n> #Publication for t1,\\n> Create Publication pub2_1 For Table t1;\\n>\\n> step3 - Node-1:\\n> Create Subscription sub1 Connection '<node-2 details>' publication\\n> pub2_1 with (origin = none);\\n>\\n> After this the data will be something like this:\\n> Node-1:\\n> 1, 2, 3, 4, 5, 6, 7, 8\\n>\\n> Node-2:\\n> 5, 6, 7, 8\\n>\\n> step4 - Node-2:\\n> Create Subscription sub2 Connection '<node-1 details>' Publication\\n> pub1_2 with (origin = none, copy_data=on);\\n> If we had allowed the create subscription to be successful with\\n> copy_data = on. After this the data will be something like this:\\n> Node-1:\\n> 1, 2, 3, 4, 5, 6, 7, 8\\n>\\n> Node-2:\\n> 1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\\n>\\n> So, you can see that data on Node-2 (5, 6, 7, 8) is duplicated. In\\n> case, table t1 has a unique key, it will lead to a unique key\\n> violation and replication won't proceed.\\n>\\n> To avoid this we will throw an error:\\n> ERROR: could not replicate table \\\"public.t1\\\"\\n> DETAIL: CREATE/ALTER SUBSCRIPTION with origin = none and copy_data =\\n> on is not allowed when the publisher has subscribed same table.\\n> HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\n>\\n> Users can then overcome this problem by using the following steps:\\n> step1 to step3 is the same as above.\\n>\\n> step4 - Node-2\\n> # Disallow truncates to be published and then truncate the table\\n> Alter Publication pub2_1 Set (publish = 'insert, update, delete');\\n> Truncate t1;\\n>\\n> After this the data will be like this:\\n> Node-1:\\n> 1, 2, 3, 4, 5, 6, 7, 8\\n>\\n> Node-2: no data\\n>\\n> step5 - Node-2\\n> Create Subscription sub2 Connection '<node-1 details>' Publication\\n> pub1_2 with (origin = none, copy_data = force);\\n>\\n> After this the data will be in sync:\\n> Node-1:\\n> 1, 2, 3, 4, 5, 6, 7, 8\\n>\\n> Node-2:\\n> 1, 2, 3, 4, 5, 6, 7, 8\\n>\\n> step6 - Node-1\\n> # Now include truncates to be published\\n> Alter Publication pub1_2 Set (publish = 'insert, update, delete, truncate');\\n>\\n> Now the replication setup between two primaries node1 and node2 is\\n> complete. Any incremental changes from node1 will be replicated to\\n> node2, and any incremental changes from node2 will be replicated to\\n> node1.\\n\\nIn the above steps, sorry that I mentioned Node-1 instead of Node-2 in\\nthe last step step6.\\nThe below:\\n step6 - Node-1\\n # Now include truncates to be published\\n Alter Publication pub1_2 Set (publish = 'insert, update, delete, truncate');\\n\\nshould be:\\nstep6 - Node-2\\n# Now include truncates to be published\\nAlter Publication pub2_1 Set (publish = 'insert, update, delete, truncate');\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Jul 2022 20:14:28 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 25, 2022 at 6:54 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n...\\n> >\\n> > Q2. parse_subscription_options\\n> >\\n> > Similarly, in the code (parse_subscription_options), I did not\\n> > understand why the checks for special name values are implemented\\n> > differently:\\n> >\\n> > The new 'origin' code is using pg_strcmpcase to check special values\\n> > (none/any), and the old 'slot_name' code uses case-sensitive strcmp to\\n> > check the special value (none).\\n> >\\n>\\n> We have a restriction for slot_names for lower case letters (aka\\n> \\\"Replication slot names may only contain lower case letters, numbers,\\n> and the underscore character.\\\") whereas there is no such restriction\\n> in origin name, that's why the check is different. So, if you try with\\n> slot_name = 'NONE', you will get the error.\\n>\\n\\n2022-07-26 09:06:06.380 AEST [3630] STATEMENT: create subscription\\nmysub connection 'host=localhost' publication mypub with\\n(slot_name='None', enabled=false, create_slot=false);\\nERROR: replication slot name \\\"None\\\" contains invalid character\\nHINT: Replication slot names may only contain lower case letters,\\nnumbers, and the underscore character.\\n\\nYou are right. Thanks for the explanation.\\n\\n(Aside: Probably that error message wording ought to say \\\"contains\\ninvalid characters\\\" instead of \\\"contains invalid character\\\")\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 09:13:17 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 25, 2022 at 7:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Firstly, I have some (case-sensitivity) questions about the previous\\n> > patch which was already pushed [1].\\n> >\\n> > Q1. create_subscription docs\\n> >\\n> > I did not understand why the docs refer to slot_name = NONE, yet the\\n> > newly added option says origin = none/any. I think that saying origin\\n> > = NONE/ANY would be more consistent with the existing usage of NONE in\\n> > this documentation.\\n>\\n> Both NONE and none are ok in the case of origin, if you want I can\\n> change it to NONE/ANY in case of origin to keep it consistent.\\n>\\n\\nI preferred the special origin values should be documented as NONE/ANY\\nfor better consistency, but let's see what others think about it.\\n\\nThere will also be associated minor changes needed for a few\\nerror/hint messages.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 09:34:11 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On 7/25/22 4:54 AM, vignesh C wrote:\\r\\n> On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\r\\n>>\\r\\n>> On 7/22/22 12:47 AM, Amit Kapila wrote:\\r\\n>>> On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\r\\n\\r\\n>>> BTW, do you have any opinion on the idea of the first remaining patch\\r\\n>>> where we accomplish two things: a) Checks and throws an error if\\r\\n>>> 'copy_data = on' and 'origin = none' but the publication tables were\\r\\n>>> also replicated from other publishers. b) Adds 'force' value for\\r\\n>>> copy_data parameter to allow copying in such a case. The primary\\r\\n>>> reason for this patch is to avoid loops or duplicate data in the\\r\\n>>> initial phase. We can't skip copying based on origin as we can do\\r\\n>>> while replicating changes from WAL. So, we detect that the publisher\\r\\n>>> already has data from some other node and doesn't allow replication\\r\\n>>> unless the user uses the 'force' option for copy_data.\\r\\n>>\\r\\n>> In general, I agree with the patch; but I'm not sure why we are calling\\r\\n>> \\\"copy_data = force\\\" in this case and how it varies from \\\"on\\\". I\\r\\n>> understand the goal is to prevent the infinite loop, but is there some\\r\\n>> technical restriction why we can't set \\\"origin = none, copy_data = on\\\"\\r\\n>> and have this work (and apologies if I missed that upthread)?\\r\\n> \\r\\n> Let's take a simple case to understand why copy_data = force is\\r\\n> required to replicate between two primaries for table t1 which has\\r\\n> data as given below:\\r\\n\\r\\n> step4 - Node-2:\\r\\n> Create Subscription sub2 Connection '<node-1 details>' Publication\\r\\n> pub1_2 with (origin = none, copy_data=on);\\r\\n> If we had allowed the create subscription to be successful with\\r\\n> copy_data = on. After this the data will be something like this:\\r\\n> Node-1:\\r\\n> 1, 2, 3, 4, 5, 6, 7, 8\\r\\n> \\r\\n> Node-2:\\r\\n> 1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\\r\\n> \\r\\n> So, you can see that data on Node-2 (5, 6, 7, 8) is duplicated. In\\r\\n> case, table t1 has a unique key, it will lead to a unique key\\r\\n> violation and replication won't proceed.\\r\\n> \\r\\n> To avoid this we will throw an error:\\r\\n> ERROR: could not replicate table \\\"public.t1\\\"\\r\\n> DETAIL: CREATE/ALTER SUBSCRIPTION with origin = none and copy_data =\\r\\n> on is not allowed when the publisher has subscribed same table.\\r\\n> HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\r\\n\\r\\nThanks for the example. I agree that it is fairly simple to reproduce.\\r\\n\\r\\nI understand that \\\"copy_data = force\\\" is meant to protect a user from \\r\\nhurting themself. I'm not convinced that this is the best way to do so.\\r\\n\\r\\nFor example today I can subscribe to multiple publications that write to \\r\\nthe same table. If I have a primary key on that table, and two of the \\r\\nsubscriptions try to write an identical ID, we conflict. We don't have \\r\\nany special flags or modes to guard against that from happening, though \\r\\nwe do have documentation on conflicts and managing them.\\r\\n\\r\\nAFAICT the same issue with \\\"copy_data\\\" also exists in the above scenario \\r\\ntoo, even without the \\\"origin\\\" attribute. However, I think this case is \\r\\nmore noticeable for \\\"origin=none\\\" because we currently default \\r\\n\\\"copy_data\\\" to \\\"true\\\" and in this case data can be copied in two \\r\\ndirections.\\r\\n\\r\\nThat said, this introduces a new restriction for this particular \\r\\nscenario that doesn't exist on other scenarios. Instead, I would \\r\\nadvocate we document how to correctly set up the two-way replication \\r\\nscenario (which we have a draft!), document the warnings around the \\r\\nconflicts, perhaps include Vignesh's instructions on how to remediate a \\r\\nconflict on initial sync, and consider throwing a WARNING as you suggested.\\r\\n\\r\\nThoughts?\\r\\n\\r\\nThanks,\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Mon, 25 Jul 2022 21:43:37 -0400\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are some review comments for the patch v38-0002:\\n\\n======\\n\\n<General> - terminology\\n\\nThere seemed to be an inconsistent alternation of the terms\\n\\\"primaries\\\" and \\\"nodes\\\"... For example \\\"Setting replication between\\ntwo primaries\\\" versus \\\"Adding a new node...\\\" (instead of \\\"Adding a new\\nprimary...\\\"?). I have included suggested minor rewording changes in\\nthe review comments below, but please check in case I miss something.\\nBecause I suggested changes to some titles maybe you will also want to\\nchange the section ids too.\\n\\n~~~\\n\\n1. Commit message\\n\\nThe documentation was recently modified to remove the term\\n\\\"bidirectional replication\\\" and replace it all with \\\"replication\\nbetween primaries\\\", so this commit message (and also the patch name\\nitself) should be similarly modified.\\n\\n~~~\\n\\n2.\\n+ <para>\\n+ Replication between primaries is useful for creating a multi-master\\n+ database environment for replicating write operations performed by any of\\n+ the member nodes. The steps to create replication between primaries in\\n+ various scenarios are given below. Note: User is responsible for designing\\n+ their schemas in a way to minimize the risk of conflicts. See\\n+ <xref linkend=\\\"logical-replication-conflicts\\\"/> for the details of logical\\n+ replication conflicts. The logical replication restrictions applies to\\n+ the replication between primaries also. See\\n+ <xref linkend=\\\"logical-replication-restrictions\\\"/> for the details of\\n+ logical replication restrictions.\\n+ </para>\\n\\n2a.\\n\\\"User\\\" -> \\\"The user\\\"\\n\\n2b.\\n\\\"The logical replication restrictions applies to...\\\" --> \\\"The logical\\nreplication restrictions apply to...\\\"\\n\\n2c.\\nThese are important notes. Instead of just being part of the text\\nblurb, perhaps these should be rendered as SGML <note> (or put them\\nboth in a single <note> if you want)\\n\\n~~~\\n\\n3. Setting replication between two primaries\\n\\n+ <title>Setting replication between two primaries</title>\\n+ <para>\\n+ The following steps demonstrate how to setup replication between two\\n+ primaries when there is no table data present on both nodes\\n+ <literal>node1</literal> and <literal>node2</literal>:\\n+ </para>\\n\\nSUGGESTED\\nThe following steps demonstrate how to set up replication between two\\nprimaries (node1 and node2) when there is no table data present on\\nboth nodes:\\n\\n~~~\\n\\n4.\\n+ <para>\\n+ Now the replication setup between two primaries <literal>node1</literal>\\n+ and <literal>node2</literal> is complete. Any incremental changes from\\n+ <literal>node1</literal> will be replicated to <literal>node2</literal>,\\n+ and any incremental changes from <literal>node2</literal> will be\\n+ replicated to <literal>node1</literal>.\\n+ </para>\\n\\n\\\"between two primaries\\\" -> \\\"between primaries\\\"\\n\\n~~~\\n\\n5. Adding a new node when there is no table data on any of the nodes\\n\\nSUGGESTION (title)\\nAdding a new primary when there is no table data on any of the primaries\\n\\n~~~\\n\\n6.\\n+ <para>\\n+ The following steps demonstrate adding a new node <literal>node3</literal>\\n+ to the existing <literal>node1</literal> and <literal>node2</literal> when\\n+ there is no <literal>t1</literal> data on any of the nodes. This requires\\n\\nSUGGESTION\\nThe following steps demonstrate adding a new primary (node3) to the\\nexisting primaries (node1 and node2) when there is no t1 data on any\\nof the nodes.\\n\\n~~~\\n\\n7. Adding a new node when table data is present on the existing nodes\\n\\nSUGGESTION (title)\\nAdding a new primary when table data is present on the existing primaries\\n\\n~~~\\n\\n8.\\n+ <para>\\n+ The following steps demonstrate adding a new node <literal>node3</literal>\\n+ which has no <literal>t1</literal> data to the existing\\n+ <literal>node1</literal> and <literal>node2</literal> where\\n+ <literal>t1</literal> data is present. This needs similar steps; the only\\n\\nSUGGESTION\\nThe following steps demonstrate adding a new primary (node3) that has\\nno t1 data to the existing primaries (node1 and node2) where t1 data\\nis present.\\n\\n~~~\\n\\n9. Adding a new node when table data is present on the new node\\n\\nSUGGESTION (title)\\nAdding a new primary that has existing table data\\n\\n~~~\\n\\n10.\\n+ <note>\\n+ <para>\\n+ Adding a new node when table data is present on the new node is not\\n+ supported.\\n+ </para>\\n+ </note>\\n\\nSUGGESTION\\nAdding a new primary that has existing table data is not supported.\\n\\n~~~\\n\\n11. Generic steps for adding a new node to an existing set of primaries\\n\\nSUGGESTION (title)\\nGeneric steps for adding a new primary to an existing set of primaries\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 11:45:52 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 11:43 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n...\\n> That said, this introduces a new restriction for this particular\\n> scenario that doesn't exist on other scenarios. Instead, I would\\n> advocate we document how to correctly set up the two-way replication\\n> scenario (which we have a draft!), document the warnings around the\\n> conflicts, perhaps include Vignesh's instructions on how to remediate a\\n> conflict on initial sync, and consider throwing a WARNING as you suggested.\\n>\\n> Thoughts?\\n\\nPerhaps a WARNING can be useful if the SUBSCRIPTION was created with\\nenabled=false because then the user still has a chance to reconsider,\\nbut otherwise, I don't see what good a warning does if the potentially\\nharmful initial copy is going to proceed anyway; isn't that like\\nputting a warning sign at the bottom of a cliff?\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 12:18:05 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 7:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> On 7/25/22 4:54 AM, vignesh C wrote:\\n> >\\n> > Let's take a simple case to understand why copy_data = force is\\n> > required to replicate between two primaries for table t1 which has\\n> > data as given below:\\n>\\n> > step4 - Node-2:\\n> > Create Subscription sub2 Connection '<node-1 details>' Publication\\n> > pub1_2 with (origin = none, copy_data=on);\\n> > If we had allowed the create subscription to be successful with\\n> > copy_data = on. After this the data will be something like this:\\n> > Node-1:\\n> > 1, 2, 3, 4, 5, 6, 7, 8\\n> >\\n> > Node-2:\\n> > 1, 2, 3, 4, 5, 6, 7, 8, 5, 6, 7, 8\\n> >\\n> > So, you can see that data on Node-2 (5, 6, 7, 8) is duplicated. In\\n> > case, table t1 has a unique key, it will lead to a unique key\\n> > violation and replication won't proceed.\\n> >\\n> > To avoid this we will throw an error:\\n> > ERROR: could not replicate table \\\"public.t1\\\"\\n> > DETAIL: CREATE/ALTER SUBSCRIPTION with origin = none and copy_data =\\n> > on is not allowed when the publisher has subscribed same table.\\n> > HINT: Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\n>\\n> Thanks for the example. I agree that it is fairly simple to reproduce.\\n>\\n> I understand that \\\"copy_data = force\\\" is meant to protect a user from\\n> hurting themself. I'm not convinced that this is the best way to do so.\\n>\\n> For example today I can subscribe to multiple publications that write to\\n> the same table. If I have a primary key on that table, and two of the\\n> subscriptions try to write an identical ID, we conflict. We don't have\\n> any special flags or modes to guard against that from happening, though\\n> we do have documentation on conflicts and managing them.\\n>\\n> AFAICT the same issue with \\\"copy_data\\\" also exists in the above scenario\\n> too, even without the \\\"origin\\\" attribute.\\n>\\n\\nThat's true but there is no parameter like origin = NONE which\\nindicates that constraint violations or duplicate data problems won't\\noccur due to replication. In the current case, I think the situation\\nis different because a user has specifically asked not to replicate\\nany remote data by specifying origin = NONE, which should be dealt\\ndifferently. Note that current users or their setup won't see any\\ndifference/change unless they specify the new parameter origin as\\nNONE.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 09:07:04 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 7:48 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Tue, Jul 26, 2022 at 11:43 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n> >\\n> ...\\n> > That said, this introduces a new restriction for this particular\\n> > scenario that doesn't exist on other scenarios. Instead, I would\\n> > advocate we document how to correctly set up the two-way replication\\n> > scenario (which we have a draft!), document the warnings around the\\n> > conflicts, perhaps include Vignesh's instructions on how to remediate a\\n> > conflict on initial sync, and consider throwing a WARNING as you suggested.\\n> >\\n> > Thoughts?\\n>\\n> Perhaps a WARNING can be useful if the SUBSCRIPTION was created with\\n> enabled=false because then the user still has a chance to reconsider,\\n>\\n\\nAgreed. I think but in that case, when the user enables it, we need to\\nensure that we won't allow replicating (during initial sync) remote\\ndata. If this is really required/preferred, it can be done as a\\nseparate enhancement.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 09:12:23 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 5:04 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Mon, Jul 25, 2022 at 7:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > >\\n> > > Firstly, I have some (case-sensitivity) questions about the previous\\n> > > patch which was already pushed [1].\\n> > >\\n> > > Q1. create_subscription docs\\n> > >\\n> > > I did not understand why the docs refer to slot_name = NONE, yet the\\n> > > newly added option says origin = none/any. I think that saying origin\\n> > > = NONE/ANY would be more consistent with the existing usage of NONE in\\n> > > this documentation.\\n> >\\n> > Both NONE and none are ok in the case of origin, if you want I can\\n> > change it to NONE/ANY in case of origin to keep it consistent.\\n> >\\n>\\n> I preferred the special origin values should be documented as NONE/ANY\\n> for better consistency, but let's see what others think about it.\\n>\\n> There will also be associated minor changes needed for a few\\n> error/hint messages.\\n>\\n\\nI am not really sure how much we gain by maintaining consistency with\\nslot_name because if due to this we have to change the error messages\\nas well then it can create an inconsistency with reserved origin\\nnames. Consider message: DETAIL: Origin names \\\"any\\\", \\\"none\\\", and\\nnames starting with \\\"pg_\\\" are reserved. Now, if we change this to\\n\\\"ANY\\\", \\\"NONE\\\" in the above message, it will look a bit odd as \\\"pg_\\\"\\nstarts with lower case letters.\\n\\n--\\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 09:39:01 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 2:09 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Jul 26, 2022 at 5:04 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > On Mon, Jul 25, 2022 at 7:33 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > > >\\n> > > > Firstly, I have some (case-sensitivity) questions about the previous\\n> > > > patch which was already pushed [1].\\n> > > >\\n> > > > Q1. create_subscription docs\\n> > > >\\n> > > > I did not understand why the docs refer to slot_name = NONE, yet the\\n> > > > newly added option says origin = none/any. I think that saying origin\\n> > > > = NONE/ANY would be more consistent with the existing usage of NONE in\\n> > > > this documentation.\\n> > >\\n> > > Both NONE and none are ok in the case of origin, if you want I can\\n> > > change it to NONE/ANY in case of origin to keep it consistent.\\n> > >\\n> >\\n> > I preferred the special origin values should be documented as NONE/ANY\\n> > for better consistency, but let's see what others think about it.\\n> >\\n> > There will also be associated minor changes needed for a few\\n> > error/hint messages.\\n> >\\n>\\n> I am not really sure how much we gain by maintaining consistency with\\n> slot_name because if due to this we have to change the error messages\\n> as well then it can create an inconsistency with reserved origin\\n> names. Consider message: DETAIL: Origin names \\\"any\\\", \\\"none\\\", and\\n> names starting with \\\"pg_\\\" are reserved. Now, if we change this to\\n> \\\"ANY\\\", \\\"NONE\\\" in the above message, it will look a bit odd as \\\"pg_\\\"\\n> starts with lower case letters.\\n>\\n\\nSure, the message looks a bit odd with the quotes like you wrote\\nabove, but I would not suggest to change it that way - I was thinking\\nmore like below (which is similar to the style the slot_name messages\\nuse)\\n\\nCURRENT\\nDETAIL: Origin names \\\"any\\\", \\\"none\\\", and names starting with \\\"pg_\\\" are reserved.\\n\\nSUGGESTED\\nDETAIL: Origin names ANY, NONE, and names starting with \\\"pg_\\\" are reserved.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 14:53:05 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sun, Jul 24, 2022 1:28 AM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> Added a note for the same and referred it to the conflicts section.\\r\\n> \\r\\n> Thanks for the comments, the attached v38 patch has the changes for the\\r\\n> same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch. A comment on the test in 0001 patch.\\r\\n\\r\\n+# Alter subscription ... refresh publication should fail when a new table is\\r\\n+# subscribing data from a different publication should fail\\r\\n+($result, $stdout, $stderr) = $node_A->psql(\\r\\n+\\t'postgres', \\\"\\r\\n+ ALTER SUBSCRIPTION tap_sub_A2 REFRESH PUBLICATION\\\");\\r\\n+like(\\r\\n+\\t$stderr,\\r\\n+\\tqr/ERROR: ( [A-Z0-9]+:)? could not replicate table \\\"public.tab_new\\\"/,\\r\\n+\\t\\\"Create subscription with origin and copy_data having replicated table in publisher\\\"\\r\\n+);\\r\\n\\r\\nThe comment says \\\"should fail\\\" twice, the latter one can be removed.\\r\\n\\r\\nBesides, \\\"Create subscription with origin and copy_data\\\" should be changed to\\r\\n\\\"Alter subscription with origin and copy_data\\\" I think.\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 08:05:05 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sun, Jul 24, 2022 1:28 AM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> Added a note for the same and referred it to the conflicts section.\\r\\n> \\r\\n> Thanks for the comments, the attached v38 patch has the changes for the same.\\r\\n\\r\\nThanks for your patches.\\r\\n\\r\\nTwo slight comments on the below message in the 0001 patch:\\r\\n\\r\\nThe error message in the function check_pub_table_subscribed().\\r\\n+\\t\\tereport(ERROR,\\r\\n+\\t\\t\\t\\terrcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\r\\n+\\t\\t\\t\\terrmsg(\\\"could not replicate table \\\\\\\"%s.%s\\\\\\\"\\\",\\r\\n+\\t\\t\\t\\t\\t nspname, relname),\\r\\n+\\t\\t\\t\\terrdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin = none and copy_data = on is not allowed when the publisher has subscribed same table.\\\"),\\r\\n+\\t\\t\\t\\terrhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\\"));\\r\\n\\r\\n1.\\r\\nI think it might be better to use true/false here than on/off.\\r\\nJust for consistency with another error message\\r\\n(in function parse_subscription_options) and the description of this parameter\\r\\nin the PG document.\\r\\n\\r\\nIf you agree with this, please also kindly consider the attached\\r\\n\\\"slight_modification.diff\\\" file.\\r\\n(This is a slight modification for the second patch, just replace \\\"off\\\" with\\r\\n\\\"false\\\" in PG document.)\\r\\n\\r\\n2.\\r\\nHow about replacing \\\"origin = none\\\" and \\\"copy_data = on\\\" in the message with\\r\\n\\\"%s\\\"? I think this might be better for translation. Just like the following\\r\\nerror message in the function parse_subscription_options:\\r\\n```\\r\\n\\t\\tif (opts->copy_data &&\\r\\n\\t\\t\\tIsSet(opts->specified_opts, SUBOPT_COPY_DATA))\\r\\n\\t\\t\\tereport(ERROR,\\r\\n\\t\\t\\t\\t\\t(errcode(ERRCODE_SYNTAX_ERROR),\\r\\n\\t\\t\\t\\t\\t errmsg(\\\"%s and %s are mutually exclusive options\\\",\\r\\n\\t\\t\\t\\t\\t\\t\\t\\\"connect = false\\\", \\\"copy_data = true/force\\\")));\\r\\n```\\r\\n\\r\\nRegards,\\r\\nWang wei\",\n \"msg_date\": \"Tue, 26 Jul 2022 08:42:09 +0000\",\n \"msg_from\": \"\\\"wangw.fnst@fujitsu.com\\\" <wangw.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Sun, Jul 24, 2022 at 10:21 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> On 7/22/22 12:47 AM, Amit Kapila wrote:\\n> > On Fri, Jul 22, 2022 at 1:39 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> >> 1. I'm concerned by calling this \\\"Bidirectional replication\\\" in the docs\\n> >> that we are overstating the current capabilities. I think this is\\n> >> accentuated int he opening paragraph:\\n> >>\\n> >> ==snip==\\n> >> Bidirectional replication is useful for creating a multi-master database\\n> >> environment for replicating read/write operations performed by any of the\\n> >> member nodes.\\n> >> ==snip==\\n> >>\\n> >> For one, we're not replicating reads, we're replicating writes. Amongst\\n> >> the writes, at this point we're only replicating DML. A reader could\\n> >> think that deploying can work for a full bidirectional solution.\\n> >>\\n> >> (Even if we're aspirationally calling this section \\\"Bidirectional\\n> >> replication\\\", that does make it sound like we're limited to two nodes,\\n> >> when we can support more than two).\\n> >>\\n> >\\n> > Right, I think the system can support N-Way replication.\\n>\\n> I did some more testing of the feature, i.e. doing 3-node and 4-node\\n> tests. While logical replication today can handle replicating between\\n> multiple nodes (N-way), the \\\"origin = none\\\" does require setting up\\n> subscribers between each of the nodes.\\n>\\n> For example, if I have 4 nodes A, B, C, D and I want to replicate the\\n> same table between all of them, I need to set up subscribers between all\\n> of them (A<=>B, A<=>C, A<=>D, B<=>C, B<=>D, C<=>D). However, each node\\n> can replicate between each other in a way that's convenient (vs. having\\n> to do something funky with partitions) so this is still a big step forward.\\n>\\n> This is a long way of saying that I do think it's fair to say we support\\n> \\\"N-way\\\" replication so long as you are set up in a mesh (vs. a ring,\\n> e.g. A=>B=>C=>D=>A).\\n>\\n> > Among the above \\\"Replicating changes between primaries\\\" sounds good to\\n> > me or simply \\\"Replication between primaries\\\". As this is a sub-section\\n> > on the Logical Replication page, I feel it is okay to not use Logical\\n> > in the title.\\n>\\n> Agreed, I think that's fine.\\n>\\n> >> At a minimum, I think we should reference the documentation we have in\\n> >> the logical replication section on conflicts. We may also want to advise\\n> >> that a user is responsible for designing their schemas in a way to\\n> >> minimize the risk of conflicts.\\n> >>\\n> >\\n> > This sounds reasonable to me.\\n> >\\n> > One more point about docs, it appears to be added as the last\\n> > sub-section on the Logical Replication page. Is there a reason for\\n> > doing so? I feel this should be third sub-section after describing\\n> > Publication and Subscription.\\n>\\n> When I first reviewed, I had not built the docs. Did so on this pass.\\n>\\n> I agree with the positioning argument, i.e. it should go after\\n> \\\"Subscription\\\" in the table of contents -- but it makes me question a\\n> couple of things:\\n>\\n> 1. The general ordering of the docs\\n> 2. How we describe that section (more on that in a sec)\\n> 3. If \\\"row filters\\\" should be part of \\\"subscription\\\" instead of its own\\n> section.\\n>\\n> If you look at the current order, \\\"Quick setup\\\" is the last section; one\\n> would think the \\\"quick\\\" portion goes first :) Given a lot of this is for\\n> the current docs, I may start a separate discussion on -docs for this part.\\n>\\n> For the time being, I agree it should be moved to the section after\\n> \\\"Subscription\\\".\\n>\\n> I think what this section describes is \\\"Configuring Replication Between\\n> Nodes\\\" as it covers a few different scenarios.\\n>\\n> I do think we need to iterate on these docs -- the examples with the\\n> commands are generally OK and easy to follow, but a few things I noticed:\\n>\\n> 1. The general description of the section needs work. We may want to\\n> refine the description of the use cases, and in the warning, link to\\n> instructions on how to take backups.\\n\\nModified\\n\\n> 2. We put the \\\"case not supported\\\" in the middle, not at the end.\\n\\nModified\\n\\n> 3. The \\\"generic steps for adding a new node...\\\" section uses a\\n> convention for steps that is not found in the docs. We also don't\\n> provide an example for this section, and this is the most complicated\\n> scenario to set up.\\n\\nModified\\n\\nThanks a lot for the suggestions, I have made the changes for the\\nsame in the v39 patch attached.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 26 Jul 2022 23:27:40 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 7:16 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some review comments for the patch v38-0002:\\n>\\n> ======\\n>\\n> <General> - terminology\\n>\\n> There seemed to be an inconsistent alternation of the terms\\n> \\\"primaries\\\" and \\\"nodes\\\"... For example \\\"Setting replication between\\n> two primaries\\\" versus \\\"Adding a new node...\\\" (instead of \\\"Adding a new\\n> primary...\\\"?). I have included suggested minor rewording changes in\\n> the review comments below, but please check in case I miss something.\\n> Because I suggested changes to some titles maybe you will also want to\\n> change the section ids too.\\n>\\n> ~~~\\n>\\n> 1. Commit message\\n>\\n> The documentation was recently modified to remove the term\\n> \\\"bidirectional replication\\\" and replace it all with \\\"replication\\n> between primaries\\\", so this commit message (and also the patch name\\n> itself) should be similarly modified.\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.\\n> + <para>\\n> + Replication between primaries is useful for creating a multi-master\\n> + database environment for replicating write operations performed by any of\\n> + the member nodes. The steps to create replication between primaries in\\n> + various scenarios are given below. Note: User is responsible for designing\\n> + their schemas in a way to minimize the risk of conflicts. See\\n> + <xref linkend=\\\"logical-replication-conflicts\\\"/> for the details of logical\\n> + replication conflicts. The logical replication restrictions applies to\\n> + the replication between primaries also. See\\n> + <xref linkend=\\\"logical-replication-restrictions\\\"/> for the details of\\n> + logical replication restrictions.\\n> + </para>\\n>\\n> 2a.\\n> \\\"User\\\" -> \\\"The user\\\"\\n\\nModified\\n\\n> 2b.\\n> \\\"The logical replication restrictions applies to...\\\" --> \\\"The logical\\n> replication restrictions apply to...\\\"\\n\\nModified\\n\\n> 2c.\\n> These are important notes. Instead of just being part of the text\\n> blurb, perhaps these should be rendered as SGML <note> (or put them\\n> both in a single <note> if you want)\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. Setting replication between two primaries\\n>\\n> + <title>Setting replication between two primaries</title>\\n> + <para>\\n> + The following steps demonstrate how to setup replication between two\\n> + primaries when there is no table data present on both nodes\\n> + <literal>node1</literal> and <literal>node2</literal>:\\n> + </para>\\n>\\n> SUGGESTED\\n> The following steps demonstrate how to set up replication between two\\n> primaries (node1 and node2) when there is no table data present on\\n> both nodes:.\\n\\nModified\\n\\n> ~~~\\n>\\n> 4.\\n> + <para>\\n> + Now the replication setup between two primaries <literal>node1</literal>\\n> + and <literal>node2</literal> is complete. Any incremental changes from\\n> + <literal>node1</literal> will be replicated to <literal>node2</literal>,\\n> + and any incremental changes from <literal>node2</literal> will be\\n> + replicated to <literal>node1</literal>.\\n> + </para>\\n>\\n> \\\"between two primaries\\\" -> \\\"between primaries\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. Adding a new node when there is no table data on any of the nodes\\n>\\n> SUGGESTION (title)\\n> Adding a new primary when there is no table data on any of the primaries\\n\\nModified\\n\\n> ~~~\\n>\\n> 6.\\n> + <para>\\n> + The following steps demonstrate adding a new node <literal>node3</literal>\\n> + to the existing <literal>node1</literal> and <literal>node2</literal> when\\n> + there is no <literal>t1</literal> data on any of the nodes. This requires\\n>\\n> SUGGESTION\\n> The following steps demonstrate adding a new primary (node3) to the\\n> existing primaries (node1 and node2) when there is no t1 data on any\\n> of the nodes.\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. Adding a new node when table data is present on the existing nodes\\n>\\n> SUGGESTION (title)\\n> Adding a new primary when table data is present on the existing primaries\\n\\nModified\\n\\n> ~~~\\n>\\n> 8.\\n> + <para>\\n> + The following steps demonstrate adding a new node <literal>node3</literal>\\n> + which has no <literal>t1</literal> data to the existing\\n> + <literal>node1</literal> and <literal>node2</literal> where\\n> + <literal>t1</literal> data is present. This needs similar steps; the only\\n>\\n> SUGGESTION\\n> The following steps demonstrate adding a new primary (node3) that has\\n> no t1 data to the existing primaries (node1 and node2) where t1 data\\n> is present.\\n\\nModified\\n\\n> ~~~\\n>\\n> 9. Adding a new node when table data is present on the new node\\n>\\n> SUGGESTION (title)\\n> Adding a new primary that has existing table data\\n\\nModified\\n\\n> ~~~\\n>\\n> 10.\\n> + <note>\\n> + <para>\\n> + Adding a new node when table data is present on the new node is not\\n> + supported.\\n> + </para>\\n> + </note>\\n>\\n> SUGGESTION\\n> Adding a new primary that has existing table data is not supported.\\n\\nModified\\n\\n> ~~~\\n>\\n> 11. Generic steps for adding a new node to an existing set of primaries\\n>\\n> SUGGESTION (title)\\n> Generic steps for adding a new primary to an existing set of primaries\\n\\nModified\\n\\nThanks for the comments, the v39 patch shared at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm2POATc_jwQ-8MBJgGCVZGdUNhnTv8zkBuGzLaY03dM%3DA%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 23:34:50 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 1:35 PM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Sun, Jul 24, 2022 1:28 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Added a note for the same and referred it to the conflicts section.\\n> >\\n> > Thanks for the comments, the attached v38 patch has the changes for the\\n> > same.\\n> >\\n>\\n> Thanks for updating the patch. A comment on the test in 0001 patch.\\n>\\n> +# Alter subscription ... refresh publication should fail when a new table is\\n> +# subscribing data from a different publication should fail\\n> +($result, $stdout, $stderr) = $node_A->psql(\\n> + 'postgres', \\\"\\n> + ALTER SUBSCRIPTION tap_sub_A2 REFRESH PUBLICATION\\\");\\n> +like(\\n> + $stderr,\\n> + qr/ERROR: ( [A-Z0-9]+:)? could not replicate table \\\"public.tab_new\\\"/,\\n> + \\\"Create subscription with origin and copy_data having replicated table in publisher\\\"\\n> +);\\n>\\n> The comment says \\\"should fail\\\" twice, the latter one can be removed.\\n\\nModified\\n\\n> Besides, \\\"Create subscription with origin and copy_data\\\" should be changed to\\n> \\\"Alter subscription with origin and copy_data\\\" I think.\\n\\nModified to \\\"Refresh publication\\\"\\n\\nThanks for the comments, the v39 patch shared at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm2POATc_jwQ-8MBJgGCVZGdUNhnTv8zkBuGzLaY03dM%3DA%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 23:37:10 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 2:12 PM wangw.fnst@fujitsu.com\\n<wangw.fnst@fujitsu.com> wrote:\\n>\\n> On Sun, Jul 24, 2022 1:28 AM vignesh C <vignesh21@gmail.com> wrote:\\n> > Added a note for the same and referred it to the conflicts section.\\n> >\\n> > Thanks for the comments, the attached v38 patch has the changes for the same.\\n>\\n> Thanks for your patches.\\n>\\n> Two slight comments on the below message in the 0001 patch:\\n>\\n> The error message in the function check_pub_table_subscribed().\\n> + ereport(ERROR,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"could not replicate table \\\\\\\"%s.%s\\\\\\\"\\\",\\n> + nspname, relname),\\n> + errdetail(\\\"CREATE/ALTER SUBSCRIPTION with origin = none and copy_data = on is not allowed when the publisher has subscribed same table.\\\"),\\n> + errhint(\\\"Use CREATE/ALTER SUBSCRIPTION with copy_data = off/force.\\\"));\\n>\\n> 1.\\n> I think it might be better to use true/false here than on/off.\\n> Just for consistency with another error message\\n> (in function parse_subscription_options) and the description of this parameter\\n> in the PG document.\\n\\nModified\\n\\n> If you agree with this, please also kindly consider the attached\\n> \\\"slight_modification.diff\\\" file.\\n> (This is a slight modification for the second patch, just replace \\\"off\\\" with\\n> \\\"false\\\" in PG document.)\\n\\nI have updated the documentation similarly\\n\\n> 2.\\n> How about replacing \\\"origin = none\\\" and \\\"copy_data = on\\\" in the message with\\n> \\\"%s\\\"? I think this might be better for translation. Just like the following\\n> error message in the function parse_subscription_options:\\n> ```\\n> if (opts->copy_data &&\\n> IsSet(opts->specified_opts, SUBOPT_COPY_DATA))\\n> ereport(ERROR,\\n> (errcode(ERRCODE_SYNTAX_ERROR),\\n> errmsg(\\\"%s and %s are mutually exclusive options\\\",\\n> \\\"connect = false\\\", \\\"copy_data = true/force\\\")));\\n> ```\\n\\nModified\\n\\nThanks for the comments, the v39 patch shared at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm2POATc_jwQ-8MBJgGCVZGdUNhnTv8zkBuGzLaY03dM%3DA%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 23:38:04 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Jul 25, 2022 at 12:58 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Firstly, I have some (case-sensitivity) questions about the previous\\n> patch which was already pushed [1].\\n>\\n> Q1. create_subscription docs\\n>\\n> I did not understand why the docs refer to slot_name = NONE, yet the\\n> newly added option says origin = none/any. I think that saying origin\\n> = NONE/ANY would be more consistent with the existing usage of NONE in\\n> this documentation.\\n>\\n> ~~~\\n>\\n> Q2. parse_subscription_options\\n>\\n> Similarly, in the code (parse_subscription_options), I did not\\n> understand why the checks for special name values are implemented\\n> differently:\\n>\\n> The new 'origin' code is using pg_strcmpcase to check special values\\n> (none/any), and the old 'slot_name' code uses case-sensitive strcmp to\\n> check the special value (none).\\n>\\n> FWIW, here I thought the new origin code is the correct one.\\n>\\n> ======\\n>\\n> Now, here are some review comments for the patch v38-0001:\\n>\\n> 1. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> @@ -1781,6 +1858,121 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\\n> table_close(rel, RowExclusiveLock);\\n> }\\n>\\n> +/*\\n> + * Check and throw an error if the publisher has subscribed to the same table\\n> + * from some other publisher. This check is required only if copydata is ON and\\n> + * the origin is local for CREATE SUBSCRIPTION and\\n> + * ALTER SUBSCRIPTION ... REFRESH statements to avoid replicating remote data\\n> + * from the publisher.\\n> + *\\n> + * This check need not be performed on the tables that are already added as\\n> + * incremental sync for such tables will happen through WAL and the origin of\\n> + * the data can be identified from the WAL records.\\n> + *\\n> + * subrel_local_oids contains the list of relation oids that are already\\n> + * present on the subscriber.\\n> + */\\n> +static void\\n> +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\\n> + CopyData copydata, char *origin,\\n> + Oid *subrel_local_oids, int subrel_count)\\n>\\n> 1a.\\n> \\\"copydata is ON\\\" --> \\\"copy_data = on\\\" (because the comment is talking\\n> about the CREATE/ALTER statements, so it seemed a bit confusing to\\n> refer to the copydata function param instead of the copy_data\\n> subscription parameter)\\n\\nModified\\n\\n> 1b.\\n> \\\"the origin is local\\\" ?? But, \\\"local\\\" was the old special name value.\\n> Now it is \\\"none\\\", so I think this part needs minor rewording.\\n\\nModified\\n\\n>\\n> ~~~\\n>\\n> 2.\\n>\\n> + if (copydata != COPY_DATA_ON || !origin ||\\n> + (pg_strcasecmp(origin, \\\"none\\\") != 0))\\n> + return;\\n>\\n> Should this be using the constant LOGICALREP_ORIGIN_NONE?\\n\\nModified\\n\\n> ~~~\\n>\\n> 3.\\n>\\n> + /*\\n> + * Throw an error if the publisher has subscribed to the same table\\n> + * from some other publisher. We cannot differentiate between the\\n> + * origin and non-origin data that is present in the HEAP during the\\n> + * initial sync. Identification of non-origin data can be done only\\n> + * from the WAL by using the origin id.\\n> + *\\n> + * XXX: For simplicity, we don't check whether the table has any data\\n> + * or not. If the table doesn't have any data then we don't need to\\n> + * distinguish between local and non-local data so we can avoid\\n> + * throwing an error in that case.\\n> + */\\n>\\n> 3a.\\n> When the special origin value changed from \\\"local\\\" to \\\"none\\\" this\\n> comment's first part seems to have got a bit lost in translation.\\n\\nModified\\n\\n> SUGGESTION:\\n> Throw an error if the publisher has subscribed to the same table from\\n> some other publisher. We cannot know the origin of data during the\\n> initial sync. Data origins can be found only from the WAL by looking\\n> at the origin id.\\n\\nModified\\n\\n> 3b.\\n> I think referring to \\\"local and non-local\\\" data in the XXX part of\\n> this comment also needs some minor rewording now that \\\"local\\\" is not a\\n> special origin name anymore.\\n\\nModified\\n\\nThanks for the comments, the v39 patch shared at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm2POATc_jwQ-8MBJgGCVZGdUNhnTv8zkBuGzLaY03dM%3DA%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 26 Jul 2022 23:43:53 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 10:23 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Tue, Jul 26, 2022 at 2:09 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> >\\n> > I am not really sure how much we gain by maintaining consistency with\\n> > slot_name because if due to this we have to change the error messages\\n> > as well then it can create an inconsistency with reserved origin\\n> > names. Consider message: DETAIL: Origin names \\\"any\\\", \\\"none\\\", and\\n> > names starting with \\\"pg_\\\" are reserved. Now, if we change this to\\n> > \\\"ANY\\\", \\\"NONE\\\" in the above message, it will look a bit odd as \\\"pg_\\\"\\n> > starts with lower case letters.\\n> >\\n>\\n> Sure, the message looks a bit odd with the quotes like you wrote\\n> above, but I would not suggest to change it that way - I was thinking\\n> more like below (which is similar to the style the slot_name messages\\n> use)\\n>\\n> CURRENT\\n> DETAIL: Origin names \\\"any\\\", \\\"none\\\", and names starting with \\\"pg_\\\" are reserved.\\n>\\n> SUGGESTED\\n> DETAIL: Origin names ANY, NONE, and names starting with \\\"pg_\\\" are reserved.\\n>\\n\\nI see your point but not sure if that is an improvement over the\\ncurrent one, so, let's wait and see if we get some other votes in\\nfavor of your suggestion.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 27 Jul 2022 08:23:19 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi Vignesh.\\n\\nFYI the v39* patch fails to apply [1]. Can you please rebase it?\\n\\n\\n[1]\\n=== Applying patches on top of PostgreSQL commit ID\\n5f858dd3bebd1f3845aef2bff7f4345bfb7b74b3 ===\\n=== applying patch\\n./v39-0001-Check-and-throw-an-error-if-publication-tables-w.patch\\npatching file doc/src/sgml/ref/alter_subscription.sgml\\npatching file doc/src/sgml/ref/create_subscription.sgml\\npatching file src/backend/commands/subscriptioncmds.c\\nHunk #10 FAILED at 886.\\n1 out of 14 hunks FAILED -- saving rejects to file\\nsrc/backend/commands/subscriptioncmds.c.rej\\npatching file src/test/regress/expected/subscription.out\\npatching file src/test/regress/sql/subscription.sql\\npatching file src/test/subscription/t/030_origin.pl\\npatching file src/tools/pgindent/typedefs.list\\n\\n------\\n[1] http://cfbot.cputube.org/patch_38_3610.log\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 28 Jul 2022 15:58:35 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jul 28, 2022 at 11:28 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Hi Vignesh.\\n>\\n> FYI the v39* patch fails to apply [1]. Can you please rebase it?\\n>\\n>\\n> [1]\\n> === Applying patches on top of PostgreSQL commit ID\\n> 5f858dd3bebd1f3845aef2bff7f4345bfb7b74b3 ===\\n> === applying patch\\n> ./v39-0001-Check-and-throw-an-error-if-publication-tables-w.patch\\n> patching file doc/src/sgml/ref/alter_subscription.sgml\\n> patching file doc/src/sgml/ref/create_subscription.sgml\\n> patching file src/backend/commands/subscriptioncmds.c\\n> Hunk #10 FAILED at 886.\\n> 1 out of 14 hunks FAILED -- saving rejects to file\\n> src/backend/commands/subscriptioncmds.c.rej\\n> patching file src/test/regress/expected/subscription.out\\n> patching file src/test/regress/sql/subscription.sql\\n> patching file src/test/subscription/t/030_origin.pl\\n> patching file src/tools/pgindent/typedefs.list\\n\\nThanks for reporting this, I will post an updated version for this soon.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Thu, 28 Jul 2022 16:15:42 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Jul 28, 2022 at 11:28 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Hi Vignesh.\\n>\\n> FYI the v39* patch fails to apply [1]. Can you please rebase it?\\n>\\n>\\n> [1]\\n> === Applying patches on top of PostgreSQL commit ID\\n> 5f858dd3bebd1f3845aef2bff7f4345bfb7b74b3 ===\\n> === applying patch\\n> ./v39-0001-Check-and-throw-an-error-if-publication-tables-w.patch\\n> patching file doc/src/sgml/ref/alter_subscription.sgml\\n> patching file doc/src/sgml/ref/create_subscription.sgml\\n> patching file src/backend/commands/subscriptioncmds.c\\n> Hunk #10 FAILED at 886.\\n> 1 out of 14 hunks FAILED -- saving rejects to file\\n> src/backend/commands/subscriptioncmds.c.rej\\n> patching file src/test/regress/expected/subscription.out\\n> patching file src/test/regress/sql/subscription.sql\\n> patching file src/test/subscription/t/030_origin.pl\\n> patching file src/tools/pgindent/typedefs.list\\n>\\n> ------\\n\\nPlease find the v40 patch attached which is rebased on top of head.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Thu, 28 Jul 2022 19:33:31 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are some comments for the patch v40-0001:\\n\\n======\\n\\n1. Commit message\\n\\nIt might be better to always use 'copy_data = true' in favour of\\n'copy_data = on' just for consistency with all the docs and the error\\nmessages.\\n\\n======\\n\\n2. doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -386,6 +401,15 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n can have non-existent publications.\\n </para>\\n\\n+ <para>\\n+ If the subscription is created with <literal>origin = NONE</literal> and\\n+ <literal>copy_data = true</literal>, it will check if the publisher has\\n+ subscribed to the same table from other publishers and, if so, throw an\\n+ error to prevent possible non-local data from being copied. The user can\\n+ override this check and continue with the copy operation by specifying\\n+ <literal>copy_data = force</literal>.\\n+ </para>\\n\\n2a.\\nIt is interesting that you changed the note to say origin = NONE.\\nPersonally, I prefer it written as you did, but I think maybe this\\nchange does not belong in this patch. The suggestion for changing from\\n\\\"none\\\" to NONE is being discussed elsewhere in this thread and\\nprobably all such changes should be done together (if at all) as a\\nseparate patch. Until then I think this patch 0001 should just stay\\nconsistent with whatever is already pushed on HEAD.\\n\\n2b.\\n\\\"possible no-local data\\\". Maybe the terminology \\\"local/non-local\\\" is a\\nhangover from back when the subscription parameter was called local\\ninstead of origin. I'm not sure if you want to change this or not, and\\nanyway I didn't have any better suggestions – so this comment is just\\nto bring it to your attention.\\n\\n======\\n\\n3. src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n\\n+ /*\\n+ * The set of strings accepted here should match up with the\\n+ * grammar's opt_boolean_or_string production.\\n+ */\\n+ if (pg_strcasecmp(sval, \\\"false\\\") == 0 ||\\n+ pg_strcasecmp(sval, \\\"off\\\") == 0)\\n+ return COPY_DATA_OFF;\\n+ if (pg_strcasecmp(sval, \\\"true\\\") == 0 ||\\n+ pg_strcasecmp(sval, \\\"on\\\") == 0)\\n+ return COPY_DATA_ON;\\n+ if (pg_strcasecmp(sval, \\\"force\\\") == 0)\\n+ return COPY_DATA_FORCE;\\n\\nI understand the intention of the comment, but it is not strictly\\ncorrect to say \\\"should match up\\\" because \\\"force\\\" is a new value.\\nPerhaps the comment should be as suggested below.\\n\\nSUGGESTION\\nThe set of strings accepted here must include all those accepted by\\nthe grammar's opt_boolean_or_string production.\\n\\n~~~\\n\\n4. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n\\n@@ -1781,6 +1858,122 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\\n table_close(rel, RowExclusiveLock);\\n }\\n\\n+/*\\n+ * Check and throw an error if the publisher has subscribed to the same table\\n+ * from some other publisher. This check is required only if \\\"copy_data = on\\\"\\n+ * and \\\"origin = NONE\\\" for CREATE SUBSCRIPTION and\\n+ * ALTER SUBSCRIPTION ... REFRESH statements to avoid the publisher from\\n+ * replicating data that has an origin.\\n+ *\\n+ * This check need not be performed on the tables that are already added as\\n+ * incremental sync for such tables will happen through WAL and the origin of\\n+ * the data can be identified from the WAL records.\\n+ *\\n+ * subrel_local_oids contains the list of relation oids that are already\\n+ * present on the subscriber.\\n+ */\\n+static void\\n+check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\\n+ CopyData copydata, char *origin,\\n+ Oid *subrel_local_oids, int subrel_count)\\n\\n4a.\\n\\\"copy_data = on\\\" -> \\\"copy_data = true\\\" (for consistency with the docs\\nand the error messages)\\n\\n4b.\\nThe same NONE/none review comment from #2a applies here too. Probably\\nit should be written as none for now unless/until *everything* changes\\nto NONE.\\n\\n4c.\\n\\\"to avoid the publisher from replicating data that has an origin.\\\" ->\\n\\\"to avoid replicating data that has an origin.\\\"\\n\\n4d.\\n+ * This check need not be performed on the tables that are already added as\\n+ * incremental sync for such tables will happen through WAL and the origin of\\n+ * the data can be identified from the WAL records.\\n\\nSUGGESTION (maybe?)\\nThis check need not be performed on the tables that are already added\\nbecause incremental sync for those tables will happen through WAL and\\nthe origin of the data can be identified from the WAL records.\\n\\n======\\n\\n5. src/test/subscription/t/030_origin.pl\\n\\n+ \\\"Refresh publication when the publisher has subscribed for the new table\\\"\\n\\nSUGGESTION (Just to mention origin = none somehow. Maybe you can\\nreword it better than this)\\nRefresh publication when the publisher has subscribed for the new\\ntable, but the subscriber-side wants origin=none\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 29 Jul 2022 13:01:30 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jul 29, 2022 at 8:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some comments for the patch v40-0001:\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> It might be better to always use 'copy_data = true' in favour of\\n> 'copy_data = on' just for consistency with all the docs and the error\\n> messages.\\n>\\n> ======\\n\\nModified\\n\\n> 2. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -386,6 +401,15 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> can have non-existent publications.\\n> </para>\\n>\\n> + <para>\\n> + If the subscription is created with <literal>origin = NONE</literal> and\\n> + <literal>copy_data = true</literal>, it will check if the publisher has\\n> + subscribed to the same table from other publishers and, if so, throw an\\n> + error to prevent possible non-local data from being copied. The user can\\n> + override this check and continue with the copy operation by specifying\\n> + <literal>copy_data = force</literal>.\\n> + </para>\\n>\\n> 2a.\\n> It is interesting that you changed the note to say origin = NONE.\\n> Personally, I prefer it written as you did, but I think maybe this\\n> change does not belong in this patch. The suggestion for changing from\\n> \\\"none\\\" to NONE is being discussed elsewhere in this thread and\\n> probably all such changes should be done together (if at all) as a\\n> separate patch. Until then I think this patch 0001 should just stay\\n> consistent with whatever is already pushed on HEAD.\\n\\nModified\\n\\n> 2b.\\n> \\\"possible no-local data\\\". Maybe the terminology \\\"local/non-local\\\" is a\\n> hangover from back when the subscription parameter was called local\\n> instead of origin. I'm not sure if you want to change this or not, and\\n> anyway I didn't have any better suggestions – so this comment is just\\n> to bring it to your attention.\\n>\\n> ======\\n\\nModified\\n\\n> 3. src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n>\\n> + /*\\n> + * The set of strings accepted here should match up with the\\n> + * grammar's opt_boolean_or_string production.\\n> + */\\n> + if (pg_strcasecmp(sval, \\\"false\\\") == 0 ||\\n> + pg_strcasecmp(sval, \\\"off\\\") == 0)\\n> + return COPY_DATA_OFF;\\n> + if (pg_strcasecmp(sval, \\\"true\\\") == 0 ||\\n> + pg_strcasecmp(sval, \\\"on\\\") == 0)\\n> + return COPY_DATA_ON;\\n> + if (pg_strcasecmp(sval, \\\"force\\\") == 0)\\n> + return COPY_DATA_FORCE;\\n>\\n> I understand the intention of the comment, but it is not strictly\\n> correct to say \\\"should match up\\\" because \\\"force\\\" is a new value.\\n> Perhaps the comment should be as suggested below.\\n>\\n> SUGGESTION\\n> The set of strings accepted here must include all those accepted by\\n> the grammar's opt_boolean_or_string production.\\n>\\n> ~~\\n\\nModified\\n\\n>\\n> 4. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> @@ -1781,6 +1858,122 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\\n> table_close(rel, RowExclusiveLock);\\n> }\\n>\\n> +/*\\n> + * Check and throw an error if the publisher has subscribed to the same table\\n> + * from some other publisher. This check is required only if \\\"copy_data = on\\\"\\n> + * and \\\"origin = NONE\\\" for CREATE SUBSCRIPTION and\\n> + * ALTER SUBSCRIPTION ... REFRESH statements to avoid the publisher from\\n> + * replicating data that has an origin.\\n> + *\\n> + * This check need not be performed on the tables that are already added as\\n> + * incremental sync for such tables will happen through WAL and the origin of\\n> + * the data can be identified from the WAL records.\\n> + *\\n> + * subrel_local_oids contains the list of relation oids that are already\\n> + * present on the subscriber.\\n> + */\\n> +static void\\n> +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\\n> + CopyData copydata, char *origin,\\n> + Oid *subrel_local_oids, int subrel_count)\\n>\\n> 4a.\\n> \\\"copy_data = on\\\" -> \\\"copy_data = true\\\" (for consistency with the docs\\n> and the error messages)\\n\\nModified\\n\\n> 4b.\\n> The same NONE/none review comment from #2a applies here too. Probably\\n> it should be written as none for now unless/until *everything* changes\\n> to NONE.\\n\\nModified\\n\\n> 4c.\\n> \\\"to avoid the publisher from replicating data that has an origin.\\\" ->\\n> \\\"to avoid replicating data that has an origin.\\\"\\n\\nModified\\n\\n> 4d.\\n> + * This check need not be performed on the tables that are already added as\\n> + * incremental sync for such tables will happen through WAL and the origin of\\n> + * the data can be identified from the WAL records.\\n>\\n> SUGGESTION (maybe?)\\n> This check need not be performed on the tables that are already added\\n> because incremental sync for those tables will happen through WAL and\\n> the origin of the data can be identified from the WAL records.\\n>\\n> ======\\n\\nModified\\n\\n>\\n> 5. src/test/subscription/t/030_origin.pl\\n>\\n> + \\\"Refresh publication when the publisher has subscribed for the new table\\\"\\n>\\n> SUGGESTION (Just to mention origin = none somehow. Maybe you can\\n> reword it better than this)\\n> Refresh publication when the publisher has subscribed for the new\\n> table, but the subscriber-side wants origin=none\\n\\nModified\\n\\nThanks for the comments, the attached v41 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Fri, 29 Jul 2022 10:51:40 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> \\r\\n> Thanks for the comments, the attached v41 patch has the changes for the\\r\\n> same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch.\\r\\n\\r\\nI wonder in the case that the publisher uses PG15 (or before), subscriber uses\\r\\nPG16, should we have this check (check if publication tables were also\\r\\nsubscribing from other publishers)? In this case, even if origin=none is\\r\\nspecified, it doesn't work because the publisher doesn't filter the origin. So\\r\\nmaybe we don't need the check for initial sync. Thoughts?\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Mon, 1 Aug 2022 05:27:30 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 1, 2022 at 3:27 PM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> >\\n> > Thanks for the comments, the attached v41 patch has the changes for the\\n> > same.\\n> >\\n>\\n> Thanks for updating the patch.\\n>\\n> I wonder in the case that the publisher uses PG15 (or before), subscriber uses\\n> PG16, should we have this check (check if publication tables were also\\n> subscribing from other publishers)? In this case, even if origin=none is\\n> specified, it doesn't work because the publisher doesn't filter the origin. So\\n> maybe we don't need the check for initial sync. Thoughts?\\n>\\n\\nIIUC for the scenario you've described (subscription origin=none and\\npublisher < PG16) the subscriber can end up getting extra data they\\ndid not want, right?\\n\\nSo instead of just \\\"don't need the check\\\", maybe this combination\\nshould throw ERROR, or at least a log a WARNING?\\n\\n------\\nKInd Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Aug 2022 18:02:21 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 1, 2022 at 1:32 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Mon, Aug 1, 2022 at 3:27 PM shiy.fnst@fujitsu.com\\n> <shiy.fnst@fujitsu.com> wrote:\\n> >\\n> > On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > >\\n> > > Thanks for the comments, the attached v41 patch has the changes for the\\n> > > same.\\n> > >\\n> >\\n> > Thanks for updating the patch.\\n> >\\n> > I wonder in the case that the publisher uses PG15 (or before), subscriber uses\\n> > PG16, should we have this check (check if publication tables were also\\n> > subscribing from other publishers)? In this case, even if origin=none is\\n> > specified, it doesn't work because the publisher doesn't filter the origin. So\\n> > maybe we don't need the check for initial sync. Thoughts?\\n> >\\n>\\n> IIUC for the scenario you've described (subscription origin=none and\\n> publisher < PG16) the subscriber can end up getting extra data they\\n> did not want, right?\\n>\\n\\nYes, because publishers won't have 'filtering based on origin' functionality.\\n\\n> So instead of just \\\"don't need the check\\\", maybe this combination\\n> should throw ERROR, or at least a log a WARNING?\\n>\\n\\nI am not sure if doing anything (ERROR or WARNING) would make sense\\nbecause anyway later during replication there won't be any filtering.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 1 Aug 2022 14:21:56 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 1, 2022 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Aug 1, 2022 at 1:32 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > On Mon, Aug 1, 2022 at 3:27 PM shiy.fnst@fujitsu.com\\n> > <shiy.fnst@fujitsu.com> wrote:\\n> > >\\n> > > On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\\n> > > >\\n> > > >\\n> > > > Thanks for the comments, the attached v41 patch has the changes for the\\n> > > > same.\\n> > > >\\n> > >\\n> > > Thanks for updating the patch.\\n> > >\\n> > > I wonder in the case that the publisher uses PG15 (or before), subscriber uses\\n> > > PG16, should we have this check (check if publication tables were also\\n> > > subscribing from other publishers)? In this case, even if origin=none is\\n> > > specified, it doesn't work because the publisher doesn't filter the origin. So\\n> > > maybe we don't need the check for initial sync. Thoughts?\\n> > >\\n> >\\n> > IIUC for the scenario you've described (subscription origin=none and\\n> > publisher < PG16) the subscriber can end up getting extra data they\\n> > did not want, right?\\n> >\\n>\\n> Yes, because publishers won't have 'filtering based on origin' functionality.\\n>\\n> > So instead of just \\\"don't need the check\\\", maybe this combination\\n> > should throw ERROR, or at least a log a WARNING?\\n> >\\n>\\n> I am not sure if doing anything (ERROR or WARNING) would make sense\\n> because anyway later during replication there won't be any filtering.\\n>\\n\\nI was suggesting stopping that replication from happening at all. If\\nthe user specifically asked for 'origin=none' but the publisher could\\nnot filter that (because < PG16) then I imagined some logic that would\\njust disable the subscription up-front. Isn't it preferable for the\\nsubscriber to get no data at all then to get data the user\\nspecifically said they did NOT want to get?\\n\\ne.g. pseudo-code for the worker code something like below:\\n\\nif (origin != ANY and publisher.server_version < PG16)\\n{\\n set subscription.option.disable_on_error = true;\\n throw ERROR (\\\"publisher does not support origin=none - disabling\\nthe subscription\\\");\\n}\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia.\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Aug 2022 12:01:15 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jul 29, 2022 1:22 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> On Fri, Jul 29, 2022 at 8:31 AM Peter Smith <smithpb2250@gmail.com>\\r\\n> wrote:\\r\\n> >\\r\\n> \\r\\n> Thanks for the comments, the attached v41 patch has the changes for the\\r\\n> same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch.\\r\\n\\r\\nA comment for 0002 patch.\\r\\n\\r\\nIn the example in section 31.11.4 (Generic steps for adding a new primary to an\\r\\nexisting set of primaries), I think there should be a Step-6, corresponding to\\r\\nthe steps mentioned before. And part of Step-5 should actually be part of\\r\\nStep-6.\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Tue, 2 Aug 2022 09:31:20 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Jul 26, 2022 at 7:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n> >\\n> > Thanks for the example. I agree that it is fairly simple to reproduce.\\n> >\\n> > I understand that \\\"copy_data = force\\\" is meant to protect a user from\\n> > hurting themself. I'm not convinced that this is the best way to do so.\\n> >\\n> > For example today I can subscribe to multiple publications that write to\\n> > the same table. If I have a primary key on that table, and two of the\\n> > subscriptions try to write an identical ID, we conflict. We don't have\\n> > any special flags or modes to guard against that from happening, though\\n> > we do have documentation on conflicts and managing them.\\n> >\\n> > AFAICT the same issue with \\\"copy_data\\\" also exists in the above scenario\\n> > too, even without the \\\"origin\\\" attribute.\\n> >\\n>\\n> That's true but there is no parameter like origin = NONE which\\n> indicates that constraint violations or duplicate data problems won't\\n> occur due to replication. In the current case, I think the situation\\n> is different because a user has specifically asked not to replicate\\n> any remote data by specifying origin = NONE, which should be dealt\\n> differently. Note that current users or their setup won't see any\\n> difference/change unless they specify the new parameter origin as\\n> NONE.\\n>\\n\\nLet me try to summarize the discussion so that it is easier for others\\nto follow. The work in this thread is to avoid loops, and duplicate\\ndata in logical replication when the operations happened on the same\\ntable in multiple nodes. It has been explained in email [1] with an\\nexample of how a logical replication setup can lead to duplicate or\\ninconsistent data.\\n\\nThe idea to solve this problem is that we don't replicate data that is\\nnot generated locally which we can normally identify based on origin\\nof data in WAL. The commit 366283961a achieves that for replication\\nbut still the problem can happen during initial sync which is\\nperformed internally via copy. We can't differentiate the data in heap\\nbased on origin. So, we decided to prohibit the subscription\\noperations that can perform initial sync (ex. Create Subscription,\\nAlter Subscription ... Refresh) by detecting that the publisher has\\nsubscribed to the same table from some other publisher.\\n\\nTo prohibit the subscription operations, the currently proposed patch\\nthrows an error. Then, it also provides a new copy_data option\\n'force' under which the user will still be able to perform the\\noperation. This could be useful when the user intentionally wants to\\nreplicate the initial data even if it contains data from multiple\\nnodes (for example, when in a multi-node setup, one decides to get the\\ninitial data from just one node and then allow replication of data to\\nproceed from each of respective nodes).\\n\\nThe other alternative discussed was to just give a warning for\\nsubscription operations and probably document the steps for users to\\navoid it. But the problem with that is once the user sees this\\nwarning, it won't be able to do anything except recreate the setup, so\\nwhy not give an error in the first place?\\n\\nThoughts?\\n\\n[1] - https://www.postgresql.org/message-id/CALDaNm1eJr6qXT9esVPzgc5Qvy4uMhV4kCCTSmxARKjf%2BMwcnw%40mail.gmail.com\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Aug 2022 17:29:36 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 2, 2022 at 8:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Tue, Jul 26, 2022 at 7:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n> > >\\n> > > Thanks for the example. I agree that it is fairly simple to reproduce.\\n> > >\\n> > > I understand that \\\"copy_data = force\\\" is meant to protect a user from\\n> > > hurting themself. I'm not convinced that this is the best way to do so.\\n> > >\\n> > > For example today I can subscribe to multiple publications that write to\\n> > > the same table. If I have a primary key on that table, and two of the\\n> > > subscriptions try to write an identical ID, we conflict. We don't have\\n> > > any special flags or modes to guard against that from happening, though\\n> > > we do have documentation on conflicts and managing them.\\n> > >\\n> > > AFAICT the same issue with \\\"copy_data\\\" also exists in the above scenario\\n> > > too, even without the \\\"origin\\\" attribute.\\n> > >\\n> >\\n> > That's true but there is no parameter like origin = NONE which\\n> > indicates that constraint violations or duplicate data problems won't\\n> > occur due to replication. In the current case, I think the situation\\n> > is different because a user has specifically asked not to replicate\\n> > any remote data by specifying origin = NONE, which should be dealt\\n> > differently. Note that current users or their setup won't see any\\n> > difference/change unless they specify the new parameter origin as\\n> > NONE.\\n> >\\n>\\n> Let me try to summarize the discussion so that it is easier for others\\n> to follow. The work in this thread is to avoid loops, and duplicate\\n> data in logical replication when the operations happened on the same\\n> table in multiple nodes. It has been explained in email [1] with an\\n> example of how a logical replication setup can lead to duplicate or\\n> inconsistent data.\\n>\\n> The idea to solve this problem is that we don't replicate data that is\\n> not generated locally which we can normally identify based on origin\\n> of data in WAL. The commit 366283961a achieves that for replication\\n> but still the problem can happen during initial sync which is\\n> performed internally via copy. We can't differentiate the data in heap\\n> based on origin. So, we decided to prohibit the subscription\\n> operations that can perform initial sync (ex. Create Subscription,\\n> Alter Subscription ... Refresh) by detecting that the publisher has\\n> subscribed to the same table from some other publisher.\\n>\\n> To prohibit the subscription operations, the currently proposed patch\\n> throws an error. Then, it also provides a new copy_data option\\n> 'force' under which the user will still be able to perform the\\n> operation. This could be useful when the user intentionally wants to\\n> replicate the initial data even if it contains data from multiple\\n> nodes (for example, when in a multi-node setup, one decides to get the\\n> initial data from just one node and then allow replication of data to\\n> proceed from each of respective nodes).\\n>\\n> The other alternative discussed was to just give a warning for\\n> subscription operations and probably document the steps for users to\\n> avoid it. But the problem with that is once the user sees this\\n> warning, it won't be able to do anything except recreate the setup, so\\n> why not give an error in the first place?\\n>\\n> Thoughts?\\n\\nThank you for the summary!\\n\\nI understand that this feature could help some cases, but I'm really\\nnot sure adding new value 'force' for them is worthwhile, and\\nconcerned it could reduce the usability.\\n\\nIIUC this feature would work only when origin = 'none' and copy_data =\\n'on', and the purpose is to prevent the data from being\\nduplicated/conflicted by the initial table sync. But there are cases\\nwhere duplication/conflict doesn't happen even if origin = 'none' and\\ncopy_data = 'on'. For instance, the table on the publisher might be\\nempty. Also, even with origin = 'any', copy_data = 'on', there is a\\npossibility of data duplication/conflict. Why do we need to address\\nonly the case where origin = 'none'? I think that using origin =\\n'none' doesn't necessarily mean using bi-directional (or N-way)\\nreplication. Even when using uni-directional logical replication with\\ntwo nodes, they may use origin = 'none'. Therefore, it seems to me\\nthat this feature works only for a narrow situation and has false\\npositives.\\n\\nSince it has been the user's responsibility not to try to make the\\ndata inconsistent by the initial table sync, I think that it might be\\nsufficient if we note the risk in the documentation.\\n\\nRegards,\\n\\n--\\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Aug 2022 22:00:53 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 4, 2022 at 6:31 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n> On Tue, Aug 2, 2022 at 8:59 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > Let me try to summarize the discussion so that it is easier for others\\n> > to follow. The work in this thread is to avoid loops, and duplicate\\n> > data in logical replication when the operations happened on the same\\n> > table in multiple nodes. It has been explained in email [1] with an\\n> > example of how a logical replication setup can lead to duplicate or\\n> > inconsistent data.\\n> >\\n> > The idea to solve this problem is that we don't replicate data that is\\n> > not generated locally which we can normally identify based on origin\\n> > of data in WAL. The commit 366283961a achieves that for replication\\n> > but still the problem can happen during initial sync which is\\n> > performed internally via copy. We can't differentiate the data in heap\\n> > based on origin. So, we decided to prohibit the subscription\\n> > operations that can perform initial sync (ex. Create Subscription,\\n> > Alter Subscription ... Refresh) by detecting that the publisher has\\n> > subscribed to the same table from some other publisher.\\n> >\\n> > To prohibit the subscription operations, the currently proposed patch\\n> > throws an error. Then, it also provides a new copy_data option\\n> > 'force' under which the user will still be able to perform the\\n> > operation. This could be useful when the user intentionally wants to\\n> > replicate the initial data even if it contains data from multiple\\n> > nodes (for example, when in a multi-node setup, one decides to get the\\n> > initial data from just one node and then allow replication of data to\\n> > proceed from each of respective nodes).\\n> >\\n> > The other alternative discussed was to just give a warning for\\n> > subscription operations and probably document the steps for users to\\n> > avoid it. But the problem with that is once the user sees this\\n> > warning, it won't be able to do anything except recreate the setup, so\\n> > why not give an error in the first place?\\n> >\\n> > Thoughts?\\n>\\n> Thank you for the summary!\\n>\\n> I understand that this feature could help some cases, but I'm really\\n> not sure adding new value 'force' for them is worthwhile, and\\n> concerned it could reduce the usability.\\n>\\n> IIUC this feature would work only when origin = 'none' and copy_data =\\n> 'on', and the purpose is to prevent the data from being\\n> duplicated/conflicted by the initial table sync. But there are cases\\n> where duplication/conflict doesn't happen even if origin = 'none' and\\n> copy_data = 'on'. For instance, the table on the publisher might be\\n> empty.\\n>\\n\\nRight, but if we want we can check the tables on publishers to ensure\\nthat. Now, another case could be where the corresponding subscription\\nwas disabled on publisher during create subscription but got enabled\\njust before copy, we can even catch that situation as we are doing for\\ncolumn lists in fetch_remote_table_info(). Are there other cases of\\nfalse positives you can think of? I see your point that we can\\ndocument that users should be careful with certain configurations to\\navoid data inconsistency but not able completely convince myself about\\nthe same. I think the main thing to decide here is how much we want to\\nask users to be careful by referring them to docs. Now, if you are not\\nconvinced with giving an ERROR here then we can probably show a\\nWARNING (that we might copy data for multiple origins during initial\\nsync in spite of the user having specified origin as NONE)?\\n\\n>\\n Also, even with origin = 'any', copy_data = 'on', there is a\\n> possibility of data duplication/conflict. Why do we need to address\\n> only the case where origin = 'none'?\\n>\\n\\nBecause the user has specifically asked not to replicate any remote\\ndata by specifying origin = NONE, which should be dealt with\\ndifferently whereas 'any' doesn't have such a requirement. Now,\\ntomorrow, if we want to support replication based on specific origin\\nnames say origin = 'node-1' then also we won't be able to identify the\\ndata during initial sync but I think 'none' is a special case where\\ngiving some intimation to user won't be a bad idea especially because\\nwe can identify the same.\\n\\n> I think that using origin =\\n> 'none' doesn't necessarily mean using bi-directional (or N-way)\\n> replication. Even when using uni-directional logical replication with\\n> two nodes, they may use origin = 'none'.\\n>\\n\\nIt is possible but still, I think it is a must for bi-directional (or\\nN-way) replication, otherwise, there is a risk of loops.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Aug 2022 09:40:02 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Jul 29, 2022 at 10:51 AM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Fri, Jul 29, 2022 at 8:31 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Here are some comments for the patch v40-0001:\\n> >\\n> > ======\\n> >\\n> > 1. Commit message\\n> >\\n> > It might be better to always use 'copy_data = true' in favour of\\n> > 'copy_data = on' just for consistency with all the docs and the error\\n> > messages.\\n> >\\n> > ======\\n>\\n> Modified\\n>\\n> > 2. doc/src/sgml/ref/create_subscription.sgml\\n> >\\n> > @@ -386,6 +401,15 @@ CREATE SUBSCRIPTION <replaceable\\n> > class=\\\"parameter\\\">subscription_name</replaceabl\\n> > can have non-existent publications.\\n> > </para>\\n> >\\n> > + <para>\\n> > + If the subscription is created with <literal>origin = NONE</literal> and\\n> > + <literal>copy_data = true</literal>, it will check if the publisher has\\n> > + subscribed to the same table from other publishers and, if so, throw an\\n> > + error to prevent possible non-local data from being copied. The user can\\n> > + override this check and continue with the copy operation by specifying\\n> > + <literal>copy_data = force</literal>.\\n> > + </para>\\n> >\\n> > 2a.\\n> > It is interesting that you changed the note to say origin = NONE.\\n> > Personally, I prefer it written as you did, but I think maybe this\\n> > change does not belong in this patch. The suggestion for changing from\\n> > \\\"none\\\" to NONE is being discussed elsewhere in this thread and\\n> > probably all such changes should be done together (if at all) as a\\n> > separate patch. Until then I think this patch 0001 should just stay\\n> > consistent with whatever is already pushed on HEAD.\\n>\\n> Modified\\n>\\n> > 2b.\\n> > \\\"possible no-local data\\\". Maybe the terminology \\\"local/non-local\\\" is a\\n> > hangover from back when the subscription parameter was called local\\n> > instead of origin. I'm not sure if you want to change this or not, and\\n> > anyway I didn't have any better suggestions – so this comment is just\\n> > to bring it to your attention.\\n> >\\n> > ======\\n>\\n> Modified\\n>\\n> > 3. src/backend/commands/subscriptioncmds.c - DefGetCopyData\\n> >\\n> > + /*\\n> > + * The set of strings accepted here should match up with the\\n> > + * grammar's opt_boolean_or_string production.\\n> > + */\\n> > + if (pg_strcasecmp(sval, \\\"false\\\") == 0 ||\\n> > + pg_strcasecmp(sval, \\\"off\\\") == 0)\\n> > + return COPY_DATA_OFF;\\n> > + if (pg_strcasecmp(sval, \\\"true\\\") == 0 ||\\n> > + pg_strcasecmp(sval, \\\"on\\\") == 0)\\n> > + return COPY_DATA_ON;\\n> > + if (pg_strcasecmp(sval, \\\"force\\\") == 0)\\n> > + return COPY_DATA_FORCE;\\n> >\\n> > I understand the intention of the comment, but it is not strictly\\n> > correct to say \\\"should match up\\\" because \\\"force\\\" is a new value.\\n> > Perhaps the comment should be as suggested below.\\n> >\\n> > SUGGESTION\\n> > The set of strings accepted here must include all those accepted by\\n> > the grammar's opt_boolean_or_string production.\\n> >\\n> > ~~\\n>\\n> Modified\\n>\\n> >\\n> > 4. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n> >\\n> > @@ -1781,6 +1858,122 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\\n> > table_close(rel, RowExclusiveLock);\\n> > }\\n> >\\n> > +/*\\n> > + * Check and throw an error if the publisher has subscribed to the same table\\n> > + * from some other publisher. This check is required only if \\\"copy_data = on\\\"\\n> > + * and \\\"origin = NONE\\\" for CREATE SUBSCRIPTION and\\n> > + * ALTER SUBSCRIPTION ... REFRESH statements to avoid the publisher from\\n> > + * replicating data that has an origin.\\n> > + *\\n> > + * This check need not be performed on the tables that are already added as\\n> > + * incremental sync for such tables will happen through WAL and the origin of\\n> > + * the data can be identified from the WAL records.\\n> > + *\\n> > + * subrel_local_oids contains the list of relation oids that are already\\n> > + * present on the subscriber.\\n> > + */\\n> > +static void\\n> > +check_pub_table_subscribed(WalReceiverConn *wrconn, List *publications,\\n> > + CopyData copydata, char *origin,\\n> > + Oid *subrel_local_oids, int subrel_count)\\n> >\\n> > 4a.\\n> > \\\"copy_data = on\\\" -> \\\"copy_data = true\\\" (for consistency with the docs\\n> > and the error messages)\\n>\\n> Modified\\n>\\n> > 4b.\\n> > The same NONE/none review comment from #2a applies here too. Probably\\n> > it should be written as none for now unless/until *everything* changes\\n> > to NONE.\\n>\\n> Modified\\n>\\n> > 4c.\\n> > \\\"to avoid the publisher from replicating data that has an origin.\\\" ->\\n> > \\\"to avoid replicating data that has an origin.\\\"\\n>\\n> Modified\\n>\\n> > 4d.\\n> > + * This check need not be performed on the tables that are already added as\\n> > + * incremental sync for such tables will happen through WAL and the origin of\\n> > + * the data can be identified from the WAL records.\\n> >\\n> > SUGGESTION (maybe?)\\n> > This check need not be performed on the tables that are already added\\n> > because incremental sync for those tables will happen through WAL and\\n> > the origin of the data can be identified from the WAL records.\\n> >\\n> > ======\\n>\\n> Modified\\n>\\n> >\\n> > 5. src/test/subscription/t/030_origin.pl\\n> >\\n> > + \\\"Refresh publication when the publisher has subscribed for the new table\\\"\\n> >\\n> > SUGGESTION (Just to mention origin = none somehow. Maybe you can\\n> > reword it better than this)\\n> > Refresh publication when the publisher has subscribed for the new\\n> > table, but the subscriber-side wants origin=none\\n>\\n> Modified\\n>\\n> Thanks for the comments, the attached v41 patch has the changes for the same.\\n\\nThe patch does not apply on Head because of the commit\\n\\\"0c20dd33db1607d6a85ffce24238c1e55e384b49\\\", I have attached a rebased\\nv41 patch on top of the HEAD. I have not yet done the changes to\\nchange the error to warning in the first patch, I will change it in\\nthe next version once it is concluded.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Mon, 8 Aug 2022 10:21:24 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tuesday, August 2, 2022 8:00 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> >\\r\\n> > On Tue, Jul 26, 2022 at 7:13 AM Jonathan S. Katz <jkatz@postgresql.org>\\r\\n> wrote:\\r\\n> > >\\r\\n> > > Thanks for the example. I agree that it is fairly simple to reproduce.\\r\\n> > >\\r\\n> > > I understand that \\\"copy_data = force\\\" is meant to protect a user\\r\\n> > > from hurting themself. I'm not convinced that this is the best way to do so.\\r\\n> > >\\r\\n> > > For example today I can subscribe to multiple publications that\\r\\n> > > write to the same table. If I have a primary key on that table, and\\r\\n> > > two of the subscriptions try to write an identical ID, we conflict.\\r\\n> > > We don't have any special flags or modes to guard against that from\\r\\n> > > happening, though we do have documentation on conflicts and managing\\r\\n> them.\\r\\n> > >\\r\\n> > > AFAICT the same issue with \\\"copy_data\\\" also exists in the above\\r\\n> > > scenario too, even without the \\\"origin\\\" attribute.\\r\\n> > >\\r\\n> >\\r\\n> > That's true but there is no parameter like origin = NONE which\\r\\n> > indicates that constraint violations or duplicate data problems won't\\r\\n> > occur due to replication. In the current case, I think the situation\\r\\n> > is different because a user has specifically asked not to replicate\\r\\n> > any remote data by specifying origin = NONE, which should be dealt\\r\\n> > differently. Note that current users or their setup won't see any\\r\\n> > difference/change unless they specify the new parameter origin as\\r\\n> > NONE.\\r\\n> >\\r\\n> \\r\\n> Let me try to summarize the discussion so that it is easier for others to follow.\\r\\n> The work in this thread is to avoid loops, and duplicate data in logical\\r\\n> replication when the operations happened on the same table in multiple nodes.\\r\\n> It has been explained in email [1] with an example of how a logical replication\\r\\n> setup can lead to duplicate or inconsistent data.\\r\\n> \\r\\n> The idea to solve this problem is that we don't replicate data that is not\\r\\n> generated locally which we can normally identify based on origin of data in\\r\\n> WAL. The commit 366283961a achieves that for replication but still the\\r\\n> problem can happen during initial sync which is performed internally via copy.\\r\\n> We can't differentiate the data in heap based on origin. So, we decided to\\r\\n> prohibit the subscription operations that can perform initial sync (ex. Create\\r\\n> Subscription, Alter Subscription ... Refresh) by detecting that the publisher has\\r\\n> subscribed to the same table from some other publisher.\\r\\n> \\r\\n> To prohibit the subscription operations, the currently proposed patch throws\\r\\n> an error. Then, it also provides a new copy_data option 'force' under which\\r\\n> the user will still be able to perform the operation. This could be useful when\\r\\n> the user intentionally wants to replicate the initial data even if it contains data\\r\\n> from multiple nodes (for example, when in a multi-node setup, one decides to\\r\\n> get the initial data from just one node and then allow replication of data to\\r\\n> proceed from each of respective nodes).\\r\\n> \\r\\n> The other alternative discussed was to just give a warning for subscription\\r\\n> operations and probably document the steps for users to avoid it. But the\\r\\n> problem with that is once the user sees this warning, it won't be able to do\\r\\n> anything except recreate the setup, so why not give an error in the first place?\\r\\n> \\r\\n> Thoughts?\\r\\n\\r\\nThanks for the summary.\\r\\n\\r\\nI think it's fine to make the user use the copy_data option more carefully to\\r\\nprevent duplicate copies by reporting an ERROR.\\r\\n\\r\\nBut I also have similar concern with Sawada-san as it's possible for user to\\r\\nreceive an ERROR in some unexpected cases.\\r\\n\\r\\nFor example I want to build bi-directional setup between two nodes:\\r\\n\\r\\nNode A: TABLE test (has actual data)\\r\\nNode B: TABLE test (empty)\\r\\n\\r\\nStep 1:\\r\\nCREATE PUBLICATION on both Node A and B.\\r\\n\\r\\nStep 2:\\r\\nCREATE SUBSCRIPTION on Node A with (copy_data = on)\\r\\n-- this is fine as there is no data on Node B\\r\\n\\r\\nStep 3:\\r\\nCREATE SUBSCRIPTION on Node B with (copy_data = on)\\r\\n-- this should be fine as user needs to copy data from Node A to Node B,\\r\\n-- but we still report an error for this case.\\r\\n\\r\\nIt looks a bit strict to report an ERROR in this case and it seems not easy to\\r\\navoid this. So, personally, I think it might be better to document the correct\\r\\nsteps to build the bi-directional replication and probably also docuemnt the\\r\\nsteps to recover if user accidently did duplicate initial copy if not\\r\\ndocumented yet.\\r\\n\\r\\nIn addition, we could also LOG some additional information about the ORIGIN and\\r\\ninitial copy which might help user to analyze if needed.\\r\\n\\r\\n\\r\\n----- Some other thoughts about the duplicate initial copy problem.\\r\\n\\r\\nActually, I feel the better way to address the possible duplicate copy problem\\r\\nis to provide a command like \\\"bi_setup\\\" which can help user build the\\r\\nbi-directional replication in all nodes and can handle the initial copy\\r\\nautomtically. But that might be too far.\\r\\n\\r\\nAnother naive idea I once thought is that maybe we can add a publication option\\r\\nlike: data_source_in_bi_group. If data_source_in_bi_group is true, we silently\\r\\nlet it succeed when we subscribe this publication\\r\\nwith(origin='NONE',copy_data=on). If data_source_in_bi_group is false, we\\r\\nreport an ERROR when we subscribe this publication\\r\\nwith(origin='NONE',copy_data=on). The slight difference in this approach is\\r\\nthat we don't do any additional check for the publisher and subscriber, we\\r\\ntrust the user and give the responsibility to choose the node with actual data\\r\\nto user. Having said that this approach seems not a good approach as it doesn't\\r\\nsolve the actual problem and can only handle the case when ORIGIN='NONE' and\\r\\ncannot handle the future case when we support ORIGIN='NODE1'.\\r\\n\\r\\nSo, I personally feel that adding document and LOG might be sufficient at this\\r\\nstage.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\",\n \"msg_date\": \"Wed, 17 Aug 2022 03:18:08 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> On Tuesday, August 2, 2022 8:00 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> Thanks for the summary.\\n>\\n> I think it's fine to make the user use the copy_data option more carefully to\\n> prevent duplicate copies by reporting an ERROR.\\n>\\n> But I also have similar concern with Sawada-san as it's possible for user to\\n> receive an ERROR in some unexpected cases.\\n>\\n> For example I want to build bi-directional setup between two nodes:\\n>\\n> Node A: TABLE test (has actual data)\\n> Node B: TABLE test (empty)\\n>\\n> Step 1:\\n> CREATE PUBLICATION on both Node A and B.\\n>\\n> Step 2:\\n> CREATE SUBSCRIPTION on Node A with (copy_data = on)\\n> -- this is fine as there is no data on Node B\\n>\\n> Step 3:\\n> CREATE SUBSCRIPTION on Node B with (copy_data = on)\\n> -- this should be fine as user needs to copy data from Node A to Node B,\\n> -- but we still report an error for this case.\\n>\\n> It looks a bit strict to report an ERROR in this case and it seems not easy to\\n> avoid this. So, personally, I think it might be better to document the correct\\n> steps to build the bi-directional replication and probably also docuemnt the\\n> steps to recover if user accidently did duplicate initial copy if not\\n> documented yet.\\n>\\n> In addition, we could also LOG some additional information about the ORIGIN and\\n> initial copy which might help user to analyze if needed.\\n>\\n\\nBut why LOG instead of WARNING? I feel in this case there is a chance\\nof inconsistent data so a WARNING like \\\"publication \\\"pub1\\\" could have\\ndata from multiple origins\\\" can be given when the user has specified\\noptions: \\\"copy_data = on, origin = NONE\\\" while creating a\\nsubscription. We give a WARNING during subscription creation when the\\ncorresponding publication doesn't exist, eg.\\n\\npostgres=# create subscription sub1 connection 'dbname = postgres'\\npublication pub1;\\nWARNING: publication \\\"pub1\\\" does not exist in the publisher\\n\\nThen, we can explain in docs how users can avoid data inconsistencies\\nwhile setting up replication.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Aug 2022 12:03:43 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 17, 2022 at 4:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n> >\\n> > On Tuesday, August 2, 2022 8:00 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > Thanks for the summary.\\n> >\\n> > I think it's fine to make the user use the copy_data option more carefully to\\n> > prevent duplicate copies by reporting an ERROR.\\n> >\\n> > But I also have similar concern with Sawada-san as it's possible for user to\\n> > receive an ERROR in some unexpected cases.\\n> >\\n> > For example I want to build bi-directional setup between two nodes:\\n> >\\n> > Node A: TABLE test (has actual data)\\n> > Node B: TABLE test (empty)\\n> >\\n> > Step 1:\\n> > CREATE PUBLICATION on both Node A and B.\\n> >\\n> > Step 2:\\n> > CREATE SUBSCRIPTION on Node A with (copy_data = on)\\n> > -- this is fine as there is no data on Node B\\n> >\\n> > Step 3:\\n> > CREATE SUBSCRIPTION on Node B with (copy_data = on)\\n> > -- this should be fine as user needs to copy data from Node A to Node B,\\n> > -- but we still report an error for this case.\\n> >\\n> > It looks a bit strict to report an ERROR in this case and it seems not easy to\\n> > avoid this. So, personally, I think it might be better to document the correct\\n> > steps to build the bi-directional replication and probably also docuemnt the\\n> > steps to recover if user accidently did duplicate initial copy if not\\n> > documented yet.\\n> >\\n> > In addition, we could also LOG some additional information about the ORIGIN and\\n> > initial copy which might help user to analyze if needed.\\n> >\\n>\\n> But why LOG instead of WARNING? I feel in this case there is a chance\\n> of inconsistent data so a WARNING like \\\"publication \\\"pub1\\\" could have\\n> data from multiple origins\\\" can be given when the user has specified\\n> options: \\\"copy_data = on, origin = NONE\\\" while creating a\\n> subscription. We give a WARNING during subscription creation when the\\n> corresponding publication doesn't exist, eg.\\n>\\n> postgres=# create subscription sub1 connection 'dbname = postgres'\\n> publication pub1;\\n> WARNING: publication \\\"pub1\\\" does not exist in the publisher\\n>\\n> Then, we can explain in docs how users can avoid data inconsistencies\\n> while setting up replication.\\n>\\n\\nI was wondering if this copy/origin case really should be a NOTICE.\\n\\nSee table [1]. It says WARNING is meant for \\\"warnings of likey\\nproblems\\\". But this is not exactly a \\\"likely\\\" problem - IIUC we really\\ndon't know if there is even any problem at all .... we only know there\\nis the *potential* for a problem, but the user has to then judge it\\nfor themselves, Perhaps WARNING may be a bit overkill for this\\nsituation - it might be unnecessarily scary to give false warnings.\\n\\nOTOH, NOTICE [1] says it is for \\\"information that might be helpful to\\nusers\\\" which seems more like what is needed here.\\n\\n------\\n[1] https://www.postgresql.org/docs/current/runtime-config-logging.html#RUNTIME-CONFIG-SEVERITY-LEVELS\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia.\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Aug 2022 17:03:50 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 17, 2022 at 12:34 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Wed, Aug 17, 2022 at 4:33 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\\n> > <houzj.fnst@fujitsu.com> wrote:\\n> > >\\n> > > On Tuesday, August 2, 2022 8:00 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > > > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > >\\n> > > Thanks for the summary.\\n> > >\\n> > > I think it's fine to make the user use the copy_data option more carefully to\\n> > > prevent duplicate copies by reporting an ERROR.\\n> > >\\n> > > But I also have similar concern with Sawada-san as it's possible for user to\\n> > > receive an ERROR in some unexpected cases.\\n> > >\\n> > > For example I want to build bi-directional setup between two nodes:\\n> > >\\n> > > Node A: TABLE test (has actual data)\\n> > > Node B: TABLE test (empty)\\n> > >\\n> > > Step 1:\\n> > > CREATE PUBLICATION on both Node A and B.\\n> > >\\n> > > Step 2:\\n> > > CREATE SUBSCRIPTION on Node A with (copy_data = on)\\n> > > -- this is fine as there is no data on Node B\\n> > >\\n> > > Step 3:\\n> > > CREATE SUBSCRIPTION on Node B with (copy_data = on)\\n> > > -- this should be fine as user needs to copy data from Node A to Node B,\\n> > > -- but we still report an error for this case.\\n> > >\\n> > > It looks a bit strict to report an ERROR in this case and it seems not easy to\\n> > > avoid this. So, personally, I think it might be better to document the correct\\n> > > steps to build the bi-directional replication and probably also docuemnt the\\n> > > steps to recover if user accidently did duplicate initial copy if not\\n> > > documented yet.\\n> > >\\n> > > In addition, we could also LOG some additional information about the ORIGIN and\\n> > > initial copy which might help user to analyze if needed.\\n> > >\\n> >\\n> > But why LOG instead of WARNING? I feel in this case there is a chance\\n> > of inconsistent data so a WARNING like \\\"publication \\\"pub1\\\" could have\\n> > data from multiple origins\\\" can be given when the user has specified\\n> > options: \\\"copy_data = on, origin = NONE\\\" while creating a\\n> > subscription. We give a WARNING during subscription creation when the\\n> > corresponding publication doesn't exist, eg.\\n> >\\n> > postgres=# create subscription sub1 connection 'dbname = postgres'\\n> > publication pub1;\\n> > WARNING: publication \\\"pub1\\\" does not exist in the publisher\\n> >\\n> > Then, we can explain in docs how users can avoid data inconsistencies\\n> > while setting up replication.\\n> >\\n>\\n> I was wondering if this copy/origin case really should be a NOTICE.\\n>\\n\\nWe usually give NOTICE for some sort of additional implicit\\ninformation, e.g., when we create a slot during CREATE SUBSCRIPTION\\ncommand: \\\"NOTICE: created replication slot \\\"sub1\\\" on publisher\\\". IMO,\\nthis is likely to be a problem of data inconsistency so I think here\\nwe can choose between WARNING and LOG. I prefer WARNING but okay with\\nLOG as well if others feel so. I think we can change this later as\\nwell if required. We do have an option to not do anything and just\\ndocument it but I feel it is better to give user some indication of\\nproblem here because not everyone reads each update of documentation.\\n\\nJonathan, Sawada-San, Hou-San, and others, what do you think is the\\nbest way to move forward here?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 18 Aug 2022 08:43:08 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thursday, August 18, 2022 11:13 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n> \\r\\n> On Wed, Aug 17, 2022 at 12:34 PM Peter Smith <smithpb2250@gmail.com>\\r\\n> wrote:\\r\\n> >\\r\\n> > On Wed, Aug 17, 2022 at 4:33 PM Amit Kapila <amit.kapila16@gmail.com>\\r\\n> wrote:\\r\\n> > >\\r\\n> > > On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\\r\\n> > > <houzj.fnst@fujitsu.com> wrote:\\r\\n> > > >\\r\\n> > > > On Tuesday, August 2, 2022 8:00 PM Amit Kapila\\r\\n> <amit.kapila16@gmail.com> wrote:\\r\\n> > > > > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila\\r\\n> <amit.kapila16@gmail.com> wrote:\\r\\n> > > >\\r\\n> > > > Thanks for the summary.\\r\\n> > > >\\r\\n> > > > I think it's fine to make the user use the copy_data option more\\r\\n> > > > carefully to prevent duplicate copies by reporting an ERROR.\\r\\n> > > >\\r\\n> > > > But I also have similar concern with Sawada-san as it's possible\\r\\n> > > > for user to receive an ERROR in some unexpected cases.\\r\\n> > > >\\r\\n> > > > For example I want to build bi-directional setup between two nodes:\\r\\n> > > >\\r\\n> > > > Node A: TABLE test (has actual data) Node B: TABLE test (empty)\\r\\n> > > >\\r\\n> > > > Step 1:\\r\\n> > > > CREATE PUBLICATION on both Node A and B.\\r\\n> > > >\\r\\n> > > > Step 2:\\r\\n> > > > CREATE SUBSCRIPTION on Node A with (copy_data = on)\\r\\n> > > > -- this is fine as there is no data on Node B\\r\\n> > > >\\r\\n> > > > Step 3:\\r\\n> > > > CREATE SUBSCRIPTION on Node B with (copy_data = on)\\r\\n> > > > -- this should be fine as user needs to copy data from Node A to\\r\\n> > > > Node B,\\r\\n> > > > -- but we still report an error for this case.\\r\\n> > > >\\r\\n> > > > It looks a bit strict to report an ERROR in this case and it seems\\r\\n> > > > not easy to avoid this. So, personally, I think it might be better\\r\\n> > > > to document the correct steps to build the bi-directional\\r\\n> > > > replication and probably also docuemnt the steps to recover if\\r\\n> > > > user accidently did duplicate initial copy if not documented yet.\\r\\n> > > >\\r\\n> > > > In addition, we could also LOG some additional information about\\r\\n> > > > the ORIGIN and initial copy which might help user to analyze if needed.\\r\\n> > > >\\r\\n> > >\\r\\n> > > But why LOG instead of WARNING? I feel in this case there is a\\r\\n> > > chance of inconsistent data so a WARNING like \\\"publication \\\"pub1\\\"\\r\\n> > > could have data from multiple origins\\\" can be given when the user\\r\\n> > > has specified\\r\\n> > > options: \\\"copy_data = on, origin = NONE\\\" while creating a\\r\\n> > > subscription. We give a WARNING during subscription creation when\\r\\n> > > the corresponding publication doesn't exist, eg.\\r\\n> > >\\r\\n> > > postgres=# create subscription sub1 connection 'dbname = postgres'\\r\\n> > > publication pub1;\\r\\n> > > WARNING: publication \\\"pub1\\\" does not exist in the publisher\\r\\n> > >\\r\\n> > > Then, we can explain in docs how users can avoid data\\r\\n> > > inconsistencies while setting up replication.\\r\\n> > >\\r\\n> >\\r\\n> > I was wondering if this copy/origin case really should be a NOTICE.\\r\\n> >\\r\\n> \\r\\n> We usually give NOTICE for some sort of additional implicit information, e.g.,\\r\\n> when we create a slot during CREATE SUBSCRIPTION\\r\\n> command: \\\"NOTICE: created replication slot \\\"sub1\\\" on publisher\\\". IMO, this is\\r\\n> likely to be a problem of data inconsistency so I think here we can choose\\r\\n> between WARNING and LOG. I prefer WARNING but okay with LOG as well if\\r\\n> others feel so. I think we can change this later as well if required. We do have an\\r\\n> option to not do anything and just document it but I feel it is better to give user\\r\\n> some indication of problem here because not everyone reads each update of\\r\\n> documentation.\\r\\n> \\r\\n> Jonathan, Sawada-San, Hou-San, and others, what do you think is the best way\\r\\n> to move forward here?\\r\\n\\r\\nI think it's fine to throw a WARNING in this case given that there is a\\r\\nchance of inconsistent data.\\r\\n\\r\\nBest regards,\\r\\nHou zj\\r\\n\\r\\n\",\n \"msg_date\": \"Mon, 22 Aug 2022 03:49:15 +0000\",\n \"msg_from\": \"\\\"houzj.fnst@fujitsu.com\\\" <houzj.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 22, 2022 at 9:19 AM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> On Thursday, August 18, 2022 11:13 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Wed, Aug 17, 2022 at 12:34 PM Peter Smith <smithpb2250@gmail.com>\\n> > wrote:\\n> > >\\n> > > On Wed, Aug 17, 2022 at 4:33 PM Amit Kapila <amit.kapila16@gmail.com>\\n> > wrote:\\n> > > >\\n> > > > On Wed, Aug 17, 2022 at 8:48 AM houzj.fnst@fujitsu.com\\n> > > > <houzj.fnst@fujitsu.com> wrote:\\n> > > > >\\n> > > > > On Tuesday, August 2, 2022 8:00 PM Amit Kapila\\n> > <amit.kapila16@gmail.com> wrote:\\n> > > > > > On Tue, Jul 26, 2022 at 9:07 AM Amit Kapila\\n> > <amit.kapila16@gmail.com> wrote:\\n> > > > >\\n> > > > > Thanks for the summary.\\n> > > > >\\n> > > > > I think it's fine to make the user use the copy_data option more\\n> > > > > carefully to prevent duplicate copies by reporting an ERROR.\\n> > > > >\\n> > > > > But I also have similar concern with Sawada-san as it's possible\\n> > > > > for user to receive an ERROR in some unexpected cases.\\n> > > > >\\n> > > > > For example I want to build bi-directional setup between two nodes:\\n> > > > >\\n> > > > > Node A: TABLE test (has actual data) Node B: TABLE test (empty)\\n> > > > >\\n> > > > > Step 1:\\n> > > > > CREATE PUBLICATION on both Node A and B.\\n> > > > >\\n> > > > > Step 2:\\n> > > > > CREATE SUBSCRIPTION on Node A with (copy_data = on)\\n> > > > > -- this is fine as there is no data on Node B\\n> > > > >\\n> > > > > Step 3:\\n> > > > > CREATE SUBSCRIPTION on Node B with (copy_data = on)\\n> > > > > -- this should be fine as user needs to copy data from Node A to\\n> > > > > Node B,\\n> > > > > -- but we still report an error for this case.\\n> > > > >\\n> > > > > It looks a bit strict to report an ERROR in this case and it seems\\n> > > > > not easy to avoid this. So, personally, I think it might be better\\n> > > > > to document the correct steps to build the bi-directional\\n> > > > > replication and probably also docuemnt the steps to recover if\\n> > > > > user accidently did duplicate initial copy if not documented yet.\\n> > > > >\\n> > > > > In addition, we could also LOG some additional information about\\n> > > > > the ORIGIN and initial copy which might help user to analyze if needed.\\n> > > > >\\n> > > >\\n> > > > But why LOG instead of WARNING? I feel in this case there is a\\n> > > > chance of inconsistent data so a WARNING like \\\"publication \\\"pub1\\\"\\n> > > > could have data from multiple origins\\\" can be given when the user\\n> > > > has specified\\n> > > > options: \\\"copy_data = on, origin = NONE\\\" while creating a\\n> > > > subscription. We give a WARNING during subscription creation when\\n> > > > the corresponding publication doesn't exist, eg.\\n> > > >\\n> > > > postgres=# create subscription sub1 connection 'dbname = postgres'\\n> > > > publication pub1;\\n> > > > WARNING: publication \\\"pub1\\\" does not exist in the publisher\\n> > > >\\n> > > > Then, we can explain in docs how users can avoid data\\n> > > > inconsistencies while setting up replication.\\n> > > >\\n> > >\\n> > > I was wondering if this copy/origin case really should be a NOTICE.\\n> > >\\n> >\\n> > We usually give NOTICE for some sort of additional implicit information, e.g.,\\n> > when we create a slot during CREATE SUBSCRIPTION\\n> > command: \\\"NOTICE: created replication slot \\\"sub1\\\" on publisher\\\". IMO, this is\\n> > likely to be a problem of data inconsistency so I think here we can choose\\n> > between WARNING and LOG. I prefer WARNING but okay with LOG as well if\\n> > others feel so. I think we can change this later as well if required. We do have an\\n> > option to not do anything and just document it but I feel it is better to give user\\n> > some indication of problem here because not everyone reads each update of\\n> > documentation.\\n> >\\n> > Jonathan, Sawada-San, Hou-San, and others, what do you think is the best way\\n> > to move forward here?\\n>\\n> I think it's fine to throw a WARNING in this case given that there is a\\n> chance of inconsistent data.\\n\\nSince there was no objections to change it to throw a warning, I have\\nmade the changes for the same.\\n\\nI have made one change for the documentation patch.\\nThe current steps states that:\\n1. Create a publication on primary3.\\n2. Lock table t1 on primary2 and primary3 in EXCLUSIVE mode until the\\nsetup is completed. There is no need to lock table t1 on primary1\\nbecause any data changes made will be synchronized while creating the\\nsubscription with copy_data = true.\\n3. Create a subscription on primary1 to subscribe to primary3.\\n4. Create a subscription on primary2 to subscribe to primary3.\\n5. Create a subscription on primary3 to subscribe to primary1. Use\\ncopy_data = true so that the existing table data is copied during\\ninitial sync.\\n6. Create a subscription on primary3 to subscribe to primary2. Use\\ncopy_data = false because the initial table data would have been\\nalready copied in the previous step.\\n7. Now the replication setup between primaries primary1, primary2 and\\nprimary3 is complete. Incremental changes made on any primary will be\\nreplicated to the other two primaries.\\n\\nWe found a problem with the proposed steps. Here the problem is that\\nwe lock table t1 on primary2/primary3 in step2, then we create a\\nsubscription on primary3 in step5. While we create subscription in\\nstep5, tablesync worker will try to lock the table in primary3 during\\nstep5 before copying, as we have already taken a lock, this process\\nwill wait.\\nEven if the lock is released in this step just before creating\\nsubscription, there is a chance that some other transaction can\\nacquire the lock and change the data which will result in inconsistent\\ndata. Here the locking solution cannot handle adding a new primary in\\na consistent way.\\n\\nI have changed the lock step to mention that user need to ensure no\\ndata changes happen until the setup is completed.\\nThe attached v42 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 24 Aug 2022 19:27:14 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 22, 2022 at 9:19 AM houzj.fnst@fujitsu.com\\n<houzj.fnst@fujitsu.com> wrote:\\n>\\n> > Jonathan, Sawada-San, Hou-San, and others, what do you think is the best way\\n> > to move forward here?\\n>\\n> I think it's fine to throw a WARNING in this case given that there is a\\n> chance of inconsistent data.\\n\\nIMHO, since the user has specifically asked for origin=NONE but we do\\nnot have any way to detect the origin during initial sync so I think\\nthis could be documented and we can also issue the WARNING. So that\\nusers notice that part and carefully set up the replication. OTOH, I\\ndo not think that giving an error is very inconvenient because we are\\nalready providing a new option \\\"origin=NONE\\\" so along with that lets\\nforce the user to choose between copy_data=off or copy_data=force and\\nwith that, there is no scope for mistakes.\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 26 Aug 2022 09:52:27 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 26, 2022 at 9:52 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n>\\n> On Mon, Aug 22, 2022 at 9:19 AM houzj.fnst@fujitsu.com\\n> <houzj.fnst@fujitsu.com> wrote:\\n> >\\n> > > Jonathan, Sawada-San, Hou-San, and others, what do you think is the best way\\n> > > to move forward here?\\n> >\\n> > I think it's fine to throw a WARNING in this case given that there is a\\n> > chance of inconsistent data.\\n>\\n> IMHO, since the user has specifically asked for origin=NONE but we do\\n> not have any way to detect the origin during initial sync so I think\\n> this could be documented and we can also issue the WARNING. So that\\n> users notice that part and carefully set up the replication. OTOH, I\\n> do not think that giving an error is very inconvenient because we are\\n> already providing a new option \\\"origin=NONE\\\" so along with that lets\\n> force the user to choose between copy_data=off or copy_data=force and\\n> with that, there is no scope for mistakes.\\n\\nSince Jonathan also had suggested to throw a warning as in [1] and\\nHou-san also had suggested to throw a warning as in [2], I have made\\nchanges to throw a warning and also documented the following contents\\nfor the same:\\n+ <para>\\n+ If the subscription is created with <literal>origin = none</literal> and\\n+ <literal>copy_data = true</literal>, it will check if the publisher has\\n+ subscribed to the same table from other publishers and, if so, throw a\\n+ warning to notify user to check the publisher tables. The user can ensure\\n+ that publisher tables do not have data which has an origin associated before\\n+ continuing with any other operations to prevent inconsistent data being\\n+ replicated.\\n+ </para>\\n\\nThe changes for the same are available in the v42 patch at [3].\\n[1] - https://www.postgresql.org/message-id/afb653ba-e2b1-33a3-a54c-849f4466e1b4%40postgresql.org\\n[2] - https://www.postgresql.org/message-id/OS0PR01MB5716C383623ADAD64CE4841194719%40OS0PR01MB5716.jpnprd01.prod.outlook.com\\n[3] - https://www.postgresql.org/message-id/CALDaNm2oLWsSYtOFLdOkbrpC94%3DJZzKMCYDJoiZaqAX_Hn%2BU9Q%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 08:24:06 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 29, 2022 at 8:24 AM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Fri, Aug 26, 2022 at 9:52 AM Dilip Kumar <dilipbalaut@gmail.com> wrote:\\n> >\\n> > IMHO, since the user has specifically asked for origin=NONE but we do\\n> > not have any way to detect the origin during initial sync so I think\\n> > this could be documented and we can also issue the WARNING. So that\\n> > users notice that part and carefully set up the replication. OTOH, I\\n> > do not think that giving an error is very inconvenient because we are\\n> > already providing a new option \\\"origin=NONE\\\" so along with that lets\\n> > force the user to choose between copy_data=off or copy_data=force and\\n> > with that, there is no scope for mistakes.\\n>\\n\\nInitially, I also thought that giving an error should be okay in this\\ncase but later multiple people voted against that idea primarily\\nbecause we won't be able to detect whether the initial sync data\\ncontains data from multiple origins. Consider that even though the\\npublisher is subscribed to some other publisher but it may not have\\ngotten any data from it by the time we try to subscribe from it. Also,\\none might have removed the subscription that led to data from multiple\\norigins on the publisher, in this case, may be one can say that we can\\nconsider that the data is now owned by the publisher but I am not\\nsure. So, considering this, I think giving a WARNING and documenting\\nhow to set up replication correctly sounds like a reasonable way\\nforward.\\n\\n> Since Jonathan also had suggested to throw a warning as in [1] and\\n> Hou-san also had suggested to throw a warning as in [2],\\n>\\n\\nAgreed that multiple seems to be in favor of that approach. We can\\nalways promote it to ERROR later if others and or users felt so.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 08:40:29 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are some review comments for patch v42-0001:\\n\\n======\\n\\n1. Commit message\\n\\nA later review comment below suggests some changes to the WARNING\\nmessage so if those changes are made then the example in this commit\\nmessage also needs to be modified.\\n\\n======\\n\\n2. doc/src/sgml/ref/alter_subscription.sgml - copy_data\\n\\nRefer to the Notes for the usage of true for copy_data parameter and\\nits interaction with the origin parameter.\\n\\nI think saying \\\"usage of true\\\" sounded a bit strange.\\n\\nSUGGESTION\\nRefer to the Notes about how copy_data = true can interact with the\\norigin parameter.\\n\\n======\\n\\n3. doc/src/sgml/ref/create_subscription.sgml - copy data\\n\\nRefer to the Notes for the usage of true for copy_data parameter and\\nits interaction with the origin parameter.\\n\\nSUGGESTION\\n(same as #2)\\n\\n~~\\n\\n4. doc/src/sgml/ref/create_subscription.sgml - origin\\n\\nRefer to the Notes for the usage of true for copy_data parameter and\\nits interaction with the origin parameter.\\n\\nSUGGESTION\\n(same as #2)\\n\\n~~~\\n\\n5. doc/src/sgml/ref/create_subscription.sgml - Notes\\n\\n+ <para>\\n+ If the subscription is created with <literal>origin = none</literal> and\\n+ <literal>copy_data = true</literal>, it will check if the publisher has\\n+ subscribed to the same table from other publishers and, if so, throw a\\n+ warning to notify user to check the publisher tables. The user can ensure\\n\\n\\\"throw a warning to notify user\\\" -> \\\"log a warning to notify the user\\\"\\n\\n~~\\n\\n6.\\n\\n+ warning to notify user to check the publisher tables. The user can ensure\\n+ that publisher tables do not have data which has an origin associated before\\n+ continuing with any other operations to prevent inconsistent data being\\n+ replicated.\\n+ </para>\\n\\n6a.\\n\\nI'm not so sure about this. IMO the warning is not about the\\nreplication – really it is about the COPY which has already happened\\nanyway. So the user can't really prevent anything from going wrong;\\ninstead, they have to take some action to clean up if anything did go\\nwrong.\\n\\nSUGGESTION\\nBefore continuing with other operations the user should check that\\npublisher tables did not have data with different origins, otherwise\\ninconsistent data may have been copied.\\n\\n~\\n\\n6b.\\n\\nI am also wondering what can the user do now. Assuming there was bad\\nCOPY then the subscriber table has already got unwanted stuff copied\\ninto it. Is there any advice we can give to help users fix this mess?\\n\\n======\\n\\n7. src/backend/commands/subscriptioncmds.c - AlterSubscription_refresh\\n\\n+ /* Check whether we can allow copy of newly added relations. */\\n+ check_pub_table_subscribed(wrconn, sub->publications, copy_data,\\n+ sub->origin, subrel_local_oids,\\n+ subrel_count);\\n\\n\\\"whether we can allow\\\" seems not quite the right wording here anymore,\\nbecause now there is no ERROR stopping this - so if there was unwanted\\ndata the COPY will proceed to copy it anyhow...\\n\\n~~~\\n\\n8. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n\\n@@ -1781,6 +1794,121 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\\n table_close(rel, RowExclusiveLock);\\n }\\n\\n+/*\\n+ * Check and throw a warning if the publisher has subscribed to the same table\\n+ * from some other publisher. This check is required only if \\\"copy_data = true\\\"\\n+ * and \\\"origin = none\\\" for CREATE SUBSCRIPTION and\\n+ * ALTER SUBSCRIPTION ... REFRESH statements to notify user that data having\\n+ * origin might have been copied.\\n\\n\\\"throw a warning\\\" -> \\\"log a warning\\\"\\n\\n\\\"to notify user\\\" -> \\\"to notify the user\\\" ?\\n\\n~~~\\n\\n9.\\n\\n+ if (copydata == false || !origin ||\\n+ (pg_strcasecmp(origin, LOGICALREP_ORIGIN_NONE) != 0))\\n+ return;\\n\\nSUGGESTION\\nif (!copydata || !origin ||\\n(pg_strcasecmp(origin, LOGICALREP_ORIGIN_NONE) != 0))\\n\\n~~~\\n\\n10. (Question)\\n\\nI can't tell just by reading the code if FOR ALL TABLES case is\\nhandled – e.g. will this recognise the case were the publisher might\\nhave table data from other origins because it has a subscription on\\nsome other node that was publishing \\\"FOR ALL TABLES\\\"?\\n\\n~~~\\n\\n11.\\n\\n+ ereport(WARNING,\\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n+ errmsg(\\\"publisher has subscribed table \\\\\\\"%s.%s\\\\\\\" from some other publisher\\\",\\n+ nspname, relname),\\n+ errdetail(\\\"Publisher might have subscribed one or more tables from\\nsome other publisher.\\\"),\\n+ errhint(\\\"Verify that these publisher tables do not have data that\\nhas an origin associated before proceeding to avoid inconsistency.\\\"));\\n+ break;\\n\\n11a.\\nI'm not sure having that \\\"break\\\" code is correct logic. Won't that\\nmean the user will only get a warning for the first potential problem\\nencountered but then other potential problems tables will have no\\nwarnings at all so the user may not be made aware of them? Perhaps you\\nneed to first gather the list of all the suspicious tables before\\nlogging a single warning which shows that list? Or perhaps you need to\\nlog multiple warnings – one warning per suspicious table?\\n\\n~\\n\\n11b.\\nThe WARNING message seems a bit inside-out because the errmsg is\\ngiving more details than the errdetail.\\n\\nSUGGESTIONS (or something similar)\\nerrmsg - \\\"subscription XXX requested origin=NONE but may have copied\\ndata that had a different origin.\\\"\\nerrdetail – Publisher YYY has subscribed table \\\\\\\"%s.%s\\\\\\\" from some\\nother publisher\\\"\\n\\n~\\n\\n11c.\\nThe errhint sentence is unusual.\\n\\nBEFORE\\n\\\"Verify that these publisher tables do not have data that has an\\norigin associated before proceeding to avoid inconsistency.\\\"\\n\\nSUGGESTION (or something similar)\\nBefore proceeding, verify that initial data copied from the publisher\\ntables did not come from other origins.\\n\\n======\\n\\n12. src/test/subscription/t/030_origin.pl\\n\\n+# have remotely originated data from node_A. We throw a warning, in this case,\\n+# to draw attention to there being possible remote data.\\n\\n\\\"throw a warning\\\" -> \\\"log a warning\\\" (this occurs 2x)\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 16:29:27 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are some review comments for patch v42-0002:\\n\\n======\\n\\n1. doc/src/sgml/logical-replication.sgml\\n\\n<literal>copy_data = true </literal>\\n\\nThere are a couple of these tags where there is a trailing space\\nbefore the </literal>. I guess it is doing no harm, but it is doing no\\ngood either, so IMO better to get rid of the space.\\n\\n~~\\n\\n2. doc/src/sgml/logical-replication.sgml - 31.3.1. Setting replication\\nbetween two primaries\\n\\nUser need to ensure that no data changes happen on table t1 on\\nprimary1 and primary2 until the setup is completed.\\n\\nSUGGESTION\\nThe user must ensure that no data changes happen on table t1 of\\nprimary1 and primary2 until the setup is completed.\\n\\n~~~\\n\\n3. doc/src/sgml/logical-replication.sgml - 31.3.2. Adding a new\\nprimary when there is no table data on any of the primaries\\n\\nUser need to ensure that no data changes happen on table t1 on all the\\nprimaries primary1, primary2 and primary3 until the setup is\\ncompleted.\\n\\nSUGGESTION\\nThe user must ensure that no data changes happen on table t1 of all\\nthe primaries primary1, primary2 and primary3 until the setup is\\ncompleted.\\n\\n~~~\\n\\n4. doc/src/sgml/logical-replication.sgml - 31.3.3. Adding a new\\nprimary when table data is present on the existing primaries\\n\\nUser need to ensure that no operations should be performed on table t1\\non primary2 and primary3 until the setup is completed.\\n\\nSUGGESTION\\nThe user must ensure that no operations are performed on table t1 of\\nprimary2 and primary3 until the setup is completed.\\n\\nHere also you changed the wording - \\\"no data changes happen\\\" versus\\n\\\"no operations should be performed\\\". Was that a deliberate difference?\\nMaybe they should all be consistent wording? If they are\\ninterchangeable IMO the \\\"no operations are performed...\\\" wording was\\nthe better of the two.\\n\\n~~~\\n\\n5. doc/src/sgml/logical-replication.sgml - 31.3.4. Generic steps for\\nadding a new primary to an existing set of primaries\\n\\nStep-2: User need to ensure that no changes happen on the required\\ntables of the new primary until the setup is complete.\\n\\nSUGGESTION\\nStep-2: The user must ensure that no changes happen on the required\\ntables of the new primary until the setup is complete.\\n\\n~~~\\n\\n6.\\n\\nStep-4. User need to ensure that no changes happen on the required\\ntables of the existing primaries except the first primary until the\\nsetup is complete.\\n\\nSUGGESTION\\nStep-4. The user must ensure that no changes happen on the required\\ntables of the existing primaries except the first primary until the\\nsetup is complete.\\n\\n~~~\\n\\n7. (the example)\\n\\n7a.\\nStep-2. User need to ensure that no changes happen on table t1 on\\nprimary4 until the setup is completed.\\n\\nSUGGESTION\\nStep-2. The user must ensure that no changes happen on table t1 of\\nprimary4 until the setup is completed.\\n\\n~\\n\\n7b.\\nStep-4. User need to ensure that no changes happen on table t1 on\\nprimary2 and primary3 until the setup is completed.\\n\\nSUGGESTION\\nStep-4. The user must ensure that no changes happen on table t1 of\\nprimary2 and primary3 until the setup is completed.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 16:31:07 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 24, 2022 at 7:27 PM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> Since there was no objections to change it to throw a warning, I have\\n> made the changes for the same.\\n>\\n\\nReview comments for v42-0001*\\n==========================\\n1. Can we improve the query in check_pub_table_subscribed() so that it\\ndoesn't fetch any table that is already part of the subscription on\\nthe subscriber? This may be better than what the patch is doing which\\nis to first fetch such information and then skip it. If forming this\\nquery turns out to be too complex then we can retain your method as\\nwell but I feel it is worth trying to optimize the query used in the\\npatch.\\n\\n2. I thought that it may be better to fetch all the tables that have\\nthe possibility to have data from more than one origin but maybe the\\nlist will be too long especially for FOR ALL TABLE types of cases. Is\\nthis the reason you have decided to give a WARNING as soon as you see\\nany such table? If so, probably adding a comment for it can be good.\\n\\n3.\\n+ $node_publisher->wait_for_catchup($sub_name);\\n+\\n+ # also wait for initial table sync to finish\\n+ $node_subscriber->poll_query_until('postgres', $synced_query)\\n+ or die \\\"Timed out while waiting for subscriber to synchronize data\\\";\\n....\\n....\\n....\\n+$node_B->safe_psql(\\n+ 'postgres', \\\"\\n+ ALTER SUBSCRIPTION $subname_BA REFRESH PUBLICATION\\\");\\n+\\n+$node_B->poll_query_until('postgres', $synced_query)\\n+ or die \\\"Timed out while waiting for subscriber to synchronize data\\\";\\n\\nYou can replace this and any similar code in the patch with a method\\nused in commit 0c20dd33db.\\n\\n4.\\n+###############################################################################\\n+# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\\n+# replication setup when the existing nodes (node_A & node_B) has pre-existing\\n+# data and the new node (node_C) does not have any data.\\n+###############################################################################\\n+$result = $node_A->safe_psql('postgres', \\\"SELECT * FROM tab ORDER BY 1;\\\");\\n+is($result, qq(), 'Check existing data');\\n+\\n+$result = $node_B->safe_psql('postgres', \\\"SELECT * FROM tab ORDER BY 1;\\\");\\n+is($result, qq(), 'Check existing data');\\n+\\n+$result = $node_C->safe_psql('postgres', \\\"SELECT * FROM tab ORDER BY 1;\\\");\\n+is($result, qq(), 'Check existing data');\\n\\nThe comments say that for this test, two of the nodes have some\\npre-existing data but I don't see the same in the test results. The\\ntest following this one will also have a similar effect. BTW, I am not\\nsure all the new three node tests added by this patch are required\\nbecause I don't see if they have additional code coverage or test\\nanything which is not tested without those. This has doubled the\\namount of test timing for this test file which is okay if these tests\\nmake any useful contribution but otherwise, it may be better to remove\\nthese. Am, I missing something?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 16:34:51 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 29, 2022 at 11:59 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some review comments for patch v42-0001:\\n>\\n> ~~\\n>\\n> 6.\\n>\\n> + warning to notify user to check the publisher tables. The user can ensure\\n> + that publisher tables do not have data which has an origin associated before\\n> + continuing with any other operations to prevent inconsistent data being\\n> + replicated.\\n> + </para>\\n>\\n> 6a.\\n>\\n> I'm not so sure about this. IMO the warning is not about the\\n> replication – really it is about the COPY which has already happened\\n> anyway. So the user can't really prevent anything from going wrong;\\n> instead, they have to take some action to clean up if anything did go\\n> wrong.\\n>\\n> SUGGESTION\\n> Before continuing with other operations the user should check that\\n> publisher tables did not have data with different origins, otherwise\\n> inconsistent data may have been copied.\\n>\\n> ~\\n\\nI would like to avoid the use of 'may' here. So, let's change it to\\nsomething like: \\\"Before continuing with other operations the user\\nshould check that publisher tables did not have data with different\\norigins to prevent data inconsistency issues on the subscriber.\\\"\\n>\\n> 6b.\\n>\\n> I am also wondering what can the user do now. Assuming there was bad\\n> COPY then the subscriber table has already got unwanted stuff copied\\n> into it. Is there any advice we can give to help users fix this mess?\\n>\\n\\nI don't think the user can do much in that case. She will probably\\nneed to re-create the replication.\\n\\n>\\n> 11.\\n>\\n> + ereport(WARNING,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"publisher has subscribed table \\\\\\\"%s.%s\\\\\\\" from some other publisher\\\",\\n> + nspname, relname),\\n> + errdetail(\\\"Publisher might have subscribed one or more tables from\\n> some other publisher.\\\"),\\n> + errhint(\\\"Verify that these publisher tables do not have data that\\n> has an origin associated before proceeding to avoid inconsistency.\\\"));\\n> + break;\\n>\\n> 11a.\\n> I'm not sure having that \\\"break\\\" code is correct logic. Won't that\\n> mean the user will only get a warning for the first potential problem\\n> encountered but then other potential problems tables will have no\\n> warnings at all so the user may not be made aware of them? Perhaps you\\n> need to first gather the list of all the suspicious tables before\\n> logging a single warning which shows that list? Or perhaps you need to\\n> log multiple warnings – one warning per suspicious table?\\n>\\n> ~\\n>\\n> 11b.\\n> The WARNING message seems a bit inside-out because the errmsg is\\n> giving more details than the errdetail.\\n>\\n> SUGGESTIONS (or something similar)\\n> errmsg - \\\"subscription XXX requested origin=NONE but may have copied\\n> data that had a different origin.\\\"\\n> errdetail – Publisher YYY has subscribed table \\\\\\\"%s.%s\\\\\\\" from some\\n> other publisher\\\"\\n>\\n\\nYour suggestion is better but we can't say 'copied' because the copy\\nmay start afterwards by tablesync worker. Also, using 'may' is not\\nadvisable in error messages. How about : \\\"subscription XXX requested\\norigin=NONE but might copy data that had a different origin.\\\"?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 17:21:44 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, 29 Aug 2022 at 11:59, Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some review comments for patch v42-0001:\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> A later review comment below suggests some changes to the WARNING\\n> message so if those changes are made then the example in this commit\\n> message also needs to be modified.\\n\\nModified\\n\\n> ======\\n>\\n> 2. doc/src/sgml/ref/alter_subscription.sgml - copy_data\\n>\\n> Refer to the Notes for the usage of true for copy_data parameter and\\n> its interaction with the origin parameter.\\n>\\n> I think saying \\\"usage of true\\\" sounded a bit strange.\\n>\\n> SUGGESTION\\n> Refer to the Notes about how copy_data = true can interact with the\\n> origin parameter.\\n\\nModified\\n\\n> ======\\n>\\n> 3. doc/src/sgml/ref/create_subscription.sgml - copy data\\n>\\n> Refer to the Notes for the usage of true for copy_data parameter and\\n> its interaction with the origin parameter.\\n>\\n> SUGGESTION\\n> (same as #2)\\n\\nModified\\n\\n> ~~\\n>\\n> 4. doc/src/sgml/ref/create_subscription.sgml - origin\\n>\\n> Refer to the Notes for the usage of true for copy_data parameter and\\n> its interaction with the origin parameter.\\n>\\n> SUGGESTION\\n> (same as #2)\\n\\nModified\\n\\n> ~~~\\n>\\n> 5. doc/src/sgml/ref/create_subscription.sgml - Notes\\n>\\n> + <para>\\n> + If the subscription is created with <literal>origin = none</literal> and\\n> + <literal>copy_data = true</literal>, it will check if the publisher has\\n> + subscribed to the same table from other publishers and, if so, throw a\\n> + warning to notify user to check the publisher tables. The user can ensure\\n>\\n> \\\"throw a warning to notify user\\\" -> \\\"log a warning to notify the user\\\"\\n\\nModified\\n\\n> ~~\\n>\\n> 6.\\n>\\n> + warning to notify user to check the publisher tables. The user can ensure\\n> + that publisher tables do not have data which has an origin associated before\\n> + continuing with any other operations to prevent inconsistent data being\\n> + replicated.\\n> + </para>\\n>\\n> 6a.\\n>\\n> I'm not so sure about this. IMO the warning is not about the\\n> replication – really it is about the COPY which has already happened\\n> anyway. So the user can't really prevent anything from going wrong;\\n> instead, they have to take some action to clean up if anything did go\\n> wrong.\\n>\\n> SUGGESTION\\n> Before continuing with other operations the user should check that\\n> publisher tables did not have data with different origins, otherwise\\n> inconsistent data may have been copied.\\n\\nModified based on the suggestion provided by Amit.\\n\\n> ~\\n>\\n> 6b.\\n>\\n> I am also wondering what can the user do now. Assuming there was bad\\n> COPY then the subscriber table has already got unwanted stuff copied\\n> into it. Is there any advice we can give to help users fix this mess?\\n\\nThere is nothing much a user can do in this case. Only option would be\\nto take a backup before the operation and restore it and then recreate\\nthe replication setup. I was not sure if we should document these\\nsteps as similar inconsistent data could get created without the\\norigin option . Thoughts?\\n\\n> ======\\n>\\n> 7. src/backend/commands/subscriptioncmds.c - AlterSubscription_refresh\\n>\\n> + /* Check whether we can allow copy of newly added relations. */\\n> + check_pub_table_subscribed(wrconn, sub->publications, copy_data,\\n> + sub->origin, subrel_local_oids,\\n> + subrel_count);\\n>\\n> \\\"whether we can allow\\\" seems not quite the right wording here anymore,\\n> because now there is no ERROR stopping this - so if there was unwanted\\n> data the COPY will proceed to copy it anyhow...\\n\\nRemoved the comment as the function details the necessary things.\\n\\n> ~~~\\n>\\n> 8. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> @@ -1781,6 +1794,121 @@ AlterSubscriptionOwner_oid(Oid subid, Oid newOwnerId)\\n> table_close(rel, RowExclusiveLock);\\n> }\\n>\\n> +/*\\n> + * Check and throw a warning if the publisher has subscribed to the same table\\n> + * from some other publisher. This check is required only if \\\"copy_data = true\\\"\\n> + * and \\\"origin = none\\\" for CREATE SUBSCRIPTION and\\n> + * ALTER SUBSCRIPTION ... REFRESH statements to notify user that data having\\n> + * origin might have been copied.\\n>\\n> \\\"throw a warning\\\" -> \\\"log a warning\\\"\\n>\\n> \\\"to notify user\\\" -> \\\"to notify the user\\\" ?\\n\\nModified\\n\\n> ~~~\\n>\\n> 9.\\n>\\n> + if (copydata == false || !origin ||\\n> + (pg_strcasecmp(origin, LOGICALREP_ORIGIN_NONE) != 0))\\n> + return;\\n>\\n> SUGGESTION\\n> if (!copydata || !origin ||\\n> (pg_strcasecmp(origin, LOGICALREP_ORIGIN_NONE) != 0))\\n\\nModified\\n\\n> ~~~\\n>\\n> 10. (Question)\\n>\\n> I can't tell just by reading the code if FOR ALL TABLES case is\\n> handled – e.g. will this recognise the case were the publisher might\\n> have table data from other origins because it has a subscription on\\n> some other node that was publishing \\\"FOR ALL TABLES\\\"?\\n\\nYes it handles it.\\n\\n> ~~~\\n>\\n> 11.\\n>\\n> + ereport(WARNING,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"publisher has subscribed table \\\\\\\"%s.%s\\\\\\\" from some other publisher\\\",\\n> + nspname, relname),\\n> + errdetail(\\\"Publisher might have subscribed one or more tables from\\n> some other publisher.\\\"),\\n> + errhint(\\\"Verify that these publisher tables do not have data that\\n> has an origin associated before proceeding to avoid inconsistency.\\\"));\\n> + break;\\n>\\n> 11a.\\n> I'm not sure having that \\\"break\\\" code is correct logic. Won't that\\n> mean the user will only get a warning for the first potential problem\\n> encountered but then other potential problems tables will have no\\n> warnings at all so the user may not be made aware of them? Perhaps you\\n> need to first gather the list of all the suspicious tables before\\n> logging a single warning which shows that list? Or perhaps you need to\\n> log multiple warnings – one warning per suspicious table?\\n\\nModified to get the list of all the tables, I have limited it to\\ndisplay 100 tables followed by ... to indicate there might be more\\ntables. I did not want to loga warning message with many 1000's of\\ntables which will not help the readability of the message.\\n\\n> ~\\n>\\n> 11b.\\n> The WARNING message seems a bit inside-out because the errmsg is\\n> giving more details than the errdetail.\\n>\\n> SUGGESTIONS (or something similar)\\n> errmsg - \\\"subscription XXX requested origin=NONE but may have copied\\n> data that had a different origin.\\\"\\n> errdetail – Publisher YYY has subscribed table \\\\\\\"%s.%s\\\\\\\" from some\\n> other publisher\\\"\\n\\nWe don't have the tables based on each publisher, it is for all the\\npublishers. I have changed the errdetail message slightly.\\n\\n> ~\\n>\\n> 11c.\\n> The errhint sentence is unusual.\\n>\\n> BEFORE\\n> \\\"Verify that these publisher tables do not have data that has an\\n> origin associated before proceeding to avoid inconsistency.\\\"\\n>\\n> SUGGESTION (or something similar)\\n> Before proceeding, verify that initial data copied from the publisher\\n> tables did not come from other origins.\\n\\nModified\\n\\n> ======\\n>\\n> 12. src/test/subscription/t/030_origin.pl\\n>\\n> +# have remotely originated data from node_A. We throw a warning, in this case,\\n> +# to draw attention to there being possible remote data.\\n>\\n> \\\"throw a warning\\\" -> \\\"log a warning\\\" (this occurs 2x)\\n\\nModified\\n\\nThe attached v43 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 30 Aug 2022 22:35:30 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, 29 Aug 2022 at 12:01, Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some review comments for patch v42-0002:\\n>\\n> ======\\n>\\n> 1. doc/src/sgml/logical-replication.sgml\\n>\\n> <literal>copy_data = true </literal>\\n>\\n> There are a couple of these tags where there is a trailing space\\n> before the </literal>. I guess it is doing no harm, but it is doing no\\n> good either, so IMO better to get rid of the space.\\n\\nModified\\n\\n> ~~\\n>\\n> 2. doc/src/sgml/logical-replication.sgml - 31.3.1. Setting replication\\n> between two primaries\\n>\\n> User need to ensure that no data changes happen on table t1 on\\n> primary1 and primary2 until the setup is completed.\\n>\\n> SUGGESTION\\n> The user must ensure that no data changes happen on table t1 of\\n> primary1 and primary2 until the setup is completed.\\n\\nModified\\n\\n> ~~~\\n>\\n> 3. doc/src/sgml/logical-replication.sgml - 31.3.2. Adding a new\\n> primary when there is no table data on any of the primaries\\n>\\n> User need to ensure that no data changes happen on table t1 on all the\\n> primaries primary1, primary2 and primary3 until the setup is\\n> completed.\\n>\\n> SUGGESTION\\n> The user must ensure that no data changes happen on table t1 of all\\n> the primaries primary1, primary2 and primary3 until the setup is\\n> completed.\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. doc/src/sgml/logical-replication.sgml - 31.3.3. Adding a new\\n> primary when table data is present on the existing primaries\\n>\\n> User need to ensure that no operations should be performed on table t1\\n> on primary2 and primary3 until the setup is completed.\\n>\\n> SUGGESTION\\n> The user must ensure that no operations are performed on table t1 of\\n> primary2 and primary3 until the setup is completed.\\n>\\n> Here also you changed the wording - \\\"no data changes happen\\\" versus\\n> \\\"no operations should be performed\\\". Was that a deliberate difference?\\n> Maybe they should all be consistent wording? If they are\\n> interchangeable IMO the \\\"no operations are performed...\\\" wording was\\n> the better of the two.\\n\\nModified, it was not a deliberate change. I have changed it to \\\"no\\noperations are performed\\\" in other places too.\\n\\n> ~~~\\n>\\n> 5. doc/src/sgml/logical-replication.sgml - 31.3.4. Generic steps for\\n> adding a new primary to an existing set of primaries\\n>\\n> Step-2: User need to ensure that no changes happen on the required\\n> tables of the new primary until the setup is complete.\\n>\\n> SUGGESTION\\n> Step-2: The user must ensure that no changes happen on the required\\n> tables of the new primary until the setup is complete.\\n\\nModified\\n\\n> ~~~\\n>\\n> 6.\\n>\\n> Step-4. User need to ensure that no changes happen on the required\\n> tables of the existing primaries except the first primary until the\\n> setup is complete.\\n>\\n> SUGGESTION\\n> Step-4. The user must ensure that no changes happen on the required\\n> tables of the existing primaries except the first primary until the\\n> setup is complete.\\n\\nModified\\n\\n> ~~~\\n>\\n> 7. (the example)\\n>\\n> 7a.\\n> Step-2. User need to ensure that no changes happen on table t1 on\\n> primary4 until the setup is completed.\\n>\\n> SUGGESTION\\n> Step-2. The user must ensure that no changes happen on table t1 of\\n> primary4 until the setup is completed.\\n\\nModified\\n\\n> ~\\n>\\n> 7b.\\n> Step-4. User need to ensure that no changes happen on table t1 on\\n> primary2 and primary3 until the setup is completed.\\n>\\n> SUGGESTION\\n> Step-4. The user must ensure that no changes happen on table t1 of\\n> primary2 and primary3 until the setup is completed.\\n\\nModified\\n\\nThanks for the comments, the changes for the same are available in the\\nv43 patch attached at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm1V%2BAvzPsUcq%3DmNYG%2BJfAyovTWBe4vWKTknOgH_ko1E0Q%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 30 Aug 2022 22:41:02 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for v43-0001.\\n\\n======\\n\\n1. Commit message\\n\\nThe initial copy phase has no way to know the origin of the row data,\\nso if 'copy_data = true' in the step 4 below, log a warning to notify user\\nthat potentially non-local data might have been copied.\\n\\n1a.\\n\\\"in the step 4\\\" -> \\\"in step 4\\\"\\n\\n~\\n\\n1b.\\n\\\"notify user\\\" -> \\\"notify the user\\\"\\n\\n======\\n\\n2. doc/src/sgml/ref/create_subscription.sgml\\n\\n+ <para>\\n+ If the subscription is created with <literal>origin = none</literal> and\\n+ <literal>copy_data = true</literal>, it will check if the publisher has\\n+ subscribed to the same table from other publishers and, if so, log a\\n+ warning to notify the user to check the publisher tables. Before continuing\\n+ with other operations the user should check that publisher tables did not\\n+ have data with different origins to prevent data inconsistency issues on the\\n+ subscriber.\\n+ </para>\\n\\nI am not sure about that wording saying \\\"to prevent data inconsistency\\nissues\\\" because I thought maybe is already too late because any\\nunwanted origin data is already copied during the initial sync. I\\nthink the user can do it try to clean up any problems caused before\\nthings become even worse (??)\\n\\n~~~\\n\\nYou asked for my thoughts (see [1] 6b):\\n\\n> There is nothing much a user can do in this case. Only option would be\\n> to take a backup before the operation and restore it and then recreate\\n> the replication setup. I was not sure if we should document these\\n> steps as similar inconsistent data could get created without the\\n> origin option . Thoughts?\\n\\nI think those cases are a bit different:\\n\\n- For a normal scenario (i.e. origin=ANY) then inconsistent data just\\nrefers to problems like maybe PK violations so I think you just get\\nwhat you get and have to deal with the errors...\\n\\n- But when the user said origin=NONE they are specifically saying they\\ndo NOT want any non-local data, yet they still might end up getting\\ndata contrary to their wishes. So I think maybe there ought to be\\nsomething documented to warn about this potentially happening. My\\nfirst impression is that all this seems like a kind of terrible\\nproblem because if it happens then cleanup could be difficult - if\\nthat subscriber in turn was also a publisher then this bad data might\\nhave propagated all over the place already! Anyway, I guess all this\\ncould be mentioned in some (new ?) section of the\\nlogical-replication.sgml file (i.e. patch 0002).\\n\\nIDEA: I was wondering if the documentation should recommend (or maybe\\nwe can even enforce this) that when using origin=NONE the user should\\nalways create the subscription with enabled=FALSE. That way if there\\nare some warnings about potential bad origin data then there it might\\nbe possible to take some action or change their mind *before* any\\nunwanted data pollutes everything.\\n\\n======\\n\\n3. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n\\n+ if (count == 0)\\n+ tablenames = makeStringInfo();\\n+ else\\n+ appendStringInfoString(tablenames, \\\", \\\");\\n+\\n+ appendStringInfo(tablenames, \\\"%s.%s\\\", nspname, relname);\\n+ count++;\\n\\nI think the quotes are not correct - each separate table name should be quoted.\\n\\n~~~\\n\\n4.\\n\\n+ /*\\n+ * There might be many tables present in this case, we will display a\\n+ * maximum of 100 tables followed by \\\"...\\\" to indicate there might be\\n+ * more tables.\\n+ */\\n+ if (count == 100)\\n+ {\\n+ appendStringInfoString(tablenames, \\\", ...\\\");\\n+ break;\\n+ }\\n\\n4a.\\nIMO this code should be part of the previous if/else\\n\\ne.g.\\nif (count == 0) ... makeStringInfo()\\nelse if (count > 100) ... append ellipsis \\\"...\\\" and break out of the loop\\nelse ... append table name to the message\\n\\nDoing it in this way means \\\"...\\\" definitely means there are more than\\n100 bad tables.\\n\\n~\\n\\n4b.\\nChanging like above (#4a) simplifies the comment because it removes\\ndoubts like \\\"might be…\\\" since you already know you found the 101st\\ntable.\\n\\nSUGGESTION (comment)\\nThe message displays a maximum of 100 tables having potentially\\nnon-local data. Any more than that will be indicated by \\\"...\\\".\\n\\n~\\n\\n4c.\\nIt still seems a bit disconcerting to me that there can be cases where\\nbad data was possibly copied but there is no record of what tables\\nhave been polluted. Perhaps when the maximum bad tables are exceeded\\nthen there can be some additional message to tell users what SQL they\\ncan use to discover what all the other bad tables were.\\n\\n~~~\\n\\n5.\\n\\n+ * XXX: For simplicity, we don't check whether the table has any data\\n+ * or not. If the table doesn't have any data then we don't need to\\n+ * distinguish between data having origin and data not having origin so\\n+ * we can avoid throwing a warning in that case.\\n\\n\\\"throwing a warning\\\" -> \\\"logging a warning\\\"\\n\\n~~~\\n\\n6.\\n\\n+ errdetail(\\\"Subscribed table(s) \\\\\\\"%s\\\\\\\" has been subscribed from some\\nother publisher.\\\",\\n+ tablenames->data),\\n\\nI don't think the table name list should be quoted in the errdetail,\\nbecause each of the individual table names should have been quoted.\\nSee previous review comment #3.\\n\\n======\\n\\n7. src/test/subscription/t/030_origin.pl\\n\\n+###############################################################################\\n+# Specify origin as 'none' which indicates that the publisher should only\\n+# replicate the changes that are generated locally from node_B, but in\\n+# this case since the node_B is also subscribing data from node_A, node_B can\\n+# have remotely originated data from node_A. We log a warning, in this case,\\n+# to draw attention to there being possible remote data.\\n+###############################################################################\\n\\nThis comment wording makes it seem like this is using different\\nsubscription parameters from the other tests that preceded it, but in\\nfact the prior tests also been using origin=none. (??)\\n\\nMaybe it just needs a trivial rewording.\\n\\nSUGGESTION\\nSpecifying origin=NONE indicates that...\\n\\n~~~\\n\\n8.\\n\\n+like(\\n+ $stderr,\\n+ qr/WARNING: ( [A-Z0-9]+:)? subscription tap_sub_a_b_2 requested\\norigin=NONE but might copy data that had a different origin/,\\n+ \\\"Create subscription with origin = none and copy_data when the\\npublisher has subscribed same table\\\"\\n+);\\n\\nIs it possible for these checks of the log file to get the errdetail\\nmessage too? Then the test will be more readable for the expected\\nresult and you will also have the bonus of testing the table-name list\\nin the warning more thoroughly.\\n\\n------\\n[1] https://www.postgresql.org/message-id/CALDaNm1V%2BAvzPsUcq%3DmNYG%2BJfAyovTWBe4vWKTknOgH_ko1E0Q%40mail.gmail.com\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 31 Aug 2022 16:05:01 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are some review comments for patch v43-0002:\\n\\n======\\n\\n1. doc/src/sgml/ref/create_subscription.sgml\\n\\n@@ -403,7 +403,9 @@ CREATE SUBSCRIPTION <replaceable\\nclass=\\\"parameter\\\">subscription_name</replaceabl\\n warning to notify the user to check the publisher tables. Before continuing\\n with other operations the user should check that publisher tables did not\\n have data with different origins to prevent data inconsistency issues on the\\n- subscriber.\\n+ subscriber. Refer to <xref linkend=\\\"replication-between-primaries\\\"/> for\\n+ how <literal>copy_data</literal> and <literal>origin</literal> can be used\\n+ to set up replication between primaries.\\n </para>\\n\\nRegarding my earlier v43-0001 review (see [1] comment #2) perhaps\\nanother pg docs section should be added in the\\nlogical-replication.sgml (e.g. \\\"Specifying origins during CREATE\\nSUBSCRIPTION\\\"), so then this Notes text also should have more added to\\nit.\\n\\nSUGGESTION\\nRefer to <XXX_REF> for details about potential initialization\\ninconsistency warnings using origin=NONE.\\nRefer to <YYY_REF> for how copy_data and origin can be used to set up\\nreplication between primaries.\\n\\n------\\n[1] https://www.postgresql.org/message-id/CAHut%2BPvonTd423-cWqoxh0w8Bd_Po3OToqqyxuR1iMNmxSLr_Q%40mail.gmail.com\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 31 Aug 2022 16:15:29 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 31, 2022 at 11:35 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for v43-0001.\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> The initial copy phase has no way to know the origin of the row data,\\n> so if 'copy_data = true' in the step 4 below, log a warning to notify user\\n> that potentially non-local data might have been copied.\\n>\\n> 1a.\\n> \\\"in the step 4\\\" -> \\\"in step 4\\\"\\n>\\n> ~\\n>\\n> 1b.\\n> \\\"notify user\\\" -> \\\"notify the user\\\"\\n>\\n> ======\\n>\\n> 2. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> + <para>\\n> + If the subscription is created with <literal>origin = none</literal> and\\n> + <literal>copy_data = true</literal>, it will check if the publisher has\\n> + subscribed to the same table from other publishers and, if so, log a\\n> + warning to notify the user to check the publisher tables. Before continuing\\n> + with other operations the user should check that publisher tables did not\\n> + have data with different origins to prevent data inconsistency issues on the\\n> + subscriber.\\n> + </para>\\n>\\n> I am not sure about that wording saying \\\"to prevent data inconsistency\\n> issues\\\" because I thought maybe is already too late because any\\n> unwanted origin data is already copied during the initial sync. I\\n> think the user can do it try to clean up any problems caused before\\n> things become even worse (??)\\n>\\n\\nOh no, it is not too late in all cases. The problem can only occur if\\nthe user will try to subscribe from all the different publications\\nwith copy_data = true. We are anyway trying to clarify in the second\\npatch the right way to accomplish this. So, I am not sure what better\\nwe can do here. The only bulletproof way is to provide some APIs where\\nusers don't need to bother about all these cases, basically something\\nsimilar to what Kuroda-San is working on in the thread [1].\\n\\n> ~~~\\n>\\n> You asked for my thoughts (see [1] 6b):\\n>\\n> > There is nothing much a user can do in this case. Only option would be\\n> > to take a backup before the operation and restore it and then recreate\\n> > the replication setup. I was not sure if we should document these\\n> > steps as similar inconsistent data could get created without the\\n> > origin option . Thoughts?\\n>\\n> I think those cases are a bit different:\\n>\\n> - For a normal scenario (i.e. origin=ANY) then inconsistent data just\\n> refers to problems like maybe PK violations so I think you just get\\n> what you get and have to deal with the errors...\\n>\\n> - But when the user said origin=NONE they are specifically saying they\\n> do NOT want any non-local data, yet they still might end up getting\\n> data contrary to their wishes. So I think maybe there ought to be\\n> something documented to warn about this potentially happening. My\\n> first impression is that all this seems like a kind of terrible\\n> problem because if it happens then cleanup could be difficult - if\\n> that subscriber in turn was also a publisher then this bad data might\\n> have propagated all over the place already! Anyway, I guess all this\\n> could be mentioned in some (new ?) section of the\\n> logical-replication.sgml file (i.e. patch 0002).\\n>\\n> IDEA: I was wondering if the documentation should recommend (or maybe\\n> we can even enforce this) that when using origin=NONE the user should\\n> always create the subscription with enabled=FALSE. That way if there\\n> are some warnings about potential bad origin data then there it might\\n> be possible to take some action or change their mind *before* any\\n> unwanted data pollutes everything.\\n>\\n\\nIf we want to give this suggestion, then we need to also say that this\\nis only valid when copy_data is true. The other drawback is if users\\nalways need to do this to use origin=NONE then this can be a burden to\\nthe user. I am not against having such a NOTE but the tone should not\\nbe to enforce users to do this.\\n\\n[1] - https://www.postgresql.org/message-id/CAHut%2BPuwRAoWY9pz%3DEubps3ooQCOBFiYPU9Yi%3DVB-U%2ByORU7OA%40mail.gmail.com\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 31 Aug 2022 16:06:40 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 31, 2022 at 8:36 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Aug 31, 2022 at 11:35 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Here are my review comments for v43-0001.\\n> >\\n...\\n> > ======\\n> >\\n> > 2. doc/src/sgml/ref/create_subscription.sgml\\n> >\\n> > + <para>\\n> > + If the subscription is created with <literal>origin = none</literal> and\\n> > + <literal>copy_data = true</literal>, it will check if the publisher has\\n> > + subscribed to the same table from other publishers and, if so, log a\\n> > + warning to notify the user to check the publisher tables. Before continuing\\n> > + with other operations the user should check that publisher tables did not\\n> > + have data with different origins to prevent data inconsistency issues on the\\n> > + subscriber.\\n> > + </para>\\n> >\\n> > I am not sure about that wording saying \\\"to prevent data inconsistency\\n> > issues\\\" because I thought maybe is already too late because any\\n> > unwanted origin data is already copied during the initial sync. I\\n> > think the user can do it try to clean up any problems caused before\\n> > things become even worse (??)\\n> >\\n>\\n> Oh no, it is not too late in all cases. The problem can only occur if\\n> the user will try to subscribe from all the different publications\\n> with copy_data = true. We are anyway trying to clarify in the second\\n> patch the right way to accomplish this. So, I am not sure what better\\n> we can do here. The only bulletproof way is to provide some APIs where\\n> users don't need to bother about all these cases, basically something\\n> similar to what Kuroda-San is working on in the thread [1].\\n>\\n\\nThe point of my review comment was only about the wording of the note\\n- specifically, you cannot \\\"prevent\\\" something (e,g, data\\ninconsistency) if it has already happened.\\n\\nMaybe modify the wording like below?\\n\\nBEFORE\\nBefore continuing with other operations the user should check that\\npublisher tables did not have data with different origins to prevent\\ndata inconsistency issues on the subscriber.\\n\\nAFTER\\nIf a publisher table with data from different origins was found to\\nhave been copied to the subscriber, then some corrective action may be\\nnecessary before continuing with other operations.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Thu, 1 Sep 2022 10:19:19 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 31, 2022 1:06 AM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> The attached v43 patch has the changes for the same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch.\\r\\n\\r\\nHere is a comment on the 0001 patch.\\r\\n\\r\\n+\\t\\tif (!isnewtable)\\r\\n+\\t\\t{\\r\\n+\\t\\t\\tpfree(nspname);\\r\\n+\\t\\t\\tpfree(relname);\\r\\n+\\t\\t\\tcontinue;\\r\\n+\\t\\t}\\r\\n\\r\\nIf it is a new table, in which case it would log a warning, should we also call\\r\\npfree()?\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Thu, 1 Sep 2022 02:31:37 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 31, 2022 at 11:35 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for v43-0001.\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> The initial copy phase has no way to know the origin of the row data,\\n> so if 'copy_data = true' in the step 4 below, log a warning to notify user\\n> that potentially non-local data might have been copied.\\n>\\n> 1a.\\n> \\\"in the step 4\\\" -> \\\"in step 4\\\"\\n>\\n> ~\\n>\\n> 1b.\\n> \\\"notify user\\\" -> \\\"notify the user\\\"\\n>\\n> ======\\n>\\n> 2. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> + <para>\\n> + If the subscription is created with <literal>origin = none</literal> and\\n> + <literal>copy_data = true</literal>, it will check if the publisher has\\n> + subscribed to the same table from other publishers and, if so, log a\\n> + warning to notify the user to check the publisher tables. Before continuing\\n> + with other operations the user should check that publisher tables did not\\n> + have data with different origins to prevent data inconsistency issues on the\\n> + subscriber.\\n> + </para>\\n>\\n> I am not sure about that wording saying \\\"to prevent data inconsistency\\n> issues\\\" because I thought maybe is already too late because any\\n> unwanted origin data is already copied during the initial sync. I\\n> think the user can do it try to clean up any problems caused before\\n> things become even worse (??)\\n>\\n> ~~~\\n>\\n> You asked for my thoughts (see [1] 6b):\\n>\\n> > There is nothing much a user can do in this case. Only option would be\\n> > to take a backup before the operation and restore it and then recreate\\n> > the replication setup. I was not sure if we should document these\\n> > steps as similar inconsistent data could get created without the\\n> > origin option . Thoughts?\\n>\\n> I think those cases are a bit different:\\n>\\n> - For a normal scenario (i.e. origin=ANY) then inconsistent data just\\n> refers to problems like maybe PK violations so I think you just get\\n> what you get and have to deal with the errors...\\n>\\n> - But when the user said origin=NONE they are specifically saying they\\n> do NOT want any non-local data, yet they still might end up getting\\n> data contrary to their wishes. So I think maybe there ought to be\\n> something documented to warn about this potentially happening. My\\n> first impression is that all this seems like a kind of terrible\\n> problem because if it happens then cleanup could be difficult - if\\n> that subscriber in turn was also a publisher then this bad data might\\n> have propagated all over the place already! Anyway, I guess all this\\n> could be mentioned in some (new ?) section of the\\n> logical-replication.sgml file (i.e. patch 0002).\\n>\\n> IDEA: I was wondering if the documentation should recommend (or maybe\\n> we can even enforce this) that when using origin=NONE the user should\\n> always create the subscription with enabled=FALSE. That way if there\\n> are some warnings about potential bad origin data then there it might\\n> be possible to take some action or change their mind *before* any\\n> unwanted data pollutes everything.\\n>\\n> ======\\n>\\n> 3. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> + if (count == 0)\\n> + tablenames = makeStringInfo();\\n> + else\\n> + appendStringInfoString(tablenames, \\\", \\\");\\n> +\\n> + appendStringInfo(tablenames, \\\"%s.%s\\\", nspname, relname);\\n> + count++;\\n>\\n> I think the quotes are not correct - each separate table name should be quoted.\\n>\\n> ~~~\\n>\\n> 4.\\n>\\n> + /*\\n> + * There might be many tables present in this case, we will display a\\n> + * maximum of 100 tables followed by \\\"...\\\" to indicate there might be\\n> + * more tables.\\n> + */\\n> + if (count == 100)\\n> + {\\n> + appendStringInfoString(tablenames, \\\", ...\\\");\\n> + break;\\n> + }\\n>\\n> 4a.\\n> IMO this code should be part of the previous if/else\\n>\\n> e.g.\\n> if (count == 0) ... makeStringInfo()\\n> else if (count > 100) ... append ellipsis \\\"...\\\" and break out of the loop\\n> else ... append table name to the message\\n>\\n> Doing it in this way means \\\"...\\\" definitely means there are more than\\n> 100 bad tables.\\n>\\n> ~\\n>\\n> 4b.\\n> Changing like above (#4a) simplifies the comment because it removes\\n> doubts like \\\"might be…\\\" since you already know you found the 101st\\n> table.\\n>\\n> SUGGESTION (comment)\\n> The message displays a maximum of 100 tables having potentially\\n> non-local data. Any more than that will be indicated by \\\"...\\\".\\n>\\n> ~\\n>\\n\\nAre we using this style of an error message anywhere else in the code?\\nIf not, then I think we should avoid it here as well as it appears a\\nbit adhoc to me in the sense that why restrict at 100 tables. Can we\\ninstead think of displaying the publications list in the message? So\\nerrdetail would be something like: Subscribed publications \\\\\\\"%s\\\\\\\" has\\nbeen subscribed from some other publisher. Then we can probably give a\\nSQL query for a user to find tables that may data from multiple\\norigins especially if that is not complicated. Also, then we can\\nchange the function name to check_publications_origin().\\n\\nFew other minor comments on v43-0001*\\n1.\\n+ ereport(WARNING,\\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n+ errmsg(\\\"subscription %s requested origin=NONE but might copy data\\nthat had a different origin.\\\",\\n+ subname),\\n\\nIn error message, we don't use full stop at the end.\\n\\n2.\\n+ ereport(WARNING,\\n+ errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n+ errmsg(\\\"subscription %s requested origin=NONE but might copy data\\nthat had a different origin.\\\",\\n+ subname),\\n+ errdetail(\\\"Subscribed table(s) \\\\\\\"%s\\\\\\\" has been subscribed from some\\nother publisher.\\\",\\n+ tablenames->data),\\n\\nIt is better to use errdetail_plural here as there could be one or\\nmore objects in this message?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 1 Sep 2022 09:19:20 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, 31 Aug 2022 at 11:35, Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for v43-0001.\\n>\\n> ======\\n>\\n> 1. Commit message\\n>\\n> The initial copy phase has no way to know the origin of the row data,\\n> so if 'copy_data = true' in the step 4 below, log a warning to notify user\\n> that potentially non-local data might have been copied.\\n>\\n> 1a.\\n> \\\"in the step 4\\\" -> \\\"in step 4\\\"\\n\\nModified\\n\\n> ~\\n>\\n> 1b.\\n> \\\"notify user\\\" -> \\\"notify the user\\\"\\n\\nModified\\n\\n> ======\\n>\\n> 2. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> + <para>\\n> + If the subscription is created with <literal>origin = none</literal> and\\n> + <literal>copy_data = true</literal>, it will check if the publisher has\\n> + subscribed to the same table from other publishers and, if so, log a\\n> + warning to notify the user to check the publisher tables. Before continuing\\n> + with other operations the user should check that publisher tables did not\\n> + have data with different origins to prevent data inconsistency issues on the\\n> + subscriber.\\n> + </para>\\n>\\n> I am not sure about that wording saying \\\"to prevent data inconsistency\\n> issues\\\" because I thought maybe is already too late because any\\n> unwanted origin data is already copied during the initial sync. I\\n> think the user can do it try to clean up any problems caused before\\n> things become even worse (??)\\n>\\n> ~~~\\n>\\n> You asked for my thoughts (see [1] 6b):\\n>\\n> > There is nothing much a user can do in this case. Only option would be\\n> > to take a backup before the operation and restore it and then recreate\\n> > the replication setup. I was not sure if we should document these\\n> > steps as similar inconsistent data could get created without the\\n> > origin option . Thoughts?\\n>\\n> I think those cases are a bit different:\\n>\\n> - For a normal scenario (i.e. origin=ANY) then inconsistent data just\\n> refers to problems like maybe PK violations so I think you just get\\n> what you get and have to deal with the errors...\\n>\\n> - But when the user said origin=NONE they are specifically saying they\\n> do NOT want any non-local data, yet they still might end up getting\\n> data contrary to their wishes. So I think maybe there ought to be\\n> something documented to warn about this potentially happening. My\\n> first impression is that all this seems like a kind of terrible\\n> problem because if it happens then cleanup could be difficult - if\\n> that subscriber in turn was also a publisher then this bad data might\\n> have propagated all over the place already! Anyway, I guess all this\\n> could be mentioned in some (new ?) section of the\\n> logical-replication.sgml file (i.e. patch 0002).\\n>\\n> IDEA: I was wondering if the documentation should recommend (or maybe\\n> we can even enforce this) that when using origin=NONE the user should\\n> always create the subscription with enabled=FALSE. That way if there\\n> are some warnings about potential bad origin data then there it might\\n> be possible to take some action or change their mind *before* any\\n> unwanted data pollutes everything.\\n\\nUpdated document based on your suggestion.\\n\\n> ======\\n>\\n> 3. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n>\\n> + if (count == 0)\\n> + tablenames = makeStringInfo();\\n> + else\\n> + appendStringInfoString(tablenames, \\\", \\\");\\n> +\\n> + appendStringInfo(tablenames, \\\"%s.%s\\\", nspname, relname);\\n> + count++;\\n>\\n> I think the quotes are not correct - each separate table name should be quoted.\\n\\nWe will not be displaying tables anymore, the comment does not apply anymore\\n\\n> ~~~\\n>\\n> 4.\\n>\\n> + /*\\n> + * There might be many tables present in this case, we will display a\\n> + * maximum of 100 tables followed by \\\"...\\\" to indicate there might be\\n> + * more tables.\\n> + */\\n> + if (count == 100)\\n> + {\\n> + appendStringInfoString(tablenames, \\\", ...\\\");\\n> + break;\\n> + }\\n>\\n> 4a.\\n> IMO this code should be part of the previous if/else\\n>\\n> e.g.\\n> if (count == 0) ... makeStringInfo()\\n> else if (count > 100) ... append ellipsis \\\"...\\\" and break out of the loop\\n> else ... append table name to the message\\n>\\n> Doing it in this way means \\\"...\\\" definitely means there are more than\\n> 100 bad tables.\\n\\nThis code is removed, the comment is no more applicable.\\n\\n> ~\\n>\\n> 4b.\\n> Changing like above (#4a) simplifies the comment because it removes\\n> doubts like \\\"might be…\\\" since you already know you found the 101st\\n> table.\\n>\\n> SUGGESTION (comment)\\n> The message displays a maximum of 100 tables having potentially\\n> non-local data. Any more than that will be indicated by \\\"...\\\".\\n\\nThis code is removed, the comment is no more applicable.\\n\\n> ~\\n>\\n> 4c.\\n> It still seems a bit disconcerting to me that there can be cases where\\n> bad data was possibly copied but there is no record of what tables\\n> have been polluted. Perhaps when the maximum bad tables are exceeded\\n> then there can be some additional message to tell users what SQL they\\n> can use to discover what all the other bad tables were.\\n\\nWe will not be displaying the tablename anymore, we will be displaying\\nthe publication names which has been subscribed from other\\npublications.\\n\\n> ~~~\\n>\\n> 5.\\n>\\n> + * XXX: For simplicity, we don't check whether the table has any data\\n> + * or not. If the table doesn't have any data then we don't need to\\n> + * distinguish between data having origin and data not having origin so\\n> + * we can avoid throwing a warning in that case.\\n>\\n> \\\"throwing a warning\\\" -> \\\"logging a warning\\\"\\n\\nModified\\n\\n> ~~~\\n>\\n> 6.\\n>\\n> + errdetail(\\\"Subscribed table(s) \\\\\\\"%s\\\\\\\" has been subscribed from some\\n> other publisher.\\\",\\n> + tablenames->data),\\n>\\n> I don't think the table name list should be quoted in the errdetail,\\n> because each of the individual table names should have been quoted.\\n> See previous review comment #3.\\n\\nModified\\n\\n> ======\\n>\\n> 7. src/test/subscription/t/030_origin.pl\\n>\\n> +###############################################################################\\n> +# Specify origin as 'none' which indicates that the publisher should only\\n> +# replicate the changes that are generated locally from node_B, but in\\n> +# this case since the node_B is also subscribing data from node_A, node_B can\\n> +# have remotely originated data from node_A. We log a warning, in this case,\\n> +# to draw attention to there being possible remote data.\\n> +###############################################################################\\n>\\n> This comment wording makes it seem like this is using different\\n> subscription parameters from the other tests that preceded it, but in\\n> fact the prior tests also been using origin=none. (??)\\n>\\n> Maybe it just needs a trivial rewording.\\n>\\n> SUGGESTION\\n> Specifying origin=NONE indicates that...\\n\\nModified\\n\\n> ~~~\\n>\\n> 8.\\n>\\n> +like(\\n> + $stderr,\\n> + qr/WARNING: ( [A-Z0-9]+:)? subscription tap_sub_a_b_2 requested\\n> origin=NONE but might copy data that had a different origin/,\\n> + \\\"Create subscription with origin = none and copy_data when the\\n> publisher has subscribed same table\\\"\\n> +);\\n>\\n> Is it possible for these checks of the log file to get the errdetail\\n> message too? Then the test will be more readable for the expected\\n> result and you will also have the bonus of testing the table-name list\\n> in the warning more thoroughly.\\n\\nI did not see any tap test doing this, I did not want to make some\\nchange which would fail in buildfarm machines. I have not done any\\nchange for this.\\n\\nThanks for the comments, the attached v44 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Fri, 2 Sep 2022 00:09:11 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, 29 Aug 2022 at 16:35, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Aug 24, 2022 at 7:27 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Since there was no objections to change it to throw a warning, I have\\n> > made the changes for the same.\\n> >\\n>\\n> Review comments for v42-0001*\\n> ==========================\\n> 1. Can we improve the query in check_pub_table_subscribed() so that it\\n> doesn't fetch any table that is already part of the subscription on\\n> the subscriber? This may be better than what the patch is doing which\\n> is to first fetch such information and then skip it. If forming this\\n> query turns out to be too complex then we can retain your method as\\n> well but I feel it is worth trying to optimize the query used in the\\n> patch.\\n\\nI'm analysing this, I will post a reply for this soon.\\n\\n> 2. I thought that it may be better to fetch all the tables that have\\n> the possibility to have data from more than one origin but maybe the\\n> list will be too long especially for FOR ALL TABLE types of cases. Is\\n> this the reason you have decided to give a WARNING as soon as you see\\n> any such table? If so, probably adding a comment for it can be good.\\n\\nYes that was the reason, now I have changed it to display the\\npublications based on your recent comment whose count will be\\nsignificantly very much lesser compared to the number of tables.\\n\\n> 3.\\n> + $node_publisher->wait_for_catchup($sub_name);\\n> +\\n> + # also wait for initial table sync to finish\\n> + $node_subscriber->poll_query_until('postgres', $synced_query)\\n> + or die \\\"Timed out while waiting for subscriber to synchronize data\\\";\\n> ....\\n> ....\\n> ....\\n> +$node_B->safe_psql(\\n> + 'postgres', \\\"\\n> + ALTER SUBSCRIPTION $subname_BA REFRESH PUBLICATION\\\");\\n> +\\n> +$node_B->poll_query_until('postgres', $synced_query)\\n> + or die \\\"Timed out while waiting for subscriber to synchronize data\\\";\\n>\\n> You can replace this and any similar code in the patch with a method\\n> used in commit 0c20dd33db.\\n\\nModified\\n\\n> 4.\\n> +###############################################################################\\n> +# Join 3rd node (node_C) to the existing 2 nodes(node_A & node_B) bidirectional\\n> +# replication setup when the existing nodes (node_A & node_B) has pre-existing\\n> +# data and the new node (node_C) does not have any data.\\n> +###############################################################################\\n> +$result = $node_A->safe_psql('postgres', \\\"SELECT * FROM tab ORDER BY 1;\\\");\\n> +is($result, qq(), 'Check existing data');\\n> +\\n> +$result = $node_B->safe_psql('postgres', \\\"SELECT * FROM tab ORDER BY 1;\\\");\\n> +is($result, qq(), 'Check existing data');\\n> +\\n> +$result = $node_C->safe_psql('postgres', \\\"SELECT * FROM tab ORDER BY 1;\\\");\\n> +is($result, qq(), 'Check existing data');\\n>\\n> The comments say that for this test, two of the nodes have some\\n> pre-existing data but I don't see the same in the test results. The\\n> test following this one will also have a similar effect. BTW, I am not\\n> sure all the new three node tests added by this patch are required\\n> because I don't see if they have additional code coverage or test\\n> anything which is not tested without those. This has doubled the\\n> amount of test timing for this test file which is okay if these tests\\n> make any useful contribution but otherwise, it may be better to remove\\n> these. Am, I missing something?\\n\\nEarlier we felt with the origin patch we could support joining of\\nnodes to existing n-wary replication setup in various scenarios, I had\\nadded the tests for the same scenarios. But as the patch evolved, I\\ncould understand that this patch cannot handle the add node scenarios.\\nI have removed these tests.\\n\\nThanks for the comments, the v44 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm0NRJ1O1cYcZD%3Df7NgynozFprb7zpJSayFN5rcaS44G6Q%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Sep 2022 00:14:09 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, 31 Aug 2022 at 11:45, Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are some review comments for patch v43-0002:\\n>\\n> ======\\n>\\n> 1. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> @@ -403,7 +403,9 @@ CREATE SUBSCRIPTION <replaceable\\n> class=\\\"parameter\\\">subscription_name</replaceabl\\n> warning to notify the user to check the publisher tables. Before continuing\\n> with other operations the user should check that publisher tables did not\\n> have data with different origins to prevent data inconsistency issues on the\\n> - subscriber.\\n> + subscriber. Refer to <xref linkend=\\\"replication-between-primaries\\\"/> for\\n> + how <literal>copy_data</literal> and <literal>origin</literal> can be used\\n> + to set up replication between primaries.\\n> </para>\\n>\\n> Regarding my earlier v43-0001 review (see [1] comment #2) perhaps\\n> another pg docs section should be added in the\\n> logical-replication.sgml (e.g. \\\"Specifying origins during CREATE\\n> SUBSCRIPTION\\\"), so then this Notes text also should have more added to\\n> it.\\n>\\n> SUGGESTION\\n> Refer to <XXX_REF> for details about potential initialization\\n> inconsistency warnings using origin=NONE.\\n> Refer to <YYY_REF> for how copy_data and origin can be used to set up\\n> replication between primaries.\\n\\nI have moved all these contents to a separate section in the\\nlogical-replication page. I have referred to this link from the\\ndocumentation of origin and copy_data parameter. I have also referred\\nto \\\"setting up replication between primaries\\\" in the newly added\\nsection. Since this new section is referred to from other places, I\\nfelt we need not provide a link from create_subscription notes. The\\nchanges for the same are available at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm0NRJ1O1cYcZD%3Df7NgynozFprb7zpJSayFN5rcaS44G6Q%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Sep 2022 00:18:47 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, 1 Sept 2022 at 08:01, shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Wed, Aug 31, 2022 1:06 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > The attached v43 patch has the changes for the same.\\n> >\\n>\\n> Thanks for updating the patch.\\n>\\n> Here is a comment on the 0001 patch.\\n>\\n> + if (!isnewtable)\\n> + {\\n> + pfree(nspname);\\n> + pfree(relname);\\n> + continue;\\n> + }\\n>\\n> If it is a new table, in which case it would log a warning, should we also call\\n> pfree()?\\n\\nYes it should be added, I have fixed the same in the v44 patch attached at [1].\\n[1] - https://www.postgresql.org/message-id/CALDaNm0NRJ1O1cYcZD%3Df7NgynozFprb7zpJSayFN5rcaS44G6Q%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Sep 2022 00:19:54 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, 1 Sept 2022 at 09:19, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Aug 31, 2022 at 11:35 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Here are my review comments for v43-0001.\\n> >\\n> > ======\\n> >\\n> > 1. Commit message\\n> >\\n> > The initial copy phase has no way to know the origin of the row data,\\n> > so if 'copy_data = true' in the step 4 below, log a warning to notify user\\n> > that potentially non-local data might have been copied.\\n> >\\n> > 1a.\\n> > \\\"in the step 4\\\" -> \\\"in step 4\\\"\\n> >\\n> > ~\\n> >\\n> > 1b.\\n> > \\\"notify user\\\" -> \\\"notify the user\\\"\\n> >\\n> > ======\\n> >\\n> > 2. doc/src/sgml/ref/create_subscription.sgml\\n> >\\n> > + <para>\\n> > + If the subscription is created with <literal>origin = none</literal> and\\n> > + <literal>copy_data = true</literal>, it will check if the publisher has\\n> > + subscribed to the same table from other publishers and, if so, log a\\n> > + warning to notify the user to check the publisher tables. Before continuing\\n> > + with other operations the user should check that publisher tables did not\\n> > + have data with different origins to prevent data inconsistency issues on the\\n> > + subscriber.\\n> > + </para>\\n> >\\n> > I am not sure about that wording saying \\\"to prevent data inconsistency\\n> > issues\\\" because I thought maybe is already too late because any\\n> > unwanted origin data is already copied during the initial sync. I\\n> > think the user can do it try to clean up any problems caused before\\n> > things become even worse (??)\\n> >\\n> > ~~~\\n> >\\n> > You asked for my thoughts (see [1] 6b):\\n> >\\n> > > There is nothing much a user can do in this case. Only option would be\\n> > > to take a backup before the operation and restore it and then recreate\\n> > > the replication setup. I was not sure if we should document these\\n> > > steps as similar inconsistent data could get created without the\\n> > > origin option . Thoughts?\\n> >\\n> > I think those cases are a bit different:\\n> >\\n> > - For a normal scenario (i.e. origin=ANY) then inconsistent data just\\n> > refers to problems like maybe PK violations so I think you just get\\n> > what you get and have to deal with the errors...\\n> >\\n> > - But when the user said origin=NONE they are specifically saying they\\n> > do NOT want any non-local data, yet they still might end up getting\\n> > data contrary to their wishes. So I think maybe there ought to be\\n> > something documented to warn about this potentially happening. My\\n> > first impression is that all this seems like a kind of terrible\\n> > problem because if it happens then cleanup could be difficult - if\\n> > that subscriber in turn was also a publisher then this bad data might\\n> > have propagated all over the place already! Anyway, I guess all this\\n> > could be mentioned in some (new ?) section of the\\n> > logical-replication.sgml file (i.e. patch 0002).\\n> >\\n> > IDEA: I was wondering if the documentation should recommend (or maybe\\n> > we can even enforce this) that when using origin=NONE the user should\\n> > always create the subscription with enabled=FALSE. That way if there\\n> > are some warnings about potential bad origin data then there it might\\n> > be possible to take some action or change their mind *before* any\\n> > unwanted data pollutes everything.\\n> >\\n> > ======\\n> >\\n> > 3. src/backend/commands/subscriptioncmds.c - check_pub_table_subscribed\\n> >\\n> > + if (count == 0)\\n> > + tablenames = makeStringInfo();\\n> > + else\\n> > + appendStringInfoString(tablenames, \\\", \\\");\\n> > +\\n> > + appendStringInfo(tablenames, \\\"%s.%s\\\", nspname, relname);\\n> > + count++;\\n> >\\n> > I think the quotes are not correct - each separate table name should be quoted.\\n> >\\n> > ~~~\\n> >\\n> > 4.\\n> >\\n> > + /*\\n> > + * There might be many tables present in this case, we will display a\\n> > + * maximum of 100 tables followed by \\\"...\\\" to indicate there might be\\n> > + * more tables.\\n> > + */\\n> > + if (count == 100)\\n> > + {\\n> > + appendStringInfoString(tablenames, \\\", ...\\\");\\n> > + break;\\n> > + }\\n> >\\n> > 4a.\\n> > IMO this code should be part of the previous if/else\\n> >\\n> > e.g.\\n> > if (count == 0) ... makeStringInfo()\\n> > else if (count > 100) ... append ellipsis \\\"...\\\" and break out of the loop\\n> > else ... append table name to the message\\n> >\\n> > Doing it in this way means \\\"...\\\" definitely means there are more than\\n> > 100 bad tables.\\n> >\\n> > ~\\n> >\\n> > 4b.\\n> > Changing like above (#4a) simplifies the comment because it removes\\n> > doubts like \\\"might be…\\\" since you already know you found the 101st\\n> > table.\\n> >\\n> > SUGGESTION (comment)\\n> > The message displays a maximum of 100 tables having potentially\\n> > non-local data. Any more than that will be indicated by \\\"...\\\".\\n> >\\n> > ~\\n> >\\n>\\n> Are we using this style of an error message anywhere else in the code?\\n> If not, then I think we should avoid it here as well as it appears a\\n> bit adhoc to me in the sense that why restrict at 100 tables. Can we\\n> instead think of displaying the publications list in the message? So\\n> errdetail would be something like: Subscribed publications \\\\\\\"%s\\\\\\\" has\\n> been subscribed from some other publisher. Then we can probably give a\\n> SQL query for a user to find tables that may data from multiple\\n> origins especially if that is not complicated. Also, then we can\\n> change the function name to check_publications_origin().\\n\\nModified as suggested.\\n\\n> Few other minor comments on v43-0001*\\n> 1.\\n> + ereport(WARNING,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"subscription %s requested origin=NONE but might copy data\\n> that had a different origin.\\\",\\n> + subname),\\n>\\n> In error message, we don't use full stop at the end.\\n\\nModified\\n\\n> 2.\\n> + ereport(WARNING,\\n> + errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),\\n> + errmsg(\\\"subscription %s requested origin=NONE but might copy data\\n> that had a different origin.\\\",\\n> + subname),\\n> + errdetail(\\\"Subscribed table(s) \\\\\\\"%s\\\\\\\" has been subscribed from some\\n> other publisher.\\\",\\n> + tablenames->data),\\n>\\n> It is better to use errdetail_plural here as there could be one or\\n> more objects in this message?\\n\\nModified the new errdetail message to errdetail_plural.\\n\\nThanks for the comments, the v44 version attached at [1] has the\\nchanges for the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm0NRJ1O1cYcZD%3Df7NgynozFprb7zpJSayFN5rcaS44G6Q%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Sep 2022 00:23:16 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for the latest patch v44-0001.\\n\\n======\\n\\n1. doc/src/sgml/logical-replication.sgml\\n\\n+ <sect1 id=\\\"specifying-origins-for-subscription\\\">\\n+ <title>Specifying origins for subscription</title>\\n\\nI thought the title is OK, but maybe can be better. OTOH, I am not\\nsure if my suggestions below are improvements or not. Anyway, even if\\nthe title says the same, the convention is to use uppercase words.\\n\\nSomething like:\\n\\\"Specifying Origins for Subscriptions\\\"\\n\\\"Specifying Data Origins for Subscriptions\\\"\\n\\\"Specifying Data Origins in CREATE SUBSCRIPTION\\\"\\n\\\"Subscription Data Origins\\\"\\n\\n~~~\\n\\n2.\\n\\nCurrently, this new section in this patch is only discussing the\\n*problems* users might encounter for using copy_data,origin=NONE, but\\nI think a section with a generic title about \\\"subscription origins\\\"\\nshould be able to stand alone. For example, it should give some brief\\nmention of what is the meaning/purpose of origin=ANY and origin=NONE.\\nAnd it should xref back to CREATE SUBSCRIPTION page.\\n\\nIMO all this content currently in the patch maybe belongs in a\\nsub-section of this new section (e.g. call it something like\\n\\\"Preventing Data Inconsistencies for copy_data,origin=NONE\\\")\\n\\n~~~\\n\\n3.\\n\\n+ To find which tables might potentially include non-local origins (due to\\n+ other subscriptions created on the publisher) try this SQL query by\\n+ specifying the publications in IN condition:\\n+<programlisting>\\n+SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\\n+FROM pg_publication P,\\n+ LATERAL pg_get_publication_tables(P.pubname) GPT\\n+ LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n+ pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n+WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\\n+ P.pubname IN (...);\\n+</programlisting>\\n+ </para>\\n\\n3a.\\nI think the \\\"To find...\\\" sentence should be a new paragraph.\\n\\n~\\n\\n3b.\\nInstead of saying \\\"specifying the publications in IN condition\\\" I\\nthink it might be simpler if those instructions are part of the SQL\\n\\nSUGGESTION\\n+<programlisting>\\n+# substitute <pub-names> below with your publication name(s) to be queried\\n+SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\\n+FROM pg_publication P,\\n+ LATERAL pg_get_publication_tables(P.pubname) GPT\\n+ LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n+ pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n+WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\\n+ P.pubname IN (<pub-names>);\\n+</programlisting>\\n\\n~\\n\\n3c.\\n</programlisting>\\n </para>\\n\\nI recall there was a commit or hackers post sometime earlier this year\\nthat reported it is better for these particular closing tags to be\\ndone together like </programlisting></para> because otherwise there is\\nsome case where the whitespace might not render correctly.\\nUnfortunately, I can't seem to find the evidence of that post/commit,\\nbut I am sure I saw something about this because I have been using\\nthat practice ever since I saw it.\\n\\n======\\n\\n4. doc/src/sgml/ref/alter_subscription.sgml\\n\\n+ <para>\\n+ Refer to the <xref\\nlinkend=\\\"specifying-origins-for-subscription\\\"/> about how\\n+ <literal>copy_data = true</literal> can interact with the\\n+ <literal>origin</literal> parameter.\\n+ </para>\\n\\nPreviously when this xref pointed to the \\\"Note\\\" section the sentence\\nlooked OK (it said, \\\"Refer to the Note about how...\\\") but now the word\\nis not \\\"Note\\\" anymore, (e.g. \\\"Refer to the Section 31.3 about how\\ncopy_data = true can interact with the origin parameter. \\\"), so I\\nthink this should be tweaked...\\n\\n\\\"Refer to the XXX about how...\\\" -> \\\"See XXX for details of how...\\\"\\n\\n======\\n\\n5. doc/src/sgml/ref/create_subscription.sgml\\n\\nSave as comment #4 (2 times)\\n\\n======\\n\\n6. src/backend/commands/subscriptioncmds.c\\n\\n+ if (bsearch(&relid, subrel_local_oids,\\n+ subrel_count, sizeof(Oid), oid_cmp))\\n+ isnewtable = false;\\n\\nSUGGESTION (search PG source - there are examples like this)\\nnewtable = bsearch(&relid, subrel_local_oids, subrel_count,\\nsizeof(Oid), oid_cmp);\\n\\n~~~\\n\\n7.\\n\\n+ if (list_length(publist))\\n+ {\\n\\nI think the proper way to test for non-empty List is\\n\\nif (publist != NIL) { ... }\\n\\nor simply\\n\\nif (publist) { ... }\\n\\n~~~\\n\\n8.\\n\\n+ errmsg(\\\"subscription \\\\\\\"%s\\\\\\\" requested origin=NONE but might copy\\ndata that had a different origin\\\",\\n\\nDo you think this should say \\\"requested copy_data with origin=NONE\\\",\\ninstead of just \\\"requested origin=NONE\\\"?\\n\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Sep 2022 18:21:20 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, Sep 2, 2022 at 6:21 PM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for the latest patch v44-0001.\\n>\\n...\\n>\\n> 6. src/backend/commands/subscriptioncmds.c\\n>\\n> + if (bsearch(&relid, subrel_local_oids,\\n> + subrel_count, sizeof(Oid), oid_cmp))\\n> + isnewtable = false;\\n>\\n> SUGGESTION (search PG source - there are examples like this)\\n> newtable = bsearch(&relid, subrel_local_oids, subrel_count,\\n> sizeof(Oid), oid_cmp);\\n>\\n\\nOops. Of course, I meant !bsearch\\n\\nSUGGESTION\\nnewtable = !bsearch(&relid, subrel_local_oids, subrel_count,\\nsizeof(Oid), oid_cmp);\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Sep 2022 18:26:08 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, 29 Aug 2022 at 16:35, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Aug 24, 2022 at 7:27 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Since there was no objections to change it to throw a warning, I have\\n> > made the changes for the same.\\n> >\\n>\\n> Review comments for v42-0001*\\n> ==========================\\n> 1. Can we improve the query in check_pub_table_subscribed() so that it\\n> doesn't fetch any table that is already part of the subscription on\\n> the subscriber? This may be better than what the patch is doing which\\n> is to first fetch such information and then skip it. If forming this\\n> query turns out to be too complex then we can retain your method as\\n> well but I feel it is worth trying to optimize the query used in the\\n> patch.\\n\\nI have modified the query to skip the tables as part of the query. The\\nattached v45 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Fri, 2 Sep 2022 15:59:01 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, 2 Sept 2022 at 13:51, Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for the latest patch v44-0001.\\n>\\n> ======\\n>\\n> 1. doc/src/sgml/logical-replication.sgml\\n>\\n> + <sect1 id=\\\"specifying-origins-for-subscription\\\">\\n> + <title>Specifying origins for subscription</title>\\n>\\n> I thought the title is OK, but maybe can be better. OTOH, I am not\\n> sure if my suggestions below are improvements or not. Anyway, even if\\n> the title says the same, the convention is to use uppercase words.\\n>\\n> Something like:\\n> \\\"Specifying Origins for Subscriptions\\\"\\n> \\\"Specifying Data Origins for Subscriptions\\\"\\n> \\\"Specifying Data Origins in CREATE SUBSCRIPTION\\\"\\n> \\\"Subscription Data Origins\\\"\\n\\nModified to \\\"Preventing Data Inconsistencies for copy_data,\\norigin=NONE\\\" to avoid confusions\\n\\n> ~~~\\n>\\n> 2.\\n>\\n> Currently, this new section in this patch is only discussing the\\n> *problems* users might encounter for using copy_data,origin=NONE, but\\n> I think a section with a generic title about \\\"subscription origins\\\"\\n> should be able to stand alone. For example, it should give some brief\\n> mention of what is the meaning/purpose of origin=ANY and origin=NONE.\\n> And it should xref back to CREATE SUBSCRIPTION page.\\n>\\n> IMO all this content currently in the patch maybe belongs in a\\n> sub-section of this new section (e.g. call it something like\\n> \\\"Preventing Data Inconsistencies for copy_data,origin=NONE\\\")\\n\\nI wanted to just document the inconsistency issue when copy_data and\\norigin is used. I felt the origin information was already present in\\ncreate subscription page. I have not documented the origin again here.\\nI have renamed the section to \\\"Preventing Data Inconsistencies for\\ncopy_data, origin=NONE\\\" to avoid any confusions.\\n\\n> ~~~\\n>\\n> 3.\\n>\\n> + To find which tables might potentially include non-local origins (due to\\n> + other subscriptions created on the publisher) try this SQL query by\\n> + specifying the publications in IN condition:\\n> +<programlisting>\\n> +SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\\n> +FROM pg_publication P,\\n> + LATERAL pg_get_publication_tables(P.pubname) GPT\\n> + LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n> + pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n> +WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\\n> + P.pubname IN (...);\\n> +</programlisting>\\n> + </para>\\n>\\n> 3a.\\n> I think the \\\"To find...\\\" sentence should be a new paragraph.\\n\\nmodified\\n\\n> ~\\n>\\n> 3b.\\n> Instead of saying \\\"specifying the publications in IN condition\\\" I\\n> think it might be simpler if those instructions are part of the SQL\\n>\\n> SUGGESTION\\n> +<programlisting>\\n> +# substitute <pub-names> below with your publication name(s) to be queried\\n> +SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\\n> +FROM pg_publication P,\\n> + LATERAL pg_get_publication_tables(P.pubname) GPT\\n> + LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n> + pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n> +WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\\n> + P.pubname IN (<pub-names>);\\n> +</programlisting>\\n\\n modified, I could not use <pub-name> since it was considering it as a\\ntag, instead I have changed it to pub-names\\n\\n> ~\\n>\\n> 3c.\\n> </programlisting>\\n> </para>\\n>\\n> I recall there was a commit or hackers post sometime earlier this year\\n> that reported it is better for these particular closing tags to be\\n> done together like </programlisting></para> because otherwise there is\\n> some case where the whitespace might not render correctly.\\n> Unfortunately, I can't seem to find the evidence of that post/commit,\\n> but I am sure I saw something about this because I have been using\\n> that practice ever since I saw it.\\n\\nModified\\n\\n> ======\\n>\\n> 4. doc/src/sgml/ref/alter_subscription.sgml\\n>\\n> + <para>\\n> + Refer to the <xref\\n> linkend=\\\"specifying-origins-for-subscription\\\"/> about how\\n> + <literal>copy_data = true</literal> can interact with the\\n> + <literal>origin</literal> parameter.\\n> + </para>\\n>\\n> Previously when this xref pointed to the \\\"Note\\\" section the sentence\\n> looked OK (it said, \\\"Refer to the Note about how...\\\") but now the word\\n> is not \\\"Note\\\" anymore, (e.g. \\\"Refer to the Section 31.3 about how\\n> copy_data = true can interact with the origin parameter. \\\"), so I\\n> think this should be tweaked...\\n>\\n> \\\"Refer to the XXX about how...\\\" -> \\\"See XXX for details of how...\\\"\\n\\nModified\\n\\n> ======\\n>\\n> 5. doc/src/sgml/ref/create_subscription.sgml\\n>\\n> Save as comment #4 (2 times)\\n\\nModified\\n\\n> ======\\n>\\n> 6. src/backend/commands/subscriptioncmds.c\\n>\\n> + if (bsearch(&relid, subrel_local_oids,\\n> + subrel_count, sizeof(Oid), oid_cmp))\\n> + isnewtable = false;\\n>\\n> SUGGESTION (search PG source - there are examples like this)\\n> newtable = bsearch(&relid, subrel_local_oids, subrel_count,\\n> sizeof(Oid), oid_cmp);\\n\\nThis code has been removed as part of another comment, the comment no\\nmore applies.\\n\\n> ~~~\\n>\\n> 7.\\n>\\n> + if (list_length(publist))\\n> + {\\n>\\n> I think the proper way to test for non-empty List is\\n>\\n> if (publist != NIL) { ... }\\n>\\n> or simply\\n>\\n> if (publist) { ... }\\n\\nModified\\n\\n> ~~~\\n>\\n> 8.\\n>\\n> + errmsg(\\\"subscription \\\\\\\"%s\\\\\\\" requested origin=NONE but might copy\\n> data that had a different origin\\\",\\n>\\n> Do you think this should say \\\"requested copy_data with origin=NONE\\\",\\n> instead of just \\\"requested origin=NONE\\\"?\\n\\nModified\\n\\nThanks for the comments, the v45 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm3fwW4mHGfpZfVLaHe_tSavOSPqntD5XPwO%2B_jwScrj_g%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Sep 2022 16:03:44 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Here are my review comments for v45-0001:\\n\\n======\\n\\n1. doc/src/sgml/logical-replication.sgml\\n\\n <para>\\n To find which tables might potentially include non-local origins (due to\\n other subscriptions created on the publisher) try this SQL query:\\n<programlisting>\\nSELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\\nFROM pg_publication P,\\n LATERAL pg_get_publication_tables(P.pubname) GPT\\n LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\nWHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\\n P.pubname IN (pub-names);\\n</programlisting></para>\\n\\n1a.\\nTo use \\\"<pub-names>\\\" with the <> then simply put meta characters in the SGML.\\ne.g.\\n<pub-names>\\n\\n~\\n\\n1b.\\nThe patch forgot to add the SQL \\\"#\\\" instruction as suggested by my\\nprevious comment (see [1] #3b)\\n\\n~~~\\n\\n2.\\n\\n <sect1 id=\\\"preventing-inconsistencies-for-copy_data-origin\\\">\\n <title>Preventing Data Inconsistencies for copy_data, origin=NONE</title>\\n\\nThe title is OK, but I think this should all be a <sect2> sub-section\\nof \\\"31.2 Subscription\\\"\\n\\n======\\n\\n3. src/backend/commands/subscriptioncmds.c - check_publications_origin\\n\\n+ initStringInfo(&cmd);\\n+ appendStringInfoString(&cmd,\\n+ \\\"SELECT DISTINCT P.pubname AS pubname\\\\n\\\"\\n+ \\\"FROM pg_publication P,\\\\n\\\"\\n+ \\\" LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n+ \\\" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n+ \\\" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n+ \\\"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\\\");\\n+ get_publications_str(publications, &cmd, true);\\n+ appendStringInfoChar(&cmd, ')');\\n+ get_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\\n\\n(see from get_skip_tables_str)\\n\\n+ appendStringInfoString(dest, \\\" AND C.oid NOT IN (SELECT C.oid FROM\\npg_class C JOIN pg_namespace N on N.oid = C.relnamespace where \\\");\\n\\n\\nIMO the way you are using get_skip_tables_str should be modified. I\\nwill show by way of example below.\\n- \\\"where\\\" -> \\\"WHERE\\\"\\n- put the SELECT at the caller instead of inside the function\\n- handle the \\\")\\\" at the caller\\n\\nAll this will also make the body of the 'get_skip_tables_str' function\\nmuch simpler (see the next review comments)\\n\\nSUGGESTION\\nif (subrel_count > 0)\\n{\\n/* TODO - put some explanatory comment here about skipping the tables */\\nappendStringInfo(&cmd,\\n\\\" AND C.oid NOT IN (\\\\n\\\"\\n\\\"SELECT C.oid FROM pg_class C\\\\n\\\"\\n\\\"JOIN pg_namespace N on N.oid = C.relnamespace WHERE \\\");\\nget_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\\nappendStringInf(&cmd, “)”);\\n}\\n\\n~~~\\n\\n4. src/backend/commands/subscriptioncmds.c - get_skip_tables_str\\n\\n+/*\\n+ * Add the table names that should be skipped.\\n+ */\\n\\nThis comment does not have enough detail to know really what the\\nfunction is for. Perhaps you only need to say that this is a helper\\nfunction for 'check_publications_origin' and then where it is called\\nyou can give a comment (e.g. see my review comment #3)\\n\\n~~\\n\\n5. get_skip_tables_str (body)\\n\\n5a. Variable 'count' is not a very good name; IMO just say 'i' for\\nindex, and it can be declared C99 style.\\n\\n~\\n\\n5b. Variable 'first' is not necessary - it's same as (i == 0)\\n\\n~\\n\\n5c. The whole \\\"SELECT\\\" part and the \\\")\\\" parts are more simply done at\\nthe caller (see the review comment #3)\\n\\n~\\n\\n5d.\\n\\n+ appendStringInfo(dest, \\\"(C.relname = '%s' and N.nspname = '%s')\\\",\\n+ tablename, schemaname);\\n\\nIt makes no difference but I thought it would feel more natural if the\\nSQL would compare the schema name *before* the table name.\\n\\n~\\n\\n5e.\\n\\\"and\\\" -> \\\"AND\\\"\\n\\n~\\n\\nDoing all 5a,b,c,d, and e means overall having a much simpler function body:\\n\\nSUGGESTION\\n+ for (int i = 0; i < subrel_count; i++)\\n+ {\\n+ Oid relid = subrel_local_oids[i];\\n+ char *schemaname = get_namespace_name(get_rel_namespace(relid));\\n+ char *tablename = get_rel_name(relid);\\n+\\n+ if (i > 0)\\n+ appendStringInfoString(dest, \\\" OR \\\");\\n+\\n+ appendStringInfo(dest, \\\"(N.nspname = '%s' AND C.relname = '%s')\\\",\\n+ schemaname, tablename);\\n+ }\\n\\n------\\n[1] https://www.postgresql.org/message-id/CAHut%2BPsku25%2BVjz7HiohWxc2WU07O_ZV4voFG%2BU7WzwKhUzpGQ%40mail.gmail.com\\n\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Mon, 5 Sep 2022 14:17:05 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 5, 2022 at 9:47 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for v45-0001:\\n>\\n> ======\\n>\\n> 1. doc/src/sgml/logical-replication.sgml\\n>\\n> <para>\\n> To find which tables might potentially include non-local origins (due to\\n> other subscriptions created on the publisher) try this SQL query:\\n> <programlisting>\\n> SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\\n> FROM pg_publication P,\\n> LATERAL pg_get_publication_tables(P.pubname) GPT\\n> LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n> pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n> WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\\n> P.pubname IN (pub-names);\\n> </programlisting></para>\\n>\\n> 1a.\\n> To use \\\"<pub-names>\\\" with the <> then simply put meta characters in the SGML.\\n> e.g.\\n> <pub-names>\\n>\\n> ~\\n>\\n> 1b.\\n> The patch forgot to add the SQL \\\"#\\\" instruction as suggested by my\\n> previous comment (see [1] #3b)\\n>\\n> ~~~\\n>\\n> 2.\\n>\\n> <sect1 id=\\\"preventing-inconsistencies-for-copy_data-origin\\\">\\n> <title>Preventing Data Inconsistencies for copy_data, origin=NONE</title>\\n>\\n> The title is OK, but I think this should all be a <sect2> sub-section\\n> of \\\"31.2 Subscription\\\"\\n>\\n> ======\\n>\\n\\nIt is recommended to create the subscription\\n+ using <literal>enabled=false</literal>, so that if the origin WARNING occurs\\n+ no copy has happened yet. Otherwise some corrective steps might be needed to\\n+ remove any unwanted data that got copied.\\n\\nI am not completely sure of this part of the docs as this can add\\nadditional steps for users while working on subscriptions even when\\nthe same is not required. I suggest for now we can remove this part.\\nLater based on some feedback on this feature, we can extend the docs\\nif required.\\n\\nAlso, instead of having it as a separate section, let's keep this as\\npart of create_subscription.sgml\\n\\n*\\n+ errhint(\\\"Verify that initial data copied from the publisher tables\\ndid not come from other origins. Some corrective action may be\\nnecessary.\\\"));\\n\\nThe second sentence in this message doesn't seem to be required.\\n\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 5 Sep 2022 15:10:41 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, 5 Sept 2022 at 09:47, Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> Here are my review comments for v45-0001:\\n>\\n> ======\\n>\\n> 1. doc/src/sgml/logical-replication.sgml\\n>\\n> <para>\\n> To find which tables might potentially include non-local origins (due to\\n> other subscriptions created on the publisher) try this SQL query:\\n> <programlisting>\\n> SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\\n> FROM pg_publication P,\\n> LATERAL pg_get_publication_tables(P.pubname) GPT\\n> LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n> pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n> WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\\n> P.pubname IN (pub-names);\\n> </programlisting></para>\\n>\\n> 1a.\\n> To use \\\"<pub-names>\\\" with the <> then simply put meta characters in the SGML.\\n> e.g.\\n> <pub-names>\\n\\nModified\\n\\n> ~\\n>\\n> 1b.\\n> The patch forgot to add the SQL \\\"#\\\" instruction as suggested by my\\n> previous comment (see [1] #3b)\\n\\nModified\\n\\n> ~~~\\n>\\n> 2.\\n>\\n> <sect1 id=\\\"preventing-inconsistencies-for-copy_data-origin\\\">\\n> <title>Preventing Data Inconsistencies for copy_data, origin=NONE</title>\\n>\\n> The title is OK, but I think this should all be a <sect2> sub-section\\n> of \\\"31.2 Subscription\\\"\\n\\nI have moved it to create subscription notes based on a recent comment\\nfrom Amit.\\n\\n> ======\\n>\\n> 3. src/backend/commands/subscriptioncmds.c - check_publications_origin\\n>\\n> + initStringInfo(&cmd);\\n> + appendStringInfoString(&cmd,\\n> + \\\"SELECT DISTINCT P.pubname AS pubname\\\\n\\\"\\n> + \\\"FROM pg_publication P,\\\\n\\\"\\n> + \\\" LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n> + \\\" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n> + \\\" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n> + \\\"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\\\");\\n> + get_publications_str(publications, &cmd, true);\\n> + appendStringInfoChar(&cmd, ')');\\n> + get_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\\n>\\n> (see from get_skip_tables_str)\\n>\\n> + appendStringInfoString(dest, \\\" AND C.oid NOT IN (SELECT C.oid FROM\\n> pg_class C JOIN pg_namespace N on N.oid = C.relnamespace where \\\");\\n>\\n>\\n> IMO the way you are using get_skip_tables_str should be modified. I\\n> will show by way of example below.\\n> - \\\"where\\\" -> \\\"WHERE\\\"\\n> - put the SELECT at the caller instead of inside the function\\n> - handle the \\\")\\\" at the caller\\n>\\n> All this will also make the body of the 'get_skip_tables_str' function\\n> much simpler (see the next review comments)\\n>\\n> SUGGESTION\\n> if (subrel_count > 0)\\n> {\\n> /* TODO - put some explanatory comment here about skipping the tables */\\n> appendStringInfo(&cmd,\\n> \\\" AND C.oid NOT IN (\\\\n\\\"\\n> \\\"SELECT C.oid FROM pg_class C\\\\n\\\"\\n> \\\"JOIN pg_namespace N on N.oid = C.relnamespace WHERE \\\");\\n> get_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\\n> appendStringInf(&cmd, “)”);\\n> }\\n\\nModified\\n\\n> ~~~\\n>\\n> 4. src/backend/commands/subscriptioncmds.c - get_skip_tables_str\\n>\\n> +/*\\n> + * Add the table names that should be skipped.\\n> + */\\n>\\n> This comment does not have enough detail to know really what the\\n> function is for. Perhaps you only need to say that this is a helper\\n> function for 'check_publications_origin' and then where it is called\\n> you can give a comment (e.g. see my review comment #3)\\n\\nModified\\n\\n> ~~\\n>\\n> 5. get_skip_tables_str (body)\\n>\\n> 5a. Variable 'count' is not a very good name; IMO just say 'i' for\\n> index, and it can be declared C99 style.\\n\\nModified\\n\\n> ~\\n>\\n> 5b. Variable 'first' is not necessary - it's same as (i == 0)\\n\\nModified\\n\\n> ~\\n>\\n> 5c. The whole \\\"SELECT\\\" part and the \\\")\\\" parts are more simply done at\\n> the caller (see the review comment #3)\\n\\nModified\\n\\n> ~\\n>\\n> 5d.\\n>\\n> + appendStringInfo(dest, \\\"(C.relname = '%s' and N.nspname = '%s')\\\",\\n> + tablename, schemaname);\\n>\\n> It makes no difference but I thought it would feel more natural if the\\n> SQL would compare the schema name *before* the table name.\\n\\nModified\\n\\n> ~\\n>\\n> 5e.\\n> \\\"and\\\" -> \\\"AND\\\"\\n\\nModified\\n\\n> ~\\n>\\n> Doing all 5a,b,c,d, and e means overall having a much simpler function body:\\n>\\n> SUGGESTION\\n> + for (int i = 0; i < subrel_count; i++)\\n> + {\\n> + Oid relid = subrel_local_oids[i];\\n> + char *schemaname = get_namespace_name(get_rel_namespace(relid));\\n> + char *tablename = get_rel_name(relid);\\n> +\\n> + if (i > 0)\\n> + appendStringInfoString(dest, \\\" OR \\\");\\n> +\\n> + appendStringInfo(dest, \\\"(N.nspname = '%s' AND C.relname = '%s')\\\",\\n> + schemaname, tablename);\\n> + }\\n\\nModified\\n\\nThanks for the comments, the attached patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Tue, 6 Sep 2022 08:43:38 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Mon, 5 Sept 2022 at 15:10, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Sep 5, 2022 at 9:47 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > Here are my review comments for v45-0001:\\n> >\\n> > ======\\n> >\\n> > 1. doc/src/sgml/logical-replication.sgml\\n> >\\n> > <para>\\n> > To find which tables might potentially include non-local origins (due to\\n> > other subscriptions created on the publisher) try this SQL query:\\n> > <programlisting>\\n> > SELECT DISTINCT N.nspname AS schemaname, C.relname AS tablename\\n> > FROM pg_publication P,\\n> > LATERAL pg_get_publication_tables(P.pubname) GPT\\n> > LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n> > pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n> > WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND\\n> > P.pubname IN (pub-names);\\n> > </programlisting></para>\\n> >\\n> > 1a.\\n> > To use \\\"<pub-names>\\\" with the <> then simply put meta characters in the SGML.\\n> > e.g.\\n> > <pub-names>\\n> >\\n> > ~\\n> >\\n> > 1b.\\n> > The patch forgot to add the SQL \\\"#\\\" instruction as suggested by my\\n> > previous comment (see [1] #3b)\\n> >\\n> > ~~~\\n> >\\n> > 2.\\n> >\\n> > <sect1 id=\\\"preventing-inconsistencies-for-copy_data-origin\\\">\\n> > <title>Preventing Data Inconsistencies for copy_data, origin=NONE</title>\\n> >\\n> > The title is OK, but I think this should all be a <sect2> sub-section\\n> > of \\\"31.2 Subscription\\\"\\n> >\\n> > ======\\n> >\\n>\\n> It is recommended to create the subscription\\n> + using <literal>enabled=false</literal>, so that if the origin WARNING occurs\\n> + no copy has happened yet. Otherwise some corrective steps might be needed to\\n> + remove any unwanted data that got copied.\\n>\\n> I am not completely sure of this part of the docs as this can add\\n> additional steps for users while working on subscriptions even when\\n> the same is not required. I suggest for now we can remove this part.\\n> Later based on some feedback on this feature, we can extend the docs\\n> if required.\\n\\nModified\\n\\n> Also, instead of having it as a separate section, let's keep this as\\n> part of create_subscription.sgml\\n\\nModified\\n\\n> *\\n> + errhint(\\\"Verify that initial data copied from the publisher tables\\n> did not come from other origins. Some corrective action may be\\n> necessary.\\\"));\\n>\\n> The second sentence in this message doesn't seem to be required.\\n\\nModified\\n\\nThanks for the comments, the v46 patch at [1] has the changes for the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm3TTGdCCkeDsN8hqtF_2z-8%2B%3D3tc9Gh5xOKAQ_BRMCkMA%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Tue, 6 Sep 2022 08:45:45 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 6, 2022 11:14 AM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> Thanks for the comments, the attached patch has the changes for the same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch. Here are some comments.\\r\\n\\r\\n1.\\r\\n+\\tif (subrel_count)\\r\\n+\\t{\\r\\n+\\t\\t/*\\r\\n+\\t\\t * In case of ALTER SUBSCRIPTION ... REFRESH, subrel_local_oids\\r\\n+\\t\\t * contains the list of relation oids that are already present on the\\r\\n+\\t\\t * subscriber. This check should be skipped for these tables.\\r\\n+\\t\\t */\\r\\n+\\t\\tappendStringInfo(&cmd, \\\"AND C.oid NOT IN (\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"SELECT C.oid \\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"FROM pg_class C\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\" JOIN pg_namespace N ON N.oid = C.relnamespace\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"WHERE \\\");\\r\\n+\\t\\tget_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\\r\\n+\\t\\tappendStringInfo(&cmd, \\\")\\\");\\r\\n+\\t}\\r\\n\\r\\nI think we can skip the tables without using a subquery. See the SQL below.\\r\\nWould it be better?\\r\\n\\r\\nSELECT DISTINCT P.pubname AS pubname\\r\\nFROM pg_publication P,\\r\\n LATERAL pg_get_publication_tables(P.pubname) GPT\\r\\n LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\r\\n pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\r\\nWHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN ('p1', 'p3')\\r\\nAND NOT (N.nspname = 'public' AND C.relname = 'tbl')\\r\\nAND NOT (N.nspname = 'public' AND C.relname = 't1');\\r\\n\\r\\n\\r\\n2.\\r\\n+ When using a subscription parameter combination of\\r\\n+ <literal>copy_data=true</literal> and <literal>origin=NONE</literal>, the\\r\\n\\r\\nCould we use \\\"copy_data = true\\\" and \\\"origin = NONE\\\" (add two spaces around the\\r\\nequals sign)? I think it looks clearer that way. And it is consistent with other\\r\\nplaces in this file.\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Tue, 6 Sep 2022 04:00:47 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tues, 6 Sept 2022 at 11:14, vignesh C <vignesh21@gmail.com> wrote:\\r\\n> Thanks for the comments, the attached patch has the changes for the same.\\r\\n\\r\\nThanks for updating the patch.\\r\\n\\r\\nHere is one comment for 0001 patch.\\r\\n1. The query in function check_publications_origin.\\r\\n+\\tappendStringInfoString(&cmd,\\r\\n+\\t\\t\\t\\t\\t\\t \\\"SELECT DISTINCT P.pubname AS pubname\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"FROM pg_publication P,\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"\\t LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"\\t LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"\\t pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\r\\n+\\t\\t\\t\\t\\t\\t \\\"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\\\");\\r\\n\\r\\nSince I found that we only use \\\"PS.srrelid\\\" in the WHERE statement by\\r\\nspecifying \\\"PS.srrelid IS NOT NULL\\\", could we just use \\\"[INNER] JOIN\\\" to join\\r\\nthe table pg_subscription_rel?\\r\\n\\r\\nRegards,\\r\\nWang wei\\r\\n\",\n \"msg_date\": \"Tue, 6 Sep 2022 04:01:04 +0000\",\n \"msg_from\": \"\\\"wangw.fnst@fujitsu.com\\\" <wangw.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Thanks for addressing my previous review comments. I have no more\\ncomments for v46*.\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Tue, 6 Sep 2022 15:11:16 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 6, 2022 at 9:31 AM wangw.fnst@fujitsu.com\\n<wangw.fnst@fujitsu.com> wrote:\\n>\\n> On Tues, 6 Sept 2022 at 11:14, vignesh C <vignesh21@gmail.com> wrote:\\n> > Thanks for the comments, the attached patch has the changes for the same.\\n>\\n> Thanks for updating the patch.\\n>\\n> Here is one comment for 0001 patch.\\n> 1. The query in function check_publications_origin.\\n> + appendStringInfoString(&cmd,\\n> + \\\"SELECT DISTINCT P.pubname AS pubname\\\\n\\\"\\n> + \\\"FROM pg_publication P,\\\\n\\\"\\n> + \\\" LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n> + \\\" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n> + \\\" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n> + \\\"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\\\");\\n>\\n> Since I found that we only use \\\"PS.srrelid\\\" in the WHERE statement by\\n> specifying \\\"PS.srrelid IS NOT NULL\\\", could we just use \\\"[INNER] JOIN\\\" to join\\n> the table pg_subscription_rel?\\n>\\n\\nI also think we can use INNER JOIN here but maybe there is a reason\\nwhy that is not used in the first place. If we change it in the code\\nthen also let's change the same in the docs section as well.\\n\\nFew minor points:\\n===============\\n1.\\nThis scenario is detected and a WARNING is logged to the\\n+ user, but the warning is only an indication of a potential problem; it is\\n+ the user responsibility to make the necessary checks to ensure the copied\\n+ data origins are really as wanted or not.\\n+ </para>\\n\\n/user/user's\\n\\n2. How about a commit message like:\\nRaise a warning if there is a possibility of data from multiple origins.\\n\\nThis commit raises a warning message for a combination of options\\n('copy_data = true' and 'origin = none') during CREATE/ALTER subscription\\noperations if the publication tables were also replicated from other\\npublishers.\\n\\nDuring replication, we can skip the data from other origins as we have that\\ninformation in WAL but that is not possible during initial sync so we raise\\na warning if there is such a possibility.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 6 Sep 2022 15:25:10 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, 6 Sept 2022 at 09:31, shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Tue, Sep 6, 2022 11:14 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached patch has the changes for the same.\\n> >\\n>\\n> Thanks for updating the patch. Here are some comments.\\n>\\n> 1.\\n> + if (subrel_count)\\n> + {\\n> + /*\\n> + * In case of ALTER SUBSCRIPTION ... REFRESH, subrel_local_oids\\n> + * contains the list of relation oids that are already present on the\\n> + * subscriber. This check should be skipped for these tables.\\n> + */\\n> + appendStringInfo(&cmd, \\\"AND C.oid NOT IN (\\\\n\\\"\\n> + \\\"SELECT C.oid \\\\n\\\"\\n> + \\\"FROM pg_class C\\\\n\\\"\\n> + \\\" JOIN pg_namespace N ON N.oid = C.relnamespace\\\\n\\\"\\n> + \\\"WHERE \\\");\\n> + get_skip_tables_str(subrel_local_oids, subrel_count, &cmd);\\n> + appendStringInfo(&cmd, \\\")\\\");\\n> + }\\n>\\n> I think we can skip the tables without using a subquery. See the SQL below.\\n> Would it be better?\\n>\\n> SELECT DISTINCT P.pubname AS pubname\\n> FROM pg_publication P,\\n> LATERAL pg_get_publication_tables(P.pubname) GPT\\n> LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\n> pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\n> WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN ('p1', 'p3')\\n> AND NOT (N.nspname = 'public' AND C.relname = 'tbl')\\n> AND NOT (N.nspname = 'public' AND C.relname = 't1');\\n\\nModified\\n\\n> 2.\\n> + When using a subscription parameter combination of\\n> + <literal>copy_data=true</literal> and <literal>origin=NONE</literal>, the\\n>\\n> Could we use \\\"copy_data = true\\\" and \\\"origin = NONE\\\" (add two spaces around the\\n> equals sign)? I think it looks clearer that way. And it is consistent with other\\n> places in this file.\\n\\nModified\\n\\nThanks for the comments, the attached v47 patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 7 Sep 2022 09:52:53 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, 6 Sept 2022 at 09:31, wangw.fnst@fujitsu.com\\n<wangw.fnst@fujitsu.com> wrote:\\n>\\n> On Tues, 6 Sept 2022 at 11:14, vignesh C <vignesh21@gmail.com> wrote:\\n> > Thanks for the comments, the attached patch has the changes for the same.\\n>\\n> Thanks for updating the patch.\\n>\\n> Here is one comment for 0001 patch.\\n> 1. The query in function check_publications_origin.\\n> + appendStringInfoString(&cmd,\\n> + \\\"SELECT DISTINCT P.pubname AS pubname\\\\n\\\"\\n> + \\\"FROM pg_publication P,\\\\n\\\"\\n> + \\\" LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n> + \\\" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n> + \\\" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n> + \\\"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\\\");\\n>\\n> Since I found that we only use \\\"PS.srrelid\\\" in the WHERE statement by\\n> specifying \\\"PS.srrelid IS NOT NULL\\\", could we just use \\\"[INNER] JOIN\\\" to join\\n> the table pg_subscription_rel?\\n\\nThanks for the comment, the v47 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm33T%3D23P-GWvy3O7cT1BfHuGV8dosAw1AVLf40MPvg2bg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 7 Sep 2022 09:54:09 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, 6 Sept 2022 at 15:25, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Sep 6, 2022 at 9:31 AM wangw.fnst@fujitsu.com\\n> <wangw.fnst@fujitsu.com> wrote:\\n> >\\n> > On Tues, 6 Sept 2022 at 11:14, vignesh C <vignesh21@gmail.com> wrote:\\n> > > Thanks for the comments, the attached patch has the changes for the same.\\n> >\\n> > Thanks for updating the patch.\\n> >\\n> > Here is one comment for 0001 patch.\\n> > 1. The query in function check_publications_origin.\\n> > + appendStringInfoString(&cmd,\\n> > + \\\"SELECT DISTINCT P.pubname AS pubname\\\\n\\\"\\n> > + \\\"FROM pg_publication P,\\\\n\\\"\\n> > + \\\" LATERAL pg_get_publication_tables(P.pubname) GPT\\\\n\\\"\\n> > + \\\" LEFT JOIN pg_subscription_rel PS ON (GPT.relid = PS.srrelid),\\\\n\\\"\\n> > + \\\" pg_class C JOIN pg_namespace N ON (N.oid = C.relnamespace)\\\\n\\\"\\n> > + \\\"WHERE C.oid = GPT.relid AND PS.srrelid IS NOT NULL AND P.pubname IN (\\\");\\n> >\\n> > Since I found that we only use \\\"PS.srrelid\\\" in the WHERE statement by\\n> > specifying \\\"PS.srrelid IS NOT NULL\\\", could we just use \\\"[INNER] JOIN\\\" to join\\n> > the table pg_subscription_rel?\\n> >\\n>\\n> I also think we can use INNER JOIN here but maybe there is a reason\\n> why that is not used in the first place. If we change it in the code\\n> then also let's change the same in the docs section as well.\\n\\nModified\\n\\n> Few minor points:\\n> ===============\\n> 1.\\n> This scenario is detected and a WARNING is logged to the\\n> + user, but the warning is only an indication of a potential problem; it is\\n> + the user responsibility to make the necessary checks to ensure the copied\\n> + data origins are really as wanted or not.\\n> + </para>\\n>\\n> /user/user's\\n\\nModified\\n\\n> 2. How about a commit message like:\\n> Raise a warning if there is a possibility of data from multiple origins.\\n>\\n> This commit raises a warning message for a combination of options\\n> ('copy_data = true' and 'origin = none') during CREATE/ALTER subscription\\n> operations if the publication tables were also replicated from other\\n> publishers.\\n>\\n> During replication, we can skip the data from other origins as we have that\\n> information in WAL but that is not possible during initial sync so we raise\\n> a warning if there is such a possibility.\\n\\nYes, this looks good. Modified\\n\\nThanks for the comment, the v47 patch attached at [1] has the changes\\nfor the same.\\n[1] - https://www.postgresql.org/message-id/CALDaNm33T%3D23P-GWvy3O7cT1BfHuGV8dosAw1AVLf40MPvg2bg%40mail.gmail.com\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Wed, 7 Sep 2022 09:55:16 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 7, 2022 12:23 PM vignesh C <vignesh21@gmail.com> wrote:\\r\\n> \\r\\n> \\r\\n> Thanks for the comments, the attached v47 patch has the changes for the\\r\\n> same.\\r\\n> \\r\\n\\r\\nThanks for updating the patch.\\r\\n\\r\\nHere is a comment.\\r\\n\\r\\n+\\tfor (i = 0; i < subrel_count; i++)\\r\\n+\\t{\\r\\n+\\t\\tOid\\t\\t\\trelid = subrel_local_oids[i];\\r\\n+\\t\\tchar\\t *schemaname = get_namespace_name(get_rel_namespace(relid));\\r\\n+\\t\\tchar\\t *tablename = get_rel_name(relid);\\r\\n+\\r\\n+\\t\\tappendStringInfo(&cmd, \\\"AND NOT (N.nspname = '%s' AND C.relname = '%s')\\\\n\\\",\\r\\n+\\t\\t\\t\\t\\t\\t schemaname, tablename);\\r\\n+\\t}\\r\\n\\r\\nWould it be better to add \\\"pfree(schemaname)\\\" and \\\"pfree(tablename)\\\" after\\r\\ncalling appendStringInfo()?\\r\\n\\r\\nRegards,\\r\\nShi yu\\r\\n\",\n \"msg_date\": \"Wed, 7 Sep 2022 07:39:23 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 7, 2022 at 1:09 PM shiy.fnst@fujitsu.com\\n<shiy.fnst@fujitsu.com> wrote:\\n>\\n> On Wed, Sep 7, 2022 12:23 PM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> >\\n> > Thanks for the comments, the attached v47 patch has the changes for the\\n> > same.\\n> >\\n>\\n> Thanks for updating the patch.\\n>\\n> Here is a comment.\\n>\\n> + for (i = 0; i < subrel_count; i++)\\n> + {\\n> + Oid relid = subrel_local_oids[i];\\n> + char *schemaname = get_namespace_name(get_rel_namespace(relid));\\n> + char *tablename = get_rel_name(relid);\\n> +\\n> + appendStringInfo(&cmd, \\\"AND NOT (N.nspname = '%s' AND C.relname = '%s')\\\\n\\\",\\n> + schemaname, tablename);\\n> + }\\n>\\n> Would it be better to add \\\"pfree(schemaname)\\\" and \\\"pfree(tablename)\\\" after\\n> calling appendStringInfo()?\\n>\\n\\nNo, I don't think we need to do retail pfree of each and every\\nallocation as these allocations are made in the portal context which\\nwill be freed by the command end.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 7 Sep 2022 13:39:14 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Sep 7, 2022 at 9:53 AM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> Thanks for the comments, the attached v47 patch has the changes for the same.\\n>\\n\\nV47-0001* looks good to me apart from below minor things. I would like\\nto commit this tomorrow unless there are more comments on it.\\n\\nFew minor suggestions:\\n==================\\n1.\\n+ list_free_deep(publist);\\n+ pfree(pubnames->data);\\n+ pfree(pubnames);\\n\\nI don't think we need to free this memory as it will automatically be\\nreleased at end of the command (this memory is allocated in portal\\ncontext). I understand there is no harm in freeing it as well but\\nbetter to leave it as there doesn't appear to be any danger of leaking\\nmemory for a longer time.\\n\\n2.\\n+ res = walrcv_exec(wrconn, cmd.data, 1, tableRow);\\n+ pfree(cmd.data);\\n\\nDon't you need to free cmd as well here? I think it is better to avoid\\nfreeing it due to reasons mentioned in the previous point.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 7 Sep 2022 14:34:02 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, 7 Sept 2022 at 14:34, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Sep 7, 2022 at 9:53 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > Thanks for the comments, the attached v47 patch has the changes for the same.\\n> >\\n>\\n> V47-0001* looks good to me apart from below minor things. I would like\\n> to commit this tomorrow unless there are more comments on it.\\n>\\n> Few minor suggestions:\\n> ==================\\n> 1.\\n> + list_free_deep(publist);\\n> + pfree(pubnames->data);\\n> + pfree(pubnames);\\n>\\n> I don't think we need to free this memory as it will automatically be\\n> released at end of the command (this memory is allocated in portal\\n> context). I understand there is no harm in freeing it as well but\\n> better to leave it as there doesn't appear to be any danger of leaking\\n> memory for a longer time.\\n\\nModified\\n\\n> 2.\\n> + res = walrcv_exec(wrconn, cmd.data, 1, tableRow);\\n> + pfree(cmd.data);\\n>\\n> Don't you need to free cmd as well here? I think it is better to avoid\\n> freeing it due to reasons mentioned in the previous point.\\n\\ncmd need not be freed here as we use initStringInfo for initialization\\nand initStringInfo allocates memory for data member only.\\n\\nThe attached patch has the changes for the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Wed, 7 Sep 2022 17:10:14 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nThere is a buildfarm failure on mylodon at [1] because of the new test\\nadded. I will analyse and share the findings for the same.\\n\\n[1] - https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mylodon&dt=2022-09-08%2001%3A40%3A27\\n\\nRegards,\\nVignesh\\n\\nOn Wed, 7 Sept 2022 at 17:10, vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Wed, 7 Sept 2022 at 14:34, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Wed, Sep 7, 2022 at 9:53 AM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > Thanks for the comments, the attached v47 patch has the changes for the same.\\n> > >\\n> >\\n> > V47-0001* looks good to me apart from below minor things. I would like\\n> > to commit this tomorrow unless there are more comments on it.\\n> >\\n> > Few minor suggestions:\\n> > ==================\\n> > 1.\\n> > + list_free_deep(publist);\\n> > + pfree(pubnames->data);\\n> > + pfree(pubnames);\\n> >\\n> > I don't think we need to free this memory as it will automatically be\\n> > released at end of the command (this memory is allocated in portal\\n> > context). I understand there is no harm in freeing it as well but\\n> > better to leave it as there doesn't appear to be any danger of leaking\\n> > memory for a longer time.\\n>\\n> Modified\\n>\\n> > 2.\\n> > + res = walrcv_exec(wrconn, cmd.data, 1, tableRow);\\n> > + pfree(cmd.data);\\n> >\\n> > Don't you need to free cmd as well here? I think it is better to avoid\\n> > freeing it due to reasons mentioned in the previous point.\\n>\\n> cmd need not be freed here as we use initStringInfo for initialization\\n> and initStringInfo allocates memory for data member only.\\n>\\n> The attached patch has the changes for the same.\\n>\\n> Regards,\\n> Vignesh\\n\\n\\n\",\n \"msg_date\": \"Thu, 8 Sep 2022 07:56:45 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 8, 2022 at 7:57 AM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> There is a buildfarm failure on mylodon at [1] because of the new test\\n> added. I will analyse and share the findings for the same.\\n>\\n> [1] - https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mylodon&dt=2022-09-08%2001%3A40%3A27\\n>\\n\\nThe log is as below:\\n\\nerror running SQL: 'psql:<stdin>:1: ERROR: could not drop relation\\nmapping for subscription \\\"tap_sub_a_b_2\\\"\\nDETAIL: Table synchronization for relation \\\"tab_new\\\" is in progress\\nand is in state \\\"s\\\".\\nHINT: Use ALTER SUBSCRIPTION ... ENABLE to enable subscription if not\\nalready enabled or use DROP SUBSCRIPTION ... to drop the\\nsubscription.'\\nwhile running 'psql -XAtq -d port=50352 host=/tmp/WMoRd6ngw2\\ndbname='postgres' -f - -v ON_ERROR_STOP=1' with sql 'DROP TABLE\\ntab_new' at /mnt/resource/bf/build/mylodon/HEAD/pgsql.build/../pgsql/src/test/perl/PostgreSQL/Test/Cluster.pm\\nline 1860.\\n### Stopping node \\\"node_A\\\" using mode immediate\\n\\nThis clearly indicates the problem. We can't drop the relation till it\\nis marked ready in pg_subscription_rel and prior to dropping, the test\\njust does $node_A->wait_for_subscription_sync($node_B, $subname_AB2);.\\nThis doesn't ensure that relation is in 'ready' state as it will\\nfinish even when the relation is in 'syncdone' state.\\n\\nSo, I think if we want to drop the tables, we need to poll for 'ready'\\nstate or otherwise, anyway, this is the end of all tests and nodes\\nwon't be reused, so we can remove the clean-up in the end. Any other\\nideas?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 8 Sep 2022 08:23:00 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, 8 Sept 2022 at 08:23, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Thu, Sep 8, 2022 at 7:57 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> > There is a buildfarm failure on mylodon at [1] because of the new test\\n> > added. I will analyse and share the findings for the same.\\n> >\\n> > [1] - https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mylodon&dt=2022-09-08%2001%3A40%3A27\\n> >\\n>\\n> The log is as below:\\n>\\n> error running SQL: 'psql:<stdin>:1: ERROR: could not drop relation\\n> mapping for subscription \\\"tap_sub_a_b_2\\\"\\n> DETAIL: Table synchronization for relation \\\"tab_new\\\" is in progress\\n> and is in state \\\"s\\\".\\n> HINT: Use ALTER SUBSCRIPTION ... ENABLE to enable subscription if not\\n> already enabled or use DROP SUBSCRIPTION ... to drop the\\n> subscription.'\\n> while running 'psql -XAtq -d port=50352 host=/tmp/WMoRd6ngw2\\n> dbname='postgres' -f - -v ON_ERROR_STOP=1' with sql 'DROP TABLE\\n> tab_new' at /mnt/resource/bf/build/mylodon/HEAD/pgsql.build/../pgsql/src/test/perl/PostgreSQL/Test/Cluster.pm\\n> line 1860.\\n> ### Stopping node \\\"node_A\\\" using mode immediate\\n>\\n> This clearly indicates the problem. We can't drop the relation till it\\n> is marked ready in pg_subscription_rel and prior to dropping, the test\\n> just does $node_A->wait_for_subscription_sync($node_B, $subname_AB2);.\\n> This doesn't ensure that relation is in 'ready' state as it will\\n> finish even when the relation is in 'syncdone' state.\\n\\nI agree with the analysis, adding wait for ready state before dropping\\nthe table approach looks good to me. I will prepare a patch for the\\nsame and share it.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Thu, 8 Sep 2022 08:57:00 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, 8 Sept 2022 at 08:57, vignesh C <vignesh21@gmail.com> wrote:\\n>\\n> On Thu, 8 Sept 2022 at 08:23, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Thu, Sep 8, 2022 at 7:57 AM vignesh C <vignesh21@gmail.com> wrote:\\n> > >\\n> > > There is a buildfarm failure on mylodon at [1] because of the new test\\n> > > added. I will analyse and share the findings for the same.\\n> > >\\n> > > [1] - https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mylodon&dt=2022-09-08%2001%3A40%3A27\\n> > >\\n> >\\n> > The log is as below:\\n> >\\n> > error running SQL: 'psql:<stdin>:1: ERROR: could not drop relation\\n> > mapping for subscription \\\"tap_sub_a_b_2\\\"\\n> > DETAIL: Table synchronization for relation \\\"tab_new\\\" is in progress\\n> > and is in state \\\"s\\\".\\n> > HINT: Use ALTER SUBSCRIPTION ... ENABLE to enable subscription if not\\n> > already enabled or use DROP SUBSCRIPTION ... to drop the\\n> > subscription.'\\n> > while running 'psql -XAtq -d port=50352 host=/tmp/WMoRd6ngw2\\n> > dbname='postgres' -f - -v ON_ERROR_STOP=1' with sql 'DROP TABLE\\n> > tab_new' at /mnt/resource/bf/build/mylodon/HEAD/pgsql.build/../pgsql/src/test/perl/PostgreSQL/Test/Cluster.pm\\n> > line 1860.\\n> > ### Stopping node \\\"node_A\\\" using mode immediate\\n> >\\n> > This clearly indicates the problem. We can't drop the relation till it\\n> > is marked ready in pg_subscription_rel and prior to dropping, the test\\n> > just does $node_A->wait_for_subscription_sync($node_B, $subname_AB2);.\\n> > This doesn't ensure that relation is in 'ready' state as it will\\n> > finish even when the relation is in 'syncdone' state.\\n>\\n> I agree with the analysis, adding wait for ready state before dropping\\n> the table approach looks good to me. I will prepare a patch for the\\n> same and share it.\\n\\nThe attached patch has the changes to handle the same.\\n\\nRegards,\\nVignesh\",\n \"msg_date\": \"Thu, 8 Sep 2022 09:31:50 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\\n>\\n>\\n> The attached patch has the changes to handle the same.\\n>\\n\\nPushed. I am not completely sure whether we want the remaining\\ndocumentation patch in this thread in its current form or by modifying\\nit. Johnathan has shown some interest in it. I feel you can start a\\nseparate thread for it to see if there is any interest in the same and\\nclose the CF entry for this work.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 9 Sep 2022 11:12:02 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >\\n> >\\n> > The attached patch has the changes to handle the same.\\n> >\\n>\\n> Pushed. I am not completely sure whether we want the remaining\\n> documentation patch in this thread in its current form or by modifying\\n> it. Johnathan has shown some interest in it. I feel you can start a\\n> separate thread for it to see if there is any interest in the same and\\n> close the CF entry for this work.\\n\\nThanks for pushing the patch. I have closed this entry in commitfest.\\nI will wait for some more time and see the response regarding the\\ndocumentation patch and then start a new thread if required.\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Sep 2022 10:53:08 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On 9/12/22 1:23 AM, vignesh C wrote:\\r\\n> On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\\r\\n>>\\r\\n>> On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\\r\\n>>>\\r\\n>>>\\r\\n>>> The attached patch has the changes to handle the same.\\r\\n>>>\\r\\n>>\\r\\n>> Pushed. I am not completely sure whether we want the remaining\\r\\n>> documentation patch in this thread in its current form or by modifying\\r\\n>> it. Johnathan has shown some interest in it. I feel you can start a\\r\\n>> separate thread for it to see if there is any interest in the same and\\r\\n>> close the CF entry for this work.\\r\\n> \\r\\n> Thanks for pushing the patch. I have closed this entry in commitfest.\\r\\n> I will wait for some more time and see the response regarding the\\r\\n> documentation patch and then start a new thread if required.\\r\\n\\r\\nI've been testing this patch in advancing of working on the \\r\\ndocumentation and came across a behavior I wanted to note. Specifically, \\r\\nI am hitting a deadlock while trying to synchronous replicate between \\r\\nthe two instances at any `synchronous_commit` level above `local`.\\r\\n\\r\\nHere is my set up. I have two instances, \\\"A\\\" and \\\"B\\\".\\r\\n\\r\\nOn A and B, run:\\r\\n\\r\\n CREATE TABLE sync (id int PRIMARY KEY, info float);\\r\\n CREATE PUBLICATION sync FOR TABLE sync;\\r\\n\\r\\nOn A, run:\\r\\n\\r\\n CREATE SUBSCRIPTION sync\\r\\n CONNECTION 'connstr-to-B'\\r\\n PUBLICATION sync\\r\\n WITH (\\r\\n streaming=true, copy_data=false,\\r\\n origin=none, synchronous_commit='on');\\r\\n\\r\\nOn B, run:\\r\\n\\r\\n CREATE SUBSCRIPTION sync\\r\\n CONNECTION 'connstr-to-A'\\r\\n PUBLICATION sync\\r\\n WITH (\\r\\n streaming=true, copy_data=false,\\r\\n origin=none, synchronous_commit='on');\\r\\n\\r\\nOn A and B, run:\\r\\n\\r\\n ALTER SYSTEM SET synchronous_standby_names TO 'sync';\\r\\n SELECT pg_reload_conf();\\r\\n\\r\\nVerify on A and B that pg_stat_replication.sync_state is set to \\\"sync\\\"\\r\\n\\r\\n SELECT application_name, sync_state = 'sync' AS is_sync\\r\\n FROM pg_stat_replication\\r\\n WHERE application_name = 'sync';\\r\\n\\r\\nThe next to commands should be run simultaneously on A and B:\\r\\n\\r\\n-- run this on A\\r\\nINSERT INTO sync\\r\\nSELECT x, random() FROM generate_series(1,2000000, 2) x;\\r\\n\\r\\n-- run this on B\\r\\nINSERT INTO sync\\r\\nSELECT x, random() FROM generate_series(2,2000000, 2) x;\\r\\n\\r\\nThis consistently created the deadlock in my testing.\\r\\n\\r\\nDiscussing with Masahiko off-list, this is due to a deadlock from 4 \\r\\nprocesses: the walsenders on A and B, and the apply workers on A and B. \\r\\nThe walsenders are waiting for feedback from the apply workers, and the \\r\\napply workers are waiting for the walsenders to synchronize (I may be \\r\\noversimplifying).\\r\\n\\r\\nHe suggested I try the above example instead with `synchronous_commit` \\r\\nset to `local`. In this case, I verified that there is no more deadlock, \\r\\nbut he informed me that we would not be able to use cascading \\r\\nsynchronous replication when \\\"origin=none\\\".\\r\\n\\r\\nIf we decide that this is a documentation issue, I'd suggest we improve \\r\\nthe guidance around using `synchronous_commit`[1] on the CREATE \\r\\nSUBSCRIPTION page, as the GUC page[2] warns against using `local`:\\r\\n\\r\\n\\\"The setting local causes commits to wait for local flush to disk, but \\r\\nnot for replication. This is usually not desirable when synchronous \\r\\nreplication is in use, but is provided for completeness.\\\"\\r\\n\\r\\nThanks,\\r\\n\\r\\nJonathan\\r\\n\\r\\n[1] https://www.postgresql.org/docs/devel/sql-createsubscription.html\\r\\n[2] \\r\\nhttps://www.postgresql.org/docs/devel/runtime-config-wal.html#GUC-SYNCHRONOUS-COMMIT\",\n \"msg_date\": \"Mon, 9 Jan 2023 21:43:15 -0500\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 10, 2023 at 8:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> On 9/12/22 1:23 AM, vignesh C wrote:\\n> > On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > Thanks for pushing the patch. I have closed this entry in commitfest.\\n> > I will wait for some more time and see the response regarding the\\n> > documentation patch and then start a new thread if required.\\n>\\n> I've been testing this patch in advancing of working on the\\n> documentation and came across a behavior I wanted to note. Specifically,\\n> I am hitting a deadlock while trying to synchronous replicate between\\n> the two instances at any `synchronous_commit` level above `local`.\\n>\\n> Here is my set up. I have two instances, \\\"A\\\" and \\\"B\\\".\\n>\\n> On A and B, run:\\n>\\n> CREATE TABLE sync (id int PRIMARY KEY, info float);\\n> CREATE PUBLICATION sync FOR TABLE sync;\\n>\\n> On A, run:\\n>\\n> CREATE SUBSCRIPTION sync\\n> CONNECTION 'connstr-to-B'\\n> PUBLICATION sync\\n> WITH (\\n> streaming=true, copy_data=false,\\n> origin=none, synchronous_commit='on');\\n>\\n> On B, run:\\n>\\n> CREATE SUBSCRIPTION sync\\n> CONNECTION 'connstr-to-A'\\n> PUBLICATION sync\\n> WITH (\\n> streaming=true, copy_data=false,\\n> origin=none, synchronous_commit='on');\\n>\\n> On A and B, run:\\n>\\n> ALTER SYSTEM SET synchronous_standby_names TO 'sync';\\n> SELECT pg_reload_conf();\\n>\\n> Verify on A and B that pg_stat_replication.sync_state is set to \\\"sync\\\"\\n>\\n> SELECT application_name, sync_state = 'sync' AS is_sync\\n> FROM pg_stat_replication\\n> WHERE application_name = 'sync';\\n>\\n> The next to commands should be run simultaneously on A and B:\\n>\\n> -- run this on A\\n> INSERT INTO sync\\n> SELECT x, random() FROM generate_series(1,2000000, 2) x;\\n>\\n> -- run this on B\\n> INSERT INTO sync\\n> SELECT x, random() FROM generate_series(2,2000000, 2) x;\\n>\\n> This consistently created the deadlock in my testing.\\n>\\n> Discussing with Masahiko off-list, this is due to a deadlock from 4\\n> processes: the walsenders on A and B, and the apply workers on A and B.\\n> The walsenders are waiting for feedback from the apply workers, and the\\n> apply workers are waiting for the walsenders to synchronize (I may be\\n> oversimplifying).\\n>\\n> He suggested I try the above example instead with `synchronous_commit`\\n> set to `local`. In this case, I verified that there is no more deadlock,\\n> but he informed me that we would not be able to use cascading\\n> synchronous replication when \\\"origin=none\\\".\\n>\\n\\nThis has nothing to do with the origin feature. I mean this should\\nhappen with origin=any or even in PG15 without using origin at all.\\nAm, I missing something? One related point to note is that in physical\\nreplication cascading replication is asynchronous. See docs [1]\\n(Cascading replication is currently asynchronous....)\\n\\n> If we decide that this is a documentation issue, I'd suggest we improve\\n> the guidance around using `synchronous_commit`[1] on the CREATE\\n> SUBSCRIPTION page, as the GUC page[2] warns against using `local`:\\n>\\n\\nYeah, but on Create Subscription page, we have mentioned that it is\\nsafe to use off for logical replication. One can use local or higher\\nfor reducing the latency for COMMIT when synchronous replication is\\nused in the publisher. Won't using 'local' while creating subscription\\nwould suffice the need to consistently replicate the data? I mean it\\nis equivalent to somebody using levels greater than local in case of\\nphysical replication. I think in the case of physical replication, we\\nwon't wait for standby to replicate to another node before sending a\\nresponse, so why to wait in the case of logical replication? If this\\nunderstanding is correct, then probably it is sufficient to support\\n'local' for a subscription.\\n\\n[1] - https://www.postgresql.org/docs/devel/warm-standby.html\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 10 Jan 2023 20:47:13 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On 1/10/23 10:17 AM, Amit Kapila wrote:\\r\\n> On Tue, Jan 10, 2023 at 8:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\r\\n\\r\\n>> This consistently created the deadlock in my testing.\\r\\n>>\\r\\n>> Discussing with Masahiko off-list, this is due to a deadlock from 4\\r\\n>> processes: the walsenders on A and B, and the apply workers on A and B.\\r\\n>> The walsenders are waiting for feedback from the apply workers, and the\\r\\n>> apply workers are waiting for the walsenders to synchronize (I may be\\r\\n>> oversimplifying).\\r\\n>>\\r\\n>> He suggested I try the above example instead with `synchronous_commit`\\r\\n>> set to `local`. In this case, I verified that there is no more deadlock,\\r\\n>> but he informed me that we would not be able to use cascading\\r\\n>> synchronous replication when \\\"origin=none\\\".\\r\\n>>\\r\\n> \\r\\n> This has nothing to do with the origin feature. I mean this should\\r\\n> happen with origin=any or even in PG15 without using origin at all.\\r\\n> Am, I missing something? One related point to note is that in physical\\r\\n> replication cascading replication is asynchronous. See docs [1]\\r\\n> (Cascading replication is currently asynchronous....)\\r\\n\\r\\nThis is not directly related to the origin feature, but the origin \\r\\nfeature makes it apparent. There is a common use-case where people want \\r\\nto replicate data synchronously between two tables, and this is enabled \\r\\nby the \\\"origin\\\" feature.\\r\\n\\r\\nTo be clear, my bigger concern is that it's fairly easy for users to \\r\\ncreate a deadlock situation based on how they set \\\"synchronous_commit\\\" \\r\\nwith the origin feature -- this is the main reason why I brought it up. \\r\\nI'm less concerned about the \\\"cascading\\\" case, though I want to try out \\r\\nsync rep between 3 instances to see what happens.\\r\\n\\r\\n>> If we decide that this is a documentation issue, I'd suggest we improve\\r\\n>> the guidance around using `synchronous_commit`[1] on the CREATE\\r\\n>> SUBSCRIPTION page, as the GUC page[2] warns against using `local`:\\r\\n>>\\r\\n> \\r\\n> Yeah, but on Create Subscription page, we have mentioned that it is\\r\\n> safe to use off for logical replication.\\r\\n\\r\\nWhile I think you can infer that it's \\\"safe\\\" for synchronous \\r\\nreplication, I don't think it's clear.\\r\\n\\r\\nWe say it's \\\"safe to use `off` for logical replication\\\", but provide a \\r\\nlengthy explanation around synchronous logical replication that \\r\\nconcludes that it's \\\"advantageous to set synchronous_commit to local or \\r\\nhigher\\\" but does not address safety. The first line in the explanation \\r\\nof the parameter links to the `synchronous_commit` GUC which \\r\\nspecifically advises against using \\\"local\\\" for synchronous replication.\\r\\n\\r\\nAdditionally, because we say \\\"local\\\" or higher, we increase the risk of \\r\\nthe aforementioned in HEAD with the origin feature.\\r\\n\\r\\nI know I'm hammering on this point a bit, but it feels like this is \\r\\nrelatively easy to misconfigure with the upcoming \\\"origin\\\" change (I did \\r\\nso myself from reading the devel docs) and we should ensure we guide our \\r\\nusers appropriately.\\r\\n\\r\\n> One can use local or higher\\r\\n> for reducing the latency for COMMIT when synchronous replication is\\r\\n> used in the publisher. Won't using 'local' while creating subscription\\r\\n> would suffice the need to consistently replicate the data? I mean it\\r\\n> is equivalent to somebody using levels greater than local in case of\\r\\n> physical replication. I think in the case of physical replication, we\\r\\n> won't wait for standby to replicate to another node before sending a\\r\\n> response, so why to wait in the case of logical replication? If this\\r\\n> understanding is correct, then probably it is sufficient to support\\r\\n> 'local' for a subscription.\\r\\n\\r\\nI think this is a good explanation. We should incorporate some version \\r\\nof this into the docs, and add some clarity around the use of \\r\\n`synchronous_commit` option in `CREATE SUBSCRIPTION` in particular with \\r\\nthe origin feature. I can make an attempt at this.\\r\\n\\r\\nPerhaps another question: based on this, should we disallow setting \\r\\nvalues of `synchronous_commit` greater than `local` when using \\r\\n\\\"origin=none\\\"? That might be too strict, but maybe we should warn around \\r\\nthe risk of deadlock either in the logs or in the docs.\\r\\n\\r\\nThanks,\\r\\n\\r\\nJonathan\",\n \"msg_date\": \"Tue, 10 Jan 2023 12:54:28 -0500\",\n \"msg_from\": \"\\\"Jonathan S. Katz\\\" <jkatz@postgresql.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 10, 2023 at 11:24 PM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> On 1/10/23 10:17 AM, Amit Kapila wrote:\\n> > On Tue, Jan 10, 2023 at 8:13 AM Jonathan S. Katz <jkatz@postgresql.org> wrote:\\n>\\n> > One can use local or higher\\n> > for reducing the latency for COMMIT when synchronous replication is\\n> > used in the publisher. Won't using 'local' while creating subscription\\n> > would suffice the need to consistently replicate the data? I mean it\\n> > is equivalent to somebody using levels greater than local in case of\\n> > physical replication. I think in the case of physical replication, we\\n> > won't wait for standby to replicate to another node before sending a\\n> > response, so why to wait in the case of logical replication? If this\\n> > understanding is correct, then probably it is sufficient to support\\n> > 'local' for a subscription.\\n>\\n> I think this is a good explanation. We should incorporate some version\\n> of this into the docs, and add some clarity around the use of\\n> `synchronous_commit` option in `CREATE SUBSCRIPTION` in particular with\\n> the origin feature. I can make an attempt at this.\\n>\\n\\nOkay, thanks!\\n\\n> Perhaps another question: based on this, should we disallow setting\\n> values of `synchronous_commit` greater than `local` when using\\n> \\\"origin=none\\\"?\\n>\\n\\nI think it should be done irrespective of the value of origin because\\nit can create a deadlock in those cases as well. I think one idea as\\nyou suggest is to block levels higher than local and the other is to\\nmake it behave like 'local' even if the level is higher than 'local'\\nwhich would be somewhat similar to our physical replication.\\n\\n> That might be too strict, but maybe we should warn around\\n> the risk of deadlock either in the logs or in the docs.\\n>\\n\\nYeah, we can mention it in docs as well. We can probably document as\\npart of the docs work for this thread but it would be better to start\\na separate thread to fix this behavior by either of the above two\\napproaches.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 11 Jan 2023 07:46:42 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On 12.09.22 07:23, vignesh C wrote:\\n> On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>>\\n>> On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\\n>>>\\n>>>\\n>>> The attached patch has the changes to handle the same.\\n>>>\\n>>\\n>> Pushed. I am not completely sure whether we want the remaining\\n>> documentation patch in this thread in its current form or by modifying\\n>> it. Johnathan has shown some interest in it. I feel you can start a\\n>> separate thread for it to see if there is any interest in the same and\\n>> close the CF entry for this work.\\n> \\n> Thanks for pushing the patch. I have closed this entry in commitfest.\\n> I will wait for some more time and see the response regarding the\\n> documentation patch and then start a new thread if required.\\n\\nThis patch added the following error message:\\n\\nerrdetail_plural(\\\"Subscribed publication %s is subscribing to other \\npublications.\\\",\\n\\\"Subscribed publications %s are subscribing to other publications.\\\",\\nlist_length(publist), pubnames->data),\\n\\nBut in PostgreSQL, a publication cannot subscribe to a publication, so \\nthis is not giving accurate information. Apparently, what it is trying \\nto say is that\\n\\nThe subscription that you are creating subscribes to publications that \\ncontain tables that are written to by other subscriptions.\\n\\nCan we get to a more accurate wording like this?\\n\\nThere is also a translatability issue there, in the way the publication \\nlist is pasted into the message.\\n\\nIs the list of affected publications really that interesting? I wonder \\nwhether the list of affected tables might be more relevant?\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Aug 2023 10:20:42 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter@eisentraut.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\\n>\\n> On 12.09.22 07:23, vignesh C wrote:\\n> > On Fri, 9 Sept 2022 at 11:12, Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >>\\n> >> On Thu, Sep 8, 2022 at 9:32 AM vignesh C <vignesh21@gmail.com> wrote:\\n> >>>\\n> >>>\\n> >>> The attached patch has the changes to handle the same.\\n> >>>\\n> >>\\n> >> Pushed. I am not completely sure whether we want the remaining\\n> >> documentation patch in this thread in its current form or by modifying\\n> >> it. Johnathan has shown some interest in it. I feel you can start a\\n> >> separate thread for it to see if there is any interest in the same and\\n> >> close the CF entry for this work.\\n> >\\n> > Thanks for pushing the patch. I have closed this entry in commitfest.\\n> > I will wait for some more time and see the response regarding the\\n> > documentation patch and then start a new thread if required.\\n>\\n> This patch added the following error message:\\n>\\n> errdetail_plural(\\\"Subscribed publication %s is subscribing to other\\n> publications.\\\",\\n> \\\"Subscribed publications %s are subscribing to other publications.\\\",\\n> list_length(publist), pubnames->data),\\n>\\n> But in PostgreSQL, a publication cannot subscribe to a publication, so\\n> this is not giving accurate information. Apparently, what it is trying\\n> to say is that\\n>\\n> The subscription that you are creating subscribes to publications that\\n> contain tables that are written to by other subscriptions.\\n>\\n> Can we get to a more accurate wording like this?\\n>\\n\\n+1 for changing the message as per your suggestion.\\n\\n> There is also a translatability issue there, in the way the publication\\n> list is pasted into the message.\\n>\\n> Is the list of affected publications really that interesting? I wonder\\n> whether the list of affected tables might be more relevant?\\n>\\n\\nIn that case, we need to specify both schema name and table name in\\nthat case. I guess the list could be very long and not sure what to do\\nfor schema publications ( Create Publication ... For Schema).\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Aug 2023 14:22:43 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 8, 2023 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\\n> >\\n> > This patch added the following error message:\\n> >\\n> > errdetail_plural(\\\"Subscribed publication %s is subscribing to other\\n> > publications.\\\",\\n> > \\\"Subscribed publications %s are subscribing to other publications.\\\",\\n> > list_length(publist), pubnames->data),\\n> >\\n> > But in PostgreSQL, a publication cannot subscribe to a publication, so\\n> > this is not giving accurate information. Apparently, what it is trying\\n> > to say is that\\n> >\\n> > The subscription that you are creating subscribes to publications that\\n> > contain tables that are written to by other subscriptions.\\n> >\\n> > Can we get to a more accurate wording like this?\\n> >\\n>\\n> +1 for changing the message as per your suggestion.\\n>\\n\\nPSA a patch to change this message text. The message now has wording\\nsimilar to the suggestion.\\n\\n> > There is also a translatability issue there, in the way the publication\\n> > list is pasted into the message.\\n> >\\n\\nThe name/list substitution is now done within parentheses, which AFAIK\\nwill be enough to eliminate any translation ambiguities.\\n\\n> > Is the list of affected publications really that interesting? I wonder\\n> > whether the list of affected tables might be more relevant?\\n> >\\n>\\n> In that case, we need to specify both schema name and table name in\\n> that case. I guess the list could be very long and not sure what to do\\n> for schema publications ( Create Publication ... For Schema).\\n\\nRight, IIUC that was the reason for not choosing to show the tables in\\nthe original patch -- i.e. the list might easily be very long with\\n100s or 1000s of tables it, and so inappropriate to substitute in the\\nmessage. OTOH, showing only problematic publications is a short list\\nbut it is still more useful than showing nothing (e.g. there other\\npublications of the subscription might be OK so those ones are not\\nlisted)\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\",\n \"msg_date\": \"Wed, 9 Aug 2023 12:50:10 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 9, 2023 at 8:20 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Tue, Aug 8, 2023 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\\n> > >\\n> > > This patch added the following error message:\\n> > >\\n> > > errdetail_plural(\\\"Subscribed publication %s is subscribing to other\\n> > > publications.\\\",\\n> > > \\\"Subscribed publications %s are subscribing to other publications.\\\",\\n> > > list_length(publist), pubnames->data),\\n> > >\\n> > > But in PostgreSQL, a publication cannot subscribe to a publication, so\\n> > > this is not giving accurate information. Apparently, what it is trying\\n> > > to say is that\\n> > >\\n> > > The subscription that you are creating subscribes to publications that\\n> > > contain tables that are written to by other subscriptions.\\n> > >\\n> > > Can we get to a more accurate wording like this?\\n> > >\\n> >\\n> > +1 for changing the message as per your suggestion.\\n> >\\n>\\n> PSA a patch to change this message text. The message now has wording\\n> similar to the suggestion.\\n>\\n> > > There is also a translatability issue there, in the way the publication\\n> > > list is pasted into the message.\\n> > >\\n>\\n> The name/list substitution is now done within parentheses, which AFAIK\\n> will be enough to eliminate any translation ambiguities.\\n>\\n\\nA similar instance as below uses \\\\\\\"%s\\\\\\\" instead. So, isn't using the\\nsame a better idea?\\nerrdetail_plural(\\\"LDAP search for filter \\\\\\\"%s\\\\\\\" on server \\\\\\\"%s\\\\\\\"\\nreturned %d entry.\\\",\\n \\\"LDAP search for filter \\\\\\\"%s\\\\\\\" on server \\\\\\\"%s\\\\\\\" returned %d entries.\\\",\\n\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 9 Aug 2023 10:14:05 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 9, 2023 at 2:44 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Aug 9, 2023 at 8:20 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > On Tue, Aug 8, 2023 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > >\\n> > > On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\\n> > > >\\n> > > > This patch added the following error message:\\n> > > >\\n> > > > errdetail_plural(\\\"Subscribed publication %s is subscribing to other\\n> > > > publications.\\\",\\n> > > > \\\"Subscribed publications %s are subscribing to other publications.\\\",\\n> > > > list_length(publist), pubnames->data),\\n> > > >\\n> > > > But in PostgreSQL, a publication cannot subscribe to a publication, so\\n> > > > this is not giving accurate information. Apparently, what it is trying\\n> > > > to say is that\\n> > > >\\n> > > > The subscription that you are creating subscribes to publications that\\n> > > > contain tables that are written to by other subscriptions.\\n> > > >\\n> > > > Can we get to a more accurate wording like this?\\n> > > >\\n> > >\\n> > > +1 for changing the message as per your suggestion.\\n> > >\\n> >\\n> > PSA a patch to change this message text. The message now has wording\\n> > similar to the suggestion.\\n> >\\n> > > > There is also a translatability issue there, in the way the publication\\n> > > > list is pasted into the message.\\n> > > >\\n> >\\n> > The name/list substitution is now done within parentheses, which AFAIK\\n> > will be enough to eliminate any translation ambiguities.\\n> >\\n>\\n> A similar instance as below uses \\\\\\\"%s\\\\\\\" instead. So, isn't using the\\n> same a better idea?\\n> errdetail_plural(\\\"LDAP search for filter \\\\\\\"%s\\\\\\\" on server \\\\\\\"%s\\\\\\\"\\n> returned %d entry.\\\",\\n> \\\"LDAP search for filter \\\\\\\"%s\\\\\\\" on server \\\\\\\"%s\\\\\\\" returned %d entries.\\\",\\n>\\n>\\n\\nHmm. I don't see any similarity -- there the plural is for the %d, not\\nfor a list of string items. And in our case the publication name (or\\nlist of names) are already quoted by the function returning that list,\\nso quoting them again doesn't really make sense.\\n\\nExample output using the patch look like this:\\n\\nSINGLULAR\\ntest_sub=# create subscription sub_test connection 'dbname=test_pub'\\npublication pub_all_at_pub with(origin=NONE);\\nWARNING: subscription \\\"sub_test\\\" requested copy_data with origin =\\nNONE but might copy data that had a different origin\\nDETAIL: The subscription that you are creating has a publication\\n(\\\"pub_all_at_pub\\\") containing tables written to by other\\nsubscriptions.\\nHINT: Verify that initial data copied from the publisher tables did\\nnot come from other origins.\\nNOTICE: created replication slot \\\"sub_test\\\" on publisher\\nCREATE SUBSCRIPTION\\n\\nPLURAL\\ntest_sub=# create subscription sub_test3 connection 'dbname=test_pub'\\npublication pub_all_at_pub,pub_s1_at_pub,pub_s2_at_pub\\nwith(origin=NONE);\\nWARNING: subscription \\\"sub_test3\\\" requested copy_data with origin =\\nNONE but might copy data that had a different origin\\nDETAIL: The subscription that you are creating has publications\\n(\\\"pub_s1_at_pub\\\", \\\"pub_all_at_pub\\\") containing tables written to by\\nother subscriptions.\\nHINT: Verify that initial data copied from the publisher tables did\\nnot come from other origins.\\nNOTICE: created replication slot \\\"sub_test3\\\" on publisher\\nCREATE SUBSCRIPTION\\n\\n~\\n\\nI thought above looked fine but if PeterE also does not like the ()\\nthen the message can be rearranged slightly like below to put the\\noffending publications at the end of the message, like this:\\n\\nDETAIL: The subscription that you are creating has the following\\npublications containing tables written to by other subscriptions:\\n\\\"pub_s1_at_pub\\\", \\\"pub_all_at_pub\\\"\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia\\n\\n\\n\",\n \"msg_date\": \"Wed, 9 Aug 2023 15:29:31 +1000\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 9, 2023 at 10:59 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> Example output using the patch look like this:\\n>\\n> SINGLULAR\\n> test_sub=# create subscription sub_test connection 'dbname=test_pub'\\n> publication pub_all_at_pub with(origin=NONE);\\n> WARNING: subscription \\\"sub_test\\\" requested copy_data with origin =\\n> NONE but might copy data that had a different origin\\n> DETAIL: The subscription that you are creating has a publication\\n> (\\\"pub_all_at_pub\\\") containing tables written to by other\\n> subscriptions.\\n> HINT: Verify that initial data copied from the publisher tables did\\n> not come from other origins.\\n> NOTICE: created replication slot \\\"sub_test\\\" on publisher\\n> CREATE SUBSCRIPTION\\n>\\n> PLURAL\\n> test_sub=# create subscription sub_test3 connection 'dbname=test_pub'\\n> publication pub_all_at_pub,pub_s1_at_pub,pub_s2_at_pub\\n> with(origin=NONE);\\n> WARNING: subscription \\\"sub_test3\\\" requested copy_data with origin =\\n> NONE but might copy data that had a different origin\\n> DETAIL: The subscription that you are creating has publications\\n> (\\\"pub_s1_at_pub\\\", \\\"pub_all_at_pub\\\") containing tables written to by\\n> other subscriptions.\\n> HINT: Verify that initial data copied from the publisher tables did\\n> not come from other origins.\\n> NOTICE: created replication slot \\\"sub_test3\\\" on publisher\\n> CREATE SUBSCRIPTION\\n>\\n> ~\\n>\\n> I thought above looked fine but if PeterE also does not like the ()\\n> then the message can be rearranged slightly like below to put the\\n> offending publications at the end of the message, like this:\\n>\\n\\nOkay, I didn't know strings were already quoted but not sure if Round\\nBrackets () exactly will address the translatability concern.\\n\\n> DETAIL: The subscription that you are creating has the following\\n> publications containing tables written to by other subscriptions:\\n> \\\"pub_s1_at_pub\\\", \\\"pub_all_at_pub\\\"\\n>\\n\\nFair enough. Peter E., do let us know what you think makes sense here?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 9 Aug 2023 11:58:23 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n },\n {\n \"msg_contents\": \"On 09.08.23 04:50, Peter Smith wrote:\\n> On Tue, Aug 8, 2023 at 6:52 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>>\\n>> On Tue, Aug 8, 2023 at 1:50 PM Peter Eisentraut <peter@eisentraut.org> wrote:\\n>>>\\n>>> This patch added the following error message:\\n>>>\\n>>> errdetail_plural(\\\"Subscribed publication %s is subscribing to other\\n>>> publications.\\\",\\n>>> \\\"Subscribed publications %s are subscribing to other publications.\\\",\\n>>> list_length(publist), pubnames->data),\\n>>>\\n>>> But in PostgreSQL, a publication cannot subscribe to a publication, so\\n>>> this is not giving accurate information. Apparently, what it is trying\\n>>> to say is that\\n>>>\\n>>> The subscription that you are creating subscribes to publications that\\n>>> contain tables that are written to by other subscriptions.\\n>>>\\n>>> Can we get to a more accurate wording like this?\\n>>>\\n>>\\n>> +1 for changing the message as per your suggestion.\\n>>\\n> \\n> PSA a patch to change this message text. The message now has wording\\n> similar to the suggestion.\\n\\ncommitted, thanks\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Aug 2023 15:50:09 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter@eisentraut.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Handle infinite recursion in logical replication setup\"\n }\n]","isConversational":false}}},{"rowIdx":53563,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello,\n\nI have created a patch that adds support for building Postgres for the\nwindows/arm64 platform using the MSVC toolchain. Following changes have\nbeen included\n\n1. Extend MSVC scripts to handle ARM64 platform.\n2. Add arm64 definition of spin_delay function.\n3. Exclude arm_acle.h import with MSVC compiler.\n\nCompilation steps are consistent and similar to other windows platforms.\n\nThe change has been tested on windows/x86_64 and windows/arm64 and all\nregression tests passes on both platforms.\n\nThanks,\nNiyas","msg_date":"Wed, 23 Feb 2022 10:09:36 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"[PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, Feb 23, 2022 at 11:09 PM Niyas Sait <niyas.sait@linaro.org> wrote:\n> I have created a patch that adds support for building Postgres for the windows/arm64 platform using the MSVC toolchain. Following changes have been included\n>\n> 1. Extend MSVC scripts to handle ARM64 platform.\n> 2. Add arm64 definition of spin_delay function.\n> 3. Exclude arm_acle.h import with MSVC compiler.\n>\n> Compilation steps are consistent and similar to other windows platforms.\n>\n> The change has been tested on windows/x86_64 and windows/arm64 and all regression tests passes on both platforms.\n\n+ # arm64 linker only supports dynamic base address\n+ my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' : 'false';\n...\n+ <RandomizedBaseAddress>$cfgrandbaseaddress</RandomizedBaseAddress>\n\nDoes that mean that you can't turn off ASLR on this arch? Does it\ncause random backend failures due to inability to map shared memory at\nthe expected address? Our experience with EXEC_BACKEND on various\nUnix systems tells us that you have to turn off ASLR if you want 100%\nsuccess rate on starting backends, but I guess it depends on the\ndetails of how the randomisation is done.\n\nAny interest in providing a build farm animal, so that we could know\nthat this works?\n\n\n","msg_date":"Sat, 19 Mar 2022 09:50:06 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Thanks, Thomas for reviewing the patch!\n\nYes, we cannot turn off ASLR for Arm64 targets with MSVC. I haven't seen\nany issues so far on win/arm64 with this option turned off. I guess it\nshould be okay. If we really need ASLR turned off, then I guess might have\nto use clang.\n\nYes, we could look into providing a build machine. Do you have any\nreference to what the CI system looks like now for PostgresSQL and how to\nadd new workers etc.?\n\nNiyas\n\nOn Fri, 18 Mar 2022 at 20:50, Thomas Munro <thomas.munro@gmail.com> wrote:\n\n> On Wed, Feb 23, 2022 at 11:09 PM Niyas Sait <niyas.sait@linaro.org> wrote:\n> > I have created a patch that adds support for building Postgres for the\n> windows/arm64 platform using the MSVC toolchain. Following changes have\n> been included\n> >\n> > 1. Extend MSVC scripts to handle ARM64 platform.\n> > 2. Add arm64 definition of spin_delay function.\n> > 3. Exclude arm_acle.h import with MSVC compiler.\n> >\n> > Compilation steps are consistent and similar to other windows platforms.\n> >\n> > The change has been tested on windows/x86_64 and windows/arm64 and all\n> regression tests passes on both platforms.\n>\n> + # arm64 linker only supports dynamic base address\n> + my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' :\n> 'false';\n> ...\n> + <RandomizedBaseAddress>$cfgrandbaseaddress</RandomizedBaseAddress>\n>\n> Does that mean that you can't turn off ASLR on this arch? Does it\n> cause random backend failures due to inability to map shared memory at\n> the expected address? Our experience with EXEC_BACKEND on various\n> Unix systems tells us that you have to turn off ASLR if you want 100%\n> success rate on starting backends, but I guess it depends on the\n> details of how the randomisation is done.\n>\n> Any interest in providing a build farm animal, so that we could know\n> that this works?\n>\n\nThanks, Thomas for reviewing the patch!Yes, we cannot turn off ASLR for Arm64 targets with MSVC. I haven't seen any issues so far on win/arm64 with this option turned off. I guess it should be okay. If we really need ASLR turned off, then I guess might have to use clang.Yes, we could look into providing a build machine. Do you have any reference to what the CI system looks like now for PostgresSQL and how to add new workers etc.?NiyasOn Fri, 18 Mar 2022 at 20:50, Thomas Munro <thomas.munro@gmail.com> wrote:On Wed, Feb 23, 2022 at 11:09 PM Niyas Sait <niyas.sait@linaro.org> wrote:\n> I have created a patch that adds support for building Postgres for the windows/arm64 platform using the MSVC toolchain. Following changes have been included\n>\n> 1. Extend MSVC scripts to handle ARM64 platform.\n> 2. Add arm64 definition of spin_delay function.\n> 3. Exclude arm_acle.h import with MSVC compiler.\n>\n> Compilation steps are consistent and similar to other windows platforms.\n>\n> The change has been tested on windows/x86_64 and windows/arm64 and all regression tests passes on both platforms.\n\n+ # arm64 linker only supports dynamic base address\n+ my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' : 'false';\n...\n+ <RandomizedBaseAddress>$cfgrandbaseaddress</RandomizedBaseAddress>\n\nDoes that mean that you can't turn off ASLR on this arch? Does it\ncause random backend failures due to inability to map shared memory at\nthe expected address? Our experience with EXEC_BACKEND on various\nUnix systems tells us that you have to turn off ASLR if you want 100%\nsuccess rate on starting backends, but I guess it depends on the\ndetails of how the randomisation is done.\n\nAny interest in providing a build farm animal, so that we could know\nthat this works?","msg_date":"Tue, 22 Mar 2022 09:37:46 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nPlease don't top-post here. See\nhttps://wiki.postgresql.org/wiki/Mailing_Lists#Email_etiquette_mechanics.\n\nOn Tue, Mar 22, 2022 at 09:37:46AM +0000, Niyas Sait wrote:\n>\n> Yes, we could look into providing a build machine. Do you have any\n> reference to what the CI system looks like now for PostgresSQL and how to\n> add new workers etc.?\n\nIt's all documented at\nhttps://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto.\n\n\n","msg_date":"Tue, 22 Mar 2022 18:30:11 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, Mar 22, 2022 at 11:30 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> On Tue, Mar 22, 2022 at 09:37:46AM +0000, Niyas Sait wrote:\n> > Yes, we could look into providing a build machine. Do you have any\n> > reference to what the CI system looks like now for PostgresSQL and how to\n> > add new workers etc.?\n>\n> It's all documented at\n> https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto.\n\nIt seems likely there will be more and more Windows/ARM users, so yeah\nhaving a machine to test that combination would be great. I wonder if\nASLR isn't breaking for you by good luck only...\n\n\n","msg_date":"Wed, 23 Mar 2022 16:30:58 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-03-23 16:30:58 +1300, Thomas Munro wrote:\n> On Tue, Mar 22, 2022 at 11:30 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> > On Tue, Mar 22, 2022 at 09:37:46AM +0000, Niyas Sait wrote:\n> > > Yes, we could look into providing a build machine. Do you have any\n> > > reference to what the CI system looks like now for PostgresSQL and how to\n> > > add new workers etc.?\n> >\n> > It's all documented at\n> > https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto.\n> \n> It seems likely there will be more and more Windows/ARM users, so yeah\n> having a machine to test that combination would be great. I wonder if\n> ASLR isn't breaking for you by good luck only...\n\nI think we've generally seen the ASLR issue become less prominent on\nwindows. Whether that's because of the silent retries we added, or because\njust about everyone moved to 64bit windows / PG, I don't know. I'd guess both,\nwith 64bit being the larger influence.\n\nWonder if it's worth adding some debug logging to the retry code and stop\ndisabling ASLR on 64bit windows... It's imo pretty crazy that we loop up to\n100 times in internal_forkexec() around CreateProcess() &&\npgwin32_ReserveSharedMemoryRegion() without, as far as I can see, a single\ndebug message.\n\nI don't think we can infer too much about the failure rate on windows from the\n!windows EXEC_BACKEND rates. The two internal_forkexec() implementations\nbehaves just too differently.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 23 Mar 2022 18:15:31 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Mar 24, 2022 at 2:15 PM Andres Freund <andres@anarazel.de> wrote:\n> I think we've generally seen the ASLR issue become less prominent on\n> windows. Whether that's because of the silent retries we added, or because\n> just about everyone moved to 64bit windows / PG, I don't know. I'd guess both,\n> with 64bit being the larger influence.\n>\n> Wonder if it's worth adding some debug logging to the retry code and stop\n> disabling ASLR on 64bit windows... It's imo pretty crazy that we loop up to\n> 100 times in internal_forkexec() around CreateProcess() &&\n> pgwin32_ReserveSharedMemoryRegion() without, as far as I can see, a single\n> debug message.\n\nYeah. I think we should commit this patch, but decree that\nWindows/aarch64 support is experimental only for now. That allows a\nbuild farm animal to be set up. Then we add a bit of extra logging\nand see how it does running our test suite over time and learn more.\n\n\n","msg_date":"Fri, 25 Mar 2022 11:38:44 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Fri, Mar 25, 2022 at 11:38:44AM +1300, Thomas Munro wrote:\n> Yeah. I think we should commit this patch, but decree that\n> Windows/aarch64 support is experimental only for now. That allows a\n> build farm animal to be set up. Then we add a bit of extra logging\n> and see how it does running our test suite over time and learn more.\n\nI don't have such a setup so my testing capabilities are limited.\nDoes anybody have one? I think that we could be flexible for this\npatch, even after feature freeze as it introduces something entirely\nnew without impacting the existing code. The patch has been moved to\nthe next CF for now.\n--\nMichael","msg_date":"Thu, 7 Apr 2022 15:36:19 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Fri, Mar 25, 2022 at 11:38:44AM +1300, Thomas Munro wrote:\n>> Yeah. I think we should commit this patch, but decree that\n>> Windows/aarch64 support is experimental only for now. That allows a\n>> build farm animal to be set up. Then we add a bit of extra logging\n>> and see how it does running our test suite over time and learn more.\n\n> I don't have such a setup so my testing capabilities are limited.\n> Does anybody have one? I think that we could be flexible for this\n> patch, even after feature freeze as it introduces something entirely\n> new without impacting the existing code. The patch has been moved to\n> the next CF for now.\n\nI dunno, the lack of any in-house capability for this makes me very\nnervous. If it causes problems down the road, how will we debug it?\nSo it seems like the sort of patch to put in at the beginning of a\ndevelopment cycle, not post-feature-freeze.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 07 Apr 2022 09:40:43 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"> On 7 Apr 2022, at 15:40, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Michael Paquier <michael@paquier.xyz> writes:\n>> On Fri, Mar 25, 2022 at 11:38:44AM +1300, Thomas Munro wrote:\n>>> Yeah. I think we should commit this patch, but decree that\n>>> Windows/aarch64 support is experimental only for now. That allows a\n>>> build farm animal to be set up. Then we add a bit of extra logging\n>>> and see how it does running our test suite over time and learn more.\n> \n>> I don't have such a setup so my testing capabilities are limited.\n>> Does anybody have one? I think that we could be flexible for this\n>> patch, even after feature freeze as it introduces something entirely\n>> new without impacting the existing code. The patch has been moved to\n>> the next CF for now.\n> \n> I dunno, the lack of any in-house capability for this makes me very\n> nervous. If it causes problems down the road, how will we debug it?\n\nIf those with an interest in such platform support cannot spare the cycles to\nat least run a buildfarm member, then it seems a stretch for us to maintain\nsuch support with any confidence.\n\n> So it seems like the sort of patch to put in at the beginning of a\n> development cycle, not post-feature-freeze.\n\n+1\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Thu, 7 Apr 2022 16:55:27 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-04-07 09:40:43 -0400, Tom Lane wrote:\n> If it causes problems down the road, how will we debug it?\n\nIf what causes problems down the road? Afaics the patch doesn't change\nanything outside of windows-on-arm, so it shouldn't cause any breakage we care\nabout until we get a buildfarm animal.\n\nWe've traditionally been somewhat relaxed about adding support for new\nplatforms, on similar notions. That said:\n\n> So it seems like the sort of patch to put in at the beginning of a\n> development cycle, not post-feature-freeze.\n\nThere doesn't seem to be a great urgency, and there's plenty stuff going on\nright now. I can see us merging it post branching off, and then backpatching\nit a bit later?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 7 Apr 2022 10:22:58 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2022-04-07 09:40:43 -0400, Tom Lane wrote:\n>> If it causes problems down the road, how will we debug it?\n\n> If what causes problems down the road? Afaics the patch doesn't change\n> anything outside of windows-on-arm, so it shouldn't cause any breakage we care\n> about until we get a buildfarm animal.\n\nDo we have a volunteer to run a buildfarm animal? I don't see much\npoint in thinking about this till there is one ready to go.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 07 Apr 2022 13:42:49 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-04-07 13:42:49 -0400, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > On 2022-04-07 09:40:43 -0400, Tom Lane wrote:\n> >> If it causes problems down the road, how will we debug it?\n>\n> > If what causes problems down the road? Afaics the patch doesn't change\n> > anything outside of windows-on-arm, so it shouldn't cause any breakage we care\n> > about until we get a buildfarm animal.\n>\n> Do we have a volunteer to run a buildfarm animal? I don't see much\n> point in thinking about this till there is one ready to go.\n\nOP said that they might be able to:\nhttps://www.postgresql.org/message-id/CAFPTBD_NZb%3Dq_6uE-QV%2BS%2Bpm%3DRc9sBKrg8CQ_Y3dc27Awrm7cQ%40mail.gmail.com\n\nNiyas, any updates on that?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 7 Apr 2022 10:53:59 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"> Niyas, any updates on that?\n\nSorry for the delay! Configuring the scripts took some time. I have\nsuccessfully run the builfarm animal script using my git repository [1]\nwhich contains the proposed patch on a Windows Arm64 machine.\n\nI made a request to add a new machine to build farm through [2].\n\nI believe we should be able to enable the build farm machine once the\nchange has been merged.\n\nThanks,\nNiyas\n\n[1] https://github.com/nsait-linaro/postgres.git\n[2] https://buildfarm.postgresql.org/cgi-bin/register-form.pl\n\n\nOn Thu, 7 Apr 2022 at 18:54, Andres Freund <andres@anarazel.de> wrote:\n\n> Hi,\n>\n> On 2022-04-07 13:42:49 -0400, Tom Lane wrote:\n> > Andres Freund <andres@anarazel.de> writes:\n> > > On 2022-04-07 09:40:43 -0400, Tom Lane wrote:\n> > >> If it causes problems down the road, how will we debug it?\n> >\n> > > If what causes problems down the road? Afaics the patch doesn't change\n> > > anything outside of windows-on-arm, so it shouldn't cause any breakage\n> we care\n> > > about until we get a buildfarm animal.\n> >\n> > Do we have a volunteer to run a buildfarm animal? I don't see much\n> > point in thinking about this till there is one ready to go.\n>\n> OP said that they might be able to:\n>\n> https://www.postgresql.org/message-id/CAFPTBD_NZb%3Dq_6uE-QV%2BS%2Bpm%3DRc9sBKrg8CQ_Y3dc27Awrm7cQ%40mail.gmail.com\n>\n> Niyas, any updates on that?\n>\n> Greetings,\n>\n> Andres Freund\n>\n\n > Niyas, any updates on that?Sorry for the delay! Configuring the scripts took some time. I have successfully run the builfarm animal script using my git repository [1] which contains the proposed patch on a Windows Arm64 machine.I made a request to add a new machine to build farm through [2].I believe we should be able to enable the build farm machine once the change has been merged.Thanks,Niyas[1] https://github.com/nsait-linaro/postgres.git[2] https://buildfarm.postgresql.org/cgi-bin/register-form.plOn Thu, 7 Apr 2022 at 18:54, Andres Freund <andres@anarazel.de> wrote:Hi,\n\nOn 2022-04-07 13:42:49 -0400, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > On 2022-04-07 09:40:43 -0400, Tom Lane wrote:\n> >> If it causes problems down the road, how will we debug it?\n>\n> > If what causes problems down the road? Afaics the patch doesn't change\n> > anything outside of windows-on-arm, so it shouldn't cause any breakage we care\n> > about until we get a buildfarm animal.\n>\n> Do we have a volunteer to run a buildfarm animal? I don't see much\n> point in thinking about this till there is one ready to go.\n\nOP said that they might be able to:\nhttps://www.postgresql.org/message-id/CAFPTBD_NZb%3Dq_6uE-QV%2BS%2Bpm%3DRc9sBKrg8CQ_Y3dc27Awrm7cQ%40mail.gmail.com\n\nNiyas, any updates on that?\n\nGreetings,\n\nAndres Freund","msg_date":"Tue, 19 Apr 2022 15:22:30 +0100","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, Apr 19, 2022 at 03:22:30PM +0100, Niyas Sait wrote:\n> Sorry for the delay! Configuring the scripts took some time. I have\n> successfully run the builfarm animal script using my git repository [1]\n> which contains the proposed patch on a Windows Arm64 machine.\n> \n> I made a request to add a new machine to build farm through [2].\n\nHave you tested with the amount of coverage provided by vcregress.pl?\n\nAnother thing I was wondering about is if it would be possible to have\nthis option in Travis, but that does not seem to be the case:\nhttps://docs.travis-ci.com/user/reference/windows/#windows-version\n\n> I believe we should be able to enable the build farm machine once the\n> change has been merged.\n\nBetter to wait for the beginning of the development cycle at this\nstage, based on all the replies received. That would bring us to the\nbeginning of July.\n\n+ if ($solution->{platform} eq 'ARM64') {\n+ push(@pgportfiles, 'pg_crc32c_armv8_choose.c');\n+ push(@pgportfiles, 'pg_crc32c_armv8.c');\n+ } else {\n+ push(@pgportfiles, 'pg_crc32c_sse42_choose.c');\n+ push(@pgportfiles, 'pg_crc32c_sse42.c');\n+ }\n\n+++ b/src/port/pg_crc32c_armv8.c\n+#ifndef _MSC_VER\n #include <arm_acle.h>\n+#endif\n[ ... ]\n+#ifdef _M_ARM64\n+ /*\n+ * arm64 way of hinting processor for spin loops optimisations\n+ * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\n+ */\n+ __isb(_ARM64_BARRIER_SY);\n+#else\nI think that such extra optimizations had better be in a separate\npatch, and we should focus on getting the build done first.\n\n+ # arm64 linker only supports dynamic base address\n+ my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' : 'false';\nThis issue is still lying around, and you may have been lucky. Would\nthere be any issues to remove this change to get a basic support in?\nAs mentioned upthread, there is a long history of Postgres with ASLR.\n\nThis would mean that a basic patch could be done with just the changes\nfor gendef.pl, and the first part of the changes inMSBuildProject.pm.\nWould that not be enough?\n--\nMichael","msg_date":"Wed, 20 Apr 2022 09:28:09 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"> Have you tested with the amount of coverage provided by vcregress.pl?\n\nI built and ran the relevant tests with the help of run_build.pl.\n\nI think following tests are executed - check, contribcheck, ecpgcheck,\ninstallcheck, isolationcheck, modulescheck, and upgradecheck.\n\n> Another thing I was wondering about is if it would be possible to have\n> this option in Travis, but that does not seem to be the case:\n> https://docs.travis-ci.com/user/reference/windows/#windows-version\n\nYes, I think Travis doesn't yet support Windows/Arm64 platform.\n\n> + if ($solution->{platform} eq 'ARM64') {\n> + push(@pgportfiles, 'pg_crc32c_armv8_choose.c');\n> + push(@pgportfiles, 'pg_crc32c_armv8.c');\n> + } else {\n> + push(@pgportfiles, 'pg_crc32c_sse42_choose.c');\n> + push(@pgportfiles, 'pg_crc32c_sse42.c');\n> + }\n>\n> +++ b/src/port/pg_crc32c_armv8.c\n> +#ifndef _MSC_VER\n> #include <arm_acle.h>\n> +#endif\n> [ ... ]\n> +#ifdef _M_ARM64\n> + /*\n> + * arm64 way of hinting processor for spin loops optimisations\n> + * ref:\nhttps://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\n> + */\n> + __isb(_ARM64_BARRIER_SY);\n> +#else\n> I think that such extra optimizations had better be in a separate\n> patch, and we should focus on getting the build done first.\n\n> This would mean that a basic patch could be done with just the changes\n> for gendef.pl, and the first part of the changes inMSBuildProject.pm.\n> Would that not be enough?\n\nI cannot build without any of the above changes. Nothing specific for\noptimization\nis added to this patch.\n\n> + # arm64 linker only supports dynamic base address\n> + my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' :\n'false';\n> This issue is still lying around, and you may have been lucky. Would\n> there be any issues to remove this change to get a basic support in?\n> As mentioned upthread, there is a long history of Postgres with ASLR.\n\nMSVC linker doesn't allow non-random base addresses for Arm64 platforms.\nIt is needed for basic support.\n\nNiyas\n\n> Have you tested with the amount of coverage provided by vcregress.pl?I built and ran the relevant tests with the help of run_build.pl.I think following tests are executed - check, contribcheck, ecpgcheck, installcheck, isolationcheck, modulescheck, and upgradecheck.> Another thing I was wondering about is if it would be possible to have> this option in Travis, but that does not seem to be the case:> https://docs.travis-ci.com/user/reference/windows/#windows-versionYes, I think Travis doesn't yet support Windows/Arm64 platform.> + if ($solution->{platform} eq 'ARM64') {> + push(@pgportfiles, 'pg_crc32c_armv8_choose.c');> + push(@pgportfiles, 'pg_crc32c_armv8.c');> + } else {> + push(@pgportfiles, 'pg_crc32c_sse42_choose.c');> + push(@pgportfiles, 'pg_crc32c_sse42.c');> + }> > +++ b/src/port/pg_crc32c_armv8.c> +#ifndef _MSC_VER> #include <arm_acle.h>> +#endif> [ ... ]> +#ifdef _M_ARM64> + /*> + * arm64 way of hinting processor for spin loops optimisations> + * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq> + */> + __isb(_ARM64_BARRIER_SY);> +#else> I think that such extra optimizations had better be in a separate> patch, and we should focus on getting the build done first.> This would mean that a basic patch could be done with just the changes> for gendef.pl, and the first part of the changes inMSBuildProject.pm.> Would that not be enough?I cannot build without any of the above changes. Nothing specific for optimizationis added to this patch.> + # arm64 linker only supports dynamic base address> + my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' : 'false';> This issue is still lying around, and you may have been lucky. Would> there be any issues to remove this change to get a basic support in?> As mentioned upthread, there is a long history of Postgres with ASLR.MSVC linker doesn't allow non-random base addresses for Arm64 platforms.It is needed for basic support.Niyas","msg_date":"Wed, 20 Apr 2022 10:43:06 +0100","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, Apr 20, 2022 at 10:43:06AM +0100, Niyas Sait wrote:\n>> This issue is still lying around, and you may have been lucky. Would\n>> there be any issues to remove this change to get a basic support in?\n>> As mentioned upthread, there is a long history of Postgres with ASLR.\n> \n> MSVC linker doesn't allow non-random base addresses for Arm64 platforms.\n> It is needed for basic support.\n\nWhy does it not allow that? Is that just a limitation of the\ncompiler? If yes, what is the error happening? This question is not\nexactly answered yet as of this thread. I may be missing a reference\nabout that in the upstream docs, but I see nowhere an explanation\nabout the reason and the why. That's also one of the first questions\nfrom Thomas upthread.\n--\nMichael","msg_date":"Thu, 21 Apr 2022 13:07:18 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"> Why does it not allow that? Is that just a limitation of the\n> compiler? If yes, what is the error happening? This question is not\n> exactly answered yet as of this thread. I may be missing a reference\n> about that in the upstream docs, but I see nowhere an explanation\n> about the reason and the why. That's also one of the first questions\n> from Thomas upthread.\n\nThe following error occurs:\n\nLINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64'\ntarget machine; link without '/DYNAMICBASE:NO\n\nThis seems to be a deliberate restriction for Arm64 targets. However, no\nreferences were provided. To clarify, I have posted a question [1] on the\ncommunity channel of Visual Studio.\n\nNiyas\n\n[1]\nhttps://developercommunity.visualstudio.com/t/LINK-:-fatal-error-LNK1246:-DYNAMICBAS/10020163\n\n\n\n\n\nOn Thu, 21 Apr 2022 at 05:07, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Wed, Apr 20, 2022 at 10:43:06AM +0100, Niyas Sait wrote:\n> >> This issue is still lying around, and you may have been lucky. Would\n> >> there be any issues to remove this change to get a basic support in?\n> >> As mentioned upthread, there is a long history of Postgres with ASLR.\n> >\n> > MSVC linker doesn't allow non-random base addresses for Arm64 platforms.\n> > It is needed for basic support.\n>\n> Why does it not allow that? Is that just a limitation of the\n> compiler? If yes, what is the error happening? This question is not\n> exactly answered yet as of this thread. I may be missing a reference\n> about that in the upstream docs, but I see nowhere an explanation\n> about the reason and the why. That's also one of the first questions\n> from Thomas upthread.\n> --\n> Michael\n>\n\n> Why does it not allow that? Is that just a limitation of the> compiler? If yes, what is the error happening? This question is not> exactly answered yet as of this thread. I may be missing a reference> about that in the upstream docs, but I see nowhere an explanation> about the reason and the why. That's also one of the first questions> from Thomas upthread.The following error occurs:LINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64' target machine; link without '/DYNAMICBASE:NOThis seems to be a deliberate restriction for Arm64 targets. However, no references were provided. To clarify, I have posted a question [1] on the community channel of Visual Studio.Niyas[1] https://developercommunity.visualstudio.com/t/LINK-:-fatal-error-LNK1246:-DYNAMICBAS/10020163On Thu, 21 Apr 2022 at 05:07, Michael Paquier <michael@paquier.xyz> wrote:On Wed, Apr 20, 2022 at 10:43:06AM +0100, Niyas Sait wrote:\n>> This issue is still lying around, and you may have been lucky. Would\n>> there be any issues to remove this change to get a basic support in?\n>> As mentioned upthread, there is a long history of Postgres with ASLR.\n> \n> MSVC linker doesn't allow non-random base addresses for Arm64 platforms.\n> It is needed for basic support.\n\nWhy does it not allow that? Is that just a limitation of the\ncompiler? If yes, what is the error happening? This question is not\nexactly answered yet as of this thread. I may be missing a reference\nabout that in the upstream docs, but I see nowhere an explanation\nabout the reason and the why. That's also one of the first questions\nfrom Thomas upthread.\n--\nMichael","msg_date":"Thu, 21 Apr 2022 10:21:04 +0100","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Apr 21, 2022 at 10:21:04AM +0100, Niyas Sait wrote:\n> The following error occurs:\n> \n> LINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64'\n> target machine; link without '/DYNAMICBASE:NO\n\nOkay, that's interesting. In light of things like 7f3e17b, that may\nbe annoying. Perhaps newer Windows versions are able to handle that\nbetter. I am wondering if it would be worth re-evaluating this\nchange, and attempt to re-enable that in environments other than\narm64. This could become interesting if we consider that there have\nbeen talks to cut the support for a bunch of Windows versions to focus\non the newer ones only.\n\n> This seems to be a deliberate restriction for Arm64 targets. However, no\n> references were provided. To clarify, I have posted a question [1] on the\n> community channel of Visual Studio.\n\nThanks for doing that! Your post is just a couple of days old, let's\nsee if we get a reply on that.\n--\nMichael","msg_date":"Mon, 25 Apr 2022 10:17:16 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"> Thanks for doing that! Your post is just a couple of days old, let's\n> see if we get a reply on that.\n\nMicrosoft updated documentation [1] and clarified that ASLR cannot be\ndisabled for Arm64 targets.\n\nBecause ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures,\n> /DYNAMICBASE:NO isn't supported for these targets.\n\n\nThanks\nNiyas\n\n[1]\nhttps://docs.microsoft.com/en-us/cpp/build/reference/dynamicbase?view=msvc-170\n\n\nOn Mon, 25 Apr 2022 at 02:17, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Thu, Apr 21, 2022 at 10:21:04AM +0100, Niyas Sait wrote:\n> > The following error occurs:\n> >\n> > LINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64'\n> > target machine; link without '/DYNAMICBASE:NO\n>\n> Okay, that's interesting. In light of things like 7f3e17b, that may\n> be annoying. Perhaps newer Windows versions are able to handle that\n> better. I am wondering if it would be worth re-evaluating this\n> change, and attempt to re-enable that in environments other than\n> arm64. This could become interesting if we consider that there have\n> been talks to cut the support for a bunch of Windows versions to focus\n> on the newer ones only.\n>\n> > This seems to be a deliberate restriction for Arm64 targets. However, no\n> > references were provided. To clarify, I have posted a question [1] on the\n> > community channel of Visual Studio.\n>\n> Thanks for doing that! Your post is just a couple of days old, let's\n> see if we get a reply on that.\n> --\n> Michael\n>\n\n> Thanks for doing that! Your post is just a couple of days old, let's> see if we get a reply on that.Microsoft updated documentation [1] and clarified that ASLR cannot be disabled for Arm64 targets.Because ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures, /DYNAMICBASE:NO isn't supported for these targets.ThanksNiyas[1] https://docs.microsoft.com/en-us/cpp/build/reference/dynamicbase?view=msvc-170On Mon, 25 Apr 2022 at 02:17, Michael Paquier <michael@paquier.xyz> wrote:On Thu, Apr 21, 2022 at 10:21:04AM +0100, Niyas Sait wrote:\n> The following error occurs:\n> \n> LINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64'\n> target machine; link without '/DYNAMICBASE:NO\n\nOkay, that's interesting. In light of things like 7f3e17b, that may\nbe annoying. Perhaps newer Windows versions are able to handle that\nbetter. I am wondering if it would be worth re-evaluating this\nchange, and attempt to re-enable that in environments other than\narm64. This could become interesting if we consider that there have\nbeen talks to cut the support for a bunch of Windows versions to focus\non the newer ones only.\n\n> This seems to be a deliberate restriction for Arm64 targets. However, no\n> references were provided. To clarify, I have posted a question [1] on the\n> community channel of Visual Studio.\n\nThanks for doing that! Your post is just a couple of days old, let's\nsee if we get a reply on that.\n--\nMichael","msg_date":"Mon, 9 May 2022 12:44:22 +0100","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Mon, May 09, 2022 at 12:44:22PM +0100, Niyas Sait wrote:\n> Microsoft updated documentation [1] and clarified that ASLR cannot be\n> disabled for Arm64 targets.\n> \n> Because ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures,\n> > /DYNAMICBASE:NO isn't supported for these targets.\n\nBetter than nothing, I guess, though this does not stand as a reason\nexplaining why this choice has been made. And it looks like we will\nnever know. We are going to need more advanced testing to check that\nDYNAMICBASE is stable enough if we begin to enable it. Perhaps we\nshould really look at that for all the build targets we currently\nsupport rather than just ARM, for the most modern Win32 environments\nif we are going to cut support for most of the oldest releases..\n--\nMichael","msg_date":"Tue, 10 May 2022 10:01:55 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hello,\n\nPlease find a new patch (no further changes) rebased on top of the master.\n\nOn Tue, 10 May 2022 at 02:02, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Mon, May 09, 2022 at 12:44:22PM +0100, Niyas Sait wrote:\n> > Microsoft updated documentation [1] and clarified that ASLR cannot be\n> > disabled for Arm64 targets.\n> >\n> > Because ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures,\n> > > /DYNAMICBASE:NO isn't supported for these targets.\n>\n> Better than nothing, I guess, though this does not stand as a reason\n> explaining why this choice has been made. And it looks like we will\n> never know. We are going to need more advanced testing to check that\n> DYNAMICBASE is stable enough if we begin to enable it. Perhaps we\n> should really look at that for all the build targets we currently\n> support rather than just ARM, for the most modern Win32 environments\n> if we are going to cut support for most of the oldest releases..\n> --\n> Michael\n>\n\n\n-- \nNiyas Sait\nLinaro Limited","msg_date":"Thu, 14 Jul 2022 09:36:13 +0100","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-05-10 10:01:55 +0900, Michael Paquier wrote:\n> On Mon, May 09, 2022 at 12:44:22PM +0100, Niyas Sait wrote:\n> > Microsoft updated documentation [1] and clarified that ASLR cannot be\n> > disabled for Arm64 targets.\n> > \n> > Because ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures,\n> > > /DYNAMICBASE:NO isn't supported for these targets.\n> \n> Better than nothing, I guess, though this does not stand as a reason\n> explaining why this choice has been made. And it looks like we will\n> never know. We are going to need more advanced testing to check that\n> DYNAMICBASE is stable enough if we begin to enable it. Perhaps we\n> should really look at that for all the build targets we currently\n> support rather than just ARM, for the most modern Win32 environments\n> if we are going to cut support for most of the oldest releases..\n\nI accidentally did a lot of testing with DYNAMICBASE - I accidentally\nmistranslated MSBuildProject.pm's <RandomizedBaseAddress>false</> to adding\n/DYNAMICBASE in the meson port. Survived several hundred CI cycles and dozens\nof local runs without observing any problems related to that.\n\nWhich doesn't surprise me, given the way we reserve space in the new process.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 25 Aug 2022 18:29:07 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Aug 25, 2022 at 06:29:07PM -0700, Andres Freund wrote:\n> I accidentally did a lot of testing with DYNAMICBASE - I accidentally\n> mistranslated MSBuildProject.pm's <RandomizedBaseAddress>false</> to adding\n> /DYNAMICBASE in the meson port. Survived several hundred CI cycles and dozens\n> of local runs without observing any problems related to that.\n> \n> Which doesn't surprise me, given the way we reserve space in the new process.\n\nStill, we had problems with that in the past and I don't recall huge\nchanges in the way we allocate shmem on WIN32. Could there be some\nstuff in Windows itself that would explain more stability? I would\nnot mind switching back to DYNAMICBASE on HEAD seeing your results.\nThat would simplify this thread's patch a bit as well.\n--\nMichael","msg_date":"Fri, 26 Aug 2022 11:09:41 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-08-26 11:09:41 +0900, Michael Paquier wrote:\n> On Thu, Aug 25, 2022 at 06:29:07PM -0700, Andres Freund wrote:\n> > I accidentally did a lot of testing with DYNAMICBASE - I accidentally\n> > mistranslated MSBuildProject.pm's <RandomizedBaseAddress>false</> to adding\n> > /DYNAMICBASE in the meson port. Survived several hundred CI cycles and dozens\n> > of local runs without observing any problems related to that.\n> > \n> > Which doesn't surprise me, given the way we reserve space in the new process.\n> \n> Still, we had problems with that in the past and I don't recall huge\n> changes in the way we allocate shmem on WIN32.\n\nIt's really not great that 7f3e17b4827 disabled randomization without even a\ncomment as to why...\n\nThere actually seems to have been a lot of work in that area. See 617dc6d299c\n/ bcbf2346d69 and as well as the retrying in 45e004fb4e39. Those weren't\nprevented by us disabling ASLR.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 25 Aug 2022 20:29:43 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Aug 25, 2022 at 08:29:43PM -0700, Andres Freund wrote:\n> It's really not great that 7f3e17b4827 disabled randomization without even a\n> comment as to why...\n\nThis story is on this thread, with some processes not able to start:\nhttps://www.postgresql.org/message-id/BD0D89EC2438455C9DE0DC94D36912F4@maumau\n\n> There actually seems to have been a lot of work in that area. See 617dc6d299c\n> / bcbf2346d69 and as well as the retrying in 45e004fb4e39. Those weren't\n> prevented by us disabling ASLR.\n\nIndeed. 45e004f looks like the most interesting bit here. FWIW, I\nwould not mind re-enabling that on HEAD, as of something like the\nattached. I have done a dozen of runs without seeing a test failure,\nand knowing that we don't support anything older than Win10 makes me\nfeel safer about this change. Any objections?\n--\nMichael","msg_date":"Sat, 27 Aug 2022 11:33:51 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> Indeed. 45e004f looks like the most interesting bit here. FWIW, I\n> would not mind re-enabling that on HEAD, as of something like the\n> attached. I have done a dozen of runs without seeing a test failure,\n> and knowing that we don't support anything older than Win10 makes me\n> feel safer about this change. Any objections?\n\nWe're early enough in the v16 cycle to have plenty of time to detect\nany problems, so I see little reason not to try it.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 26 Aug 2022 23:02:58 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-08-27 11:33:51 +0900, Michael Paquier wrote:\n> FWIW, I would not mind re-enabling that on HEAD, as of something like the\n> attached. I have done a dozen of runs without seeing a test failure, and\n> knowing that we don't support anything older than Win10 makes me feel safer\n> about this change. Any objections?\n\n+1. We don't have a choice about it on other architectures and it seems like a\nbad idea to disable on its own.\n\nI checked, and it looks like we didn't add --disable-dynamicbase, so ASLR\nwasn't disabled for mingw builds.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sat, 27 Aug 2022 12:27:57 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Sat, Aug 27, 2022 at 12:27:57PM -0700, Andres Freund wrote:\n> I checked, and it looks like we didn't add --disable-dynamicbase, so ASLR\n> wasn't disabled for mingw builds.\n\nTrue enough. I have applied the patch to re-enable that on HEAD.\nLet's now see what happens in the next couple of days. Popcorn is\nready here.\n--\nMichael","msg_date":"Sun, 28 Aug 2022 16:23:08 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> True enough. I have applied the patch to re-enable that on HEAD.\n> Let's now see what happens in the next couple of days. Popcorn is\n> ready here.\n\nHmm:\n\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-28%2010%3A30%3A29\n\nMaybe that's just unrelated noise, but it sure looks weird: after\npassing the core regression tests in \"make check\", it failed in\npg_upgrade's run with\n\ndiff -w -U3 H:/prog/bf/root/HEAD/pgsql.build/src/test/regress/expected/xml.out H:/prog/bf/root/HEAD/pgsql.build/src/bin/pg_upgrade/tmp_check/results/xml.out\n--- H:/prog/bf/root/HEAD/pgsql.build/src/test/regress/expected/xml.out\tWed May 18 18:30:12 2022\n+++ H:/prog/bf/root/HEAD/pgsql.build/src/bin/pg_upgrade/tmp_check/results/xml.out\tSun Aug 28 06:55:14 2022\n@@ -1557,7 +1557,7 @@\n \\\\x\n SELECT * FROM XMLTABLE('*' PASSING '<e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>' COLUMNS x xml PATH 'node()', y xml PATH '/');\n -[ RECORD 1 ]-----------------------------------------------------------\n-x | pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\n+x | pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\n y | <e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>+\n | \n \n\nI don't recall ever seeing a failure quite like this.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 28 Aug 2022 10:09:53 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-08-28 10:09:53 -0400, Tom Lane wrote:\n> Michael Paquier <michael@paquier.xyz> writes:\n> > True enough. I have applied the patch to re-enable that on HEAD.\n> > Let's now see what happens in the next couple of days. Popcorn is\n> > ready here.\n>\n> Hmm:\n>\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-28%2010%3A30%3A29\n>\n> Maybe that's just unrelated noise, but it sure looks weird: after\n> passing the core regression tests in \"make check\", it failed in\n> pg_upgrade's run with\n>\n> diff -w -U3 H:/prog/bf/root/HEAD/pgsql.build/src/test/regress/expected/xml.out H:/prog/bf/root/HEAD/pgsql.build/src/bin/pg_upgrade/tmp_check/results/xml.out\n> --- H:/prog/bf/root/HEAD/pgsql.build/src/test/regress/expected/xml.out\tWed May 18 18:30:12 2022\n> +++ H:/prog/bf/root/HEAD/pgsql.build/src/bin/pg_upgrade/tmp_check/results/xml.out\tSun Aug 28 06:55:14 2022\n> @@ -1557,7 +1557,7 @@\n> \\\\x\n> SELECT * FROM XMLTABLE('*' PASSING '<e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>' COLUMNS x xml PATH 'node()', y xml PATH '/');\n> -[ RECORD 1 ]-----------------------------------------------------------\n> -x | pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\n> +x | pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\n> y | <e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>+\n> |\n\nPretty weird, agreed. But hard to see how it could be caused by the\nrandomization change, except that perhaps it could highlight a preexisting\nbug?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sun, 28 Aug 2022 08:41:54 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2022-08-28 10:09:53 -0400, Tom Lane wrote:\n>> -x | pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\n>> +x | pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\n\n> Pretty weird, agreed. But hard to see how it could be caused by the\n> randomization change, except that perhaps it could highlight a preexisting\n> bug?\n\nI have no idea either. I agree there *shouldn't* be any connection,\nso if ASLR is somehow triggering this then whatever is failing is\nalmost certainly buggy on its own terms. But there's a lot of\nmoving parts here (mumble libxml mumble). I'm going to wait to see\nif it reproduces before spending much effort.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 28 Aug 2022 11:51:19 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Sun, Aug 28, 2022 at 11:51:19AM -0400, Tom Lane wrote:\n> I have no idea either. I agree there *shouldn't* be any connection,\n> so if ASLR is somehow triggering this then whatever is failing is\n> almost certainly buggy on its own terms. But there's a lot of\n> moving parts here (mumble libxml mumble). I'm going to wait to see\n> if it reproduces before spending much effort.\n\nI have noticed that yesterday, but cannot think much about it. This\nbasically changes the position of \"<!--c1-->\" for the first record,\nleaving the second one untouched:\n<!--c1--><?pi arg?><![CDATA[&ent1]]>\n<?pi arg?><![CDATA[&ent1]]><!--c1-->\n\nI am not used to xmltable(), but I wonder if there is something in one\nof these support functions in xml.c that gets influenced by the\nrandomization. That sounds a bit hairy as make check passed in\nbowerbird, and I have noticed at least two other Windows hosts running\nTAP that passed. Or that's just something with libxml itself.\n--\nMichael","msg_date":"Mon, 29 Aug 2022 10:12:20 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Mon, Aug 29, 2022 at 10:12:20AM +0900, Michael Paquier wrote:\n> I have noticed that yesterday, but cannot think much about it. This\n> basically changes the position of \"<!--c1-->\" for the first record,\n> leaving the second one untouched:\n> <!--c1--><?pi arg?><![CDATA[&ent1]]>\n> <?pi arg?><![CDATA[&ent1]]><!--c1-->\n> \n> I am not used to xmltable(), but I wonder if there is something in one\n> of these support functions in xml.c that gets influenced by the\n> randomization. That sounds a bit hairy as make check passed in\n> bowerbird, and I have noticed at least two other Windows hosts running\n> TAP that passed. Or that's just something with libxml itself.\n\nThis is amazing. The issue has showed up a second time in a row in\nbowerbird, as of:\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-29%2013%3A30%3A32\n\nI don't know what to think about ASLR that manipulates the comment in\nthis XML object under VS 2017 (perhaps a compiler issue?), but it\nshould be possible to go back to green simply by removing \"<!--c1-->\"\nfrom the input string. Creating an extra output pattern here would be\nvery costly, as xml_2.out and xml.out have outputs for --with-libxml.\n\nWould people object if I do that for now?\n--\nMichael","msg_date":"Tue, 30 Aug 2022 08:48:38 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> This is amazing. The issue has showed up a second time in a row in\n> bowerbird, as of:\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-29%2013%3A30%3A32\n\nThat is fascinating.\n\n> I don't know what to think about ASLR that manipulates the comment in\n> this XML object under VS 2017 (perhaps a compiler issue?), but it\n> should be possible to go back to green simply by removing \"<!--c1-->\"\n> from the input string. Creating an extra output pattern here would be\n> very costly, as xml_2.out and xml.out have outputs for --with-libxml.\n\n> Would people object if I do that for now?\n\nLet's let it go for a few more runs. I want to know whether it\nreproduces 100% or only sometimes.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 29 Aug 2022 19:58:31 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-08-30 08:48:38 +0900, Michael Paquier wrote:\n> On Mon, Aug 29, 2022 at 10:12:20AM +0900, Michael Paquier wrote:\n> > I have noticed that yesterday, but cannot think much about it. This\n> > basically changes the position of \"<!--c1-->\" for the first record,\n> > leaving the second one untouched:\n> > <!--c1--><?pi arg?><![CDATA[&ent1]]>\n> > <?pi arg?><![CDATA[&ent1]]><!--c1-->\n> > \n> > I am not used to xmltable(), but I wonder if there is something in one\n> > of these support functions in xml.c that gets influenced by the\n> > randomization. That sounds a bit hairy as make check passed in\n> > bowerbird, and I have noticed at least two other Windows hosts running\n> > TAP that passed. Or that's just something with libxml itself.\n> \n> This is amazing. The issue has showed up a second time in a row in\n> bowerbird, as of:\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-29%2013%3A30%3A32\n> \n> I don't know what to think about ASLR that manipulates the comment in\n> this XML object under VS 2017 (perhaps a compiler issue?), but it\n> should be possible to go back to green simply by removing \"<!--c1-->\"\n> from the input string. Creating an extra output pattern here would be\n> very costly, as xml_2.out and xml.out have outputs for --with-libxml.\n> \n> Would people object if I do that for now?\n\nI do. Clearly something is wrong, what do we gain by averting our eyes? An\nalternative output file strikes me as an equally bad idea - it's not like this\nis a case where there are legitimately different outputs?\n\nBowerbird is the only animal testing libxml with visual studio afaics. And CI\ndoesn't have libxml either. So we have no idea if it's the msvc version that\nmatters.\n\nAndrew, what version of libxml is installed? Afaict we don't see that anywhere\nwith msvc.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 29 Aug 2022 17:09:54 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-08-29 19:58:31 -0400, Tom Lane wrote:\n> Michael Paquier <michael@paquier.xyz> writes:\n> > This is amazing. The issue has showed up a second time in a row in\n> > bowerbird, as of:\n> > https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-29%2013%3A30%3A32\n> \n> That is fascinating.\n\n> > I don't know what to think about ASLR that manipulates the comment in\n> > this XML object under VS 2017 (perhaps a compiler issue?), but it\n> > should be possible to go back to green simply by removing \"<!--c1-->\"\n> > from the input string. Creating an extra output pattern here would be\n> > very costly, as xml_2.out and xml.out have outputs for --with-libxml.\n> \n> > Would people object if I do that for now?\n> \n> Let's let it go for a few more runs. I want to know whether it\n> reproduces 100% or only sometimes.\n\nThe weirdest part is that it only happens as part of the the pg_upgrade test.\n\nI just tested it in my test windows vm (win 10, vs 2019), with a build of\nlibxml I had around (2.9.7), and the regression tests passed both \"normally\"\nand within pg_upgrade.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 29 Aug 2022 17:33:56 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Mon, Aug 29, 2022 at 05:33:56PM -0700, Andres Freund wrote:\n> The weirdest part is that it only happens as part of the the pg_upgrade test.\n\nmake check has just failed:\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-30%2001%3A15%3A13\n\n> I just tested it in my test windows vm (win 10, vs 2019), with a build of\n> libxml I had around (2.9.7), and the regression tests passed both \"normally\"\n> and within pg_upgrade.\n\nbowerbird is feeling from c:\\\\prog\\\\3p64\\\\include\\\\libxml2 and\nc:\\\\prog\\\\3p64\\\\lib\\\\libxml2.lib. I am not sure which version of\nlibxml this is, and the other animals of MSVC don't use libxml so it\nis not possible to correlate that only to VS 2017, either.\n--\nMichael","msg_date":"Tue, 30 Aug 2022 10:35:35 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Mon, Aug 29, 2022 at 05:33:56PM -0700, Andres Freund wrote:\n>> The weirdest part is that it only happens as part of the the pg_upgrade test.\n\n> make check has just failed:\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-30%2001%3A15%3A13\n\nSo it *is* probabilistic, which is pretty much what you'd expect\nif ASLR triggers it. That brings us no closer to understanding\nwhat the mechanism is, though.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 29 Aug 2022 22:00:10 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Mon, Aug 29, 2022 at 10:00:10PM -0400, Tom Lane wrote:\n> Michael Paquier <michael@paquier.xyz> writes:\n> > On Mon, Aug 29, 2022 at 05:33:56PM -0700, Andres Freund wrote:\n> >> The weirdest part is that it only happens as part of the the pg_upgrade test.\n> \n> > make check has just failed:\n> > https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-30%2001%3A15%3A13\n> \n> So it *is* probabilistic, which is pretty much what you'd expect\n> if ASLR triggers it. That brings us no closer to understanding\n> what the mechanism is, though.\n\nThere have been more failures, always switching the input from\n\"pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\"\nto \"pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\".\n\nUsing a PATH of node() influences the output. I am not verse unto\nXMLTABLE, but could it be an issue where each node is parsed and we\nhave something like a qsort() applied on the pointer addresses for\neach part or something like that, causing the output to become\nunstable?\n--\nMichael","msg_date":"Wed, 31 Aug 2022 08:36:01 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> Using a PATH of node() influences the output. I am not verse unto\n> XMLTABLE, but could it be an issue where each node is parsed and we\n> have something like a qsort() applied on the pointer addresses for\n> each part or something like that, causing the output to become\n> unstable?\n\nI'm not sure. I dug into this enough to find that the output from\nthis query is generated in xml.c lines 4635ff:\n\n /* Concatenate serialized values */\n initStringInfo(&str);\n for (int i = 0; i < count; i++)\n {\n textstr =\n xml_xmlnodetoxmltype(xpathobj->nodesetval->nodeTab[i],\n xtCxt->xmlerrcxt);\n\n appendStringInfoText(&str, textstr);\n }\n cstr = str.data;\n\nSo evidently, the problem occurs because the elements of the\nnodesetval->nodeTab[] array are in a different order than we expect.\n\nI looked into <xpath.h> and found the definition of the \"nodesetval\"\nstruct, and the comments are eye-opening to say the least:\n\n/*\n * A node-set (an unordered collection of nodes without duplicates).\n */\ntypedef struct _xmlNodeSet xmlNodeSet;\ntypedef xmlNodeSet *xmlNodeSetPtr;\nstruct _xmlNodeSet {\n int nodeNr;\t\t\t/* number of nodes in the set */\n int nodeMax;\t\t/* size of the array as allocated */\n xmlNodePtr *nodeTab;\t/* array of nodes in no particular order */\n /* @@ with_ns to check wether namespace nodes should be looked at @@ */\n};\n\nIt seems like maybe we're relying on an ordering we should not.\nYet surely the ordering of the pieces of this output is meaningful?\nAre we using the wrong libxml API to create this result?\n\nAnyway, I'm now suspicious that we've accidentally exposed a logic\nbug in the XMLTABLE code, rather than anything wrong with the\nASLR stuff.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 30 Aug 2022 20:07:27 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, Aug 31, 2022 at 08:36:01AM +0900, Michael Paquier wrote:\n> There have been more failures, always switching the input from\n> \"pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\"\n> to \"pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\".\n> \n> Using a PATH of node() influences the output. I am not verse unto\n> XMLTABLE, but could it be an issue where each node is parsed and we\n> have something like a qsort() applied on the pointer addresses for\n> each part or something like that, causing the output to become\n> unstable?\n\nHmm. I think that I may have an idea here after looking at our xml.c\nand xpath.c in libxml2/. From what I understand, we process the PATH\nthrough XmlTableGetValue() that builds a XML node path in\nxmlXPathCompiledEval(). The interesting part comes from libxml2's\nxmlXPathCompiledEvalInternal(), where I think we don't apply a sort on\nthe contents generated. Hence, I am wondering if the solution here\nwould be to do one xmlXPathNodeSetSort(xpathobj->nodesetval) after\ncompiling the path with xmlXPathCompiledEval() in XmlTableGetValue().\nThis should ensure that the items are ordered even if ASLR mixes if\nthe pointer positions.\n\nA complete solution would involve more code paths, but we'd need only\none change in XmlTableGetValue() for the current regression tests to\nwork. I don't have an environment where I can reproduce that, so that\nwould be up to the buildfarm to stress this solution..\n\nThoughts?\n--\nMichael","msg_date":"Wed, 31 Aug 2022 09:15:00 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, Aug 30, 2022 at 08:07:27PM -0400, Tom Lane wrote:\n> It seems like maybe we're relying on an ordering we should not.\n> Yet surely the ordering of the pieces of this output is meaningful?\n> Are we using the wrong libxml API to create this result?\n> \n> Anyway, I'm now suspicious that we've accidentally exposed a logic\n> bug in the XMLTABLE code, rather than anything wrong with the\n> ASLR stuff.\n\nFunny. I have reached basically the same conclusion as you a couple\nof minutes ago, but I also think that I have found what we need to do\nhere to ensure the ordering of the nodes generated by the libxml\ncode.\n--\nMichael","msg_date":"Wed, 31 Aug 2022 09:16:50 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> Hence, I am wondering if the solution here\n> would be to do one xmlXPathNodeSetSort(xpathobj->nodesetval) after\n> compiling the path with xmlXPathCompiledEval() in XmlTableGetValue().\n\nHmm, I see that function declared in <xpathInternals.h>, which\nsure doesn't give me a warm feeling that we're supposed to call it\ndirectly. But maybe there's an approved way to get the result.\n\nOr perhaps this test case is wrong, and instead of \"node()\" we\nneed to write something that specifies a sorted result?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 30 Aug 2022 20:28:58 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, Aug 30, 2022 at 08:28:58PM -0400, Tom Lane wrote:\n> Hmm, I see that function declared in <xpathInternals.h>, which\n> sure doesn't give me a warm feeling that we're supposed to call it\n> directly. But maybe there's an approved way to get the result.\n> \n> Or perhaps this test case is wrong, and instead of \"node()\" we\n> need to write something that specifies a sorted result?\n\nI am no specialist on this matter.. Still, looking around, this looks\nto be the kind of job that xsl:sort would do? We cannot use it here\nfor an XML-only context, though, and it does not seem like there is a\ndirect way to enforce the ordering per the nature of the XPath\nlanguage, either. I may be missing something of course.\n\nPlease note that I am also getting a bit the cold feet with the idea\nof enforcing a sorting of the elements in the generated output,\nactually, as this could have a performance penalty for all users\nparticularly on large blobs of data.\n\nAt the end, I'd like to think that it would be wiser to just remove\nentirely the test for node() or reduce the contents of the string to\nbe able to strictly control the output order (say a simple\n'<e>pre<![CDATA[&ent1]]>post</e>' would do the trick). I cannot think\nabout a better idea now. Note that removing the test case where we\nhave node() has no impact on the coverage of xml.c.\n--\nMichael","msg_date":"Wed, 31 Aug 2022 15:09:16 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> At the end, I'd like to think that it would be wiser to just remove\n> entirely the test for node() or reduce the contents of the string to\n> be able to strictly control the output order (say a simple\n> '<e>pre<![CDATA[&ent1]]>post</e>' would do the trick). I cannot think\n> about a better idea now. Note that removing the test case where we\n> have node() has no impact on the coverage of xml.c.\n\nYeah, I confirm that: local code-coverage testing shows no change\nin the number of lines reached in xml.c when I remove that column:\n\n-SELECT * FROM XMLTABLE('*' PASSING '<e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>' COLUMNS x xml PATH 'node()', y xml PATH '/');\n+SELECT * FROM XMLTABLE('*' PASSING '<e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>' COLUMNS y xml PATH '/');\n\nDropping the query altogether does result in a reduction in the\nnumber of lines hit.\n\nI did not check the other XMLTABLE infrastructure, so what probably\nis best to do is keep the two output columns but change 'node()'\nto something with a more stable result; any preferences?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 31 Aug 2022 13:33:53 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, Aug 31, 2022 at 01:33:53PM -0400, Tom Lane wrote:\n> I did not check the other XMLTABLE infrastructure, so what probably\n> is best to do is keep the two output columns but change 'node()'\n> to something with a more stable result; any preferences?\n\nThe input object could also be reworked so as we would not have any\nordering issues, say \"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\".\nChanging only the path, you could use \"/e/n2\" instead of \"node()\", so\nas we'd always get the most internal member.\n--\nMichael","msg_date":"Thu, 1 Sep 2022 09:32:38 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> The input object could also be reworked so as we would not have any\n> ordering issues, say \"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\".\n> Changing only the path, you could use \"/e/n2\" instead of \"node()\", so\n> as we'd always get the most internal member.\n\nDone that way.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 31 Aug 2022 22:22:06 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, Aug 31, 2022 at 10:22:06PM -0400, Tom Lane wrote:\n> Michael Paquier <michael@paquier.xyz> writes:\n> > The input object could also be reworked so as we would not have any\n> > ordering issues, say \"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\".\n> > Changing only the path, you could use \"/e/n2\" instead of \"node()\", so\n> > as we'd always get the most internal member.\n> \n> Done that way.\n\nCool, thanks. bowerbird looks happy after its first run.\n\nAn argument in favor of backpatching is if one decides to build the\ncode with MSVC and patch the scripts to enable ASLR. However, nobody\nhas complained about that yet, so fine by me to leave this change only\non HEAD for now.\n--\nMichael","msg_date":"Thu, 1 Sep 2022 13:41:59 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"2022年9月1日(木) 13:42 Michael Paquier <michael@paquier.xyz>:\n>\n> On Wed, Aug 31, 2022 at 10:22:06PM -0400, Tom Lane wrote:\n> > Michael Paquier <michael@paquier.xyz> writes:\n> > > The input object could also be reworked so as we would not have any\n> > > ordering issues, say \"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\".\n> > > Changing only the path, you could use \"/e/n2\" instead of \"node()\", so\n> > > as we'd always get the most internal member.\n> >\n> > Done that way.\n>\n> Cool, thanks. bowerbird looks happy after its first run.\n>\n> An argument in favor of backpatching is if one decides to build the\n> code with MSVC and patch the scripts to enable ASLR. However, nobody\n> has complained about that yet, so fine by me to leave this change only\n> on HEAD for now.\n\nApologies for the thread bump, but there is an entry for this thread\nin the current CF\npreviously marked \"Needs review\":\n\n https://commitfest.postgresql.org/40/3561/\n\nI've gone ahead and marked it as \"Committed\", as that appears to have happened\nback in August as 36389a060c.\n\nIf for whatever reason that was the Wrong Thing To Do, please let me know.\n\nRegards\n\nIan Barwick\n\n\n","msg_date":"Thu, 3 Nov 2022 17:38:09 +0900","msg_from":"Ian Lawrence Barwick <barwick@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"> I've gone ahead and marked it as \"Committed\", as that appears to have\nhappened\n> back in August as 36389a060c.\n> If for whatever reason that was the Wrong Thing To Do, please let me know.\n\n36389a060c commit enables ASLR for windows but does not include other\nchanges required for Arm64.\n\nI've attached a new version of the patch which excludes the already merged\nASLR changes and add\nsmall changes to handle latest changes in the build scripts.\n\nOn Thu, 3 Nov 2022 at 08:38, Ian Lawrence Barwick <barwick@gmail.com> wrote:\n\n> 2022年9月1日(木) 13:42 Michael Paquier <michael@paquier.xyz>:\n> >\n> > On Wed, Aug 31, 2022 at 10:22:06PM -0400, Tom Lane wrote:\n> > > Michael Paquier <michael@paquier.xyz> writes:\n> > > > The input object could also be reworked so as we would not have any\n> > > > ordering issues, say\n> \"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\".\n> > > > Changing only the path, you could use \"/e/n2\" instead of \"node()\", so\n> > > > as we'd always get the most internal member.\n> > >\n> > > Done that way.\n> >\n> > Cool, thanks. bowerbird looks happy after its first run.\n> >\n> > An argument in favor of backpatching is if one decides to build the\n> > code with MSVC and patch the scripts to enable ASLR. However, nobody\n> > has complained about that yet, so fine by me to leave this change only\n> > on HEAD for now.\n>\n> Apologies for the thread bump, but there is an entry for this thread\n> in the current CF\n> previously marked \"Needs review\":\n>\n> https://commitfest.postgresql.org/40/3561/\n>\n> I've gone ahead and marked it as \"Committed\", as that appears to have\n> happened\n> back in August as 36389a060c.\n>\n> If for whatever reason that was the Wrong Thing To Do, please let me know.\n>\n> Regards\n>\n> Ian Barwick\n>\n\n\n-- \nNiyas Sait\nLinaro Limited","msg_date":"Thu, 3 Nov 2022 11:06:46 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Nov 03, 2022 at 11:06:46AM +0000, Niyas Sait wrote:\n> > I've gone ahead and marked it as \"Committed\", as that appears to have\n> happened\n> > back in August as 36389a060c.\n> > If for whatever reason that was the Wrong Thing To Do, please let me know.\n> \n> 36389a060c commit enables ASLR for windows but does not include other\n> changes required for Arm64.\n\nYes, marking this patch as committed is incorrect. That's just me\nlagging behind for the review :)\n\nI have switched that back as \"Needs review\" for the moment.\n\n> I've attached a new version of the patch which excludes the already merged\n> ASLR changes and add\n> small changes to handle latest changes in the build scripts.\n\nThanks.\n--\nMichael","msg_date":"Thu, 3 Nov 2022 20:43:04 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-11-03 11:06:46 +0000, Niyas Sait wrote:\n> I've attached a new version of the patch which excludes the already merged\n> ASLR changes and add\n> small changes to handle latest changes in the build scripts.\n\nNote that we're planning to remove the custom windows build scripts before the\nnext release, relying on the meson build instead.\n\n\n> diff --git a/src/include/storage/s_lock.h b/src/include/storage/s_lock.h\n> index 8b19ab160f..bf6a6dba35 100644\n> --- a/src/include/storage/s_lock.h\n> +++ b/src/include/storage/s_lock.h\n> @@ -708,13 +708,21 @@ typedef LONG slock_t;\n> #define SPIN_DELAY() spin_delay()\n> \n> /* If using Visual C++ on Win64, inline assembly is unavailable.\n> - * Use a _mm_pause intrinsic instead of rep nop.\n> + * Use _mm_pause (x64) or __isb(arm64) intrinsic instead of rep nop.\n> */\n> #if defined(_WIN64)\n> static __forceinline void\n> spin_delay(void)\n> {\n> +#ifdef _M_ARM64\n> +\t/*\n> +\t * arm64 way of hinting processor for spin loops optimisations\n> +\t * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\n> +\t*/\n> +\t__isb(_ARM64_BARRIER_SY);\n> +#else\n> \t_mm_pause();\n> +#endif\n> }\n> #else\n> static __forceinline void\n\nI think we should just apply this, there seems very little risk of making\nanything worse, given the gating to _WIN64 && _M_ARM64.\n\n\n> diff --git a/src/port/pg_crc32c_armv8.c b/src/port/pg_crc32c_armv8.c\n> index 9e301f96f6..981718752f 100644\n> --- a/src/port/pg_crc32c_armv8.c\n> +++ b/src/port/pg_crc32c_armv8.c\n> @@ -14,7 +14,9 @@\n> */\n> #include \"c.h\"\n> \n> +#ifndef _MSC_VER\n> #include <arm_acle.h>\n> +#endif\n> \n> #include \"port/pg_crc32c.h\"\n\nThis won't suffice with the meson build, since the relevant configure test\nalso uses arm_acle.h:\nelif host_cpu == 'arm' or host_cpu == 'aarch64'\n\n prog = '''\n#include <arm_acle.h>\n\nint main(void)\n{\n unsigned int crc = 0;\n crc = __crc32cb(crc, 0);\n crc = __crc32ch(crc, 0);\n crc = __crc32cw(crc, 0);\n crc = __crc32cd(crc, 0);\n\n /* return computed value, to prevent the above being optimized away */\n return crc == 0;\n}\n'''\n\n if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n args: test_c_args)\n # Use ARM CRC Extension unconditionally\n cdata.set('USE_ARMV8_CRC32C', 1)\n have_optimized_crc = true\n elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\n args: test_c_args + ['-march=armv8-a+crc'])\n # Use ARM CRC Extension, with runtime check\n cflags_crc += '-march=armv8-a+crc'\n cdata.set('USE_ARMV8_CRC32C', false)\n cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n have_optimized_crc = true\n endif\nendif\n\nThe meson checking logic is used both for msvc and other compilers, so this\nwill need to work with both.\n\n\n> diff --git a/src/tools/msvc/gendef.pl b/src/tools/msvc/gendef.pl\n> index d6bed1ce15..f1e0ff446b 100644\n> --- a/src/tools/msvc/gendef.pl\n> +++ b/src/tools/msvc/gendef.pl\n> @@ -120,9 +120,9 @@ sub writedef\n> \t{\n> \t\tmy $isdata = $def->{$f} eq 'data';\n> \n> -\t\t# Strip the leading underscore for win32, but not x64\n> +\t\t# Strip the leading underscore for win32, but not x64 and arm64\n> \t\t$f =~ s/^_//\n> -\t\t unless ($arch eq \"x86_64\");\n> +\t\t unless ($arch ne \"x86\");\n> \n> \t\t# Emit just the name if it's a function symbol, or emit the name\n> \t\t# decorated with the DATA option for variables.\n> @@ -143,7 +143,7 @@ sub writedef\n> sub usage\n> {\n> \tdie(\"Usage: gendef.pl --arch <arch> --deffile <deffile> --tempdir <tempdir> files-or-directories\\n\"\n> -\t\t . \" arch: x86 | x86_64\\n\"\n> +\t\t . \" arch: x86 | x86_64 | arm64 \\n\"\n> \t\t . \" deffile: path of the generated file\\n\"\n> \t\t . \" tempdir: directory for temporary files\\n\"\n> \t\t . \" files or directories: object files or directory containing object files\\n\"\n> @@ -160,7 +160,7 @@ GetOptions(\n> \t'tempdir:s' => \\$tempdir,) or usage();\n> \n> usage(\"arch: $arch\")\n> - unless ($arch eq 'x86' || $arch eq 'x86_64');\n> + unless ($arch eq 'x86' || $arch eq 'x86_64' || $arch eq 'arm64');\n> \n> my @files;\n> \n\nSeems reasonable.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sat, 5 Nov 2022 11:31:36 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Sat, Nov 05, 2022 at 11:31:36AM -0700, Andres Freund wrote:\n> On 2022-11-03 11:06:46 +0000, Niyas Sait wrote:\n> Note that we're planning to remove the custom windows build scripts before the\n> next release, relying on the meson build instead.\n\nYoupi.\n\n>> #if defined(_WIN64)\n>> static __forceinline void\n>> spin_delay(void)\n>> {\n>> +#ifdef _M_ARM64\n>> +\t/*\n>> +\t * arm64 way of hinting processor for spin loops optimisations\n>> +\t * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\n>> +\t*/\n>> +\t__isb(_ARM64_BARRIER_SY);\n>> +#else\n>> \t_mm_pause();\n>> +#endif\n>> }\n>> #else\n>> static __forceinline void\n> \n> I think we should just apply this, there seems very little risk of making\n> anything worse, given the gating to _WIN64 && _M_ARM64.\n\nSeems so. Hmm, where does _ARM64_BARRIER_SY come from? Perhaps it\nwould be better to have a comment referring to it from a different\nplace than the forums of arm, like some actual docs?\n\n> The meson checking logic is used both for msvc and other compilers, so this\n> will need to work with both.\n\nThe patch has been switched as waiting on author for now.\n--\nMichael","msg_date":"Mon, 7 Nov 2022 15:58:13 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Mon, Nov 07, 2022 at 03:58:13PM +0900, Michael Paquier wrote:\n> The patch has been switched as waiting on author for now.\n\nSeeing no replies after three weeks, I have marked the patch as\nreturned with feedback for now.\n--\nMichael","msg_date":"Thu, 1 Dec 2022 16:42:19 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"\nOn 05/11/2022 18:31, Andres Freund wrote:\n> On 2022-11-03 11:06:46 +0000, Niyas Sait wrote:\n>> I've attached a new version of the patch which excludes the already merged\n>> ASLR changes and add\n>> small changes to handle latest changes in the build scripts.\n> Note that we're planning to remove the custom windows build scripts before the\n> next release, relying on the meson build instead.\n> \n\nThanks. I will add changes to add meson build support.\n\n> This won't suffice with the meson build, since the relevant configure test\n> also uses arm_acle.h:\n> elif host_cpu == 'arm' or host_cpu == 'aarch64'\n> \n> prog = '''\n> #include <arm_acle.h>\n> \n> int main(void)\n> {\n> unsigned int crc = 0;\n> crc = __crc32cb(crc, 0);\n> crc = __crc32ch(crc, 0);\n> crc = __crc32cw(crc, 0);\n> crc = __crc32cd(crc, 0);\n> \n> /* return computed value, to prevent the above being optimized away */\n> return crc == 0;\n> }\n> '''\n> \n> if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n> args: test_c_args)\n> # Use ARM CRC Extension unconditionally\n> cdata.set('USE_ARMV8_CRC32C', 1)\n> have_optimized_crc = true\n> elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\n> args: test_c_args + ['-march=armv8-a+crc'])\n> # Use ARM CRC Extension, with runtime check\n> cflags_crc += '-march=armv8-a+crc'\n> cdata.set('USE_ARMV8_CRC32C', false)\n> cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> have_optimized_crc = true\n> endif\n> endif\n> \n> The meson checking logic is used both for msvc and other compilers, so this\n> will need to work with both.\n\nYes, will handle that.\n\n-- \nNiyas\n\n\n","msg_date":"Thu, 1 Dec 2022 17:17:53 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"\nOn 07/11/2022 06:58, Michael Paquier wrote:\n>>> #if defined(_WIN64)\n>>> static __forceinline void\n>>> spin_delay(void)\n>>> {\n>>> +#ifdef _M_ARM64\n>>> +\t/*\n>>> +\t * arm64 way of hinting processor for spin loops optimisations\n>>> +\t * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\n>>> +\t*/\n>>> +\t__isb(_ARM64_BARRIER_SY);\n>>> +#else\n>>> \t_mm_pause();\n>>> +#endif\n>>> }\n>>> #else\n>>> static __forceinline void\n>>\n>> I think we should just apply this, there seems very little risk of making\n>> anything worse, given the gating to _WIN64 && _M_ARM64.\n> \n> Seems so. Hmm, where does _ARM64_BARRIER_SY come from? Perhaps it\n> would be better to have a comment referring to it from a different\n> place than the forums of arm, like some actual docs?\n\n\n_ARM64_BARRIER_SY is defined in Microsoft Arm64 intrinsic documentation \n- \nhttps://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-170#BarrierRestrictions\n\nI couldn't find something more official for the sse2neon library part.\n\n-- \nNiyas\n\n\n","msg_date":"Thu, 1 Dec 2022 17:20:20 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hello,\n\nI've attached a new revision of the patch (v4) and includes following \nchanges,\n\n1. Add support for meson build system\n2. Extend MSVC scripts to handle ARM64 platform.\n3. Add arm64 definition of spin_delay function.\n4. Exclude arm_acle.h import with MSVC compiler.\n\nV3->V4: Add support for meson build system\nV2->V3: Removed ASLR enablement and rebased on master.\nV1->V2: Rebased on top of master\n\n-- \nNiyas","msg_date":"Thu, 1 Dec 2022 17:53:47 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Dec 01, 2022 at 05:53:47PM +0000, Niyas Sait wrote:\n> I've attached a new revision of the patch (v4) and includes following\n> changes,\n> \n> 1. Add support for meson build system\n> 2. Extend MSVC scripts to handle ARM64 platform.\n> 3. Add arm64 definition of spin_delay function.\n> 4. Exclude arm_acle.h import with MSVC compiler.\n> \n> V3->V4: Add support for meson build system\n> V2->V3: Removed ASLR enablement and rebased on master.\n> V1->V2: Rebased on top of master\n\nThanks for the updated version. I have been looking at it closely and\nit looks like it should be able to do the job (no arm64 machine for\nWindows here, sigh).\n\nI have one tiny comment about this part:\n\n- USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => undef,\n+ USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => 1,\n\nShouldn't we only enable this flag when we are under aarch64?\nSimilarly, I don't think that it is a good idea to switch on\nUSE_SSE42_CRC32C_WITH_RUNTIME_CHECK all the time. We should only set\nit when building under x86 and x86_64.\n\nI would also add your link [1] in s_lock.h.\n\n[1]: https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics#BarrierRestrictions\n--\nMichael","msg_date":"Fri, 2 Dec 2022 14:02:54 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Fri, Dec 2, 2022 at 12:20 AM Niyas Sait <niyas.sait@linaro.org> wrote:\n>\n>\n> On 07/11/2022 06:58, Michael Paquier wrote:\n> > Seems so. Hmm, where does _ARM64_BARRIER_SY come from? Perhaps it\n> > would be better to have a comment referring to it from a different\n> > place than the forums of arm, like some actual docs?\n>\n>\n> _ARM64_BARRIER_SY is defined in Microsoft Arm64 intrinsic documentation\n> -\n>\nhttps://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-170#BarrierRestrictions\n\nIn particular, at the bottom of that section:\n\n\"For the __isb intrinsic, the only restriction that is currently valid is\n_ARM64_BARRIER_SY; all other values are reserved by the architecture.\"\n\nThis corresponds to\n\nhttps://developer.arm.com/documentation/ddi0596/2021-06/Base-Instructions/ISB--Instruction-Synchronization-Barrier-\n\nwhich says\n\n\"SY Full system barrier operation, encoded as CRm = 0b1111. Can be omitted.\"\n\n> I couldn't find something more official for the sse2neon library part.\n\nNot quite sure what this is referring to, but it seems we can just point to\nthe __aarch64__ section in the same file, which uses the same instruction:\n\nspin_delay(void)\n{\n __asm__ __volatile__(\n \" isb; \\n\");\n}\n\n...and which already explains the choice with a comment.\n\nAbout v4:\n\n+ * Use _mm_pause (x64) or __isb(arm64) intrinsic instead of rep nop.\n\nNeed a space here after __isb.\n\n+ if cc.get_id() == 'msvc'\n+ cdata.set('USE_ARMV8_CRC32C', false)\n+ cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n+ have_optimized_crc = true\n+ else\n\nThat seems like a heavy-handed way to force it. Could we just use the same\ngating in the test program that the patch puts in the code of interest?\nNamely:\n\n+#ifndef _MSC_VER\n #include <arm_acle.h>\n+#endif\n\n--\nJohn Naylor\nEDB: http://www.enterprisedb.com\n\nOn Fri, Dec 2, 2022 at 12:20 AM Niyas Sait <niyas.sait@linaro.org> wrote:>>> On 07/11/2022 06:58, Michael Paquier wrote:> > Seems so. Hmm, where does _ARM64_BARRIER_SY come from? Perhaps it> > would be better to have a comment referring to it from a different> > place than the forums of arm, like some actual docs?>>> _ARM64_BARRIER_SY is defined in Microsoft Arm64 intrinsic documentation> -> https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-170#BarrierRestrictionsIn particular, at the bottom of that section:\"For the __isb intrinsic, the only restriction that is currently valid is _ARM64_BARRIER_SY; all other values are reserved by the architecture.\"This corresponds tohttps://developer.arm.com/documentation/ddi0596/2021-06/Base-Instructions/ISB--Instruction-Synchronization-Barrier-which says\"SY Full system barrier operation, encoded as CRm = 0b1111. Can be omitted.\"> I couldn't find something more official for the sse2neon library part.Not quite sure what this is referring to, but it seems we can just point to the __aarch64__ section in the same file, which uses the same instruction:spin_delay(void){ __asm__ __volatile__( \" isb;\t\t\t\t\\n\");}...and which already explains the choice with a comment.About v4:+ * Use _mm_pause (x64) or __isb(arm64) intrinsic instead of rep nop.Need a space here after __isb.+ if cc.get_id() == 'msvc'+ cdata.set('USE_ARMV8_CRC32C', false)+ cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)+ have_optimized_crc = true+ elseThat seems like a heavy-handed way to force it. Could we just use the same gating in the test program that the patch puts in the code of interest? Namely:+#ifndef _MSC_VER #include <arm_acle.h>+#endif--John NaylorEDB: http://www.enterprisedb.com","msg_date":"Fri, 2 Dec 2022 12:41:06 +0700","msg_from":"John Naylor <john.naylor@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"\nOn 02/12/2022 05:02, Michael Paquier wrote:\n> Thanks for the updated version. I have been looking at it closely and\n> it looks like it should be able to do the job (no arm64 machine for\n> Windows here, sigh).\n> \n> I have one tiny comment about this part:\n> \n> - USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => undef,\n> + USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => 1,\n> \n> Shouldn't we only enable this flag when we are under aarch64?\n> Similarly, I don't think that it is a good idea to switch on\n> USE_SSE42_CRC32C_WITH_RUNTIME_CHECK all the time. We should only set\n> it when building under x86 and x86_64.\n> \n\nOk, I will fix ARM64 specific one in the next revision.\n\n-- \nNiyas\n\n\n","msg_date":"Fri, 2 Dec 2022 09:37:43 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 02/12/2022 05:41, John Naylor wrote:\n>> I couldn't find something more official for the sse2neon library part.\n> Not quite sure what this is referring to, but it seems we can just point to\n> the __aarch64__ section in the same file, which uses the same instruction:\n> \n> spin_delay(void)\n> {\n> __asm__ __volatile__(\n> \" isb; \\n\");\n> }\n> \n> ...and which already explains the choice with a comment.\n\nGood point. Will add the comment.\n\n> + if cc.get_id() == 'msvc'\n> + cdata.set('USE_ARMV8_CRC32C', false)\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> + have_optimized_crc = true\n> + else\n> \n> That seems like a heavy-handed way to force it. Could we just use the same\n> gating in the test program that the patch puts in the code of interest?\n> Namely:\n> \n> +#ifndef _MSC_VER\n> #include <arm_acle.h>\n> +#endif\nI took a similar approach as x86 MSVC code. I don't think the test \nprogram would work with MSVC. The compiler options are not MSVC friendly.\n\n-- \nNiyas\n\n\n","msg_date":"Fri, 2 Dec 2022 09:50:20 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hello,\n\nI've attached a new revision of the patch (v5) and includes following \nchanges,\n\n1. Add support for meson build system\n2. Extend MSVC scripts to handle ARM64 platform.\n3. Add arm64 definition of spin_delay function.\n4. Exclude arm_acle.h import with MSVC compiler.\n\nV4->V5: * Added reference to Microsoft arm64 intrinsic documentation\n * Conditionally enable USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK\n * Fixed styling issue spotted by Ian Lawerence Barwick\nV3->V4: Add support for meson build system\nV2->V3: Removed ASLR enablement and rebased on master.\nV1->V2: Rebased on top of master\n\n-- \nNiyas","msg_date":"Fri, 2 Dec 2022 11:09:15 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Fri, Dec 02, 2022 at 11:09:15AM +0000, Niyas Sait wrote:\n> I've attached a new revision of the patch (v5) and includes following\n> changes,\n> \n> 1. Add support for meson build system\n> 2. Extend MSVC scripts to handle ARM64 platform.\n> 3. Add arm64 definition of spin_delay function.\n> 4. Exclude arm_acle.h import with MSVC compiler.\n\nHmm. There are still a few things that need some attention here:\n- USE_SSE42_CRC32C_WITH_RUNTIME_CHECK should not be set for aarch64.\n- This is missing updates for ICU. Looking at the upstream code,\nBuild.Windows.ProjectConfiguration.props uses libARM64 and binARM64\nfor the output library and binary paths.\n- This is missing updates for krb5. For this case, I am seeing no\ntraces of packages for aarch64, so I guess that we could just fail\nhard until someone cares enough to ping us about what to do here.\n- There were zero changes in the docs, but we need to update at least\nthe section about architectures supported for the 64-bit builds.\n- Last comes OpenSSL, that supports amd64_arm64 as build target (see\nNOTES-WINDOWS.md), and the library names are libssl.lib and\nlibcrypto.lib. Looking at\nhttps://slproweb.com/products/Win32OpenSSL.html, there are\nexperimental builds for arm64 with OpenSSL 3.0. Niyas or somebody\nelse, could you look at the contents of lib/VC/ and see what we could\nrely on for the debug builds after installing this MSI? We should\nrely on something like lib/VC/sslcrypto64MD.lib or\nlib/VC/sslcrypto32MD.lib, but for arm64.\n\nWith meson gaining in maturity, perhaps that's not the most urgent\nthing as we will likely remove src/tools/msvc/ soon but I'd rather do\nthat right anyway as much as I can to avoid an incorrect state in the\ntree at any time in its history.\n\n- USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => undef,\n+ USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => $self->{platform} eq \"ARM64\" ? : 1 : undef,\nDid you actually test this patch? This won't work at all with perl,\nper se the double colon after the question mark.\n\nFor now, please find attached an updated patch with all the fixes I\ncould come up with.\n--\nMichael","msg_date":"Mon, 5 Dec 2022 14:12:41 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"\nOn 05/12/2022 05:12, Michael Paquier wrote:\n> - Last comes OpenSSL, that supports amd64_arm64 as build target (see\n> NOTES-WINDOWS.md), and the library names are libssl.lib and\n> libcrypto.lib. Looking at\n> https://slproweb.com/products/Win32OpenSSL.html, there are\n> experimental builds for arm64 with OpenSSL 3.0. Niyas or somebody\n> else, could you look at the contents of lib/VC/ and see what we could\n> rely on for the debug builds after installing this MSI? We should\n> rely on something like lib/VC/sslcrypto64MD.lib or\n> lib/VC/sslcrypto32MD.lib, but for arm64.\n\nI tried that installer. And I can see following libraries installed in \nlib/VC location.\n\nlibcryptoarm64MD.lib\nlibcryptoarm64MDd.lib\nlibcryptoarm64MT.lib\nlibcryptoarm64MTd.lib\nlibsslarm64MD.lib\nlibsslarm64MDd.lib\nlibsslarm64MT.lib\nlibsslarm64MTd.lib\n\n> - USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => undef,\n> + USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => $self->{platform} eq \"ARM64\" ? : 1 : undef,\n> Did you actually test this patch? This won't work at all with perl,\n> per se the double colon after the question mark.\n\n\nYes I did a full build. I am not sure why I didn't see any error with \nthat. My perl skills are very limited and I started with something bit \nmore naive like \"$self->{platform} == \"ARM64\" ? : 1 : undef\" But that \ndidn't work and I changed to using eq and the compilation was fine. \nThanks for fixing the patch.\n\n > For now, please find attached an updated patch with all the fixes I\n > could come up with.\n\nThanks. I did a quick sanity build with your updated patch and looks fine.\n\n-- \nNiyas\n\n\n","msg_date":"Mon, 5 Dec 2022 08:54:52 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-12-05 14:12:41 +0900, Michael Paquier wrote:\n> With meson gaining in maturity, perhaps that's not the most urgent\n> thing as we will likely remove src/tools/msvc/ soon but I'd rather do\n> that right anyway as much as I can to avoid an incorrect state in the\n> tree at any time in its history.\n\nI'd actually argue that we should just not add win32 support to\nsrc/tools/msvc/.\n\n\n> --- a/src/include/storage/s_lock.h\n> +++ b/src/include/storage/s_lock.h\n> @@ -708,13 +708,21 @@ typedef LONG slock_t;\n> #define SPIN_DELAY() spin_delay()\n> \n> /* If using Visual C++ on Win64, inline assembly is unavailable.\n> - * Use a _mm_pause intrinsic instead of rep nop.\n> + * Use _mm_pause (x64) or __isb (arm64) intrinsic instead of rep nop.\n> */\n> #if defined(_WIN64)\n> static __forceinline void\n> spin_delay(void)\n> {\n> +#ifdef _M_ARM64\n> +\t/*\n> +\t * See spin_delay aarch64 inline assembly definition above for details\n> +\t * ref: https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics#BarrierRestrictions\n> +\t*/\n> +\t__isb(_ARM64_BARRIER_SY);\n> +#else\n> \t_mm_pause();\n> +#endif\n> }\n> #else\n> static __forceinline void\n\nThis looks somewhat wrong to me. We end up with some ifdefs on the function\ndefintion level, and some others inside the function body. I think it should\nbe either or.\n\n\n\n> diff --git a/meson.build b/meson.build\n> index 725e10d815..e354ad7650 100644\n> --- a/meson.build\n> +++ b/meson.build\n> @@ -1944,7 +1944,13 @@ int main(void)\n> \n> elif host_cpu == 'arm' or host_cpu == 'aarch64'\n> \n> - prog = '''\n> + if cc.get_id() == 'msvc'\n> + cdata.set('USE_ARMV8_CRC32C', false)\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> + have_optimized_crc = true\n> + else\n> +\n> + prog = '''\n> #include <arm_acle.h>\n> \n> int main(void)\n\nWhy does this need to be hardcoded? The compiler probe should just work for\nmsvc.\n\n\n> @@ -1960,18 +1966,19 @@ int main(void)\n> }\n> '''\n> \n> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n> - args: test_c_args)\n> - # Use ARM CRC Extension unconditionally\n> - cdata.set('USE_ARMV8_CRC32C', 1)\n> - have_optimized_crc = true\n> - elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\n> - args: test_c_args + ['-march=armv8-a+crc'])\n> - # Use ARM CRC Extension, with runtime check\n> - cflags_crc += '-march=armv8-a+crc'\n> - cdata.set('USE_ARMV8_CRC32C', false)\n> - cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> - have_optimized_crc = true\n> + if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n> + args: test_c_args)\n> + # Use ARM CRC Extension unconditionally\n> + cdata.set('USE_ARMV8_CRC32C', 1)\n> + have_optimized_crc = true\n> + elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\n> + args: test_c_args + ['-march=armv8-a+crc'])\n> + # Use ARM CRC Extension, with runtime check\n> + cflags_crc += '-march=armv8-a+crc'\n> + cdata.set('USE_ARMV8_CRC32C', false)\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> + have_optimized_crc = true\n> + endif\n> endif\n> endif\n\nAnd then this reindent wouldn't be needed.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 5 Dec 2022 10:14:49 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Mon, Dec 05, 2022 at 10:14:49AM -0800, Andres Freund wrote:\n> On 2022-12-05 14:12:41 +0900, Michael Paquier wrote:\n>> With meson gaining in maturity, perhaps that's not the most urgent\n>> thing as we will likely remove src/tools/msvc/ soon but I'd rather do\n>> that right anyway as much as I can to avoid an incorrect state in the\n>> tree at any time in its history.\n> \n> I'd actually argue that we should just not add win32 support to\n> src/tools/msvc/.\n\nAre you arguing about ripping out the existing win32 or do nothing for\narm64? I guess the later, but these words also sound like the former\nto me.\n--\nMichael","msg_date":"Tue, 6 Dec 2022 08:31:16 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-12-06 08:31:16 +0900, Michael Paquier wrote:\n> On Mon, Dec 05, 2022 at 10:14:49AM -0800, Andres Freund wrote:\n> > On 2022-12-05 14:12:41 +0900, Michael Paquier wrote:\n> >> With meson gaining in maturity, perhaps that's not the most urgent\n> >> thing as we will likely remove src/tools/msvc/ soon but I'd rather do\n> >> that right anyway as much as I can to avoid an incorrect state in the\n> >> tree at any time in its history.\n> > \n> > I'd actually argue that we should just not add win32 support to\n> > src/tools/msvc/.\n> \n> Are you arguing about ripping out the existing win32 or do nothing for\n> arm64? I guess the later, but these words also sound like the former\n> to me.\n\nYes, that's a typo. I meant that we shouldn't add arm64 support to\nsrc/tools/msvc/.\n\nI do think we should rip out src/tools/msvc/ soon-ish, but we need\nbuildfarm support first.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 5 Dec 2022 15:48:45 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Mon, Dec 05, 2022 at 03:48:45PM -0800, Andres Freund wrote:\n> Yes, that's a typo. I meant that we shouldn't add arm64 support to\n> src/tools/msvc/.\n\nOkay, thanks.\n\n> I do think we should rip out src/tools/msvc/ soon-ish, but we need\n> buildfarm support first.\n\nSure, let's see where it goes.\n--\nMichael","msg_date":"Tue, 6 Dec 2022 09:16:17 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 05/12/2022 18:14, Andres Freund wrote:\n\n>> With meson gaining in maturity, perhaps that's not the most urgent\n>> thing as we will likely remove src/tools/msvc/ soon but I'd rather do\n>> that right anyway as much as I can to avoid an incorrect state in the\n>> tree at any time in its history.\n> \n> I'd actually argue that we should just not add win32 support to\n> src/tools/msvc/.\n> \n> \n\nI think the old build system specific part is really minimal in the \npatch. I can strip out those if that's preferred.\n\n\n>> --- a/src/include/storage/s_lock.h\n>> +++ b/src/include/storage/s_lock.h\n>> @@ -708,13 +708,21 @@ typedef LONG slock_t;\n>> #define SPIN_DELAY() spin_delay()\n>> \n>> /* If using Visual C++ on Win64, inline assembly is unavailable.\n>> - * Use a _mm_pause intrinsic instead of rep nop.\n>> + * Use _mm_pause (x64) or __isb (arm64) intrinsic instead of rep nop.\n>> */\n>> #if defined(_WIN64)\n>> static __forceinline void\n>> spin_delay(void)\n>> {\n>> +#ifdef _M_ARM64\n>> +\t/*\n>> +\t * See spin_delay aarch64 inline assembly definition above for details\n>> +\t * ref: https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics#BarrierRestrictions\n>> +\t*/\n>> +\t__isb(_ARM64_BARRIER_SY);\n>> +#else\n>> \t_mm_pause();\n>> +#endif\n>> }\n>> #else\n>> static __forceinline void\n> \n> This looks somewhat wrong to me. We end up with some ifdefs on the function\n> defintion level, and some others inside the function body. I think it should\n> be either or.\n> \n\nOk, I can add an MSVC/ARM64 specific function.\n\n>> diff --git a/meson.build b/meson.build\n>> index 725e10d815..e354ad7650 100644\n>> --- a/meson.build\n>> +++ b/meson.build\n>> @@ -1944,7 +1944,13 @@ int main(void)\n>> \n>> elif host_cpu == 'arm' or host_cpu == 'aarch64'\n>> \n>> - prog = '''\n>> + if cc.get_id() == 'msvc'\n>> + cdata.set('USE_ARMV8_CRC32C', false)\n>> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n>> + have_optimized_crc = true\n>> + else\n>> +\n>> + prog = '''\n>> #include <arm_acle.h>\n>> \n>> int main(void)\n> Why does this need to be hardcoded? The compiler probe should just work for\n> msvc.\n> \n\nThere are couple of minor issues in the code probe with MSVC such as \narm_acle.h needs to be removed and requires an explicit import of \nintrin.h. But even with those fixes, USE_ARMV8_CRC32C would be set and \nno runtime CRC extension check will be performed. Since CRC extension is \noptional in ARMv8, It would be better to use the CRC variant with \nruntime check. So I end up following the x64 variant and hardcoded the \nflags in case of ARM64 and MSVC.\n\n\n-- \nNiyas\n\n\n","msg_date":"Mon, 12 Dec 2022 13:38:37 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Mon, Dec 12, 2022 at 01:38:37PM +0000, Niyas Sait wrote:\n> On 05/12/2022 18:14, Andres Freund wrote:\n> I think the old build system specific part is really minimal in the patch. I\n> can strip out those if that's preferred.\n\nRemoving all the changes from src/tools/msvc/ is an approach that\nworks for me.\n\n>> Why does this need to be hardcoded? The compiler probe should just work for\n>> msvc.\n> \n> There are couple of minor issues in the code probe with MSVC such as\n> arm_acle.h needs to be removed and requires an explicit import of intrin.h.\n> But even with those fixes, USE_ARMV8_CRC32C would be set and no runtime CRC\n> extension check will be performed. Since CRC extension is optional in ARMv8,\n> It would be better to use the CRC variant with runtime check. So I end up\n> following the x64 variant and hardcoded the flags in case of ARM64 and MSVC.\n\nHm.. Andres, do you have access to Windows hosts with ARM64\nunderneath to validate any of that? No way to blame Cirrus for not\nhaving this option, they already do a lot :)\n--\nMichael","msg_date":"Tue, 13 Dec 2022 08:56:33 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 12/12/2022 23:56, Michael Paquier wrote:\n> On Mon, Dec 12, 2022 at 01:38:37PM +0000, Niyas Sait wrote:\n>> On 05/12/2022 18:14, Andres Freund wrote:\n>> I think the old build system specific part is really minimal in the patch. I\n>> can strip out those if that's preferred.\n> Removing all the changes from src/tools/msvc/ is an approach that\n> works for me.\n> \n\nI've attached a new version (v7) that removes changes to support old \nMSVC build system.\n\n\n-- \nNiyas","msg_date":"Fri, 16 Dec 2022 10:52:23 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"\nOn 16/12/2022 10:52, Niyas Sait wrote:\n> On 12/12/2022 23:56, Michael Paquier wrote:\n>> On Mon, Dec 12, 2022 at 01:38:37PM +0000, Niyas Sait wrote:\n>>> On 05/12/2022 18:14, Andres Freund wrote:\n>>> I think the old build system specific part is really minimal in the \n>>> patch. I\n>>> can strip out those if that's preferred.\n>> Removing all the changes from src/tools/msvc/ is an approach that\n>> works for me.\n>>\n> \n> I've attached a new version (v7) that removes changes to support old \n> MSVC build system.\n\n\nHello,\n\nCan I get review for the latest patch please ?\n\n-- \nNiyas\n\n\n","msg_date":"Tue, 17 Jan 2023 09:19:58 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2022-12-16 10:52:23 +0000, Niyas Sait wrote:\n> Subject: [PATCH v7] Enable postgres native build for windows-arm64 platform\n\n> elif host_cpu == 'arm' or host_cpu == 'aarch64'\n> \n> - prog = '''\n> + if cc.get_id() == 'msvc'\n> + cdata.set('USE_ARMV8_CRC32C', false)\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> + have_optimized_crc = true\n\nI dimly recall that windows might actually require the relevant extension on\narm?\n\n\n> + else\n> + prog = '''\n> #include <arm_acle.h>\n\nI'd just make this include #ifdef _MSV_VER (or whatever it is).\n\n\n> int main(void)\n> @@ -1960,18 +1966,19 @@ int main(void)\n> }\n> '''\n> \n> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n> - args: test_c_args)\n> - # Use ARM CRC Extension unconditionally\n> - cdata.set('USE_ARMV8_CRC32C', 1)\n> - have_optimized_crc = true\n> - elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\n> - args: test_c_args + ['-march=armv8-a+crc'])\n> - # Use ARM CRC Extension, with runtime check\n> - cflags_crc += '-march=armv8-a+crc'\n> - cdata.set('USE_ARMV8_CRC32C', false)\n> - cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> - have_optimized_crc = true\n> + if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n> + args: test_c_args)\n\nSeems like it'd be easier to read if you don't re-indent this, but just have\nthe cc.get_id() == 'msvc' part of this if/else-if.\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 17 Jan 2023 14:51:31 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"\n\nOn 17/01/2023 22:51, Andres Freund wrote:\n> Hi,\n> \n> On 2022-12-16 10:52:23 +0000, Niyas Sait wrote:\n>> Subject: [PATCH v7] Enable postgres native build for windows-arm64 platform\n> \n>> elif host_cpu == 'arm' or host_cpu == 'aarch64'\n>> \n>> - prog = '''\n>> + if cc.get_id() == 'msvc'\n>> + cdata.set('USE_ARMV8_CRC32C', false)\n>> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n>> + have_optimized_crc = true\n> \n> I dimly recall that windows might actually require the relevant extension on\n> arm?\n\nDo you mean we don't need the runtime checks for CRC ?\n\nCRC is an optional extension for ARMv8 base architecture. I am not sure \nif windows make it mandatory to have this implementation.\n\n> \n>> + else\n>> + prog = '''\n>> #include <arm_acle.h>\n> \n> I'd just make this include #ifdef _MSV_VER (or whatever it is).\n\nThe code snippet is not used for MSVC part. I am not sure why we need to \nadd the #ifdef _MSC_VER.\n\n>> int main(void)\n>> @@ -1960,18 +1966,19 @@ int main(void)\n>> }\n>> '''\n>> \n>> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n>> - args: test_c_args)\n>> - # Use ARM CRC Extension unconditionally\n>> - cdata.set('USE_ARMV8_CRC32C', 1)\n>> - have_optimized_crc = true\n>> - elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\n>> - args: test_c_args + ['-march=armv8-a+crc'])\n>> - # Use ARM CRC Extension, with runtime check\n>> - cflags_crc += '-march=armv8-a+crc'\n>> - cdata.set('USE_ARMV8_CRC32C', false)\n>> - cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n>> - have_optimized_crc = true\n>> + if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n>> + args: test_c_args)\n> \n> Seems like it'd be easier to read if you don't re-indent this, but just have\n> the cc.get_id() == 'msvc' part of this if/else-if.\n> \n\nYes that looks better. will do it in next patch.\n\nThanks Andres for the review.\n\n-- \nNiyas\n\n\n","msg_date":"Thu, 19 Jan 2023 10:09:01 +0000","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 19/01/2023 10:09, Niyas Sait wrote:\n> \n> \n> On 17/01/2023 22:51, Andres Freund wrote:\n\n>>> int main(void)\n>>> @@ -1960,18 +1966,19 @@ int main(void)\n>>> }\n>>> '''\n>>> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and \n>>> __crc32cd without -march=armv8-a+crc',\n>>> - args: test_c_args)\n>>> - # Use ARM CRC Extension unconditionally\n>>> - cdata.set('USE_ARMV8_CRC32C', 1)\n>>> - have_optimized_crc = true\n>>> - elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and \n>>> __crc32cd with -march=armv8-a+crc',\n>>> - args: test_c_args + ['-march=armv8-a+crc'])\n>>> - # Use ARM CRC Extension, with runtime check\n>>> - cflags_crc += '-march=armv8-a+crc'\n>>> - cdata.set('USE_ARMV8_CRC32C', false)\n>>> - cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n>>> - have_optimized_crc = true\n>>> + if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and \n>>> __crc32cd without -march=armv8-a+crc',\n>>> + args: test_c_args)\n>>\n>> Seems like it'd be easier to read if you don't re-indent this, but \n>> just have\n>> the cc.get_id() == 'msvc' part of this if/else-if.\n>>\n> \n\nI've attached a new version (v8) to fix the above indentation issue.\n\nCould someone please help with the review ?\n\n-- \nNiyas","msg_date":"Tue, 2 May 2023 08:51:09 +0100","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, May 02, 2023 at 08:51:09AM +0100, Niyas Sait wrote:\n> I've attached a new version (v8) to fix the above indentation issue.\n> \n> Could someone please help with the review ?\n\nSeeing how simple this has become, I would be really tempted to get\nthat applied, especially if there's a buildfarm machine able to follow\nup.. On the one hand, we are in a stability period for v16, so it may\nnot be the best moment ever. On the other hand, waiting one more year\nsounds like a waste for a patch that just adds new code paths.\n\nRMT members, any thoughts?\n--\nMichael","msg_date":"Sat, 6 May 2023 12:09:22 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> Seeing how simple this has become, I would be really tempted to get\n> that applied, especially if there's a buildfarm machine able to follow\n> up.. On the one hand, we are in a stability period for v16, so it may\n> not be the best moment ever. On the other hand, waiting one more year\n> sounds like a waste for a patch that just adds new code paths.\n\nIndeed, there's not much in this patch ... but it's impossible to see\nhow \"run on an entirely new platform\" isn't a new feature. Moreover,\nit's a platform that very few of us will have any ability to support\nor debug for. I can't see how it's a good idea to shove this in\npost-feature-freeze. Let's plan to push it in when v17 opens.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 06 May 2023 00:35:40 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Sat, May 06, 2023 at 12:35:40AM -0400, Tom Lane wrote:\n> Indeed, there's not much in this patch ... but it's impossible to see\n> how \"run on an entirely new platform\" isn't a new feature. Moreover,\n> it's a platform that very few of us will have any ability to support\n> or debug for. I can't see how it's a good idea to shove this in\n> post-feature-freeze. Let's plan to push it in when v17 opens.\n\nFine by me. I have added an entry in the CF app to remember about\nthis stuff as there was nothing present yet:\nhttps://commitfest.postgresql.org/43/4309/\n--\nMichael","msg_date":"Sun, 7 May 2023 09:37:50 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2023-05-06 00:35:40 -0400, Tom Lane wrote:\n> Michael Paquier <michael@paquier.xyz> writes:\n> > Seeing how simple this has become, I would be really tempted to get\n> > that applied, especially if there's a buildfarm machine able to follow\n> > up.. On the one hand, we are in a stability period for v16, so it may\n> > not be the best moment ever. On the other hand, waiting one more year\n> > sounds like a waste for a patch that just adds new code paths.\n> \n> Indeed, there's not much in this patch ... but it's impossible to see\n> how \"run on an entirely new platform\" isn't a new feature. Moreover,\n> it's a platform that very few of us will have any ability to support\n> or debug for. I can't see how it's a good idea to shove this in\n> post-feature-freeze. Let's plan to push it in when v17 opens.\n\nI don't really have feelings either way - but haven't we gone further and even\nbackpatched things like spinlock support for new arches in the past?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 8 May 2023 12:17:31 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> I don't really have feelings either way - but haven't we gone further and even\n> backpatched things like spinlock support for new arches in the past?\n\nMmmm ... don't really think those cases were comparable. We weren't\nadding support for a whole new OS. Now, you might argue that Windows\non arm64 will be just like Windows on x86_64, but I think the jury\nis still out on that. Microsoft was so Intel-only for so many years\nthat I bet they've had to change quite a bit to make it go on ARM.\n\nAlso, the cases of back-patched spinlock support that I can find\nin the last few years were pretty low-risk. I'll grant that\nc32fcac56 was a bit blue-sky because hardly anybody had RISC-V\nat that point, but by the same token anybody relying on it at the\ntime would be dealing with a beta-grade OS too. On the other hand,\n1c72d82c2 was immediately testable in the buildfarm, and f3bd00c01\nwas importing code already verified by our OpenBSD packagers.\n\nAs I said upthread, this seems like something to put in at the\nbeginning of a dev cycle, not post-feature-freeze.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 08 May 2023 15:58:41 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2023-05-08 Mo 15:58, Tom Lane wrote:\n> Andres Freund<andres@anarazel.de> writes:\n>> I don't really have feelings either way - but haven't we gone further and even\n>> backpatched things like spinlock support for new arches in the past?\n> Mmmm ... don't really think those cases were comparable. We weren't\n> adding support for a whole new OS. Now, you might argue that Windows\n> on arm64 will be just like Windows on x86_64, but I think the jury\n> is still out on that. Microsoft was so Intel-only for so many years\n> that I bet they've had to change quite a bit to make it go on ARM.\n>\n> Also, the cases of back-patched spinlock support that I can find\n> in the last few years were pretty low-risk. I'll grant that\n> c32fcac56 was a bit blue-sky because hardly anybody had RISC-V\n> at that point, but by the same token anybody relying on it at the\n> time would be dealing with a beta-grade OS too. On the other hand,\n> 1c72d82c2 was immediately testable in the buildfarm, and f3bd00c01\n> was importing code already verified by our OpenBSD packagers.\n>\n> As I said upthread, this seems like something to put in at the\n> beginning of a dev cycle, not post-feature-freeze.\n> \t\t\t\n\n\nWe will definitely want buildfarm support. I don't have such a machine \nand am not likely to have one any time soon. I do run drongo and \nfairywren on an EC2 instance, but AWS doesn't seem to have support for \nWindows on ARM. Maybe it's available on Azure (maybe then some of our \ncolleagues working at Microsoft could arrange such a beast for me to set \nup potential buildfarm animal, or else run one themselves.)\n\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2023-05-08 Mo 15:58, Tom Lane wrote:\n\n\nAndres Freund <andres@anarazel.de> writes:\n\n\nI don't really have feelings either way - but haven't we gone further and even\nbackpatched things like spinlock support for new arches in the past?\n\n\n\nMmmm ... don't really think those cases were comparable. We weren't\nadding support for a whole new OS. Now, you might argue that Windows\non arm64 will be just like Windows on x86_64, but I think the jury\nis still out on that. Microsoft was so Intel-only for so many years\nthat I bet they've had to change quite a bit to make it go on ARM.\n\nAlso, the cases of back-patched spinlock support that I can find\nin the last few years were pretty low-risk. I'll grant that\nc32fcac56 was a bit blue-sky because hardly anybody had RISC-V\nat that point, but by the same token anybody relying on it at the\ntime would be dealing with a beta-grade OS too. On the other hand,\n1c72d82c2 was immediately testable in the buildfarm, and f3bd00c01\nwas importing code already verified by our OpenBSD packagers.\n\nAs I said upthread, this seems like something to put in at the\nbeginning of a dev cycle, not post-feature-freeze.\n\t\t\t\n\n\n\nWe will definitely want buildfarm support. I don't have such a\n machine and am not likely to have one any time soon. I do run\n drongo and fairywren on an EC2 instance, but AWS doesn't seem to\n have support for Windows on ARM. Maybe it's available on Azure\n (maybe then some of our colleagues working at Microsoft could\n arrange such a beast for me to set up potential buildfarm animal,\n or else run one themselves.)\n\n\n\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Wed, 10 May 2023 09:24:39 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, May 10, 2023 at 09:24:39AM -0400, Andrew Dunstan wrote:\n> We will definitely want buildfarm support. I don't have such a machine and\n> am not likely to have one any time soon. I do run drongo and fairywren on an\n> EC2 instance, but AWS doesn't seem to have support for Windows on ARM. Maybe\n> it's available on Azure (maybe then some of our colleagues working at\n> Microsoft could arrange such a beast for me to set up potential buildfarm\n> animal, or else run one themselves.)\n\nUnfortunately AWS does not offer ARM on Windows with an EC2, that's\nsomething I have also looked at setting up :/\n\nAnyway, Niyat has mentioned me that he was working on that, but it\nseems that it was off-list. Niyat, could you confirm this part?\n--\nMichael","msg_date":"Thu, 11 May 2023 08:34:29 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, May 11, 2023 at 11:34 AM Michael Paquier <michael@paquier.xyz> wrote:\n> On Wed, May 10, 2023 at 09:24:39AM -0400, Andrew Dunstan wrote:\n> > We will definitely want buildfarm support. I don't have such a machine and\n> > am not likely to have one any time soon. I do run drongo and fairywren on an\n> > EC2 instance, but AWS doesn't seem to have support for Windows on ARM. Maybe\n> > it's available on Azure (maybe then some of our colleagues working at\n> > Microsoft could arrange such a beast for me to set up potential buildfarm\n> > animal, or else run one themselves.)\n>\n> Unfortunately AWS does not offer ARM on Windows with an EC2, that's\n> something I have also looked at setting up :/\n\nAzure does have an image \"Microsoft Windows 11 Preview arm64\" to run\non Ampere CPUs, which says it's a pre-release version intended for\nvalidation, which sounds approximately like what we want. I will try\nto find out more.\n\n\n","msg_date":"Thu, 11 May 2023 12:17:25 +1200","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, May 11, 2023 at 12:17:25PM +1200, Thomas Munro wrote:\n> Azure does have an image \"Microsoft Windows 11 Preview arm64\" to run\n> on Ampere CPUs, which says it's a pre-release version intended for\n> validation, which sounds approximately like what we want. I will try\n> to find out more.\n\nInteresting. Thanks for mentioning it.\n--\nMichael","msg_date":"Thu, 11 May 2023 13:19:54 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 11/05/2023 00:34, Michael Paquier wrote:\n> On Wed, May 10, 2023 at 09:24:39AM -0400, Andrew Dunstan wrote:\n>> We will definitely want buildfarm support. I don't have such a machine and\n>> am not likely to have one any time soon. I do run drongo and fairywren on an\n>> EC2 instance, but AWS doesn't seem to have support for Windows on ARM. Maybe\n>> it's available on Azure (maybe then some of our colleagues working at\n>> Microsoft could arrange such a beast for me to set up potential buildfarm\n>> animal, or else run one themselves.)\n> Unfortunately AWS does not offer ARM on Windows with an EC2, that's\n> something I have also looked at setting up :/\n> \n> Anyway, Niyat has mentioned me that he was working on that, but it\n> seems that it was off-list. Niyat, could you confirm this part?\n> --\n> Michael\n\nIf we have Azure VM credits, we could use that to deploy buildfarm machines.\n\nOtherwise, I can try to allocate a machine from Linaro Windows Lab for \nthe buildbot.\n\n-- \nNiyas\n\n\n","msg_date":"Thu, 11 May 2023 08:57:57 +0100","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hello,\n\nJust wanted to give a heads up that I am moving to a new role outside \nLinaro and as a result wouldn't be able to continue contributing.\n\nAnthony Roberts from Linaro is going to support the enablement work.\n\n-- \nNiyas\n\n\n","msg_date":"Wed, 14 Jun 2023 13:25:17 +0100","msg_from":"Niyas Sait <niyas.sait@linaro.org>","msg_from_op":true,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, May 11, 2023 at 01:19:54PM +0900, Michael Paquier wrote:\n> On Thu, May 11, 2023 at 12:17:25PM +1200, Thomas Munro wrote:\n>> Azure does have an image \"Microsoft Windows 11 Preview arm64\" to run\n>> on Ampere CPUs, which says it's a pre-release version intended for\n>> validation, which sounds approximately like what we want. I will try\n>> to find out more.\n> \n> Interesting. Thanks for mentioning it.\n\nNow that v17 is open, I was looking at v8 posted at [1] and I don't\nhave much more to add about it. The problem about the lack of\nbuildfarm machine is still annoying, but I don't see a reason not to\nmove forward and let people play with this stuff on HEAD. At least\nthat would be progress. Any thoughts?\n\nThomas, what's the state of ARM support for Windows on Azure? Is that\nstill in preview?\n\n[1]: https://www.postgresql.org/message-id/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56%40linaro.org\n--\nMichael","msg_date":"Tue, 4 Jul 2023 16:34:36 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi All,\n\nJust following up on this, has there been any movement?\n\nI did see another message go into the thread here with no reply: \nhttps://www.postgresql.org/message-id/ZKPLjLjIjwN2lxkg%40paquier.xyz\n\nThanks,\nAnthony\n\nOn 14/06/2023 13:25, Niyas Sait wrote:\n> Hello,\n>\n> Just wanted to give a heads up that I am moving to a new role outside \n> Linaro and as a result wouldn't be able to continue contributing.\n>\n> Anthony Roberts from Linaro is going to support the enablement work.\n>\n\n\n","msg_date":"Thu, 17 Aug 2023 09:41:44 +0100","msg_from":"Anthony Roberts <anthony.roberts@linaro.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Aug 17, 2023 at 09:41:44AM +0100, Anthony Roberts wrote:\n> Just following up on this, has there been any movement?\n> \n> I did see another message go into the thread here with no reply:\n> https://www.postgresql.org/message-id/ZKPLjLjIjwN2lxkg%40paquier.xyz\n\nI don't have an environment to test the patch, but I don't object to\nit per se. However, I don't really want to move forward completely\nblindly as well in the long-term.\n\nAs mentioned to Niyas, could it be possible to provide to the\ncommunity a buildfarm machine that would be able to test this\nenvironment? I am not sure what's your status on that. Perhaps this\nis already set up and you are just waiting for the patch to be merged?\n--\nMichael","msg_date":"Fri, 18 Aug 2023 07:28:04 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nWhich of these are you looking for us to provide:\n\n* A machine for you to directly access (via a VPN)\n* A machine we just run a worker script on that automatically picks up \nthe builds as required\n* Us to do downstream CI\n\nAll are possible, but preferably not option 1, as it would mean straight \nup pulling out a machine from our build farm, and it has to go through \nall sorts of approvals internally. If it's the only way forward I can \nkick it up the chain though.\n\nOption 2 and 3 are ones we do for various other projects (ie. 2 - CMake, \n3 - OpenSSL)\n\nThanks,\nAnthony\n\nOn 17/08/2023 23:28, Michael Paquier wrote:\n> On Thu, Aug 17, 2023 at 09:41:44AM +0100, Anthony Roberts wrote:\n>> Just following up on this, has there been any movement?\n>>\n>> I did see another message go into the thread here with no reply:\n>> https://www.postgresql.org/message-id/ZKPLjLjIjwN2lxkg%40paquier.xyz\n> I don't have an environment to test the patch, but I don't object to\n> it per se. However, I don't really want to move forward completely\n> blindly as well in the long-term.\n>\n> As mentioned to Niyas, could it be possible to provide to the\n> community a buildfarm machine that would be able to test this\n> environment? I am not sure what's your status on that. Perhaps this\n> is already set up and you are just waiting for the patch to be merged?\n> --\n> Michael\n\n\n","msg_date":"Fri, 25 Aug 2023 10:40:39 +0100","msg_from":"Anthony Roberts <anthony.roberts@linaro.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Fri, Aug 25, 2023 at 10:40:39AM +0100, Anthony Roberts wrote:\n> Which of these are you looking for us to provide:\n> \n> * A machine for you to directly access (via a VPN)\n> * A machine we just run a worker script on that automatically picks up the\n> builds as required\n> * Us to do downstream CI\n> \n> All are possible, but preferably not option 1, as it would mean straight up\n> pulling out a machine from our build farm, and it has to go through all\n> sorts of approvals internally. If it's the only way forward I can kick it up\n> the chain though.\n> \n> Option 2 and 3 are ones we do for various other projects (ie. 2 - CMake, 3 -\n> OpenSSL)\n\nThe community has its own CI facility. Here is a link of how to set\nup a machine:\nhttps://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto\n\nAnd all the results are published by the maintainers of the machines\non a periodic basis where the buildfarm client code is set:\nhttps://buildfarm.postgresql.org/cgi-bin/show_status.pl\n\nThe final result would be for you to maintain the machine so as we are\nable to see if anything breaks with it. Adding extra dependencies to\nthe build like OpenSSL would be nice, but based on the contents of the\npatch to add this port that does not seem mandatory to me, either.\n\nOnce the machine is ready, you will need to request a buildfarm\nmachine name and a key to be able to begin publishing the reports of\nthe builds to the community buildfarm. Once the machine is ready to\ngo, just let me know and I'd be OK to merge the patch (I still want to\ndo a final review of it in case I've missed something, but I can move\non with that before the buildfarm machine is set up).\n--\nMichael","msg_date":"Fri, 25 Aug 2023 19:17:27 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Thanks for the link - that looks basically the same as what we do for \nCMake (option 2), so should be relatively easy.\n\nI will have a chat to relevant people about setting a machine up \nproperly for it.\n\nThanks,\nAnthony\n\nOn 25/08/2023 11:17, Michael Paquier wrote:\n> On Fri, Aug 25, 2023 at 10:40:39AM +0100, Anthony Roberts wrote:\n>> Which of these are you looking for us to provide:\n>>\n>> * A machine for you to directly access (via a VPN)\n>> * A machine we just run a worker script on that automatically picks up the\n>> builds as required\n>> * Us to do downstream CI\n>>\n>> All are possible, but preferably not option 1, as it would mean straight up\n>> pulling out a machine from our build farm, and it has to go through all\n>> sorts of approvals internally. If it's the only way forward I can kick it up\n>> the chain though.\n>>\n>> Option 2 and 3 are ones we do for various other projects (ie. 2 - CMake, 3 -\n>> OpenSSL)\n> The community has its own CI facility. Here is a link of how to set\n> up a machine:\n> https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto\n>\n> And all the results are published by the maintainers of the machines\n> on a periodic basis where the buildfarm client code is set:\n> https://buildfarm.postgresql.org/cgi-bin/show_status.pl\n>\n> The final result would be for you to maintain the machine so as we are\n> able to see if anything breaks with it. Adding extra dependencies to\n> the build like OpenSSL would be nice, but based on the contents of the\n> patch to add this port that does not seem mandatory to me, either.\n>\n> Once the machine is ready, you will need to request a buildfarm\n> machine name and a key to be able to begin publishing the reports of\n> the builds to the community buildfarm. Once the machine is ready to\n> go, just let me know and I'd be OK to merge the patch (I still want to\n> do a final review of it in case I've missed something, but I can move\n> on with that before the buildfarm machine is set up).\n> --\n> Michael\n\n\n","msg_date":"Fri, 25 Aug 2023 11:34:18 +0100","msg_from":"Anthony Roberts <anthony.roberts@linaro.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"> On 25 Aug 2023, at 11:40, Anthony Roberts <anthony.roberts@linaro.org> wrote:\n\n> * A machine for you to directly access (via a VPN)\n\nThat could be very helpful to provide to developers who are trying to debug\nvery wicked bugs which requires access for analysis. Such things are on a case\nby case basis of course at your discretion, but it would most likely be very\nappreciated when/if the need exists.\n\n> * A machine we just run a worker script on that automatically picks up the builds as required\n\nThis sounds like what we have for the buildfarm, where machines independently\nbuilds the branches from the Git repo by using the buildfarm client as worker.\n\n> * Us to do downstream CI\n\nThis sounds like the CFBot, where we (currently) use Cirrus with our own\nworkers for building and testing the patches which are being worked on.\n\n\thttp://cfbot.cputube.org/\n\nBoth the two latter options are very helpful for the community, and serve two\ndistincly different purposes. Option 2 is required for a platform to be\nconsidered supported whereas option 3 assists developers in being proactive\nrather than just reactive.\n\n--\nDaniel Gustafsson\n\n\n\n","msg_date":"Fri, 25 Aug 2023 13:48:21 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nJust a follow-up: having spoken to the relevant people, sadly, right now,\nwe do not have the capacity to be pulling machines out of our day-to-day CI\ntasks to dedicate to specific projects.\n\nI can, however, offer a few alternatives:\n\n1. If you have capacity on your linux x64 machines, it is possible to\nemulate WoA, and run tests that way:\nhttps://www.linaro.org/blog/emulate-windows-on-arm/\nThis is something that works well for projects that we have contributed it\nto (ie, sse2neon uses it via github actions) - if you run into any issues\nwith this, we are able to try and fix them on our side, as likely more than\npostgres would benefit.\n\n2. If there is any scope for it on your side at all, a Samsung Galaxy Book\nGo 14\" can be had from ebay for ~£140.\nUnderstandably, this might not be an option for you.\n\nIf it helps, if this is merged, we do have the possibility of running our\nown downstream CI builds nightly for the tip of your main branch here:\nhttps://gitlab.com/Linaro/windowsonarm/nightly\nThis is something we do for a few projects that do not have full upstream\nCI.\n\nThanks,\nAnthony\n\n\nOn Fri, 25 Aug 2023 at 11:34, Anthony Roberts <anthony.roberts@linaro.org>\nwrote:\n\n> Thanks for the link - that looks basically the same as what we do for\n> CMake (option 2), so should be relatively easy.\n>\n> I will have a chat to relevant people about setting a machine up\n> properly for it.\n>\n> Thanks,\n> Anthony\n>\n> On 25/08/2023 11:17, Michael Paquier wrote:\n> > On Fri, Aug 25, 2023 at 10:40:39AM +0100, Anthony Roberts wrote:\n> >> Which of these are you looking for us to provide:\n> >>\n> >> * A machine for you to directly access (via a VPN)\n> >> * A machine we just run a worker script on that automatically picks up\n> the\n> >> builds as required\n> >> * Us to do downstream CI\n> >>\n> >> All are possible, but preferably not option 1, as it would mean\n> straight up\n> >> pulling out a machine from our build farm, and it has to go through all\n> >> sorts of approvals internally. If it's the only way forward I can kick\n> it up\n> >> the chain though.\n> >>\n> >> Option 2 and 3 are ones we do for various other projects (ie. 2 -\n> CMake, 3 -\n> >> OpenSSL)\n> > The community has its own CI facility. Here is a link of how to set\n> > up a machine:\n> > https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto\n> >\n> > And all the results are published by the maintainers of the machines\n> > on a periodic basis where the buildfarm client code is set:\n> > https://buildfarm.postgresql.org/cgi-bin/show_status.pl\n> >\n> > The final result would be for you to maintain the machine so as we are\n> > able to see if anything breaks with it. Adding extra dependencies to\n> > the build like OpenSSL would be nice, but based on the contents of the\n> > patch to add this port that does not seem mandatory to me, either.\n> >\n> > Once the machine is ready, you will need to request a buildfarm\n> > machine name and a key to be able to begin publishing the reports of\n> > the builds to the community buildfarm. Once the machine is ready to\n> > go, just let me know and I'd be OK to merge the patch (I still want to\n> > do a final review of it in case I've missed something, but I can move\n> > on with that before the buildfarm machine is set up).\n> > --\n> > Michael\n>\n\nHi,Just a follow-up: having spoken to the relevant people, sadly, right now, we do not have the capacity to be pulling machines out of our day-to-day CI tasks to dedicate to specific projects.I can, however, offer a few alternatives:1. If you have capacity on your linux x64 machines, it is possible to emulate WoA, and run tests that way: https://www.linaro.org/blog/emulate-windows-on-arm/This is something that works well for projects that we have contributed it to (ie, sse2neon uses it via github actions) - if you run into any issues with this, we are able to try and fix them on our side, as likely more than postgres would benefit.2. If there is any scope for it on your side at all, a Samsung Galaxy Book Go 14\" can be had from ebay for ~£140.Understandably, this might not be an option for you.If it helps, if this is merged, we do have the possibility of running our own downstream CI builds nightly for the tip of your main branch here: https://gitlab.com/Linaro/windowsonarm/nightlyThis is something we do for a few projects that do not have full upstream CI.Thanks,AnthonyOn Fri, 25 Aug 2023 at 11:34, Anthony Roberts <anthony.roberts@linaro.org> wrote:Thanks for the link - that looks basically the same as what we do for \nCMake (option 2), so should be relatively easy.\n\nI will have a chat to relevant people about setting a machine up \nproperly for it.\n\nThanks,\nAnthony\n\nOn 25/08/2023 11:17, Michael Paquier wrote:\n> On Fri, Aug 25, 2023 at 10:40:39AM +0100, Anthony Roberts wrote:\n>> Which of these are you looking for us to provide:\n>>\n>> * A machine for you to directly access (via a VPN)\n>> * A machine we just run a worker script on that automatically picks up the\n>> builds as required\n>> * Us to do downstream CI\n>>\n>> All are possible, but preferably not option 1, as it would mean straight up\n>> pulling out a machine from our build farm, and it has to go through all\n>> sorts of approvals internally. If it's the only way forward I can kick it up\n>> the chain though.\n>>\n>> Option 2 and 3 are ones we do for various other projects (ie. 2 - CMake, 3 -\n>> OpenSSL)\n> The community has its own CI facility. Here is a link of how to set\n> up a machine:\n> https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto\n>\n> And all the results are published by the maintainers of the machines\n> on a periodic basis where the buildfarm client code is set:\n> https://buildfarm.postgresql.org/cgi-bin/show_status.pl\n>\n> The final result would be for you to maintain the machine so as we are\n> able to see if anything breaks with it. Adding extra dependencies to\n> the build like OpenSSL would be nice, but based on the contents of the\n> patch to add this port that does not seem mandatory to me, either.\n>\n> Once the machine is ready, you will need to request a buildfarm\n> machine name and a key to be able to begin publishing the reports of\n> the builds to the community buildfarm. Once the machine is ready to\n> go, just let me know and I'd be OK to merge the patch (I still want to\n> do a final review of it in case I've missed something, but I can move\n> on with that before the buildfarm machine is set up).\n> --\n> Michael","msg_date":"Wed, 30 Aug 2023 16:12:16 +0100","msg_from":"Anthony Roberts <anthony.roberts@linaro.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, Aug 30, 2023 at 04:12:16PM +0100, Anthony Roberts wrote:\n> Just a follow-up: having spoken to the relevant people, sadly, right now,\n> we do not have the capacity to be pulling machines out of our day-to-day CI\n> tasks to dedicate to specific projects.\n\nOkay, thanks for letting us know.\n\n> I can, however, offer a few alternatives:\n> \n> 1. If you have capacity on your linux x64 machines, it is possible to\n> emulate WoA, and run tests that way:\n> https://www.linaro.org/blog/emulate-windows-on-arm/\n> This is something that works well for projects that we have contributed it\n> to (ie, sse2neon uses it via github actions) - if you run into any issues\n> with this, we are able to try and fix them on our side, as likely more than\n> postgres would benefit.\n> \n> 2. If there is any scope for it on your side at all, a Samsung Galaxy Book\n> Go 14\" can be had from ebay for ~£140.\n> Understandably, this might not be an option for you.\n> \n> If it helps, if this is merged, we do have the possibility of running our\n> own downstream CI builds nightly for the tip of your main branch here:\n> https://gitlab.com/Linaro/windowsonarm/nightly\n> This is something we do for a few projects that do not have full upstream\n> CI.\n\nHonestly, I am not sure how to proceed here. Having something in the\nbuildfarm is necessary IMO, because this is the central place that\ncommunity members look at when it comes to monitoring the stability of\na patch committed.\n\nAny opinions from the others?\n--\nMichael","msg_date":"Thu, 31 Aug 2023 13:22:16 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Wed, Aug 30, 2023 at 04:12:16PM +0100, Anthony Roberts wrote:\n>> Just a follow-up: having spoken to the relevant people, sadly, right now,\n>> we do not have the capacity to be pulling machines out of our day-to-day CI\n>> tasks to dedicate to specific projects.\n\n> Okay, thanks for letting us know.\n\nToo bad.\n\n> Honestly, I am not sure how to proceed here. Having something in the\n> buildfarm is necessary IMO, because this is the central place that\n> community members look at when it comes to monitoring the stability of\n> a patch committed.\n> Any opinions from the others?\n\nI agree. I'm really uncomfortable with claiming support for\nWindows-on-ARM if we don't have a buildfarm member testing it.\nFor other platforms that have a track record of multiple\nhardware support, it might not be a stretch ... but Windows was\nso resolutely Intel-only for so long that \"it works on ARM\" is\na proposition that I won't trust without hard evidence. There\nare too many bits of that system that might not have gotten the\nword yet, or at least not gotten sufficient testing.\n\nMy vote for this is we don't commit without a buildfarm member.\n\n(As a comparison point, I'd still be unsure about our support\nfor macOS-on-ARM if we didn't have buildfarm support for that.\nThat exists, and is being paid for out of my own pocket.\nI do not have much sympathy for a corporation that claims\nit can't afford to support one measly buildfarm animal.)\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 31 Aug 2023 00:44:07 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Aug 31, 2023 at 4:22 PM Michael Paquier <michael@paquier.xyz> wrote:\n> Honestly, I am not sure how to proceed here. Having something in the\n> buildfarm is necessary IMO, because this is the central place that\n> community members look at when it comes to monitoring the stability of\n> a patch committed.\n\nIt's tricky, because usually the org that wants the thing to work\nsupplies the thing, and I personally think that provides a meaningful\n\"demand\" signal too. For example, we have external Solaris\nrepresentation, but no users of AIX represented at this point, and I\nconsider that to be quite interesting information. Plenty of people\nwant PG to work well on their OS or architecture.\n\nBut, supposing Azure credits were suddenly available (just between\nyou, me and the mailing list, I heard that this should be imminent as\nsome kind of grant to PG US, as a result of the Cirrus thing), do we\nhave a Windows-savvy volunteer to feed and water an animal? I feel\nlike this case is inevitable anyway because people in our own\ncommunity are probably going to be using Windows ARM laptops sooner or\nlater...\n\n\n","msg_date":"Thu, 31 Aug 2023 16:57:05 +1200","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Thomas Munro <thomas.munro@gmail.com> writes:\n> But, supposing Azure credits were suddenly available (just between\n> you, me and the mailing list, I heard that this should be imminent as\n> some kind of grant to PG US, as a result of the Cirrus thing), do we\n> have a Windows-savvy volunteer to feed and water an animal? I feel\n> like this case is inevitable anyway because people in our own\n> community are probably going to be using Windows ARM laptops sooner or\n> later...\n\nThat's a good point. The hardware resources to support a buildfarm\nanimal are really pretty trivial these days. What is important is\nan animal owner who is willing to help chase down problems when they\narise.\n\nMaybe we just need to wait for some community members to acquire\nthose laptops.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 31 Aug 2023 01:06:53 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Aug 31, 2023 at 01:06:53AM -0400, Tom Lane wrote:\n> That's a good point. The hardware resources to support a buildfarm\n> animal are really pretty trivial these days. What is important is\n> an animal owner who is willing to help chase down problems when they\n> arise.\n\nUnfortunately I don't see myself running that at home. It is too hot\nand humid in Summer so the machine would not last. Something in the\ncloud would be OK for me, but AWS has no option for a Windows host\nwith ARM yet.\n--\nMichael","msg_date":"Thu, 31 Aug 2023 14:33:56 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 02.05.23 09:51, Niyas Sait wrote:\n> diff --git a/doc/src/sgml/install-windows.sgml b/doc/src/sgml/install-windows.sgml\n> index cbc70a039c..2ecd5fcf38 100644\n> --- a/doc/src/sgml/install-windows.sgml\n> +++ b/doc/src/sgml/install-windows.sgml\n> @@ -352,7 +352,8 @@ $ENV{MSBFLAGS}=\"/m\";\n> <title>Special Considerations for 64-Bit Windows</title>\n> \n> <para>\n> - PostgreSQL will only build for the x64 architecture on 64-bit Windows.\n> + PostgreSQL will only build for the x64 and ARM64 architectures on 64-bit\n> + Windows.\n> </para>\n\nAre there other possible architectures besides x64 and ARM64? Itanium? \nOr should we just delete this sentence?\n\n> diff --git a/meson.build b/meson.build\n> index 096044628c..6cca212fae 100644\n> --- a/meson.build\n> +++ b/meson.build\n> @@ -2045,8 +2045,11 @@ int main(void)\n> elif host_cpu == 'arm' or host_cpu == 'aarch64'\n> \n> prog = '''\n> +#ifdef _MSC_VER\n> +#include <intrin.h>\n> +#else\n> #include <arm_acle.h>\n> -\n> +#endif\n\nMaybe keep the whitespace here.\n\n> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n> + if cc.get_id() == 'msvc'\n> + cdata.set('USE_ARMV8_CRC32C', false)\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> + have_optimized_crc = true\n> + elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\n\nAdd a comment here.\n\n> diff --git a/src/port/pg_crc32c_armv8.c b/src/port/pg_crc32c_armv8.c\n> index d8fae510cf..3d7eb748ff 100644\n> --- a/src/port/pg_crc32c_armv8.c\n> +++ b/src/port/pg_crc32c_armv8.c\n> @@ -14,7 +14,9 @@\n> */\n> #include \"c.h\"\n> \n> +#ifndef _MSC_VER\n> #include <arm_acle.h>\n> +#endif\n\nAlso add a comment here.\n\n> \n> #include \"port/pg_crc32c.h\"\n> \n> diff --git a/src/tools/msvc/gendef.pl b/src/tools/msvc/gendef.pl\n> index e7cbefcbc3..934dc17b96 100644\n> --- a/src/tools/msvc/gendef.pl\n> +++ b/src/tools/msvc/gendef.pl\n> @@ -120,9 +120,9 @@ sub writedef\n> \t{\n> \t\tmy $isdata = $def->{$f} eq 'data';\n> \n> -\t\t# Strip the leading underscore for win32, but not x64\n> +\t\t# Strip the leading underscore for win32, but not x64 and aarch64\n\nIs x64 the opposite of win32? Does this make sense? Should we reverse \nthe logic here and single out the one variant where the stripping is \nnecessary?\n\n\n\n","msg_date":"Wed, 13 Sep 2023 21:09:12 +0200","msg_from":"Peter Eisentraut <peter@eisentraut.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 31.08.23 06:44, Tom Lane wrote:\n> I agree. I'm really uncomfortable with claiming support for\n> Windows-on-ARM if we don't have a buildfarm member testing it.\n> For other platforms that have a track record of multiple\n> hardware support, it might not be a stretch ... but Windows was\n> so resolutely Intel-only for so long that \"it works on ARM\" is\n> a proposition that I won't trust without hard evidence. There\n> are too many bits of that system that might not have gotten the\n> word yet, or at least not gotten sufficient testing.\n> \n> My vote for this is we don't commit without a buildfarm member.\n\nI think we can have a multi-tiered approach, where we can commit support \nbut consider it experimental until we have buildfarm coverage.\n\n\n","msg_date":"Wed, 13 Sep 2023 21:12:46 +0200","msg_from":"Peter Eisentraut <peter@eisentraut.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2023-08-31 14:33:56 +0900, Michael Paquier wrote:\n> On Thu, Aug 31, 2023 at 01:06:53AM -0400, Tom Lane wrote:\n> > That's a good point. The hardware resources to support a buildfarm\n> > animal are really pretty trivial these days. What is important is\n> > an animal owner who is willing to help chase down problems when they\n> > arise.\n> \n> Unfortunately I don't see myself running that at home. It is too hot\n> and humid in Summer so the machine would not last. Something in the\n> cloud would be OK for me, but AWS has no option for a Windows host\n> with ARM yet.\n\nFWIW, we could run ARM windows as part of CI. It'd be a bit of scripting work,\nas google doesn't yet have readymade VM images with windows for ARM. Right now\nthe CI VM images for windows are based on top of the google images. Scripting\ngeneration of ARM images would be some one-off work, but after that...\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 13 Sep 2023 12:32:46 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"> On 13 Sep 2023, at 21:12, Peter Eisentraut <peter@eisentraut.org> wrote:\n> \n> On 31.08.23 06:44, Tom Lane wrote:\n>> I agree. I'm really uncomfortable with claiming support for\n>> Windows-on-ARM if we don't have a buildfarm member testing it.\n>> For other platforms that have a track record of multiple\n>> hardware support, it might not be a stretch ... but Windows was\n>> so resolutely Intel-only for so long that \"it works on ARM\" is\n>> a proposition that I won't trust without hard evidence. There\n>> are too many bits of that system that might not have gotten the\n>> word yet, or at least not gotten sufficient testing.\n>> My vote for this is we don't commit without a buildfarm member.\n> \n> I think we can have a multi-tiered approach, where we can commit support but consider it experimental until we have buildfarm coverage.\n\nIf it's experimental it should probably be behind an opt-in flag in\nautoconf/meson, or be reverted by the time REL_17_STABLE branches unless\ncoverage has materialized by then.\n\n--\nDaniel Gustafsson\n\n\n\n","msg_date":"Thu, 14 Sep 2023 11:39:20 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 14.09.23 11:39, Daniel Gustafsson wrote:\n>> On 13 Sep 2023, at 21:12, Peter Eisentraut <peter@eisentraut.org> wrote:\n>>\n>> On 31.08.23 06:44, Tom Lane wrote:\n>>> I agree. I'm really uncomfortable with claiming support for\n>>> Windows-on-ARM if we don't have a buildfarm member testing it.\n>>> For other platforms that have a track record of multiple\n>>> hardware support, it might not be a stretch ... but Windows was\n>>> so resolutely Intel-only for so long that \"it works on ARM\" is\n>>> a proposition that I won't trust without hard evidence. There\n>>> are too many bits of that system that might not have gotten the\n>>> word yet, or at least not gotten sufficient testing.\n>>> My vote for this is we don't commit without a buildfarm member.\n>>\n>> I think we can have a multi-tiered approach, where we can commit support but consider it experimental until we have buildfarm coverage.\n> \n> If it's experimental it should probably be behind an opt-in flag in\n> autoconf/meson, or be reverted by the time REL_17_STABLE branches unless\n> coverage has materialized by then.\n\nThe author's email is bouncing now, due to job change, so it's unlikely \nwe will see any progress on this anymore. I am setting it to returned \nwith feedback.\n\n\n\n","msg_date":"Tue, 19 Sep 2023 10:33:48 +0200","msg_from":"Peter Eisentraut <peter@eisentraut.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nThis was covered earlier in the thread - I have taken this on in Niyas' \nstead.\n\nWas there an explicit request for something there? I was under the \nimpression that this was all just suggestion/theory at the moment.\n\nThanks,\nAnthony\n\nOn 19/09/2023 09:33, Peter Eisentraut wrote:\n> On 14.09.23 11:39, Daniel Gustafsson wrote:\n>>> On 13 Sep 2023, at 21:12, Peter Eisentraut <peter@eisentraut.org> \n>>> wrote:\n>>>\n>>> On 31.08.23 06:44, Tom Lane wrote:\n>>>> I agree. I'm really uncomfortable with claiming support for\n>>>> Windows-on-ARM if we don't have a buildfarm member testing it.\n>>>> For other platforms that have a track record of multiple\n>>>> hardware support, it might not be a stretch ... but Windows was\n>>>> so resolutely Intel-only for so long that \"it works on ARM\" is\n>>>> a proposition that I won't trust without hard evidence. There\n>>>> are too many bits of that system that might not have gotten the\n>>>> word yet, or at least not gotten sufficient testing.\n>>>> My vote for this is we don't commit without a buildfarm member.\n>>>\n>>> I think we can have a multi-tiered approach, where we can commit \n>>> support but consider it experimental until we have buildfarm coverage.\n>>\n>> If it's experimental it should probably be behind an opt-in flag in\n>> autoconf/meson, or be reverted by the time REL_17_STABLE branches unless\n>> coverage has materialized by then.\n>\n> The author's email is bouncing now, due to job change, so it's \n> unlikely we will see any progress on this anymore. I am setting it to \n> returned with feedback.\n>\n\n\n","msg_date":"Tue, 19 Sep 2023 13:35:17 +0100","msg_from":"Anthony Roberts <anthony.roberts@linaro.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, Sep 19, 2023 at 01:35:17PM +0100, Anthony Roberts wrote:\n> Was there an explicit request for something there? I was under the\n> impression that this was all just suggestion/theory at the moment.\n\nYes. The addition of a buildfarm member registered into the community\nfacilities, so as it is possible to provide back to developers dynamic\nfeedback to the new configuration setup you would like to see\nsupported in PostgreSQL. This has been mentioned a few times on this\nthread, around these places:\nhttps://www.postgresql.org/message-id/20220322103011.i6z2tuj4hle23wgx@jrouhaud\nhttps://www.postgresql.org/message-id/DBD80715-E56B-40EB-84AA-ACE70198E7AD@yesql.se\nhttps://www.postgresql.org/message-id/6D1E2719-FA49-42A5-A6FF-0B0072BF631D@yesql.se\n--\nMichael","msg_date":"Wed, 20 Sep 2023 07:59:53 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, 19 Sept 2023 at 23:48, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Tue, Sep 19, 2023 at 01:35:17PM +0100, Anthony Roberts wrote:\n> > Was there an explicit request for something there? I was under the\n> > impression that this was all just suggestion/theory at the moment.\n>\n> Yes. The addition of a buildfarm member registered into the community\n> facilities, so as it is possible to provide back to developers dynamic\n> feedback to the new configuration setup you would like to see\n> supported in PostgreSQL. This has been mentioned a few times on this\n> thread, around these places:\n>\n> https://www.postgresql.org/message-id/20220322103011.i6z2tuj4hle23wgx@jrouhaud\n>\n> https://www.postgresql.org/message-id/DBD80715-E56B-40EB-84AA-ACE70198E7AD@yesql.se\n>\n> https://www.postgresql.org/message-id/6D1E2719-FA49-42A5-A6FF-0B0072BF631D@yesql.se\n> --\n> Michael\n>\n\nThe attached patch works with v17. I will work on getting a buildfarm\nanimal up shortly","msg_date":"Tue, 23 Jan 2024 08:46:37 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks> wrote:\n\n>\n>\n> On Tue, 19 Sept 2023 at 23:48, Michael Paquier <michael@paquier.xyz>\n> wrote:\n>\n>> On Tue, Sep 19, 2023 at 01:35:17PM +0100, Anthony Roberts wrote:\n>> > Was there an explicit request for something there? I was under the\n>> > impression that this was all just suggestion/theory at the moment.\n>>\n>> Yes. The addition of a buildfarm member registered into the community\n>> facilities, so as it is possible to provide back to developers dynamic\n>> feedback to the new configuration setup you would like to see\n>> supported in PostgreSQL. This has been mentioned a few times on this\n>> thread, around these places:\n>>\n>> https://www.postgresql.org/message-id/20220322103011.i6z2tuj4hle23wgx@jrouhaud\n>>\n>> https://www.postgresql.org/message-id/DBD80715-E56B-40EB-84AA-ACE70198E7AD@yesql.se\n>>\n>> https://www.postgresql.org/message-id/6D1E2719-FA49-42A5-A6FF-0B0072BF631D@yesql.se\n>> --\n>> Michael\n>>\n>\n> The attached patch works with v17. I will work on getting a buildfarm\n> animal up shortly\n>\n>\nI can build using meson manually, but for some reason building with the\nbuildfarm fails.\n\nI'm not sure which files to attach to this to help. Andrew, what files do\nyou need ?\n\nDave\n\n>\n>\n\nOn Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks> wrote:On Tue, 19 Sept 2023 at 23:48, Michael Paquier <michael@paquier.xyz> wrote:On Tue, Sep 19, 2023 at 01:35:17PM +0100, Anthony Roberts wrote:\n> Was there an explicit request for something there? I was under the\n> impression that this was all just suggestion/theory at the moment.\n\nYes. The addition of a buildfarm member registered into the community\nfacilities, so as it is possible to provide back to developers dynamic\nfeedback to the new configuration setup you would like to see\nsupported in PostgreSQL. This has been mentioned a few times on this\nthread, around these places:\nhttps://www.postgresql.org/message-id/20220322103011.i6z2tuj4hle23wgx@jrouhaud\nhttps://www.postgresql.org/message-id/DBD80715-E56B-40EB-84AA-ACE70198E7AD@yesql.se\nhttps://www.postgresql.org/message-id/6D1E2719-FA49-42A5-A6FF-0B0072BF631D@yesql.se\n--\nMichaelThe attached patch works with v17. I will work on getting a buildfarm animal up shortlyI can build using meson manually, but for some reason building with the buildfarm fails.I'm not sure which files to attach to this to help. Andrew, what files do you need ?Dave","msg_date":"Tue, 23 Jan 2024 16:13:05 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, Jan 23, 2024 at 04:13:05PM -0500, Dave Cramer wrote:\n> On Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks> wrote:\n>> The attached patch works with v17. I will work on getting a buildfarm\n>> animal up shortly\n\nThanks for doing that!\n\n> I can build using meson manually, but for some reason building with the\n> buildfarm fails.\n> \n> I'm not sure which files to attach to this to help. Andrew, what files do\n> you need ?\n\nProbably build.ninja and the contents of meson-logs/ would offer\nenough information?\n--\nMichael","msg_date":"Wed, 24 Jan 2024 08:32:08 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, 23 Jan 2024 at 18:32, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Tue, Jan 23, 2024 at 04:13:05PM -0500, Dave Cramer wrote:\n> > On Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks>\n> wrote:\n> >> The attached patch works with v17. I will work on getting a buildfarm\n> >> animal up shortly\n>\n> Thanks for doing that!\n>\n> > I can build using meson manually, but for some reason building with the\n> > buildfarm fails.\n> >\n> > I'm not sure which files to attach to this to help. Andrew, what files do\n> > you need ?\n>\n> Probably build.ninja and the contents of meson-logs/ would offer\n> enough information?\n> --\n> Michael\n>\n\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\nadd some options, but the patch above needs to be applied to build it.\n\nDave\n\nOn Tue, 23 Jan 2024 at 18:32, Michael Paquier <michael@paquier.xyz> wrote:On Tue, Jan 23, 2024 at 04:13:05PM -0500, Dave Cramer wrote:\n> On Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks> wrote:\n>> The attached patch works with v17. I will work on getting a buildfarm\n>> animal up shortly\n\nThanks for doing that!\n\n> I can build using meson manually, but for some reason building with the\n> buildfarm fails.\n> \n> I'm not sure which files to attach to this to help. Andrew, what files do\n> you need ?\n\nProbably build.ninja and the contents of meson-logs/ would offer\nenough information?\n--\nMichaelI managed to get it to build the vcvarsall arch needs to be x64. I need to add some options, but the patch above needs to be applied to build it.Dave","msg_date":"Wed, 24 Jan 2024 06:45:21 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\n> add some options, but the patch above needs to be applied to build it.\n\nNice. If I may ask, what kind of host and/or configuration have you\nused to reach a state where the code can be compiled and run tests\nwith meson? If you have found specific steps, it may be a good thing\nto document that on the wiki, say around [1].\n\nPerhaps you have not included TAP? It may be fine in terms of runtime\nchecks and coverage.\n\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\n--\nMichael","msg_date":"Thu, 25 Jan 2024 09:02:30 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, 24 Jan 2024 at 19:03, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n> > I managed to get it to build the vcvarsall arch needs to be x64. I need\n> to\n> > add some options, but the patch above needs to be applied to build it.\n>\n> Nice. If I may ask, what kind of host and/or configuration have you\n> used to reach a state where the code can be compiled and run tests\n> with meson? If you have found specific steps, it may be a good thing\n> to document that on the wiki, say around [1].\n>\n\nI am running a mac-mini M2 with parallels running windows pro 11\nThe only thing required is this patch. Running in a native developer\nprompt\n\nmeson setup build --prefix=c:\\postgres\ncd build\nninja\nninja install\nninja test\n\nall run without errors\nI also have buildfarm running and will run that today\n\nDave\n\n>\n> Perhaps you have not included TAP? It may be fine in terms of runtime\n> checks and coverage.\n>\n\nDoes TAP have to be explicitly added to the meson build ?\n\nDave\n\nOn Wed, 24 Jan 2024 at 19:03, Michael Paquier <michael@paquier.xyz> wrote:On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\n> add some options, but the patch above needs to be applied to build it.\n\nNice. If I may ask, what kind of host and/or configuration have you\nused to reach a state where the code can be compiled and run tests\nwith meson? If you have found specific steps, it may be a good thing\nto document that on the wiki, say around [1].I am running a mac-mini M2 with parallels running windows pro 11The only thing required is this patch. Running in a native developer prompt meson setup build --prefix=c:\\postgrescd buildninja ninja installninja testall run without errorsI also have buildfarm running and will run that todayDave\n\nPerhaps you have not included TAP? It may be fine in terms of runtime\nchecks and coverage.Does TAP have to be explicitly added to the meson build ?Dave","msg_date":"Thu, 25 Jan 2024 08:31:07 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, 25 Jan 2024 at 08:31, Dave Cramer <davecramer@postgres.rocks> wrote:\n\n>\n>\n> On Wed, 24 Jan 2024 at 19:03, Michael Paquier <michael@paquier.xyz> wrote:\n>\n>> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n>> > I managed to get it to build the vcvarsall arch needs to be x64. I need\n>> to\n>> > add some options, but the patch above needs to be applied to build it.\n>>\n>> Nice. If I may ask, what kind of host and/or configuration have you\n>> used to reach a state where the code can be compiled and run tests\n>> with meson? If you have found specific steps, it may be a good thing\n>> to document that on the wiki, say around [1].\n>>\n>\n> I am running a mac-mini M2 with parallels running windows pro 11\n> The only thing required is this patch. Running in a native developer\n> prompt\n>\n> meson setup build --prefix=c:\\postgres\n> cd build\n> ninja\n> ninja install\n> ninja test\n>\n> all run without errors\n> I also have buildfarm running and will run that today\n>\n> Dave\n>\n>>\n>> Perhaps you have not included TAP? It may be fine in terms of runtime\n>> checks and coverage.\n>>\n>\n> Does TAP have to be explicitly added to the meson build ?\n>\n> Dave\n>\n\n\nI tried running my buildfarm using my git repo and my branch, but get the\nfollowing error\nStatus Line: 492 bad branch parameter\nContent:\nbad branch parameter fix_arm\n\nWeb txn failed with status: 1\n\nOn Thu, 25 Jan 2024 at 08:31, Dave Cramer <davecramer@postgres.rocks> wrote:On Wed, 24 Jan 2024 at 19:03, Michael Paquier <michael@paquier.xyz> wrote:On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\n> add some options, but the patch above needs to be applied to build it.\n\nNice. If I may ask, what kind of host and/or configuration have you\nused to reach a state where the code can be compiled and run tests\nwith meson? If you have found specific steps, it may be a good thing\nto document that on the wiki, say around [1].I am running a mac-mini M2 with parallels running windows pro 11The only thing required is this patch. Running in a native developer prompt meson setup build --prefix=c:\\postgrescd buildninja ninja installninja testall run without errorsI also have buildfarm running and will run that todayDave\n\nPerhaps you have not included TAP? It may be fine in terms of runtime\nchecks and coverage.Does TAP have to be explicitly added to the meson build ?DaveI tried running my buildfarm using my git repo and my branch, but get the following errorStatus Line: 492 bad branch parameterContent:bad branch parameter fix_armWeb txn failed with status: 1","msg_date":"Thu, 25 Jan 2024 08:45:52 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-01-24 We 19:02, Michael Paquier wrote:\n> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n>> I managed to get it to build the vcvarsall arch needs to be x64. I need to\n>> add some options, but the patch above needs to be applied to build it.\n> Nice. If I may ask, what kind of host and/or configuration have you\n> used to reach a state where the code can be compiled and run tests\n> with meson? If you have found specific steps, it may be a good thing\n> to document that on the wiki, say around [1].\n>\n> Perhaps you have not included TAP? It may be fine in terms of runtime\n> checks and coverage.\n>\n> [1]:https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\n\n\n\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think we \nreally want to build with x64_arm64, i.e. to generate native arm64 \nbinaries. Setting just x64 will not do that, AIUI.\n\nI tried that with the buidfarm, setting that in the config file's call \nto PGBuild::VSenv::getenv().\n\nThat upset msvc_gendef.pl, so I added this there to keep it happy:\n\n $arch = 'x86_64' if $arch eq 'aarch64';\n\nAfter that things went ok until I got this:\n\n[1453/2088] \"link\" @src/backend/postgres.exe.rsp\nFAILED: src/backend/postgres.exe src/backend/postgres.pdb\n\"link\" @src/backend/postgres.exe.rsp\n Creating library src\\backend\\postgres.exe.lib\nstorage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol \n_mm_pause referenced in function perform_spin_delay\nsrc\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n\n\nI haven't made further progress, but I will return to it in the next day \nor so.\n\nWhile this will be nice to have, I think it won't really matter until \nthere is ARM64 support in released versions of Windows Server. AFAICT \nthey still only sell versions for x86_64\n\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-01-24 We 19:02, Michael Paquier\n wrote:\n\n\nOn Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n\n\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\nadd some options, but the patch above needs to be applied to build it.\n\n\n\nNice. If I may ask, what kind of host and/or configuration have you\nused to reach a state where the code can be compiled and run tests\nwith meson? If you have found specific steps, it may be a good thing\nto document that on the wiki, say around [1].\n\nPerhaps you have not included TAP? It may be fine in terms of runtime\nchecks and coverage.\n\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\n\n\n\n\n\n\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think\n we really want to build with x64_arm64, i.e. to generate native\n arm64 binaries. Setting just x64 will not do that, AIUI.\nI tried that with the buidfarm, setting that in the config file's\n call to PGBuild::VSenv::getenv().\nThat upset msvc_gendef.pl, so I added this there to keep it\n happy:\n\n$arch = 'x86_64' if $arch eq 'aarch64';\n\nAfter that things went ok until I got this:\n[1453/2088] \"link\" @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library src\\backend\\postgres.exe.lib\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\n symbol _mm_pause referenced in function perform_spin_delay\n src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved\n externals \n\n\n\nI haven't made further progress, but I will return to it in the\n next day or so.\nWhile this will be nice to have, I think it won't really matter\n until there is ARM64 support in released versions of Windows\n Server. AFAICT they still only sell versions for x86_64\n\n\n\n\n\ncheers\n\n\nandrew\n\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Thu, 25 Jan 2024 12:30:55 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-01-24 We 19:02, Michael Paquier wrote:\n>\n> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n>\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\n> add some options, but the patch above needs to be applied to build it.\n>\n> Nice. If I may ask, what kind of host and/or configuration have you\n> used to reach a state where the code can be compiled and run tests\n> with meson? If you have found specific steps, it may be a good thing\n> to document that on the wiki, say around [1].\n>\n> Perhaps you have not included TAP? It may be fine in terms of runtime\n> checks and coverage.\n>\n> [1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\n>\n>\n>\n> I now have an ARM64 machine, so I set up a W11 ARM64 VM. I think we really\n> want to build with x64_arm64, i.e. to generate native arm64 binaries.\n> Setting just x64 will not do that, AIUI.\n>\n> I tried that with the buidfarm, setting that in the config file's call to\n> PGBuild::VSenv::getenv().\n>\n> That upset msvc_gendef.pl, so I added this there to keep it happy:\n>\n> $arch = 'x86_64' if $arch eq 'aarch64';\n>\n> After that things went ok until I got this:\n>\n> [1453/2088] \"link\" @src/backend/postgres.exe.rsp\n> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n> \"link\" @src/backend/postgres.exe.rsp\n> Creating library src\\backend\\postgres.exe.lib\n> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\n> _mm_pause referenced in function perform_spin_delay\n> src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n>\n>\n> I haven't made further progress, but I will return to it in the next day\n> or so.\n>\n> While this will be nice to have, I think it won't really matter until\n> there is ARM64 support in released versions of Windows Server. AFAICT they\n> still only sell versions for x86_64\n>\n\nI need it to build clients. The clients need arm64 libraries to link against\n\nDave\n\nOn Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-01-24 We 19:02, Michael Paquier\n wrote:\n\n\nOn Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n\n\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\nadd some options, but the patch above needs to be applied to build it.\n\n\nNice. If I may ask, what kind of host and/or configuration have you\nused to reach a state where the code can be compiled and run tests\nwith meson? If you have found specific steps, it may be a good thing\nto document that on the wiki, say around [1].\n\nPerhaps you have not included TAP? It may be fine in terms of runtime\nchecks and coverage.\n\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\n\n\n\n\n\n\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think\n we really want to build with x64_arm64, i.e. to generate native\n arm64 binaries. Setting just x64 will not do that, AIUI.\nI tried that with the buidfarm, setting that in the config file's\n call to PGBuild::VSenv::getenv().\nThat upset msvc_gendef.pl, so I added this there to keep it\n happy:\n\n$arch = 'x86_64' if $arch eq 'aarch64';\n\nAfter that things went ok until I got this:\n[1453/2088] \"link\" @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library src\\backend\\postgres.exe.lib\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\n symbol _mm_pause referenced in function perform_spin_delay\n src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved\n externals \n\n\n\nI haven't made further progress, but I will return to it in the\n next day or so.\nWhile this will be nice to have, I think it won't really matter\n until there is ARM64 support in released versions of Windows\n Server. AFAICT they still only sell versions for x86_64I need it to build clients. The clients need arm64 libraries to link againstDave","msg_date":"Thu, 25 Jan 2024 14:07:11 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-01-25 Th 08:45, Dave Cramer wrote:\n>\n>\n>\n>\n> I tried running my buildfarm using my git repo and my branch, but get \n> the following error\n> Status Line: 492 bad branch parameter\n> Content:\n> bad branch parameter fix_arm\n>\n> Web txn failed with status: 1\n\n\n\nYou can't use your own branch with the buildfarm, you need to let it set \nup its own branches.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-01-25 Th 08:45, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nI tried running my buildfarm using my git repo and my\n branch, but get the following error\nStatus Line: 492 bad branch parameter\nContent:\nbad branch parameter fix_arm\n\n\nWeb txn failed with status: 1 \n\n\n\n\n\n\n\n\n\nYou can't use your own branch with the buildfarm, you need to let\n it set up its own branches.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Thu, 25 Jan 2024 14:31:43 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, 25 Jan 2024 at 14:31, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-01-25 Th 08:45, Dave Cramer wrote:\n>\n>\n>\n>\n>\n> I tried running my buildfarm using my git repo and my branch, but get the\n> following error\n> Status Line: 492 bad branch parameter\n> Content:\n> bad branch parameter fix_arm\n>\n> Web txn failed with status: 1\n>\n>\n>\n> You can't use your own branch with the buildfarm, you need to let it set\n> up its own branches.\n>\n\nSo I guess I have to wait until this patch is merged in ?\n\nDave\n\nOn Thu, 25 Jan 2024 at 14:31, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-01-25 Th 08:45, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nI tried running my buildfarm using my git repo and my\n branch, but get the following error\nStatus Line: 492 bad branch parameter\nContent:\nbad branch parameter fix_arm\n\n\nWeb txn failed with status: 1 \n\n\n\n\n\n\n\n\n\nYou can't use your own branch with the buildfarm, you need to let\n it set up its own branches.So I guess I have to wait until this patch is merged in ?Dave","msg_date":"Thu, 25 Jan 2024 15:56:20 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-01-24 We 19:02, Michael Paquier wrote:\n>\n> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n>\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\n> add some options, but the patch above needs to be applied to build it.\n>\n> Nice. If I may ask, what kind of host and/or configuration have you\n> used to reach a state where the code can be compiled and run tests\n> with meson? If you have found specific steps, it may be a good thing\n> to document that on the wiki, say around [1].\n>\n> Perhaps you have not included TAP? It may be fine in terms of runtime\n> checks and coverage.\n>\n> [1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\n>\n>\n>\n> I now have an ARM64 machine, so I set up a W11 ARM64 VM. I think we really\n> want to build with x64_arm64, i.e. to generate native arm64 binaries.\n> Setting just x64 will not do that, AIUI.\n>\n> I tried that with the buidfarm, setting that in the config file's call to\n> PGBuild::VSenv::getenv().\n>\n> That upset msvc_gendef.pl, so I added this there to keep it happy:\n>\n> $arch = 'x86_64' if $arch eq 'aarch64';\n>\n> After that things went ok until I got this:\n>\n> [1453/2088] \"link\" @src/backend/postgres.exe.rsp\n> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n> \"link\" @src/backend/postgres.exe.rsp\n> Creating library src\\backend\\postgres.exe.lib\n> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\n> _mm_pause referenced in function perform_spin_delay\n> src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n>\n>\n> I haven't made further progress, but I will return to it in the next day\n> or so.\n>\n> While this will be nice to have, I think it won't really matter until\n> there is ARM64 support in released versions of Windows Server. AFAICT they\n> still only sell versions for x86_64\n>\n\nI've tried it with my patch attached previously and x64_arm64 and it works\nfine. builds using the buildfarm as well.\n\nIs there a definitive way to figure out if the binaries are x64_arm64\n\nDave\n\nOn Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-01-24 We 19:02, Michael Paquier\n wrote:\n\n\nOn Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n\n\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\nadd some options, but the patch above needs to be applied to build it.\n\n\nNice. If I may ask, what kind of host and/or configuration have you\nused to reach a state where the code can be compiled and run tests\nwith meson? If you have found specific steps, it may be a good thing\nto document that on the wiki, say around [1].\n\nPerhaps you have not included TAP? It may be fine in terms of runtime\nchecks and coverage.\n\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\n\n\n\n\n\n\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think\n we really want to build with x64_arm64, i.e. to generate native\n arm64 binaries. Setting just x64 will not do that, AIUI.\nI tried that with the buidfarm, setting that in the config file's\n call to PGBuild::VSenv::getenv().\nThat upset msvc_gendef.pl, so I added this there to keep it\n happy:\n\n$arch = 'x86_64' if $arch eq 'aarch64';\n\nAfter that things went ok until I got this:\n[1453/2088] \"link\" @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library src\\backend\\postgres.exe.lib\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\n symbol _mm_pause referenced in function perform_spin_delay\n src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved\n externals \n\n\n\nI haven't made further progress, but I will return to it in the\n next day or so.\nWhile this will be nice to have, I think it won't really matter\n until there is ARM64 support in released versions of Windows\n Server. AFAICT they still only sell versions for x86_64I've tried it with my patch attached previously and x64_arm64 and it works fine. builds using the buildfarm as well.Is there a definitive way to figure out if the binaries are x64_arm64Dave","msg_date":"Thu, 25 Jan 2024 16:01:28 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi David,\n\nUnix \"file\" or \"dumpbin /headers\" in vcvarsall are your best bets.\n\nThanks,\nAnthony\n\nOn Thu, 25 Jan 2024, 21:01 Dave Cramer, <davecramer@postgres.rocks> wrote:\n\n>\n>\n> On Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>>\n>> On 2024-01-24 We 19:02, Michael Paquier wrote:\n>>\n>> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n>>\n>> I managed to get it to build the vcvarsall arch needs to be x64. I need to\n>> add some options, but the patch above needs to be applied to build it.\n>>\n>> Nice. If I may ask, what kind of host and/or configuration have you\n>> used to reach a state where the code can be compiled and run tests\n>> with meson? If you have found specific steps, it may be a good thing\n>> to document that on the wiki, say around [1].\n>>\n>> Perhaps you have not included TAP? It may be fine in terms of runtime\n>> checks and coverage.\n>>\n>> [1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\n>>\n>>\n>>\n>> I now have an ARM64 machine, so I set up a W11 ARM64 VM. I think we\n>> really want to build with x64_arm64, i.e. to generate native arm64\n>> binaries. Setting just x64 will not do that, AIUI.\n>>\n>> I tried that with the buidfarm, setting that in the config file's call to\n>> PGBuild::VSenv::getenv().\n>>\n>> That upset msvc_gendef.pl, so I added this there to keep it happy:\n>>\n>> $arch = 'x86_64' if $arch eq 'aarch64';\n>>\n>> After that things went ok until I got this:\n>>\n>> [1453/2088] \"link\" @src/backend/postgres.exe.rsp\n>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n>> \"link\" @src/backend/postgres.exe.rsp\n>> Creating library src\\backend\\postgres.exe.lib\n>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\n>> _mm_pause referenced in function perform_spin_delay\n>> src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n>>\n>>\n>> I haven't made further progress, but I will return to it in the next day\n>> or so.\n>>\n>> While this will be nice to have, I think it won't really matter until\n>> there is ARM64 support in released versions of Windows Server. AFAICT they\n>> still only sell versions for x86_64\n>>\n>\n> I've tried it with my patch attached previously and x64_arm64 and it works\n> fine. builds using the buildfarm as well.\n>\n> Is there a definitive way to figure out if the binaries are x64_arm64\n>\n> Dave\n>\n\nHi David,Unix \"file\" or \"dumpbin /headers\" in vcvarsall are your best bets.Thanks,AnthonyOn Thu, 25 Jan 2024, 21:01 Dave Cramer, <davecramer@postgres.rocks> wrote:On Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-01-24 We 19:02, Michael Paquier\n wrote:\n\n\nOn Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\n\n\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\nadd some options, but the patch above needs to be applied to build it.\n\n\nNice. If I may ask, what kind of host and/or configuration have you\nused to reach a state where the code can be compiled and run tests\nwith meson? If you have found specific steps, it may be a good thing\nto document that on the wiki, say around [1].\n\nPerhaps you have not included TAP? It may be fine in terms of runtime\nchecks and coverage.\n\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\n\n\n\n\n\n\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think\n we really want to build with x64_arm64, i.e. to generate native\n arm64 binaries. Setting just x64 will not do that, AIUI.\nI tried that with the buidfarm, setting that in the config file's\n call to PGBuild::VSenv::getenv().\nThat upset msvc_gendef.pl, so I added this there to keep it\n happy:\n\n$arch = 'x86_64' if $arch eq 'aarch64';\n\nAfter that things went ok until I got this:\n[1453/2088] \"link\" @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library src\\backend\\postgres.exe.lib\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\n symbol _mm_pause referenced in function perform_spin_delay\n src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved\n externals \n\n\n\nI haven't made further progress, but I will return to it in the\n next day or so.\nWhile this will be nice to have, I think it won't really matter\n until there is ARM64 support in released versions of Windows\n Server. AFAICT they still only sell versions for x86_64I've tried it with my patch attached previously and x64_arm64 and it works fine. builds using the buildfarm as well.Is there a definitive way to figure out if the binaries are x64_arm64Dave","msg_date":"Thu, 25 Jan 2024 21:03:51 +0000","msg_from":"Anthony Roberts <anthony.roberts@linaro.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-01-25 Th 15:56, Dave Cramer wrote:\n>\n>\n> On Thu, 25 Jan 2024 at 14:31, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>\n> On 2024-01-25 Th 08:45, Dave Cramer wrote:\n>>\n>>\n>>\n>>\n>> I tried running my buildfarm using my git repo and my branch, but\n>> get the following error\n>> Status Line: 492 bad branch parameter\n>> Content:\n>> bad branch parameter fix_arm\n>>\n>> Web txn failed with status: 1\n>\n>\n>\n> You can't use your own branch with the buildfarm, you need to let\n> it set up its own branches.\n>\n>\n> So I guess I have to wait until this patch is merged in ?\n>\n>\n\nNot quite.\n\nThere's a switch that lets you test your own code. If you say \n--from-source /path/to/repo it will run in whatever state the repo is \nin. It won't care what branch you are on, and it won't try to update the \nrepo. It won't report the results to the server, but it will build and \ntest everything just like in a regular run. (On the \"eat your own dog \nfood\" principle I use this mode a lot.) If you use that mode you \nprobably also want to specify --verbose as well.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-01-25 Th 15:56, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\nOn Thu, 25 Jan 2024 at\n 14:31, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-01-25 Th 08:45, Dave Cramer wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nI tried running my buildfarm using my git\n repo and my branch, but get the following error\nStatus Line: 492 bad branch parameter\nContent:\nbad branch parameter fix_arm\n\n\nWeb txn failed with status: 1 \n\n\n\n\n\n\n\n\n\nYou can't use your own branch with the buildfarm, you\n need to let it set up its own branches.\n\n\n\n\nSo I guess I have to wait until this patch is merged in ?\n\n\n\n\n\n\n\n\n\nNot quite. \n\nThere's a switch that lets you test your own code. If you say\n --from-source /path/to/repo it will run in whatever state the\n repo is in. It won't care what branch you are on, and it won't try\n to update the repo. It won't report the results to the server, but\n it will build and test everything just like in a regular run. (On\n the \"eat your own dog food\" principle I use this mode a lot.) If\n you use that mode you probably also want to specify --verbose as\n well.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Thu, 25 Jan 2024 16:14:25 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, 25 Jan 2024 at 16:04, Anthony Roberts <anthony.roberts@linaro.org>\nwrote:\n\n> Hi David,\n>\n> Unix \"file\" or \"dumpbin /headers\" in vcvarsall are your best bets.\n>\n> Thanks,\n> Anthony\n>\n\n\nSo there is another way, select the file in Windows Explorer and right\nclick, in the compatibility tab if the \"Windows on ARM\" is greyed out it is\nan arm binary.\n\nSo far mine are not :(\n\nThanks,\n\nDave\n\n>\n\nOn Thu, 25 Jan 2024 at 16:04, Anthony Roberts <anthony.roberts@linaro.org> wrote:Hi David,Unix \"file\" or \"dumpbin /headers\" in vcvarsall are your best bets.Thanks,AnthonySo there is another way, select the file in Windows Explorer and right click, in the compatibility tab if the \"Windows on ARM\" is greyed out it is an arm binary.So far mine are not :(Thanks,Dave","msg_date":"Thu, 25 Jan 2024 16:17:15 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>\n>\n> On Thu, 25 Jan 2024 at 16:04, Anthony Roberts \n> <anthony.roberts@linaro.org> wrote:\n>\n> Hi David,\n>\n> Unix \"file\" or \"dumpbin /headers\" in vcvarsall are your best bets.\n>\n> Thanks,\n> Anthony\n>\n>\n>\n> So there is another way, select the file in Windows Explorer and right \n> click, in the compatibility tab if the \"Windows on ARM\" is greyed out \n> it is an arm binary.\n>\n> So far mine are not :(\n>\n\nYeah, I think the default Developer Command Prompt for VS2022 is set up \nfor x86 builds. AIUI you should start by executing \"vcvarsall x64_arm64\".\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-01-25 Th 16:17, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\nOn Thu, 25 Jan 2024 at\n 16:04, Anthony Roberts <anthony.roberts@linaro.org>\n wrote:\n\n\nHi David,\n \n\nUnix \"file\" or \"dumpbin /headers\" in\n vcvarsall are your best bets.\n\n\nThanks,\nAnthony\n\n\n\n\n\n\nSo there is another way, select the file in Windows\n Explorer and right click, in the compatibility tab if the\n \"Windows on ARM\" is greyed out it is an arm binary.\n\n\nSo far mine are not :(\n\n\n\n\n\n\n\nYeah, I think the default Developer Command Prompt for VS2022 is\n set up for x86 builds. AIUI you should start by executing\n \"vcvarsall x64_arm64\".\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Thu, 25 Jan 2024 16:32:39 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>\n>\n>\n> On Thu, 25 Jan 2024 at 16:04, Anthony Roberts <anthony.roberts@linaro.org>\n> wrote:\n>\n>> Hi David,\n>>\n>> Unix \"file\" or \"dumpbin /headers\" in vcvarsall are your best bets.\n>>\n>> Thanks,\n>> Anthony\n>>\n>\n>\n> So there is another way, select the file in Windows Explorer and right\n> click, in the compatibility tab if the \"Windows on ARM\" is greyed out it is\n> an arm binary.\n>\n> So far mine are not :(\n>\n>\n> Yeah, I think the default Developer Command Prompt for VS2022 is set up\n> for x86 builds. AIUI you should start by executing \"vcvarsall x64_arm64\".\n>\n\nYup, now I'm in the same state you are\n\nDave\n\nOn Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-01-25 Th 16:17, Dave Cramer\n wrote:\n\n\n\n\n\n\n\nOn Thu, 25 Jan 2024 at\n 16:04, Anthony Roberts <anthony.roberts@linaro.org>\n wrote:\n\n\nHi David,\n \n\nUnix \"file\" or \"dumpbin /headers\" in\n vcvarsall are your best bets.\n\n\nThanks,\nAnthony\n\n\n\n\n\n\nSo there is another way, select the file in Windows\n Explorer and right click, in the compatibility tab if the\n \"Windows on ARM\" is greyed out it is an arm binary.\n\n\nSo far mine are not :(\n\n\n\n\n\n\n\nYeah, I think the default Developer Command Prompt for VS2022 is\n set up for x86 builds. AIUI you should start by executing\n \"vcvarsall x64_arm64\".Yup, now I'm in the same state you are Dave","msg_date":"Thu, 25 Jan 2024 16:52:30 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net> wrote:\n>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>> Yeah, I think the default Developer Command Prompt for VS2022 is set up\n>> for x86 builds. AIUI you should start by executing \"vcvarsall x64_arm64\".\n> \n> Yup, now I'm in the same state you are\n\nWait a minute here. Based on [1], x64_arm64 means you can use a x64\nhost and you'll be able to produce ARM64 builds, still these will not\nbe able to run on the host where they were built. How much of the\npatch posted upthread is required to produce such builds? Basically \neverything from it, I guess, so as build dependencies can be\nsatisfied?\n\n[1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n--\nMichael","msg_date":"Fri, 26 Jan 2024 10:32:48 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"\nOn 2024-01-25 Th 20:32, Michael Paquier wrote:\n> On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n>> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>>> Yeah, I think the default Developer Command Prompt for VS2022 is set up\n>>> for x86 builds. AIUI you should start by executing \"vcvarsall x64_arm64\".\n>> Yup, now I'm in the same state you are\n> Wait a minute here. Based on [1], x64_arm64 means you can use a x64\n> host and you'll be able to produce ARM64 builds, still these will not\n> be able to run on the host where they were built. How much of the\n> patch posted upthread is required to produce such builds? Basically\n> everything from it, I guess, so as build dependencies can be\n> satisfied?\n>\n> [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n\n\nIf you look at the table here x86 and x64 are the only supported host \narchitectures. But that's OK, the x64 binaries will run on arm64 (W11 \nARM64 has x64 emulation builtin). If that didn't work Dave and I would \nnot have got as far as we have. But you want the x64_arm64 argument to \nvcvarsall so you will get ARM64 output.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Fri, 26 Jan 2024 07:36:25 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\n> wrote:\n> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n> >>> Yeah, I think the default Developer Command Prompt for VS2022 is set up\n> >>> for x86 builds. AIUI you should start by executing \"vcvarsall\n> x64_arm64\".\n> >> Yup, now I'm in the same state you are\n> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\n> > host and you'll be able to produce ARM64 builds, still these will not\n> > be able to run on the host where they were built. How much of the\n> > patch posted upthread is required to produce such builds? Basically\n> > everything from it, I guess, so as build dependencies can be\n> > satisfied?\n> >\n> > [1]:\n> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>\n>\n> If you look at the table here x86 and x64 are the only supported host\n> architectures. But that's OK, the x64 binaries will run on arm64 (W11\n> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\n> not have got as far as we have. But you want the x64_arm64 argument to\n> vcvarsall so you will get ARM64 output.\n>\n\nI've rebuilt it using x64_arm64 and with the attached (very naive patch)\nand I still get an x64 binary :(\n\n>\n>","msg_date":"Fri, 26 Jan 2024 09:18:50 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-01-26 Fr 09:18, Dave Cramer wrote:\n>\n>\n> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>\n> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\n> <andrew@dunslane.net> wrote:\n> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n> >>> Yeah, I think the default Developer Command Prompt for VS2022\n> is set up\n> >>> for x86 builds. AIUI you should start by executing \"vcvarsall\n> x64_arm64\".\n> >> Yup, now I'm in the same state you are\n> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\n> > host and you'll be able to produce ARM64 builds, still these\n> will not\n> > be able to run on the host where they were built. How much of the\n> > patch posted upthread is required to produce such builds? Basically\n> > everything from it, I guess, so as build dependencies can be\n> > satisfied?\n> >\n> > [1]:\n> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>\n>\n> If you look at the table here x86 and x64 are the only supported host\n> architectures. But that's OK, the x64 binaries will run on arm64 (W11\n> ARM64 has x64 emulation builtin). If that didn't work Dave and I\n> would\n> not have got as far as we have. But you want the x64_arm64\n> argument to\n> vcvarsall so you will get ARM64 output.\n>\n>\n> I've rebuilt it using x64_arm64 and with the attached (very naive \n> patch) and I still get an x64 binary :(\n\n\nI am definitely getting ARM64 binaries (e.g. for initdb.exe the Windows \non ARM compatibility setting is greyed out)\n\n\nI'll try your patch.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-01-26 Fr 09:18, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\n\nOn Fri, 26 Jan 2024 at\n 07:36, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n On 2024-01-25 Th 20:32, Michael Paquier wrote:\n > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer\n wrote:\n >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n >>> Yeah, I think the default Developer Command\n Prompt for VS2022 is set up\n >>> for x86 builds. AIUI you should start by\n executing \"vcvarsall x64_arm64\".\n >> Yup, now I'm in the same state you are\n > Wait a minute here. Based on [1], x64_arm64 means\n you can use a x64\n > host and you'll be able to produce ARM64 builds,\n still these will not\n > be able to run on the host where they were built. \n How much of the\n > patch posted upthread is required to produce such\n builds? Basically\n > everything from it, I guess, so as build dependencies\n can be\n > satisfied?\n >\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n\n\n If you look at the table here x86 and x64 are the only\n supported host \n architectures. But that's OK, the x64 binaries will run on\n arm64 (W11 \n ARM64 has x64 emulation builtin). If that didn't work Dave\n and I would \n not have got as far as we have. But you want the x64_arm64\n argument to \n vcvarsall so you will get ARM64 output.\n\n\n\nI've rebuilt it using x64_arm64 and with the attached\n (very naive patch) and I still get an x64 binary :(\n\n\n\n\n\n\nI am definitely getting ARM64 binaries (e.g. for initdb.exe the\n Windows on ARM compatibility setting is greyed out)\n\n\nI'll try your patch.\n\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Mon, 29 Jan 2024 10:35:51 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-01-26 Fr 09:18, Dave Cramer wrote:\n>\n>\n> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>\n> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\n> <andrew@dunslane.net> wrote:\n> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n> >>> Yeah, I think the default Developer Command Prompt for VS2022\n> is set up\n> >>> for x86 builds. AIUI you should start by executing \"vcvarsall\n> x64_arm64\".\n> >> Yup, now I'm in the same state you are\n> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\n> > host and you'll be able to produce ARM64 builds, still these\n> will not\n> > be able to run on the host where they were built. How much of the\n> > patch posted upthread is required to produce such builds? Basically\n> > everything from it, I guess, so as build dependencies can be\n> > satisfied?\n> >\n> > [1]:\n> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>\n>\n> If you look at the table here x86 and x64 are the only supported host\n> architectures. But that's OK, the x64 binaries will run on arm64 (W11\n> ARM64 has x64 emulation builtin). If that didn't work Dave and I\n> would\n> not have got as far as we have. But you want the x64_arm64\n> argument to\n> vcvarsall so you will get ARM64 output.\n>\n>\n> I've rebuilt it using x64_arm64 and with the attached (very naive \n> patch) and I still get an x64 binary :(\n\n\nWith this patch I still get a build error, but it's different :-)\n\n\n[1406/2088] \"link\" @src/backend/postgres.exe.rsp\nFAILED: src/backend/postgres.exe src/backend/postgres.pdb\n\"link\" @src/backend/postgres.exe.rsp\n Creating library src\\backend\\postgres.exe.lib\n\nstorage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol \nspin_delay referenced in function perform_spin_delay\n\nsrc\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-01-26 Fr 09:18, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\n\nOn Fri, 26 Jan 2024 at\n 07:36, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n On 2024-01-25 Th 20:32, Michael Paquier wrote:\n > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer\n wrote:\n >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n >>> Yeah, I think the default Developer Command\n Prompt for VS2022 is set up\n >>> for x86 builds. AIUI you should start by\n executing \"vcvarsall x64_arm64\".\n >> Yup, now I'm in the same state you are\n > Wait a minute here. Based on [1], x64_arm64 means\n you can use a x64\n > host and you'll be able to produce ARM64 builds,\n still these will not\n > be able to run on the host where they were built. \n How much of the\n > patch posted upthread is required to produce such\n builds? Basically\n > everything from it, I guess, so as build dependencies\n can be\n > satisfied?\n >\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n\n\n If you look at the table here x86 and x64 are the only\n supported host \n architectures. But that's OK, the x64 binaries will run on\n arm64 (W11 \n ARM64 has x64 emulation builtin). If that didn't work Dave\n and I would \n not have got as far as we have. But you want the x64_arm64\n argument to \n vcvarsall so you will get ARM64 output.\n\n\n\nI've rebuilt it using x64_arm64 and with the attached\n (very naive patch) and I still get an x64 binary :(\n\n\n\n\n\n\nWith this patch I still get a build error, but it's different :-)\n\n\n[1406/2088] \"link\" @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library src\\backend\\postgres.exe.lib\n\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\n symbol spin_delay referenced in function perform_spin_delay\n\n src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved\n externals\n\n\n\ncheers\n\n\nandrew\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Mon, 29 Jan 2024 11:16:41 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Dave Cramer\nwww.postgres.rocks\n\n\nOn Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\n>\n>\n>\n> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>>\n>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\n>> wrote:\n>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>> >>> Yeah, I think the default Developer Command Prompt for VS2022 is set\n>> up\n>> >>> for x86 builds. AIUI you should start by executing \"vcvarsall\n>> x64_arm64\".\n>> >> Yup, now I'm in the same state you are\n>> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\n>> > host and you'll be able to produce ARM64 builds, still these will not\n>> > be able to run on the host where they were built. How much of the\n>> > patch posted upthread is required to produce such builds? Basically\n>> > everything from it, I guess, so as build dependencies can be\n>> > satisfied?\n>> >\n>> > [1]:\n>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>>\n>>\n>> If you look at the table here x86 and x64 are the only supported host\n>> architectures. But that's OK, the x64 binaries will run on arm64 (W11\n>> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\n>> not have got as far as we have. But you want the x64_arm64 argument to\n>> vcvarsall so you will get ARM64 output.\n>>\n>\n> I've rebuilt it using x64_arm64 and with the attached (very naive patch)\n> and I still get an x64 binary :(\n>\n>\n> With this patch I still get a build error, but it's different :-)\n>\n>\n> [1406/2088] \"link\" @src/backend/postgres.exe.rsp\n> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n> \"link\" @src/backend/postgres.exe.rsp\n> Creating library src\\backend\\postgres.exe.lib\n>\n> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\n> spin_delay referenced in function perform_spin_delay\n>\n> src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n>\n\nDid you add the latest lock.patch ?\n\nDave\n\nDave Cramerwww.postgres.rocksOn Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-01-26 Fr 09:18, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\nOn Fri, 26 Jan 2024 at\n 07:36, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n On 2024-01-25 Th 20:32, Michael Paquier wrote:\n > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer\n wrote:\n >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n >>> Yeah, I think the default Developer Command\n Prompt for VS2022 is set up\n >>> for x86 builds. AIUI you should start by\n executing \"vcvarsall x64_arm64\".\n >> Yup, now I'm in the same state you are\n > Wait a minute here. Based on [1], x64_arm64 means\n you can use a x64\n > host and you'll be able to produce ARM64 builds,\n still these will not\n > be able to run on the host where they were built. \n How much of the\n > patch posted upthread is required to produce such\n builds? Basically\n > everything from it, I guess, so as build dependencies\n can be\n > satisfied?\n >\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n\n\n If you look at the table here x86 and x64 are the only\n supported host \n architectures. But that's OK, the x64 binaries will run on\n arm64 (W11 \n ARM64 has x64 emulation builtin). If that didn't work Dave\n and I would \n not have got as far as we have. But you want the x64_arm64\n argument to \n vcvarsall so you will get ARM64 output.\n\n\n\nI've rebuilt it using x64_arm64 and with the attached\n (very naive patch) and I still get an x64 binary :(\n\n\n\n\n\n\nWith this patch I still get a build error, but it's different :-)\n\n\n[1406/2088] \"link\" @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library src\\backend\\postgres.exe.lib\n\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\n symbol spin_delay referenced in function perform_spin_delay\n\n src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved\n externalsDid you add the latest lock.patch ?Dave","msg_date":"Mon, 29 Jan 2024 11:20:18 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-01-29 Mo 11:20, Dave Cramer wrote:\n>\n> Dave Cramer\n> www.postgres.rocks\n>\n>\n> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>\n> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\n>>\n>>\n>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan\n>> <andrew@dunslane.net> wrote:\n>>\n>>\n>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\n>> <andrew@dunslane.net> wrote:\n>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>> >>> Yeah, I think the default Developer Command Prompt for\n>> VS2022 is set up\n>> >>> for x86 builds. AIUI you should start by executing\n>> \"vcvarsall x64_arm64\".\n>> >> Yup, now I'm in the same state you are\n>> > Wait a minute here. Based on [1], x64_arm64 means you can\n>> use a x64\n>> > host and you'll be able to produce ARM64 builds, still\n>> these will not\n>> > be able to run on the host where they were built. How much\n>> of the\n>> > patch posted upthread is required to produce such builds? \n>> Basically\n>> > everything from it, I guess, so as build dependencies can be\n>> > satisfied?\n>> >\n>> > [1]:\n>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>>\n>>\n>> If you look at the table here x86 and x64 are the only\n>> supported host\n>> architectures. But that's OK, the x64 binaries will run on\n>> arm64 (W11\n>> ARM64 has x64 emulation builtin). If that didn't work Dave\n>> and I would\n>> not have got as far as we have. But you want the x64_arm64\n>> argument to\n>> vcvarsall so you will get ARM64 output.\n>>\n>>\n>> I've rebuilt it using x64_arm64 and with the attached (very\n>> naive patch) and I still get an x64 binary :(\n>\n>\n> With this patch I still get a build error, but it's different :-)\n>\n>\n> [1406/2088] \"link\" @src/backend/postgres.exe.rsp\n> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n> \"link\" @src/backend/postgres.exe.rsp\n> Creating library src\\backend\\postgres.exe.lib\n>\n> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\n> symbol spin_delay referenced in function perform_spin_delay\n>\n> src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n>\n>\n> Did you add the latest lock.patch ?\n>\n>\n\n\nI'm a bit confused about exactly what needs to be applied. Can you \nsupply a complete patch to be applied to a pristine checkout that will \nlet me build?\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-01-29 Mo 11:20, Dave Cramer\n wrote:\n\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Mon, 29 Jan 2024 at\n 11:16, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-01-26 Fr 09:18, Dave Cramer wrote:\n\n\n\n\n\n\n\n\nOn Fri, 26 Jan\n 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n On 2024-01-25 Th 20:32, Michael Paquier wrote:\n > On Thu, Jan 25, 2024 at 04:52:30PM -0500,\n Dave Cramer wrote:\n >> On Thu, 25 Jan 2024 at 16:32, Andrew\n Dunstan <andrew@dunslane.net>\n wrote:\n >>> On 2024-01-25 Th 16:17, Dave Cramer\n wrote:\n >>> Yeah, I think the default Developer\n Command Prompt for VS2022 is set up\n >>> for x86 builds. AIUI you should\n start by executing \"vcvarsall x64_arm64\".\n >> Yup, now I'm in the same state you are\n > Wait a minute here. Based on [1],\n x64_arm64 means you can use a x64\n > host and you'll be able to produce ARM64\n builds, still these will not\n > be able to run on the host where they were\n built. How much of the\n > patch posted upthread is required to\n produce such builds? Basically\n > everything from it, I guess, so as build\n dependencies can be\n > satisfied?\n >\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n\n\n If you look at the table here x86 and x64 are\n the only supported host \n architectures. But that's OK, the x64 binaries\n will run on arm64 (W11 \n ARM64 has x64 emulation builtin). If that didn't\n work Dave and I would \n not have got as far as we have. But you want the\n x64_arm64 argument to \n vcvarsall so you will get ARM64 output.\n\n\n\nI've rebuilt it using x64_arm64 and with the\n attached (very naive patch) and I still get an\n x64 binary :(\n\n\n\n\n\n\nWith this patch I still get a build error, but it's\n different :-)\n\n\n[1406/2088] \"link\" @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe\n src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library src\\backend\\postgres.exe.lib\n\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved\n external symbol spin_delay referenced in function\n perform_spin_delay\n\n src\\backend\\postgres.exe : fatal error LNK1120: 1\n unresolved externals\n\n\n\n\n\nDid you add the latest lock.patch ?\n\n\n\n\n\n\n\n\n\n\n\nI'm a bit confused about exactly what needs to be applied. Can\n you supply a complete patch to be applied to a pristine checkout\n that will let me build?\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Tue, 30 Jan 2024 08:38:40 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\n>\n>\n> Dave Cramer\n> www.postgres.rocks\n>\n>\n> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>>\n>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\n>>\n>>\n>>\n>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>\n>>>\n>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\n>>> wrote:\n>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>>> >>> Yeah, I think the default Developer Command Prompt for VS2022 is set\n>>> up\n>>> >>> for x86 builds. AIUI you should start by executing \"vcvarsall\n>>> x64_arm64\".\n>>> >> Yup, now I'm in the same state you are\n>>> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\n>>> > host and you'll be able to produce ARM64 builds, still these will not\n>>> > be able to run on the host where they were built. How much of the\n>>> > patch posted upthread is required to produce such builds? Basically\n>>> > everything from it, I guess, so as build dependencies can be\n>>> > satisfied?\n>>> >\n>>> > [1]:\n>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>>>\n>>>\n>>> If you look at the table here x86 and x64 are the only supported host\n>>> architectures. But that's OK, the x64 binaries will run on arm64 (W11\n>>> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\n>>> not have got as far as we have. But you want the x64_arm64 argument to\n>>> vcvarsall so you will get ARM64 output.\n>>>\n>>\n>> I've rebuilt it using x64_arm64 and with the attached (very naive patch)\n>> and I still get an x64 binary :(\n>>\n>>\n>> With this patch I still get a build error, but it's different :-)\n>>\n>>\n>> [1406/2088] \"link\" @src/backend/postgres.exe.rsp\n>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n>> \"link\" @src/backend/postgres.exe.rsp\n>> Creating library src\\backend\\postgres.exe.lib\n>>\n>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\n>> spin_delay referenced in function perform_spin_delay\n>>\n>> src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n>>\n>\n> Did you add the latest lock.patch ?\n>\n>\n>\n>\n> I'm a bit confused about exactly what needs to be applied. Can you supply\n> a complete patch to be applied to a pristine checkout that will let me\n> build?\n>\n>\n> cheers\n>\n\nSee attached.\n\nDave","msg_date":"Tue, 30 Jan 2024 09:50:10 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-01-30 Tu 09:50, Dave Cramer wrote:\n>\n>\n> On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>\n> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\n>>\n>> Dave Cramer\n>> www.postgres.rocks <http://www.postgres.rocks>\n>>\n>>\n>> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan\n>> <andrew@dunslane.net> wrote:\n>>\n>>\n>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\n>>>\n>>>\n>>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan\n>>> <andrew@dunslane.net> wrote:\n>>>\n>>>\n>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer\n>>> wrote:\n>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\n>>> <andrew@dunslane.net> wrote:\n>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>>> >>> Yeah, I think the default Developer Command Prompt\n>>> for VS2022 is set up\n>>> >>> for x86 builds. AIUI you should start by executing\n>>> \"vcvarsall x64_arm64\".\n>>> >> Yup, now I'm in the same state you are\n>>> > Wait a minute here. Based on [1], x64_arm64 means you\n>>> can use a x64\n>>> > host and you'll be able to produce ARM64 builds, still\n>>> these will not\n>>> > be able to run on the host where they were built. How\n>>> much of the\n>>> > patch posted upthread is required to produce such\n>>> builds? Basically\n>>> > everything from it, I guess, so as build dependencies\n>>> can be\n>>> > satisfied?\n>>> >\n>>> > [1]:\n>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>>>\n>>>\n>>> If you look at the table here x86 and x64 are the only\n>>> supported host\n>>> architectures. But that's OK, the x64 binaries will run\n>>> on arm64 (W11\n>>> ARM64 has x64 emulation builtin). If that didn't work\n>>> Dave and I would\n>>> not have got as far as we have. But you want the\n>>> x64_arm64 argument to\n>>> vcvarsall so you will get ARM64 output.\n>>>\n>>>\n>>> I've rebuilt it using x64_arm64 and with the attached (very\n>>> naive patch) and I still get an x64 binary :(\n>>\n>>\n>> With this patch I still get a build error, but it's different :-)\n>>\n>>\n>> [1406/2088] \"link\" @src/backend/postgres.exe.rsp\n>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n>> \"link\" @src/backend/postgres.exe.rsp\n>> Creating library src\\backend\\postgres.exe.lib\n>>\n>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved\n>> external symbol spin_delay referenced in function\n>> perform_spin_delay\n>>\n>> src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved\n>> externals\n>>\n>>\n>> Did you add the latest lock.patch ?\n>>\n>>\n>\n>\n> I'm a bit confused about exactly what needs to be applied. Can you\n> supply a complete patch to be applied to a pristine checkout that\n> will let me build?\n>\n>\n> cheers\n>\n>\n> See attached.\n>\n>\n\nNo, that is what is giving me the error shown above (just tried again to \nbe certain). And it's not surprising, as patch 2 #ifdef's out the \ndefinition of spin_delay().\n\nIf you can get a complete build with these patches then I suspect you're \nnot doing a proper ARM64 build.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-01-30 Tu 09:50, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\nOn Tue, 30 Jan 2024 at\n 08:38, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-01-29 Mo 11:20, Dave Cramer wrote:\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Mon, 29 Jan\n 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-01-26 Fr 09:18, Dave Cramer wrote:\n\n\n\n\n\n\n\n\nOn\n Fri, 26 Jan 2024 at 07:36, Andrew\n Dunstan <andrew@dunslane.net>\n wrote:\n\n\n On 2024-01-25 Th 20:32, Michael\n Paquier wrote:\n > On Thu, Jan 25, 2024 at\n 04:52:30PM -0500, Dave Cramer wrote:\n >> On Thu, 25 Jan 2024 at 16:32,\n Andrew Dunstan <andrew@dunslane.net>\n wrote:\n >>> On 2024-01-25 Th 16:17,\n Dave Cramer wrote:\n >>> Yeah, I think the default\n Developer Command Prompt for VS2022 is\n set up\n >>> for x86 builds. AIUI you\n should start by executing \"vcvarsall\n x64_arm64\".\n >> Yup, now I'm in the same\n state you are\n > Wait a minute here. Based on\n [1], x64_arm64 means you can use a x64\n > host and you'll be able to\n produce ARM64 builds, still these will\n not\n > be able to run on the host where\n they were built. How much of the\n > patch posted upthread is required\n to produce such builds? Basically\n > everything from it, I guess, so\n as build dependencies can be\n > satisfied?\n >\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n\n\n If you look at the table here x86 and\n x64 are the only supported host \n architectures. But that's OK, the x64\n binaries will run on arm64 (W11 \n ARM64 has x64 emulation builtin). If\n that didn't work Dave and I would \n not have got as far as we have. But\n you want the x64_arm64 argument to \n vcvarsall so you will get ARM64\n output.\n\n\n\nI've rebuilt it using x64_arm64\n and with the attached (very naive\n patch) and I still get an x64 binary\n :(\n\n\n\n\n\n\nWith this patch I still get a build error,\n but it's different :-)\n\n\n[1406/2088] \"link\"\n @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe\n src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library\n src\\backend\\postgres.exe.lib\n\n storage_lmgr_s_lock.c.obj : error LNK2019:\n unresolved external symbol spin_delay\n referenced in function perform_spin_delay\n\n src\\backend\\postgres.exe : fatal error\n LNK1120: 1 unresolved externals\n\n\n\n\n\nDid you add the latest lock.patch ?\n\n\n\n\n\n\n\n\n\n\n\nI'm a bit confused about exactly what needs to be\n applied. Can you supply a complete patch to be applied\n to a pristine checkout that will let me build?\n\n\ncheers\n\n\n\n\nSee attached.\n\n\n\n\n\n\n\n\n\nNo, that is what is giving me the error shown above (just tried\n again to be certain). And it's not surprising, as patch 2 #ifdef's\n out the definition of spin_delay().\nIf you can get a complete build with these patches then I suspect\n you're not doing a proper ARM64 build.\n\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Tue, 30 Jan 2024 16:56:19 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, Jan 30, 2024 at 4:56 PM Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-01-30 Tu 09:50, Dave Cramer wrote:\n>\n>\n>\n> On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>>\n>> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\n>>\n>>\n>> Dave Cramer\n>> www.postgres.rocks\n>>\n>>\n>> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>\n>>>\n>>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\n>>>\n>>>\n>>>\n>>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net>\n>>> wrote:\n>>>\n>>>>\n>>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n>>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n>>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\n>>>> wrote:\n>>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>>>> >>> Yeah, I think the default Developer Command Prompt for VS2022 is\n>>>> set up\n>>>> >>> for x86 builds. AIUI you should start by executing \"vcvarsall\n>>>> x64_arm64\".\n>>>> >> Yup, now I'm in the same state you are\n>>>> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\n>>>> > host and you'll be able to produce ARM64 builds, still these will not\n>>>> > be able to run on the host where they were built. How much of the\n>>>> > patch posted upthread is required to produce such builds? Basically\n>>>> > everything from it, I guess, so as build dependencies can be\n>>>> > satisfied?\n>>>> >\n>>>> > [1]:\n>>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>>>>\n>>>>\n>>>> If you look at the table here x86 and x64 are the only supported host\n>>>> architectures. But that's OK, the x64 binaries will run on arm64 (W11\n>>>> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\n>>>> not have got as far as we have. But you want the x64_arm64 argument to\n>>>> vcvarsall so you will get ARM64 output.\n>>>>\n>>>\n>>> I've rebuilt it using x64_arm64 and with the attached (very naive\n>>> patch) and I still get an x64 binary :(\n>>>\n>>>\n>>> With this patch I still get a build error, but it's different :-)\n>>>\n>>>\n>>> [1406/2088] \"link\" @src/backend/postgres.exe.rsp\n>>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n>>> \"link\" @src/backend/postgres.exe.rsp\n>>> Creating library src\\backend\\postgres.exe.lib\n>>>\n>>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\n>>> spin_delay referenced in function perform_spin_delay\n>>>\n>>> src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n>>>\n>>\n>> Did you add the latest lock.patch ?\n>>\n>>\n>>\n>>\n>> I'm a bit confused about exactly what needs to be applied. Can you supply\n>> a complete patch to be applied to a pristine checkout that will let me\n>> build?\n>>\n>>\n>> cheers\n>>\n>\n> See attached.\n>\n>\n>\n> No, that is what is giving me the error shown above (just tried again to\n> be certain). And it's not surprising, as patch 2 #ifdef's out the\n> definition of spin_delay().\n>\n> If you can get a complete build with these patches then I suspect you're\n> not doing a proper ARM64 build.\n>\n>\nOkay I will look when I get home in a week\n\nDave\n\n>\n> cheers\n>\n>\n> andrew\n>\n> --\n> Andrew Dunstan\n> EDB: https://www.enterprisedb.com\n>\n>\n\nOn Tue, Jan 30, 2024 at 4:56 PM Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-01-30 Tu 09:50, Dave Cramer\n wrote:\n\n\n\n\n\n\n\nOn Tue, 30 Jan 2024 at\n 08:38, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-01-29 Mo 11:20, Dave Cramer wrote:\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Mon, 29 Jan\n 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-01-26 Fr 09:18, Dave Cramer wrote:\n\n\n\n\n\n\n\n\nOn\n Fri, 26 Jan 2024 at 07:36, Andrew\n Dunstan <andrew@dunslane.net>\n wrote:\n\n\n On 2024-01-25 Th 20:32, Michael\n Paquier wrote:\n > On Thu, Jan 25, 2024 at\n 04:52:30PM -0500, Dave Cramer wrote:\n >> On Thu, 25 Jan 2024 at 16:32,\n Andrew Dunstan <andrew@dunslane.net>\n wrote:\n >>> On 2024-01-25 Th 16:17,\n Dave Cramer wrote:\n >>> Yeah, I think the default\n Developer Command Prompt for VS2022 is\n set up\n >>> for x86 builds. AIUI you\n should start by executing \"vcvarsall\n x64_arm64\".\n >> Yup, now I'm in the same\n state you are\n > Wait a minute here. Based on\n [1], x64_arm64 means you can use a x64\n > host and you'll be able to\n produce ARM64 builds, still these will\n not\n > be able to run on the host where\n they were built. How much of the\n > patch posted upthread is required\n to produce such builds? Basically\n > everything from it, I guess, so\n as build dependencies can be\n > satisfied?\n >\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n\n\n If you look at the table here x86 and\n x64 are the only supported host \n architectures. But that's OK, the x64\n binaries will run on arm64 (W11 \n ARM64 has x64 emulation builtin). If\n that didn't work Dave and I would \n not have got as far as we have. But\n you want the x64_arm64 argument to \n vcvarsall so you will get ARM64\n output.\n\n\n\nI've rebuilt it using x64_arm64\n and with the attached (very naive\n patch) and I still get an x64 binary\n :(\n\n\n\n\n\n\nWith this patch I still get a build error,\n but it's different :-)\n\n\n[1406/2088] \"link\"\n @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe\n src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library\n src\\backend\\postgres.exe.lib\n\n storage_lmgr_s_lock.c.obj : error LNK2019:\n unresolved external symbol spin_delay\n referenced in function perform_spin_delay\n\n src\\backend\\postgres.exe : fatal error\n LNK1120: 1 unresolved externals\n\n\n\n\n\nDid you add the latest lock.patch ?\n\n\n\n\n\n\n\n\n\n\n\nI'm a bit confused about exactly what needs to be\n applied. Can you supply a complete patch to be applied\n to a pristine checkout that will let me build?\n\n\ncheers\n\n\n\n\nSee attached.\n\n\n\n\n\n\n\n\n\nNo, that is what is giving me the error shown above (just tried\n again to be certain). And it's not surprising, as patch 2 #ifdef's\n out the definition of spin_delay().\nIf you can get a complete build with these patches then I suspect\n you're not doing a proper ARM64 build.\n\nOkay I will look when I get home in a weekDave \n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Tue, 30 Jan 2024 17:54:01 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-01-30 Tu 17:54, Dave Cramer wrote:\n>\n>\n>\n> On Tue, Jan 30, 2024 at 4:56 PM Andrew Dunstan <andrew@dunslane.net> \n> wrote:\n>\n>\n> On 2024-01-30 Tu 09:50, Dave Cramer wrote:\n>>\n>>\n>> On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan\n>> <andrew@dunslane.net> wrote:\n>>\n>>\n>> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\n>>>\n>>> Dave Cramer\n>>> www.postgres.rocks <http://www.postgres.rocks>\n>>>\n>>>\n>>> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan\n>>> <andrew@dunslane.net> wrote:\n>>>\n>>>\n>>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\n>>>>\n>>>>\n>>>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan\n>>>> <andrew@dunslane.net> wrote:\n>>>>\n>>>>\n>>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n>>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave\n>>>> Cramer wrote:\n>>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\n>>>> <andrew@dunslane.net> wrote:\n>>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>>>> >>> Yeah, I think the default Developer Command\n>>>> Prompt for VS2022 is set up\n>>>> >>> for x86 builds. AIUI you should start by\n>>>> executing \"vcvarsall x64_arm64\".\n>>>> >> Yup, now I'm in the same state you are\n>>>> > Wait a minute here. Based on [1], x64_arm64 means\n>>>> you can use a x64\n>>>> > host and you'll be able to produce ARM64 builds,\n>>>> still these will not\n>>>> > be able to run on the host where they were built.\n>>>> How much of the\n>>>> > patch posted upthread is required to produce such\n>>>> builds? Basically\n>>>> > everything from it, I guess, so as build\n>>>> dependencies can be\n>>>> > satisfied?\n>>>> >\n>>>> > [1]:\n>>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>>>>\n>>>>\n>>>> If you look at the table here x86 and x64 are the\n>>>> only supported host\n>>>> architectures. But that's OK, the x64 binaries will\n>>>> run on arm64 (W11\n>>>> ARM64 has x64 emulation builtin). If that didn't\n>>>> work Dave and I would\n>>>> not have got as far as we have. But you want the\n>>>> x64_arm64 argument to\n>>>> vcvarsall so you will get ARM64 output.\n>>>>\n>>>>\n>>>> I've rebuilt it using x64_arm64 and with the attached\n>>>> (very naive patch) and I still get an x64 binary :(\n>>>\n>>>\n>>> With this patch I still get a build error, but it's\n>>> different :-)\n>>>\n>>>\n>>> [1406/2088] \"link\" @src/backend/postgres.exe.rsp\n>>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n>>> \"link\" @src/backend/postgres.exe.rsp\n>>> Creating library src\\backend\\postgres.exe.lib\n>>>\n>>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved\n>>> external symbol spin_delay referenced in function\n>>> perform_spin_delay\n>>>\n>>> src\\backend\\postgres.exe : fatal error LNK1120: 1\n>>> unresolved externals\n>>>\n>>>\n>>> Did you add the latest lock.patch ?\n>>>\n>>>\n>>\n>>\n>> I'm a bit confused about exactly what needs to be applied.\n>> Can you supply a complete patch to be applied to a pristine\n>> checkout that will let me build?\n>>\n>>\n>> cheers\n>>\n>>\n>> See attached.\n>>\n>>\n>\n> No, that is what is giving me the error shown above (just tried\n> again to be certain). And it's not surprising, as patch 2 #ifdef's\n> out the definition of spin_delay().\n>\n> If you can get a complete build with these patches then I suspect\n> you're not doing a proper ARM64 build.\n>\n>\n> Okay I will look when I get home in a week\n\n\nI made some progress. The attached is mostly taken from \n<https://postgr.es/m/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56@linaro.org>\n\nWith it applied I was able to get a successful build using the buildfarm \nclient. However, there are access violations when running some tests, so \nthere is still some work to do, apparently.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com","msg_date":"Wed, 31 Jan 2024 10:20:54 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 31.01.24 16:20, Andrew Dunstan wrote:\n> - PostgreSQL will only build for the x64 architecture on 64-bit Windows. \n> + PostgreSQL will only build for the x64 and ARM64 architecture on \n> 64-bit Windows.\n\nAre there any other 64-bit architectures for Windows?\n\nPossibly, the original sentence was meant to communicate that ARM was \nnot supported, in which case it could now be removed?\n\n\n\n","msg_date":"Wed, 31 Jan 2024 16:34:13 +0100","msg_from":"Peter Eisentraut <peter@eisentraut.org>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"\nOn 2024-01-31 We 10:34, Peter Eisentraut wrote:\n> On 31.01.24 16:20, Andrew Dunstan wrote:\n>> - PostgreSQL will only build for the x64 architecture on 64-bit \n>> Windows. + PostgreSQL will only build for the x64 and ARM64 \n>> architecture on 64-bit Windows.\n>\n> Are there any other 64-bit architectures for Windows?\n>\n> Possibly, the original sentence was meant to communicate that ARM was \n> not supported, in which case it could now be removed?\n\n\nx86? That is in fact the default setting for VS even on ARM64.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 31 Jan 2024 10:41:57 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Wed, 31 Jan 2024 at 10:21, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-01-30 Tu 17:54, Dave Cramer wrote:\n>\n>\n>\n>\n> On Tue, Jan 30, 2024 at 4:56 PM Andrew Dunstan <andrew@dunslane.net>\n> wrote:\n>\n>>\n>> On 2024-01-30 Tu 09:50, Dave Cramer wrote:\n>>\n>>\n>>\n>> On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>\n>>>\n>>> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\n>>>\n>>>\n>>> Dave Cramer\n>>> www.postgres.rocks\n>>>\n>>>\n>>> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net>\n>>> wrote:\n>>>\n>>>>\n>>>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\n>>>>\n>>>>\n>>>>\n>>>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net>\n>>>> wrote:\n>>>>\n>>>>>\n>>>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\n>>>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\n>>>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\n>>>>> wrote:\n>>>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\n>>>>> >>> Yeah, I think the default Developer Command Prompt for VS2022 is\n>>>>> set up\n>>>>> >>> for x86 builds. AIUI you should start by executing \"vcvarsall\n>>>>> x64_arm64\".\n>>>>> >> Yup, now I'm in the same state you are\n>>>>> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\n>>>>> > host and you'll be able to produce ARM64 builds, still these will not\n>>>>> > be able to run on the host where they were built. How much of the\n>>>>> > patch posted upthread is required to produce such builds? Basically\n>>>>> > everything from it, I guess, so as build dependencies can be\n>>>>> > satisfied?\n>>>>> >\n>>>>> > [1]:\n>>>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n>>>>>\n>>>>>\n>>>>> If you look at the table here x86 and x64 are the only supported host\n>>>>> architectures. But that's OK, the x64 binaries will run on arm64 (W11\n>>>>> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\n>>>>> not have got as far as we have. But you want the x64_arm64 argument to\n>>>>> vcvarsall so you will get ARM64 output.\n>>>>>\n>>>>\n>>>> I've rebuilt it using x64_arm64 and with the attached (very naive\n>>>> patch) and I still get an x64 binary :(\n>>>>\n>>>>\n>>>> With this patch I still get a build error, but it's different :-)\n>>>>\n>>>>\n>>>> [1406/2088] \"link\" @src/backend/postgres.exe.rsp\n>>>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\n>>>> \"link\" @src/backend/postgres.exe.rsp\n>>>> Creating library src\\backend\\postgres.exe.lib\n>>>>\n>>>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\n>>>> spin_delay referenced in function perform_spin_delay\n>>>>\n>>>> src\\backend\\postgres.exe : fatal error LNK1120: 1 unresolved externals\n>>>>\n>>>\n>>> Did you add the latest lock.patch ?\n>>>\n>>>\n>>>\n>>>\n>>> I'm a bit confused about exactly what needs to be applied. Can you\n>>> supply a complete patch to be applied to a pristine checkout that will let\n>>> me build?\n>>>\n>>>\n>>> cheers\n>>>\n>>\n>> See attached.\n>>\n>>\n>>\n>> No, that is what is giving me the error shown above (just tried again to\n>> be certain). And it's not surprising, as patch 2 #ifdef's out the\n>> definition of spin_delay().\n>>\n>> If you can get a complete build with these patches then I suspect you're\n>> not doing a proper ARM64 build.\n>>\n>\n> Okay I will look when I get home in a week\n>\n>\n> I made some progress. The attached is mostly taken from\n> <https://postgr.es/m/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56@linaro.org>\n> <https://postgr.es/m/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56@linaro.org>\n>\n> With it applied I was able to get a successful build using the buildfarm\n> client. However, there are access violations when running some tests, so\n> there is still some work to do, apparently.\n>\n>\n> cheers\n>\n>\n> andrew\n>\n> --\n> Andrew Dunstan\n> EDB: https://www.enterprisedb.com\n>\n>\n\nThanks, this patch works and\ntesting with meon passes.\n\nI'll try the buildfarm next.\n\nDave\n\nOn Wed, 31 Jan 2024 at 10:21, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-01-30 Tu 17:54, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\nOn Tue, Jan 30, 2024 at\n 4:56 PM Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-01-30 Tu 09:50, Dave Cramer wrote:\n\n\n\n\n\n\n\nOn Tue, 30 Jan\n 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-01-29 Mo 11:20, Dave Cramer wrote:\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Mon,\n 29 Jan 2024 at 11:16, Andrew Dunstan\n <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-01-26 Fr 09:18, Dave\n Cramer wrote:\n\n\n\n\n\n\n\n\nOn Fri,\n 26 Jan 2024 at 07:36, Andrew\n Dunstan <andrew@dunslane.net>\n wrote:\n\n\n On 2024-01-25 Th 20:32,\n Michael Paquier wrote:\n > On Thu, Jan 25, 2024 at\n 04:52:30PM -0500, Dave\n Cramer wrote:\n >> On Thu, 25 Jan 2024\n at 16:32, Andrew Dunstan\n <andrew@dunslane.net>\n wrote:\n >>> On 2024-01-25\n Th 16:17, Dave Cramer wrote:\n >>> Yeah, I think\n the default Developer\n Command Prompt for VS2022 is\n set up\n >>> for x86 builds.\n AIUI you should start by\n executing \"vcvarsall\n x64_arm64\".\n >> Yup, now I'm in the\n same state you are\n > Wait a minute here. \n Based on [1], x64_arm64\n means you can use a x64\n > host and you'll be able\n to produce ARM64 builds,\n still these will not\n > be able to run on the\n host where they were built. \n How much of the\n > patch posted upthread\n is required to produce such\n builds? Basically\n > everything from it, I\n guess, so as build\n dependencies can be\n > satisfied?\n >\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\n\n\n If you look at the table\n here x86 and x64 are the\n only supported host \n architectures. But that's\n OK, the x64 binaries will\n run on arm64 (W11 \n ARM64 has x64 emulation\n builtin). If that didn't\n work Dave and I would \n not have got as far as we\n have. But you want the\n x64_arm64 argument to \n vcvarsall so you will get\n ARM64 output.\n\n\n\nI've rebuilt it using \n x64_arm64 and with the\n attached (very naive patch)\n and I still get an x64\n binary :(\n\n\n\n\n\n\nWith this patch I still get a build\n error, but it's different :-)\n\n\n[1406/2088] \"link\"\n @src/backend/postgres.exe.rsp\n FAILED: src/backend/postgres.exe\n src/backend/postgres.pdb\n \"link\" @src/backend/postgres.exe.rsp\n Creating library\n src\\backend\\postgres.exe.lib\n\n storage_lmgr_s_lock.c.obj : error\n LNK2019: unresolved external symbol\n spin_delay referenced in function\n perform_spin_delay\n\n src\\backend\\postgres.exe : fatal\n error LNK1120: 1 unresolved\n externals\n\n\n\n\n\nDid you add the latest lock.patch ?\n\n\n\n\n\n\n\n\n\n\n\nI'm a bit confused about exactly what needs\n to be applied. Can you supply a complete patch\n to be applied to a pristine checkout that will\n let me build?\n\n\ncheers\n\n\n\n\nSee attached.\n\n\n\n\n\n\n\n\n\n\n\nNo, that is what is giving me the error shown above\n (just tried again to be certain). And it's not\n surprising, as patch 2 #ifdef's out the definition of\n spin_delay().\nIf you can get a complete build with these patches then\n I suspect you're not doing a proper ARM64 build.\n\n\n \n\n\n\nOkay I will look when I get home in a week\n\n\n\n\n\nI made some progress. The attached is mostly taken from \n<https://postgr.es/m/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56@linaro.org>\nWith it applied I was able to get a successful build using the\n buildfarm client. However, there are access violations when\n running some tests, so there is still some work to do, apparently.\n\n\ncheers\n\n\nandrew\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.comThanks, this patch works and testing with meon passes. I'll try the buildfarm next.Dave","msg_date":"Tue, 6 Feb 2024 07:01:49 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n> Thanks, this patch works and\n> testing with meson passes.\n\nOnly with the version posted at [1]? Interesting, that's the same\ncontents as v8 posted upthread, minus src/tools/ because we don't need\nto care about them anymore.\n\nAndrew, what's happening on the test side? It does not seem you've\nmentioned any details about what is going wrong, or I have just missed\nthem.\n\n> I'll try the buildfarm next.\n\n[1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n--\nMichael","msg_date":"Fri, 9 Feb 2024 14:26:34 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n> > Thanks, this patch works and\n> > testing with meson passes.\n>\n> Only with the version posted at [1]? Interesting, that's the same\n> contents as v8 posted upthread, minus src/tools/ because we don't need\n> to care about them anymore.\n>\n> Andrew, what's happening on the test side? It does not seem you've\n> mentioned any details about what is going wrong, or I have just missed\n> them.\n>\n> > I'll try the buildfarm next.\n>\n> [1]:\n> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n\n\ninterestingly meson test does not produce any error\nThe buildfarm produces the following error for me:\n\n-SELECT relname, attname, coltypes, get_columns_length(coltypes)\n- FROM check_columns\n- WHERE get_columns_length(coltypes) % 8 != 0 OR\n- 'name'::regtype::oid = ANY(coltypes);\n- relname | attname | coltypes | get_columns_length\n----------+---------+----------+--------------------\n-(0 rows)\n-\n+server closed the connection unexpectedly\n+ This probably means the server terminated abnormally\n+ before or while processing the request.\n+connection to server was lost\n\nDave\n\nOn Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n> Thanks, this patch works and\n> testing with meson passes.\n\nOnly with the version posted at [1]? Interesting, that's the same\ncontents as v8 posted upthread, minus src/tools/ because we don't need\nto care about them anymore.\n\nAndrew, what's happening on the test side? It does not seem you've\nmentioned any details about what is going wrong, or I have just missed\nthem.\n\n> I'll try the buildfarm next.\n\n[1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.netinterestingly meson test does not produce any errorThe buildfarm produces the following error for me:-SELECT relname, attname, coltypes, get_columns_length(coltypes)- FROM check_columns- WHERE get_columns_length(coltypes) % 8 != 0 OR- 'name'::regtype::oid = ANY(coltypes);- relname | attname | coltypes | get_columns_length ----------+---------+----------+---------------------(0 rows)-+server closed the connection unexpectedly+ This probably means the server terminated abnormally+ before or while processing the request.+connection to server was lost Dave","msg_date":"Fri, 9 Feb 2024 07:18:13 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Dave Cramer\nwww.postgres.rocks\n\n\nOn Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks> wrote:\n\n>\n>\n>\n>\n> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:\n>\n>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n>> > Thanks, this patch works and\n>> > testing with meson passes.\n>>\n>> Only with the version posted at [1]? Interesting, that's the same\n>> contents as v8 posted upthread, minus src/tools/ because we don't need\n>> to care about them anymore.\n>>\n>> Andrew, what's happening on the test side? It does not seem you've\n>> mentioned any details about what is going wrong, or I have just missed\n>> them.\n>>\n>> > I'll try the buildfarm next.\n>>\n>> [1]:\n>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n>\n>\n> interestingly meson test does not produce any error\n> The buildfarm produces the following error for me:\n>\n> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\n> - FROM check_columns\n> - WHERE get_columns_length(coltypes) % 8 != 0 OR\n> - 'name'::regtype::oid = ANY(coltypes);\n> - relname | attname | coltypes | get_columns_length\n> ----------+---------+----------+--------------------\n> -(0 rows)\n> -\n> +server closed the connection unexpectedly\n> + This probably means the server terminated abnormally\n> + before or while processing the request.\n> +connection to server was lost\n>\n\nActually digging some more, here is the actual error\n\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\n11204) was terminated by exception 0xC0000005\n2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\nrunning: VACUUM;\n2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\n\"ntstatus.h\" for a description of the hexadecimal value.\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other\nactive server processes\n2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes\nterminated; reinitializing\n2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was\ninterrupted; last known up at 2024-02-09 13:31:01 -05\n\n\nDave\n\n>\n> Dave\n>\n\nDave Cramerwww.postgres.rocksOn Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks> wrote:On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n> Thanks, this patch works and\n> testing with meson passes.\n\nOnly with the version posted at [1]? Interesting, that's the same\ncontents as v8 posted upthread, minus src/tools/ because we don't need\nto care about them anymore.\n\nAndrew, what's happening on the test side? It does not seem you've\nmentioned any details about what is going wrong, or I have just missed\nthem.\n\n> I'll try the buildfarm next.\n\n[1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.netinterestingly meson test does not produce any errorThe buildfarm produces the following error for me:-SELECT relname, attname, coltypes, get_columns_length(coltypes)- FROM check_columns- WHERE get_columns_length(coltypes) % 8 != 0 OR- 'name'::regtype::oid = ANY(coltypes);- relname | attname | coltypes | get_columns_length ----------+---------+----------+---------------------(0 rows)-+server closed the connection unexpectedly+ This probably means the server terminated abnormally+ before or while processing the request.+connection to server was lost Actually digging some more, here is the actual error2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \"ntstatus.h\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05Dave Dave","msg_date":"Fri, 9 Feb 2024 14:23:46 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2024-02-09 14:23:46 -0500, Dave Cramer wrote:\n> > interestingly meson test does not produce any error\n> > The buildfarm produces the following error for me:\n> >\n> > -SELECT relname, attname, coltypes, get_columns_length(coltypes)\n> > - FROM check_columns\n> > - WHERE get_columns_length(coltypes) % 8 != 0 OR\n> > - 'name'::regtype::oid = ANY(coltypes);\n> > - relname | attname | coltypes | get_columns_length\n> > ----------+---------+----------+--------------------\n> > -(0 rows)\n> > -\n> > +server closed the connection unexpectedly\n> > + This probably means the server terminated abnormally\n> > + before or while processing the request.\n> > +connection to server was lost\n> >\n> \n> Actually digging some more, here is the actual error\n> \n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\n> 11204) was terminated by exception 0xC0000005\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\n> running: VACUUM;\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\n> \"ntstatus.h\" for a description of the hexadecimal value.\n\nThat's something like a segfault.\n\nOne suspicion I have is that src/port/pg_crc32c_armv8_choose.c possibly\ndoesn't properly support msvc. It seems to assume that SIGILL can be trapped,\nbut that IIRC doesn't work on windows.\n\nI'd check if the problem persists if you change\ncdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\nto\ncdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 0)\n\n\nAlso, yikes, that's an ugly way of doing hardware detection. Jumping out of a\nsignal handler into normal code. Brrr.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 9 Feb 2024 11:36:18 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Fri, 9 Feb 2024 at 14:36, Andres Freund <andres@anarazel.de> wrote:\n\n> Hi,\n>\n> On 2024-02-09 14:23:46 -0500, Dave Cramer wrote:\n> > > interestingly meson test does not produce any error\n> > > The buildfarm produces the following error for me:\n> > >\n> > > -SELECT relname, attname, coltypes, get_columns_length(coltypes)\n> > > - FROM check_columns\n> > > - WHERE get_columns_length(coltypes) % 8 != 0 OR\n> > > - 'name'::regtype::oid = ANY(coltypes);\n> > > - relname | attname | coltypes | get_columns_length\n> > > ----------+---------+----------+--------------------\n> > > -(0 rows)\n> > > -\n> > > +server closed the connection unexpectedly\n> > > + This probably means the server terminated abnormally\n> > > + before or while processing the request.\n> > > +connection to server was lost\n> > >\n> >\n> > Actually digging some more, here is the actual error\n> >\n> > 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\n> > 11204) was terminated by exception 0xC0000005\n> > 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\n> > running: VACUUM;\n> > 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\n> > \"ntstatus.h\" for a description of the hexadecimal value.\n>\n> That's something like a segfault.\n>\n> One suspicion I have is that src/port/pg_crc32c_armv8_choose.c possibly\n> doesn't properly support msvc. It seems to assume that SIGILL can be\n> trapped,\n> but that IIRC doesn't work on windows.\n>\n> I'd check if the problem persists if you change\n> cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> to\n> cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 0)\n>\n\nThis results in\n\nFAILED: src/bin/pg_checksums/pg_checksums.exe\nsrc/bin/pg_checksums/pg_checksums.pdb\n\"link\" /MACHINE:ARM64 /OUT:src/bin/pg_checksums/pg_checksums.exe\nsrc/bin/pg_checksums/pg_checksums.exe.p/win32ver.res\nsrc/bin/pg_checksums/pg_checksums.exe.p/pg_checksums.c.obj \"/release\"\n\"/nologo\" \"/DEBUG\" \"/PDB:src\\bin\\pg_checksums\\pg_checksums.pdb\"\n\"/INCREMENTAL:NO\" \"/STACK:4194304\" \"/NOEXP\" \"src/fe_utils/libpgfeutils.a\"\n\"src/common/libpgcommon.a\" \"src/port/libpgport.a\" \"ws2_32.lib\" \"ws2_32.lib\"\n\"ws2_32.lib\" \"ws2_32.lib\" \"/SUBSYSTEM:CONSOLE\" \"kernel32.lib\" \"user32.lib\"\n\"gdi32.lib\" \"winspool.lib\" \"shell32.lib\" \"ole32.lib\" \"oleaut32.lib\"\n\"uuid.lib\" \"comdlg32.lib\" \"advapi32.lib\"\nlibpgcommon.a(controldata_utils.c.obj) : error LNK2001: unresolved external\nsymbol pg_comp_crc32c\n\n\nDave\n\n\n>\n>\n> Also, yikes, that's an ugly way of doing hardware detection. Jumping out\n> of a\n> signal handler into normal code. Brrr.\n>\n> Greetings,\n>\n> Andres Freund\n>\n\nOn Fri, 9 Feb 2024 at 14:36, Andres Freund <andres@anarazel.de> wrote:Hi,\n\nOn 2024-02-09 14:23:46 -0500, Dave Cramer wrote:\n> > interestingly meson test does not produce any error\n> > The buildfarm produces the following error for me:\n> >\n> > -SELECT relname, attname, coltypes, get_columns_length(coltypes)\n> > - FROM check_columns\n> > - WHERE get_columns_length(coltypes) % 8 != 0 OR\n> > - 'name'::regtype::oid = ANY(coltypes);\n> > - relname | attname | coltypes | get_columns_length\n> > ----------+---------+----------+--------------------\n> > -(0 rows)\n> > -\n> > +server closed the connection unexpectedly\n> > + This probably means the server terminated abnormally\n> > + before or while processing the request.\n> > +connection to server was lost\n> >\n> \n> Actually digging some more, here is the actual error\n> \n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\n> 11204) was terminated by exception 0xC0000005\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\n> running: VACUUM;\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\n> \"ntstatus.h\" for a description of the hexadecimal value.\n\nThat's something like a segfault.\n\nOne suspicion I have is that src/port/pg_crc32c_armv8_choose.c possibly\ndoesn't properly support msvc. It seems to assume that SIGILL can be trapped,\nbut that IIRC doesn't work on windows.\n\nI'd check if the problem persists if you change\ncdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\nto\ncdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 0)This results in FAILED: src/bin/pg_checksums/pg_checksums.exe src/bin/pg_checksums/pg_checksums.pdb \"link\" /MACHINE:ARM64 /OUT:src/bin/pg_checksums/pg_checksums.exe src/bin/pg_checksums/pg_checksums.exe.p/win32ver.res src/bin/pg_checksums/pg_checksums.exe.p/pg_checksums.c.obj \"/release\" \"/nologo\" \"/DEBUG\" \"/PDB:src\\bin\\pg_checksums\\pg_checksums.pdb\" \"/INCREMENTAL:NO\" \"/STACK:4194304\" \"/NOEXP\" \"src/fe_utils/libpgfeutils.a\" \"src/common/libpgcommon.a\" \"src/port/libpgport.a\" \"ws2_32.lib\" \"ws2_32.lib\" \"ws2_32.lib\" \"ws2_32.lib\" \"/SUBSYSTEM:CONSOLE\" \"kernel32.lib\" \"user32.lib\" \"gdi32.lib\" \"winspool.lib\" \"shell32.lib\" \"ole32.lib\" \"oleaut32.lib\" \"uuid.lib\" \"comdlg32.lib\" \"advapi32.lib\"libpgcommon.a(controldata_utils.c.obj) : error LNK2001: unresolved external symbol pg_comp_crc32cDave \n\n\nAlso, yikes, that's an ugly way of doing hardware detection. Jumping out of a\nsignal handler into normal code. Brrr.\n\nGreetings,\n\nAndres Freund","msg_date":"Fri, 9 Feb 2024 15:32:10 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-02-09 Fr 14:23, Dave Cramer wrote:\n>\n> Dave Cramer\n> www.postgres.rocks\n>\n>\n> On Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks> \n> wrote:\n>\n>\n>\n>\n>\n> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz>\n> wrote:\n>\n> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n> > Thanks, this patch works and\n> > testing with meson passes.\n>\n> Only with the version posted at [1]? Interesting, that's the same\n> contents as v8 posted upthread, minus src/tools/ because we\n> don't need\n> to care about them anymore.\n>\n> Andrew, what's happening on the test side? It does not seem\n> you've\n> mentioned any details about what is going wrong, or I have\n> just missed\n> them.\n>\n> > I'll try the buildfarm next.\n>\n> [1]:\n> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n>\n>\n> interestingly meson test does not produce any error\n> The buildfarm produces the following error for me:\n>\n> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\n> - FROM check_columns\n> - WHERE get_columns_length(coltypes) % 8 != 0 OR\n> - 'name'::regtype::oid = ANY(coltypes);\n> - relname | attname | coltypes | get_columns_length\n> ----------+---------+----------+--------------------\n> -(0 rows)\n> -\n> +server closed the connection unexpectedly\n> +This probably means the server terminated abnormally\n> +before or while processing the request.\n> +connection to server was lost\n>\n>\n> Actually digging some more, here is the actual error\n>\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process \n> (PID 11204) was terminated by exception 0xC0000005\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process \n> was running: VACUUM;\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include \n> file \"ntstatus.h\" for a description of the hexadecimal value.\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any \n> other active server processes\n> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server \n> processes terminated; reinitializing\n> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was \n> interrupted; last known up at 2024-02-09 13:31:01 -05\n>\n>\n>\n\nYes, this is pretty much what I saw.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-02-09 Fr 14:23, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Fri, 9 Feb 2024 at\n 07:18, Dave Cramer <davecramer@postgres.rocks>\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nOn Fri, 9 Feb 2024\n at 00:26, Michael Paquier <michael@paquier.xyz>\n wrote:\n\nOn Tue, Feb 06,\n 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n > Thanks, this patch works and\n > testing with meson passes.\n\n Only with the version posted at [1]? Interesting,\n that's the same\n contents as v8 posted upthread, minus src/tools/\n because we don't need\n to care about them anymore.\n\n Andrew, what's happening on the test side? It\n does not seem you've\n mentioned any details about what is going wrong,\n or I have just missed\n them.\n\n > I'll try the buildfarm next.\n\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n\n\ninterestingly meson test does not produce any\n error\nThe buildfarm produces the following error for\n me:\n\n\n-SELECT relname, attname, coltypes,\n get_columns_length(coltypes)\n- FROM check_columns\n- WHERE get_columns_length(coltypes) % 8 != 0\n OR\n- 'name'::regtype::oid = ANY(coltypes);\n- relname | attname | coltypes |\n get_columns_length \n----------+---------+----------+--------------------\n-(0 rows)\n-\n+server closed the connection unexpectedly\n+ This\n probably means the server terminated abnormally\n+ before\n or while processing the request.\n+connection to server was lost \n\n\n\n\n\n\nActually digging some more, here is the actual error\n\n\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \"ntstatus.h\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\n\n\n\n\n\n\n\n\n\n\n\nYes, this is pretty much what I saw.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Sat, 10 Feb 2024 11:19:21 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\n>\n>\n> Dave Cramer\n> www.postgres.rocks\n>\n>\n> On Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\n> <davecramer@postgres.rocks> wrote:\n>\n>>\n>>\n>>\n>>\n>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:\n>>\n>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n>>> > Thanks, this patch works and\n>>> > testing with meson passes.\n>>>\n>>> Only with the version posted at [1]? Interesting, that's the same\n>>> contents as v8 posted upthread, minus src/tools/ because we don't need\n>>> to care about them anymore.\n>>>\n>>> Andrew, what's happening on the test side? It does not seem you've\n>>> mentioned any details about what is going wrong, or I have just missed\n>>> them.\n>>>\n>>> > I'll try the buildfarm next.\n>>>\n>>> [1]:\n>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n>>\n>>\n>> interestingly meson test does not produce any error\n>> The buildfarm produces the following error for me:\n>>\n>> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\n>> - FROM check_columns\n>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\n>> - 'name'::regtype::oid = ANY(coltypes);\n>> - relname | attname | coltypes | get_columns_length\n>> ----------+---------+----------+--------------------\n>> -(0 rows)\n>> -\n>> +server closed the connection unexpectedly\n>> + This probably means the server terminated abnormally\n>> + before or while processing the request.\n>> +connection to server was lost\n>>\n>\n> Actually digging some more, here is the actual error\n>\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\n> 11204) was terminated by exception 0xC0000005\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\n> running: VACUUM;\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\n> \"ntstatus.h\" for a description of the hexadecimal value.\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other\n> active server processes\n> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes\n> terminated; reinitializing\n> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was\n> interrupted; last known up at 2024-02-09 13:31:01 -05\n>\n>\n>\n>\n>\nSo how does one debug this ?\n\nAlso if I `run meson test` I don't see this error. What does the buildfarm\ndo differently?\n\nDave\n\n> Yes, this is pretty much what I saw.\n>\n>\n> cheers\n>\n>\n> andrew\n>\n> --\n> Andrew Dunstan\n> EDB: https://www.enterprisedb.com\n>\n>\n\nOn Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-02-09 Fr 14:23, Dave Cramer\n wrote:\n\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Fri, 9 Feb 2024 at\n 07:18, Dave Cramer <davecramer@postgres.rocks>\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nOn Fri, 9 Feb 2024\n at 00:26, Michael Paquier <michael@paquier.xyz>\n wrote:\n\nOn Tue, Feb 06,\n 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n > Thanks, this patch works and\n > testing with meson passes.\n\n Only with the version posted at [1]? Interesting,\n that's the same\n contents as v8 posted upthread, minus src/tools/\n because we don't need\n to care about them anymore.\n\n Andrew, what's happening on the test side? It\n does not seem you've\n mentioned any details about what is going wrong,\n or I have just missed\n them.\n\n > I'll try the buildfarm next.\n\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n\n\ninterestingly meson test does not produce any\n error\nThe buildfarm produces the following error for\n me:\n\n\n-SELECT relname, attname, coltypes,\n get_columns_length(coltypes)\n- FROM check_columns\n- WHERE get_columns_length(coltypes) % 8 != 0\n OR\n- 'name'::regtype::oid = ANY(coltypes);\n- relname | attname | coltypes |\n get_columns_length \n----------+---------+----------+--------------------\n-(0 rows)\n-\n+server closed the connection unexpectedly\n+ This\n probably means the server terminated abnormally\n+ before\n or while processing the request.\n+connection to server was lost \n\n\n\n\n\n\nActually digging some more, here is the actual error\n\n\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \"ntstatus.h\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\n\n\n\n\n\n\n\n\n\nSo how does one debug this ?Also if I `run meson test` I don't see this error. What does the buildfarm do differently?Dave\n\nYes, this is pretty much what I saw.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Sat, 10 Feb 2024 12:20:12 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-02-10 Sa 12:20, Dave Cramer wrote:\n>\n>\n> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>\n> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\n>>\n>> Dave Cramer\n>> www.postgres.rocks <http://www.postgres.rocks>\n>>\n>>\n>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer\n>> <davecramer@postgres.rocks> <mailto:davecramer@postgres.rocks> wrote:\n>>\n>>\n>>\n>>\n>>\n>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier\n>> <michael@paquier.xyz> wrote:\n>>\n>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n>> > Thanks, this patch works and\n>> > testing with meson passes.\n>>\n>> Only with the version posted at [1]? Interesting, that's\n>> the same\n>> contents as v8 posted upthread, minus src/tools/ because\n>> we don't need\n>> to care about them anymore.\n>>\n>> Andrew, what's happening on the test side? It does not\n>> seem you've\n>> mentioned any details about what is going wrong, or I\n>> have just missed\n>> them.\n>>\n>> > I'll try the buildfarm next.\n>>\n>> [1]:\n>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n>>\n>>\n>> interestingly meson test does not produce any error\n>> The buildfarm produces the following error for me:\n>>\n>> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\n>> - FROM check_columns\n>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\n>> - 'name'::regtype::oid = ANY(coltypes);\n>> - relname | attname | coltypes | get_columns_length\n>> ----------+---------+----------+--------------------\n>> -(0 rows)\n>> -\n>> +server closed the connection unexpectedly\n>> +This probably means the server terminated abnormally\n>> +before or while processing the request.\n>> +connection to server was lost\n>>\n>>\n>> Actually digging some more, here is the actual error\n>>\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server\n>> process (PID 11204) was terminated by exception 0xC0000005\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed\n>> process was running: VACUUM;\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C\n>> include file \"ntstatus.h\" for a description of the hexadecimal value.\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating\n>> any other active server processes\n>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server\n>> processes terminated; reinitializing\n>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system\n>> was interrupted; last known up at 2024-02-09 13:31:01 -05\n>>\n>>\n>>\n>\n>\n> So how does one debug this ?\n>\n> Also if I `run meson test` I don't see this error. What does the \n> buildfarm do differently?\n\n\nFirst it does this:\n\n\n meson test -C $pgsql --no-rebuild --suite setup\n\n\nThen it does this (jflag is for the number of jobs):\n\n\n meson test -t $meson_test_timeout $jflag -C $pgsql --logbase\n checklog --print-errorlogs --no-rebuild --suite regress\n --test-args=--no-locale\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-02-10 Sa 12:20, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\nOn Sat, 10 Feb 2024 at\n 11:19, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-02-09 Fr 14:23, Dave Cramer wrote:\n\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Fri, 9 Feb\n 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nOn Fri,\n 9 Feb 2024 at 00:26, Michael Paquier\n <michael@paquier.xyz>\n wrote:\n\nOn\n Tue, Feb 06, 2024 at 07:01:49AM -0500,\n Dave Cramer wrote:\n > Thanks, this patch works and\n > testing with meson passes.\n\n Only with the version posted at [1]? \n Interesting, that's the same\n contents as v8 posted upthread, minus\n src/tools/ because we don't need\n to care about them anymore.\n\n Andrew, what's happening on the test\n side? It does not seem you've\n mentioned any details about what is\n going wrong, or I have just missed\n them.\n\n > I'll try the buildfarm next.\n\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n\n\ninterestingly meson test does not\n produce any error\nThe buildfarm produces the following\n error for me:\n\n\n-SELECT relname, attname, coltypes,\n get_columns_length(coltypes)\n- FROM check_columns\n- WHERE get_columns_length(coltypes)\n % 8 != 0 OR\n- 'name'::regtype::oid =\n ANY(coltypes);\n- relname | attname | coltypes |\n get_columns_length \n----------+---------+----------+--------------------\n-(0 rows)\n-\n+server closed the connection\n unexpectedly\n+ This\n probably means the server terminated\n abnormally\n+ before\n or while processing the request.\n+connection to server was lost \n\n\n\n\n\n\nActually digging some more, here is the\n actual error\n\n\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \"ntstatus.h\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nSo how does one debug this ?\n\n\nAlso if I `run meson test` I don't see this error. What\n does the buildfarm do differently?\n\n\n\n\n\nFirst it does this:\n\n\n\nmeson test -C $pgsql --no-rebuild --suite setup\n\n\n\nThen it does this (jflag is for the number of jobs):\n\n\n\nmeson test -t $meson_test_timeout $jflag -C $pgsql --logbase\n checklog --print-errorlogs --no-rebuild --suite regress\n --test-args=--no-locale\n\n\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Sat, 10 Feb 2024 13:28:29 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Sat, 10 Feb 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-02-10 Sa 12:20, Dave Cramer wrote:\n>\n>\n>\n> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>>\n>> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\n>>\n>>\n>> Dave Cramer\n>> www.postgres.rocks\n>>\n>>\n>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\n>> <davecramer@postgres.rocks> wrote:\n>>\n>>>\n>>>\n>>>\n>>>\n>>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz>\n>>> wrote:\n>>>\n>>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n>>>> > Thanks, this patch works and\n>>>> > testing with meson passes.\n>>>>\n>>>> Only with the version posted at [1]? Interesting, that's the same\n>>>> contents as v8 posted upthread, minus src/tools/ because we don't need\n>>>> to care about them anymore.\n>>>>\n>>>> Andrew, what's happening on the test side? It does not seem you've\n>>>> mentioned any details about what is going wrong, or I have just missed\n>>>> them.\n>>>>\n>>>> > I'll try the buildfarm next.\n>>>>\n>>>> [1]:\n>>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n>>>\n>>>\n>>> interestingly meson test does not produce any error\n>>> The buildfarm produces the following error for me:\n>>>\n>>> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\n>>> - FROM check_columns\n>>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\n>>> - 'name'::regtype::oid = ANY(coltypes);\n>>> - relname | attname | coltypes | get_columns_length\n>>> ----------+---------+----------+--------------------\n>>> -(0 rows)\n>>> -\n>>> +server closed the connection unexpectedly\n>>> + This probably means the server terminated abnormally\n>>> + before or while processing the request.\n>>> +connection to server was lost\n>>>\n>>\n>> Actually digging some more, here is the actual error\n>>\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\n>> 11204) was terminated by exception 0xC0000005\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process\n>> was running: VACUUM;\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\n>> \"ntstatus.h\" for a description of the hexadecimal value.\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any\n>> other active server processes\n>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes\n>> terminated; reinitializing\n>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was\n>> interrupted; last known up at 2024-02-09 13:31:01 -05\n>>\n>>\n>>\n>>\n>>\n> So how does one debug this ?\n>\n> Also if I `run meson test` I don't see this error. What does the buildfarm\n> do differently?\n>\n>\n> First it does this:\n>\n>\n> meson test -C $pgsql --no-rebuild --suite setup\n>\n>\n> Then it does this (jflag is for the number of jobs):\n>\n>\n> meson test -t $meson_test_timeout $jflag -C $pgsql --logbase checklog\n> --print-errorlogs --no-rebuild --suite regress --test-args=--no-locale\n>\n>\n\nrunning the above manually produces no errors ??\n\nDave\n\nOn Sat, 10 Feb 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-02-10 Sa 12:20, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\n\n\n\nOn Sat, 10 Feb 2024 at\n 11:19, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-02-09 Fr 14:23, Dave Cramer wrote:\n\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Fri, 9 Feb\n 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nOn Fri,\n 9 Feb 2024 at 00:26, Michael Paquier\n <michael@paquier.xyz>\n wrote:\n\nOn\n Tue, Feb 06, 2024 at 07:01:49AM -0500,\n Dave Cramer wrote:\n > Thanks, this patch works and\n > testing with meson passes.\n\n Only with the version posted at [1]? \n Interesting, that's the same\n contents as v8 posted upthread, minus\n src/tools/ because we don't need\n to care about them anymore.\n\n Andrew, what's happening on the test\n side? It does not seem you've\n mentioned any details about what is\n going wrong, or I have just missed\n them.\n\n > I'll try the buildfarm next.\n\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n\n\ninterestingly meson test does not\n produce any error\nThe buildfarm produces the following\n error for me:\n\n\n-SELECT relname, attname, coltypes,\n get_columns_length(coltypes)\n- FROM check_columns\n- WHERE get_columns_length(coltypes)\n % 8 != 0 OR\n- 'name'::regtype::oid =\n ANY(coltypes);\n- relname | attname | coltypes |\n get_columns_length \n----------+---------+----------+--------------------\n-(0 rows)\n-\n+server closed the connection\n unexpectedly\n+ This\n probably means the server terminated\n abnormally\n+ before\n or while processing the request.\n+connection to server was lost \n\n\n\n\n\n\nActually digging some more, here is the\n actual error\n\n\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \"ntstatus.h\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nSo how does one debug this ?\n\n\nAlso if I `run meson test` I don't see this error. What\n does the buildfarm do differently?\n\n\n\n\n\nFirst it does this:\n\n\n\nmeson test -C $pgsql --no-rebuild --suite setup\n\n\n\nThen it does this (jflag is for the number of jobs):\n\n\n\nmeson test -t $meson_test_timeout $jflag -C $pgsql --logbase\n checklog --print-errorlogs --no-rebuild --suite regress\n --test-args=--no-localerunning the above manually produces no errors ??Dave","msg_date":"Mon, 12 Feb 2024 08:51:45 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-02-12 Mo 08:51, Dave Cramer wrote:\n>\n>\n> On Sat, 10 Feb 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>\n> On 2024-02-10 Sa 12:20, Dave Cramer wrote:\n>>\n>>\n>> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan\n>> <andrew@dunslane.net> wrote:\n>>\n>>\n>> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\n>>>\n>>> Dave Cramer\n>>> www.postgres.rocks <http://www.postgres.rocks>\n>>>\n>>>\n>>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer\n>>> <davecramer@postgres.rocks>\n>>> <mailto:davecramer@postgres.rocks> wrote:\n>>>\n>>>\n>>>\n>>>\n>>>\n>>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier\n>>> <michael@paquier.xyz> wrote:\n>>>\n>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave\n>>> Cramer wrote:\n>>> > Thanks, this patch works and\n>>> > testing with meson passes.\n>>>\n>>> Only with the version posted at [1]? Interesting,\n>>> that's the same\n>>> contents as v8 posted upthread, minus src/tools/\n>>> because we don't need\n>>> to care about them anymore.\n>>>\n>>> Andrew, what's happening on the test side? It does\n>>> not seem you've\n>>> mentioned any details about what is going wrong, or\n>>> I have just missed\n>>> them.\n>>>\n>>> > I'll try the buildfarm next.\n>>>\n>>> [1]:\n>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n>>>\n>>>\n>>> interestingly meson test does not produce any error\n>>> The buildfarm produces the following error for me:\n>>>\n>>> -SELECT relname, attname, coltypes,\n>>> get_columns_length(coltypes)\n>>> - FROM check_columns\n>>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\n>>> - 'name'::regtype::oid = ANY(coltypes);\n>>> - relname | attname | coltypes | get_columns_length\n>>> ----------+---------+----------+--------------------\n>>> -(0 rows)\n>>> -\n>>> +server closed the connection unexpectedly\n>>> +This probably means the server terminated abnormally\n>>> +before or while processing the request.\n>>> +connection to server was lost\n>>>\n>>>\n>>> Actually digging some more, here is the actual error\n>>>\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server\n>>> process (PID 11204) was terminated by exception 0xC0000005\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL:\n>>> Failed process was running: VACUUM;\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C\n>>> include file \"ntstatus.h\" for a description of the\n>>> hexadecimal value.\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG:\n>>> terminating any other active server processes\n>>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all\n>>> server processes terminated; reinitializing\n>>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database\n>>> system was interrupted; last known up at 2024-02-09 13:31:01 -05\n>>>\n>>>\n>>>\n>>\n>>\n>> So how does one debug this ?\n>>\n>> Also if I `run meson test` I don't see this error. What does the\n>> buildfarm do differently?\n>\n>\n> First it does this:\n>\n>\n> meson test -C $pgsql --no-rebuild --suite setup\n>\n>\n> Then it does this (jflag is for the number of jobs):\n>\n>\n> meson test -t $meson_test_timeout $jflag -C $pgsql --logbase\n> checklog --print-errorlogs --no-rebuild --suite regress\n> --test-args=--no-locale\n>\n>\n>\n> running the above manually produces no errors ??\n>\n>\n\nNot for me. I get the error I previously reported, It's an access \nviolation error.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-02-12 Mo 08:51, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\nOn Sat, 10 Feb 2024 at\n 13:28, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-02-10 Sa 12:20, Dave Cramer wrote:\n\n\n\n\n\n\n\n\n\n\n\nOn Sat, 10 Feb\n 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-02-09 Fr 14:23, Dave Cramer wrote:\n\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn\n Fri, 9 Feb 2024 at 07:18, Dave Cramer\n <davecramer@postgres.rocks>\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nOn Fri, 9\n Feb 2024 at 00:26, Michael\n Paquier <michael@paquier.xyz>\n wrote:\n\nOn\n Tue, Feb 06, 2024 at\n 07:01:49AM -0500, Dave Cramer\n wrote:\n > Thanks, this patch works\n and\n > testing with meson\n passes.\n\n Only with the version posted\n at [1]? Interesting, that's\n the same\n contents as v8 posted\n upthread, minus src/tools/\n because we don't need\n to care about them anymore.\n\n Andrew, what's happening on\n the test side? It does not\n seem you've\n mentioned any details about\n what is going wrong, or I have\n just missed\n them.\n\n > I'll try the buildfarm\n next.\n\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n\n\ninterestingly meson test\n does not produce any error\nThe buildfarm produces the\n following error for me:\n\n\n-SELECT relname, attname,\n coltypes,\n get_columns_length(coltypes)\n- FROM check_columns\n- WHERE\n get_columns_length(coltypes) %\n 8 != 0 OR\n- \n 'name'::regtype::oid =\n ANY(coltypes);\n- relname | attname |\n coltypes | get_columns_length \n----------+---------+----------+--------------------\n-(0 rows)\n-\n+server closed the\n connection unexpectedly\n+ This\n probably means the server\n terminated abnormally\n+ before\n or while processing the\n request.\n+connection to server was\n lost \n\n\n\n\n\n\nActually digging some more, here is\n the actual error\n\n\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \"ntstatus.h\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nSo how does one debug this ?\n\n\nAlso if I `run meson test` I don't see this\n error. What does the buildfarm do differently?\n\n\n\n\n\nFirst it does this:\n\n\n\nmeson test -C $pgsql --no-rebuild --suite setup\n\n\n\nThen it does this (jflag is for the number of jobs):\n\n\n\nmeson test -t $meson_test_timeout $jflag -C $pgsql\n --logbase checklog --print-errorlogs --no-rebuild\n --suite regress --test-args=--no-locale\n\n\n\n\n\n\n\n\nrunning the above manually produces no errors ??\n\n\n\n\n\n\n\n\n\nNot for me. I get the error I previously reported, It's an access\n violation error.\n\n\ncheers\n\n\nandrew\n\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Mon, 12 Feb 2024 09:19:26 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Dave Cramer\nwww.postgres.rocks\n\n\nOn Mon, 12 Feb 2024 at 09:19, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n>\n> On 2024-02-12 Mo 08:51, Dave Cramer wrote:\n>\n>\n>\n> On Sat, 10 Feb 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>>\n>> On 2024-02-10 Sa 12:20, Dave Cramer wrote:\n>>\n>>\n>>\n>> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\n>>\n>>>\n>>> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\n>>>\n>>>\n>>> Dave Cramer\n>>> www.postgres.rocks\n>>>\n>>>\n>>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\n>>> <davecramer@postgres.rocks> wrote:\n>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz>\n>>>> wrote:\n>>>>\n>>>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\n>>>>> > Thanks, this patch works and\n>>>>> > testing with meson passes.\n>>>>>\n>>>>> Only with the version posted at [1]? Interesting, that's the same\n>>>>> contents as v8 posted upthread, minus src/tools/ because we don't need\n>>>>> to care about them anymore.\n>>>>>\n>>>>> Andrew, what's happening on the test side? It does not seem you've\n>>>>> mentioned any details about what is going wrong, or I have just missed\n>>>>> them.\n>>>>>\n>>>>> > I'll try the buildfarm next.\n>>>>>\n>>>>> [1]:\n>>>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n>>>>\n>>>>\n>>>> interestingly meson test does not produce any error\n>>>> The buildfarm produces the following error for me:\n>>>>\n>>>> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\n>>>> - FROM check_columns\n>>>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\n>>>> - 'name'::regtype::oid = ANY(coltypes);\n>>>> - relname | attname | coltypes | get_columns_length\n>>>> ----------+---------+----------+--------------------\n>>>> -(0 rows)\n>>>> -\n>>>> +server closed the connection unexpectedly\n>>>> + This probably means the server terminated abnormally\n>>>> + before or while processing the request.\n>>>> +connection to server was lost\n>>>>\n>>>\n>>> Actually digging some more, here is the actual error\n>>>\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\n>>> 11204) was terminated by exception 0xC0000005\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process\n>>> was running: VACUUM;\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\n>>> \"ntstatus.h\" for a description of the hexadecimal value.\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any\n>>> other active server processes\n>>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server\n>>> processes terminated; reinitializing\n>>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was\n>>> interrupted; last known up at 2024-02-09 13:31:01 -05\n>>>\n>>>\n>>>\n>>>\n>>>\n>> So how does one debug this ?\n>>\n>> Also if I `run meson test` I don't see this error. What does the\n>> buildfarm do differently?\n>>\n>>\n>> First it does this:\n>>\n>>\n>> meson test -C $pgsql --no-rebuild --suite setup\n>>\n>>\n>> Then it does this (jflag is for the number of jobs):\n>>\n>>\n>> meson test -t $meson_test_timeout $jflag -C $pgsql --logbase checklog\n>> --print-errorlogs --no-rebuild --suite regress --test-args=--no-locale\n>>\n>>\n>\n> running the above manually produces no errors ??\n>\n>\n>\n> Not for me. I get the error I previously reported, It's an access\n> violation error.\n>\n>\n> cheers\n>\n>\n> andrew\n>\n\nOK, so I have managed to get a debugger attached to postgres.exe when it\nfaults and the fault occurs at\nhttps://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\nspan is pointing to 0x0\n\nDave\n\nDave Cramerwww.postgres.rocksOn Mon, 12 Feb 2024 at 09:19, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n\n\nOn 2024-02-12 Mo 08:51, Dave Cramer\n wrote:\n\n\n\n\n\n\n\nOn Sat, 10 Feb 2024 at\n 13:28, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-02-10 Sa 12:20, Dave Cramer wrote:\n\n\n\n\n\n\n\n\n\n\n\nOn Sat, 10 Feb\n 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-02-09 Fr 14:23, Dave Cramer wrote:\n\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn\n Fri, 9 Feb 2024 at 07:18, Dave Cramer\n <davecramer@postgres.rocks>\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nOn Fri, 9\n Feb 2024 at 00:26, Michael\n Paquier <michael@paquier.xyz>\n wrote:\n\nOn\n Tue, Feb 06, 2024 at\n 07:01:49AM -0500, Dave Cramer\n wrote:\n > Thanks, this patch works\n and\n > testing with meson\n passes.\n\n Only with the version posted\n at [1]? Interesting, that's\n the same\n contents as v8 posted\n upthread, minus src/tools/\n because we don't need\n to care about them anymore.\n\n Andrew, what's happening on\n the test side? It does not\n seem you've\n mentioned any details about\n what is going wrong, or I have\n just missed\n them.\n\n > I'll try the buildfarm\n next.\n\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n\n\ninterestingly meson test\n does not produce any error\nThe buildfarm produces the\n following error for me:\n\n\n-SELECT relname, attname,\n coltypes,\n get_columns_length(coltypes)\n- FROM check_columns\n- WHERE\n get_columns_length(coltypes) %\n 8 != 0 OR\n- \n 'name'::regtype::oid =\n ANY(coltypes);\n- relname | attname |\n coltypes | get_columns_length \n----------+---------+----------+--------------------\n-(0 rows)\n-\n+server closed the\n connection unexpectedly\n+ This\n probably means the server\n terminated abnormally\n+ before\n or while processing the\n request.\n+connection to server was\n lost \n\n\n\n\n\n\nActually digging some more, here is\n the actual error\n\n\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \"ntstatus.h\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nSo how does one debug this ?\n\n\nAlso if I `run meson test` I don't see this\n error. What does the buildfarm do differently?\n\n\n\n\n\nFirst it does this:\n\n\n\nmeson test -C $pgsql --no-rebuild --suite setup\n\n\n\nThen it does this (jflag is for the number of jobs):\n\n\n\nmeson test -t $meson_test_timeout $jflag -C $pgsql\n --logbase checklog --print-errorlogs --no-rebuild\n --suite regress --test-args=--no-locale\n\n\n\n\n\n\n\n\nrunning the above manually produces no errors ??\n\n\n\n\n\n\n\n\n\nNot for me. I get the error I previously reported, It's an access\n violation error.\n\n\ncheers\n\n\nandrewOK, so I have managed to get a debugger attached to postgres.exe when it faults and the fault occurs athttps://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869span is pointing to 0x0Dave","msg_date":"Mon, 12 Feb 2024 11:44:23 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-02-12 Mo 11:44, Dave Cramer wrote:\n>\n> Dave Cramer\n> www.postgres.rocks\n>\n>\n> On Mon, 12 Feb 2024 at 09:19, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>\n> On 2024-02-12 Mo 08:51, Dave Cramer wrote:\n>>\n>>\n>> On Sat, 10 Feb 2024 at 13:28, Andrew Dunstan\n>> <andrew@dunslane.net> wrote:\n>>\n>>\n>> On 2024-02-10 Sa 12:20, Dave Cramer wrote:\n>>>\n>>>\n>>> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan\n>>> <andrew@dunslane.net> wrote:\n>>>\n>>>\n>>> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\n>>>>\n>>>> Dave Cramer\n>>>> www.postgres.rocks <http://www.postgres.rocks>\n>>>>\n>>>>\n>>>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer\n>>>> <davecramer@postgres.rocks>\n>>>> <mailto:davecramer@postgres.rocks> wrote:\n>>>>\n>>>>\n>>>>\n>>>>\n>>>>\n>>>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier\n>>>> <michael@paquier.xyz> wrote:\n>>>>\n>>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave\n>>>> Cramer wrote:\n>>>> > Thanks, this patch works and\n>>>> > testing with meson passes.\n>>>>\n>>>> Only with the version posted at [1]? \n>>>> Interesting, that's the same\n>>>> contents as v8 posted upthread, minus\n>>>> src/tools/ because we don't need\n>>>> to care about them anymore.\n>>>>\n>>>> Andrew, what's happening on the test side? It\n>>>> does not seem you've\n>>>> mentioned any details about what is going\n>>>> wrong, or I have just missed\n>>>> them.\n>>>>\n>>>> > I'll try the buildfarm next.\n>>>>\n>>>> [1]:\n>>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n>>>>\n>>>>\n>>>> interestingly meson test does not produce any error\n>>>> The buildfarm produces the following error for me:\n>>>>\n>>>> -SELECT relname, attname, coltypes,\n>>>> get_columns_length(coltypes)\n>>>> - FROM check_columns\n>>>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\n>>>> - 'name'::regtype::oid = ANY(coltypes);\n>>>> - relname | attname | coltypes | get_columns_length\n>>>> ----------+---------+----------+--------------------\n>>>> -(0 rows)\n>>>> -\n>>>> +server closed the connection unexpectedly\n>>>> +This probably means the server terminated abnormally\n>>>> +before or while processing the request.\n>>>> +connection to server was lost\n>>>>\n>>>>\n>>>> Actually digging some more, here is the actual error\n>>>>\n>>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG:\n>>>> server process (PID 11204) was terminated by exception\n>>>> 0xC0000005\n>>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL:\n>>>> Failed process was running: VACUUM;\n>>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT:\n>>>> See C include file \"ntstatus.h\" for a description of\n>>>> the hexadecimal value.\n>>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG:\n>>>> terminating any other active server processes\n>>>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all\n>>>> server processes terminated; reinitializing\n>>>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG:\n>>>> database system was interrupted; last known up at\n>>>> 2024-02-09 13:31:01 -05\n>>>>\n>>>>\n>>>>\n>>>\n>>>\n>>> So how does one debug this ?\n>>>\n>>> Also if I `run meson test` I don't see this error. What does\n>>> the buildfarm do differently?\n>>\n>>\n>> First it does this:\n>>\n>>\n>> meson test -C $pgsql --no-rebuild --suite setup\n>>\n>>\n>> Then it does this (jflag is for the number of jobs):\n>>\n>>\n>> meson test -t $meson_test_timeout $jflag -C $pgsql\n>> --logbase checklog --print-errorlogs --no-rebuild --suite\n>> regress --test-args=--no-locale\n>>\n>>\n>>\n>> running the above manually produces no errors ??\n>>\n>>\n>\n> Not for me. I get the error I previously reported, It's an access\n> violation error.\n>\n>\n>\n>\n> OK, so I have managed to get a debugger attached to postgres.exe when \n> it faults and the fault occurs at\n> https://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\n> span is pointing to 0x0\n>\n>\n\nFurther data point. If I select a build type of 'debug' instead of \n'debugoptimized' the error disappears.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-02-12 Mo 11:44, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\nDave Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Mon, 12 Feb 2024 at\n 09:19, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-02-12 Mo 08:51, Dave Cramer wrote:\n\n\n\n\n\n\n\nOn Sat, 10 Feb\n 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-02-10 Sa 12:20, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\n\n\n\nOn\n Sat, 10 Feb 2024 at 11:19, Andrew\n Dunstan <andrew@dunslane.net>\n wrote:\n\n\n\n\n\nOn 2024-02-09 Fr 14:23, Dave\n Cramer wrote:\n\n\n\n\n\n\n\nDave\n Cramer\n www.postgres.rocks\n\n\n\n\n\n\n\nOn Fri,\n 9 Feb 2024 at 07:18, Dave\n Cramer <davecramer@postgres.rocks>\n wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nOn\n Fri, 9 Feb 2024 at\n 00:26, Michael\n Paquier <michael@paquier.xyz>\n wrote:\n\nOn\n Tue, Feb 06, 2024\n at 07:01:49AM\n -0500, Dave Cramer\n wrote:\n > Thanks, this\n patch works and\n > testing with\n meson passes.\n\n Only with the\n version posted at\n [1]? Interesting,\n that's the same\n contents as v8\n posted upthread,\n minus src/tools/\n because we don't\n need\n to care about them\n anymore.\n\n Andrew, what's\n happening on the\n test side? It\n does not seem\n you've\n mentioned any\n details about what\n is going wrong, or\n I have just missed\n them.\n\n > I'll try the\n buildfarm next.\n\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\n\n\ninterestingly\n meson test does\n not produce any\n error\nThe buildfarm\n produces the\n following error\n for me:\n\n\n-SELECT\n relname, attname,\n coltypes,\n get_columns_length(coltypes)\n- FROM\n check_columns\n- WHERE\n get_columns_length(coltypes)\n % 8 != 0 OR\n- \n 'name'::regtype::oid\n = ANY(coltypes);\n- relname |\n attname | coltypes\n |\n get_columns_length \n----------+---------+----------+--------------------\n-(0 rows)\n-\n+server closed\n the connection\n unexpectedly\n+ This\n probably means the\n server terminated\n abnormally\n+ before\n or while\n processing the\n request.\n+connection to\n server was lost \n\n\n\n\n\n\nActually digging some\n more, here is the actual\n error\n\n\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \"ntstatus.h\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nSo how does one debug this ?\n\n\nAlso if I `run meson test` I don't\n see this error. What does the\n buildfarm do differently?\n\n\n\n\n\nFirst it does this:\n\n\n\nmeson test -C $pgsql --no-rebuild --suite\n setup\n\n\n\nThen it does this (jflag is for the number\n of jobs):\n\n\n\nmeson test -t $meson_test_timeout $jflag\n -C $pgsql --logbase checklog\n --print-errorlogs --no-rebuild --suite\n regress --test-args=--no-locale\n\n\n\n\n\n\n\n\nrunning the above manually produces no errors\n ??\n\n\n\n\n\n\n\n\n\nNot for me. I get the error I previously reported,\n It's an access violation error.\n\n\n\n\n\n\n\nOK, so I have managed to get a debugger attached to\n postgres.exe when it faults and the fault occurs at\nhttps://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\nspan is pointing to 0x0\n\n\n\n\n\n\n\n\n\n\nFurther data point. If I select a build type of 'debug' instead\n of 'debugoptimized' the error disappears.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Mon, 12 Feb 2024 13:28:40 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\n> On 2024-02-12 Mo 11:44, Dave Cramer wrote:\n> > OK, so I have managed to get a debugger attached to postgres.exe when it\n> > faults and the fault occurs at\n> > https://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\n> > span is pointing to 0x0\n> \n> Further data point. If I select a build type of 'debug' instead of\n> 'debugoptimized' the error disappears.\n\nOh, this is quite interesting. Dave, could you post the backtrace?\n\nI wonder if this indicates that we are either missing memory barriers\nsomewhere or that the memory barriers we end up with on msvc + arm aren't\ncorrect? Either could explain why the problem doesn't occur when building\nwith optimizations.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 12 Feb 2024 12:50:12 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2024-02-09 15:32:10 -0500, Dave Cramer wrote:\n> On Fri, 9 Feb 2024 at 14:36, Andres Freund <andres@anarazel.de> wrote:\n> > That's something like a segfault.\n> >\n> > One suspicion I have is that src/port/pg_crc32c_armv8_choose.c possibly\n> > doesn't properly support msvc. It seems to assume that SIGILL can be\n> > trapped,\n> > but that IIRC doesn't work on windows.\n> >\n> > I'd check if the problem persists if you change\n> > cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\n> > to\n> > cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 0)\n> >\n> \n> This results in\n> \n> FAILED: src/bin/pg_checksums/pg_checksums.exe\n> src/bin/pg_checksums/pg_checksums.pdb\n> \"link\" /MACHINE:ARM64 /OUT:src/bin/pg_checksums/pg_checksums.exe\n> src/bin/pg_checksums/pg_checksums.exe.p/win32ver.res\n> src/bin/pg_checksums/pg_checksums.exe.p/pg_checksums.c.obj \"/release\"\n> \"/nologo\" \"/DEBUG\" \"/PDB:src\\bin\\pg_checksums\\pg_checksums.pdb\"\n> \"/INCREMENTAL:NO\" \"/STACK:4194304\" \"/NOEXP\" \"src/fe_utils/libpgfeutils.a\"\n> \"src/common/libpgcommon.a\" \"src/port/libpgport.a\" \"ws2_32.lib\" \"ws2_32.lib\"\n> \"ws2_32.lib\" \"ws2_32.lib\" \"/SUBSYSTEM:CONSOLE\" \"kernel32.lib\" \"user32.lib\"\n> \"gdi32.lib\" \"winspool.lib\" \"shell32.lib\" \"ole32.lib\" \"oleaut32.lib\"\n> \"uuid.lib\" \"comdlg32.lib\" \"advapi32.lib\"\n> libpgcommon.a(controldata_utils.c.obj) : error LNK2001: unresolved external\n> symbol pg_comp_crc32c\n\nThat's because have_optimized_crc ends up being set.\n\n\nI'm somewhat unhappy about the msvc specific path in meson.build. Shouldn't at\nthe very least the \"Use ARM CRC Extension unconditionally\" path be first. I\nguess it's not really this patch's fault that the runtime detection for armv8\ncrc is implemented this shoddily :(.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 12 Feb 2024 12:53:52 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Mon, 12 Feb 2024 at 15:50, Andres Freund <andres@anarazel.de> wrote:\n\n> Hi,\n>\n> On 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\n> > On 2024-02-12 Mo 11:44, Dave Cramer wrote:\n> > > OK, so I have managed to get a debugger attached to postgres.exe when\n> it\n> > > faults and the fault occurs at\n> > >\n> https://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\n> > > span is pointing to 0x0\n> >\n> > Further data point. If I select a build type of 'debug' instead of\n> > 'debugoptimized' the error disappears.\n>\n> Oh, this is quite interesting. Dave, could you post the backtrace?\n>\n\nAndres,\n\nI am using Visual Studio as the debugger. Here is what I have.\n\n> postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C\n postgres.exe!resize(dshash_table * hash_table, unsigned __int64\nnew_size_log2) Line 879 C\n postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const void\n* key, bool * found) Line 480 C\n postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int dboid,\nunsigned int objoid, bool create, bool * created_entry) Line 455 C\n postgres.exe!pgstat_prep_pending_entry(PgStat_Kind kind, unsigned int\ndboid, unsigned int objoid, bool * created_entry) Line 1123 C\n [Inline Frame] postgres.exe!pgstat_prep_relation_pending(unsigned int)\nLine 904 C\n postgres.exe!pgstat_assoc_relation(RelationData * rel) Line 139 C\n [Inline Frame] postgres.exe!ReadBufferExtended(RelationData *) Line 802 C\n postgres.exe!ReadBuffer(RelationData * reln, unsigned int blockNum) Line\n737 C\n postgres.exe!read_seq_tuple(RelationData * rel, int * buf, HeapTupleData\n* seqdatatuple) Line 1196 C\n postgres.exe!AlterSequence(ParseState * pstate, AlterSeqStmt * stmt) Line\n481 C\n postgres.exe!ProcessUtilitySlow(ParseState * pstate, PlannedStmt * pstmt,\nconst char * queryString, ProcessUtilityContext context, ParamListInfoData\n* params, QueryEnvironment * queryEnv, _DestReceiver * dest,\nQueryCompletion * qc) Line 1679 C\n postgres.exe!standard_ProcessUtility(PlannedStmt * pstmt, const char *\nqueryString, bool readOnlyTree, ProcessUtilityContext context,\nParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver *\ndest, QueryCompletion * qc) Line 1080 C\n postgres.exe!ProcessUtility(PlannedStmt * pstmt, const char *\nqueryString, bool readOnlyTree, ProcessUtilityContext context,\nParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver *\ndest, QueryCompletion * qc) Line 530 C\n postgres.exe!ProcessUtilitySlow(ParseState * pstate, PlannedStmt * pstmt,\nconst char * queryString, ProcessUtilityContext context, ParamListInfoData\n* params, QueryEnvironment * queryEnv, _DestReceiver * dest,\nQueryCompletion * qc) Line 1263 C\n postgres.exe!standard_ProcessUtility(PlannedStmt * pstmt, const char *\nqueryString, bool readOnlyTree, ProcessUtilityContext context,\nParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver *\ndest, QueryCompletion * qc) Line 1080 C\n\n\nif there is a better debugger to use, please let me know\n\nDave\n\n\n\n>\n> I wonder if this indicates that we are either missing memory barriers\n> somewhere or that the memory barriers we end up with on msvc + arm aren't\n> correct? Either could explain why the problem doesn't occur when building\n> with optimizations.\n>\n> Greetings,\n>\n> Andres Freund\n>\n\nOn Mon, 12 Feb 2024 at 15:50, Andres Freund <andres@anarazel.de> wrote:Hi,\n\nOn 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\n> On 2024-02-12 Mo 11:44, Dave Cramer wrote:\n> > OK, so I have managed to get a debugger attached to postgres.exe when it\n> > faults and the fault occurs at\n> > https://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\n> > span is pointing to 0x0\n> \n> Further data point. If I select a build type of 'debug' instead of\n> 'debugoptimized' the error disappears.\n\nOh, this is quite interesting. Dave, could you post the backtrace?Andres,I am using Visual Studio as the debugger. Here is what I have.> postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C postgres.exe!resize(dshash_table * hash_table, unsigned __int64 new_size_log2) Line 879 C postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const void * key, bool * found) Line 480 C postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int dboid, unsigned int objoid, bool create, bool * created_entry) Line 455 C postgres.exe!pgstat_prep_pending_entry(PgStat_Kind kind, unsigned int dboid, unsigned int objoid, bool * created_entry) Line 1123 C [Inline Frame] postgres.exe!pgstat_prep_relation_pending(unsigned int) Line 904 C postgres.exe!pgstat_assoc_relation(RelationData * rel) Line 139 C [Inline Frame] postgres.exe!ReadBufferExtended(RelationData *) Line 802 C postgres.exe!ReadBuffer(RelationData * reln, unsigned int blockNum) Line 737 C postgres.exe!read_seq_tuple(RelationData * rel, int * buf, HeapTupleData * seqdatatuple) Line 1196 C postgres.exe!AlterSequence(ParseState * pstate, AlterSeqStmt * stmt) Line 481 C postgres.exe!ProcessUtilitySlow(ParseState * pstate, PlannedStmt * pstmt, const char * queryString, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 1679 C postgres.exe!standard_ProcessUtility(PlannedStmt * pstmt, const char * queryString, bool readOnlyTree, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 1080 C postgres.exe!ProcessUtility(PlannedStmt * pstmt, const char * queryString, bool readOnlyTree, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 530 C postgres.exe!ProcessUtilitySlow(ParseState * pstate, PlannedStmt * pstmt, const char * queryString, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 1263 C postgres.exe!standard_ProcessUtility(PlannedStmt * pstmt, const char * queryString, bool readOnlyTree, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 1080 Cif there is a better debugger to use, please let me knowDave \n\nI wonder if this indicates that we are either missing memory barriers\nsomewhere or that the memory barriers we end up with on msvc + arm aren't\ncorrect? Either could explain why the problem doesn't occur when building\nwith optimizations.\n\nGreetings,\n\nAndres Freund","msg_date":"Mon, 12 Feb 2024 16:01:01 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2024-02-12 12:50:12 -0800, Andres Freund wrote:\n> On 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\n> I wonder if this indicates that we are either missing memory barriers\n> somewhere or that the memory barriers we end up with on msvc + arm aren't\n> correct? Either could explain why the problem doesn't occur when building\n> with optimizations.\n\nI think I might have been on to something - if my human emulation of a\npreprocessor isn't wrong, we'd end up with\n\n#define S_UNLOCK(lock)\t\\\n\tdo { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n\non msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\nlimits *compiler* level reordering, not CPU level reordering. I think it's\neven insufficient on x86[-64], but it's definitely insufficient on arm.\n\n\n\nAnother issue I see is that we have a bunch of code that's dependant on\n__aarch64__ being set - but it doesn't look to me that it is set on msvc. One\nobvious consequence of that is that 64bit atomics won't be used (see\nsrc/include/port/atomics/arch-arm.h) - but that shouldn't cause this issue.\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 12 Feb 2024 13:19:15 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Sat, Feb 10, 2024 at 8:36 AM Andres Freund <andres@anarazel.de> wrote:\n> Also, yikes, that's an ugly way of doing hardware detection. Jumping out of a\n> signal handler into normal code. Brrr.\n\nMaybe it's a little baroque but what's actually wrong with it?\nOpenSSL does something similar during initialisation as a fallback\n(see armcap.c), and I borrowed that particular idea because I didn't\nwant the rat's nest of unportable strategies required to detect the\nfeature otherwise. As people who ran OpenSSL-linked stuff under gdb\non 32 bit ARM discover (on e.g. early 32 bit RPis and early iOS etc\nsystems the instruction was missing, and the signal dropped you into\nthe debugger, so there are lots of reports of that if you google this\ntopic). FWIW psql also long-jumps out of signal handlers, arguably\nmore questionably because it's an async signal and the program counter\nmay be inside readline...\n\nAs for Windows, let's see... SIGILL doesn't even work (the identifier\nis defined for compatibility, but illegal instruction exception is not\nconnected up to it), not to mention that in the backend we redirect\nsigaction() to our own fake co-operative signal system so this\ntechnique wouldn't work even if SIGILL did. But I would be surprised\nif anyone actually has a platform that can run Windows that doesn't\nhave this instruction (if they did I think we'd see a crash with\nuntrapped illegal instruction exception 0xC000001D, let me put that\nhere to help google to find this message if it ever happens...).\nCRC32C is required by ARMv8.1[1], and was optional but AFAIK present\nin general purpose ARMv8.0 cores including lowly RPis. In other words\nanything < 8 years old[2] has it certainly, and older stuff probably\ndoes too if it is a 64 bit server/PC/laptop/RPi? If you look for\nreports of systems without it they seem to be pre-ARMv8, or early\nphones/tablets (?)?\n\nIt looks like Windows 11 requires ARMv8.1[3] (\"the Windows kernel\ndropped ARMv8.0 support shortly after builds 22621 [= 2022], now\nrequiring and enforcing a minimum of ARMv8.1 and thus the new atomics\ninstructions\", that atomic/lock stuff being a pretty solid reason to\nwant to do so as discussed on this list before...). What if we just\nassumed that the same must be effectively true in practice for Windows\n10 until we hear otherwise? If a reasonable non-hypothetical system\nshows up that rejects this instruction in the short time before we\ndisclaim support for 10 (whose EOL is next year?), we can decide\nwhether to just turn off the CRC32 fast path for that OS version, or\nfigure out the runtime feature-probe patch. Clearly we need to\nminimise Windows magic in this project due to lack of hackers, so\nwriting code proactively for imaginary users seems like a bad plan...\n\n[1] https://developer.arm.com/documentation/ddi0597/2023-12/Base-Instructions/CRC32C--CRC32C-\n[2] https://en.wikipedia.org/wiki/Comparison_of_ARM_processors#ARMv8-A\n[3] http://www.emulators.com/docs/abc_history_of_woa.htm\n\n\n","msg_date":"Tue, 13 Feb 2024 15:07:17 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Dave Cramer\nwww.postgres.rocks\n\n\nOn Mon, 12 Feb 2024 at 16:19, Andres Freund <andres@anarazel.de> wrote:\n\n> Hi,\n>\n> On 2024-02-12 12:50:12 -0800, Andres Freund wrote:\n> > On 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\n> > I wonder if this indicates that we are either missing memory barriers\n> > somewhere or that the memory barriers we end up with on msvc + arm aren't\n> > correct? Either could explain why the problem doesn't occur when\n> building\n> > with optimizations.\n>\n> I think I might have been on to something - if my human emulation of a\n> preprocessor isn't wrong, we'd end up with\n>\n> #define S_UNLOCK(lock) \\\n> do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n>\n> on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\n> limits *compiler* level reordering, not CPU level reordering. I think it's\n> even insufficient on x86[-64], but it's definitely insufficient on arm.\n>\nIn fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\nLearn\n<https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\n\nI did try using atomic_thread_fence as per atomic_thread_fence -\ncppreference.com\n<https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n\n\nHowever for some reason including #include <stdatomic.h>\ncauses a bunch of compiler errors.\n\nc:\\Program Files\\Microsoft Visual\nStudio\\2022\\Community\\VC\\Tools\\MSVC\\14.38.33130\\include\\vcruntime_c11_stdatomic.h(36):\nerror C2061: syntax error: identifier 'atomic_bool'\n\nDave\n\nDave Cramerwww.postgres.rocksOn Mon, 12 Feb 2024 at 16:19, Andres Freund <andres@anarazel.de> wrote:Hi,\n\nOn 2024-02-12 12:50:12 -0800, Andres Freund wrote:\n> On 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\n> I wonder if this indicates that we are either missing memory barriers\n> somewhere or that the memory barriers we end up with on msvc + arm aren't\n> correct? Either could explain why the problem doesn't occur when building\n> with optimizations.\n\nI think I might have been on to something - if my human emulation of a\npreprocessor isn't wrong, we'd end up with\n\n#define S_UNLOCK(lock) \\\n do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n\non msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\nlimits *compiler* level reordering, not CPU level reordering. I think it's\neven insufficient on x86[-64], but it's definitely insufficient on arm.In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft LearnI did try using atomic_thread_fence as per atomic_thread_fence - cppreference.comHowever for some reason including #include <stdatomic.h>causes a bunch of compiler errors.c:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\VC\\Tools\\MSVC\\14.38.33130\\include\\vcruntime_c11_stdatomic.h(36): error C2061: syntax error: identifier 'atomic_bool'Dave","msg_date":"Tue, 13 Feb 2024 12:49:33 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"Hi,\n\nOn 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\n> > I think I might have been on to something - if my human emulation of a\n> > preprocessor isn't wrong, we'd end up with\n> >\n> > #define S_UNLOCK(lock) \\\n> > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n> >\n> > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\n> > limits *compiler* level reordering, not CPU level reordering. I think it's\n> > even insufficient on x86[-64], but it's definitely insufficient on arm.\n> >\n> In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\n> Learn\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\n\nI'd just ignore that, that's just pushing towards more modern stuff that's\nmore applicable to C++ than C.\n\n\n> I did try using atomic_thread_fence as per atomic_thread_fence -\n> cppreference.com\n> <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n\nThe semantics of atomic_thread_fence are, uh, very odd. I'd just use\nMemoryBarrier().\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 13 Feb 2024 09:52:17 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:\n\n> Hi,\n>\n> On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\n> > > I think I might have been on to something - if my human emulation of a\n> > > preprocessor isn't wrong, we'd end up with\n> > >\n> > > #define S_UNLOCK(lock) \\\n> > > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n> > >\n> > > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier()\n> just\n> > > limits *compiler* level reordering, not CPU level reordering. I think\n> it's\n> > > even insufficient on x86[-64], but it's definitely insufficient on arm.\n> > >\n> > In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier |\n> Microsoft\n> > Learn\n> > <\n> https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170\n> >\n>\n> I'd just ignore that, that's just pushing towards more modern stuff that's\n> more applicable to C++ than C.\n>\n>\n> > I did try using atomic_thread_fence as per atomic_thread_fence -\n> > cppreference.com\n> > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n>\n> The semantics of atomic_thread_fence are, uh, very odd. I'd just use\n> MemoryBarrier().\n>\n> #define S_UNLOCK(lock) \\\n do { MemoryBarrier(); (*(lock)) = 0; } while (0)\n\n#endif\n\nHas no effect.\n\nI have no idea if that is what you meant that I should do ?\n\nDave\n\nOn Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:Hi,\n\nOn 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\n> > I think I might have been on to something - if my human emulation of a\n> > preprocessor isn't wrong, we'd end up with\n> >\n> > #define S_UNLOCK(lock) \\\n> > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n> >\n> > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\n> > limits *compiler* level reordering, not CPU level reordering. I think it's\n> > even insufficient on x86[-64], but it's definitely insufficient on arm.\n> >\n> In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\n> Learn\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\n\nI'd just ignore that, that's just pushing towards more modern stuff that's\nmore applicable to C++ than C.\n\n\n> I did try using atomic_thread_fence as per atomic_thread_fence -\n> cppreference.com\n> <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n\nThe semantics of atomic_thread_fence are, uh, very odd. I'd just use\nMemoryBarrier().\n#define S_UNLOCK(lock) \\ do { MemoryBarrier(); (*(lock)) = 0; } while (0)#endif Has no effect.I have no idea if that is what you meant that I should do ?Dave","msg_date":"Tue, 13 Feb 2024 16:28:49 -0500","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks> wrote:\n\n>\n>\n>\n> On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:\n>\n>> Hi,\n>>\n>> On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\n>> > > I think I might have been on to something - if my human emulation of a\n>> > > preprocessor isn't wrong, we'd end up with\n>> > >\n>> > > #define S_UNLOCK(lock) \\\n>> > > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n>> > >\n>> > > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier()\n>> just\n>> > > limits *compiler* level reordering, not CPU level reordering. I\n>> think it's\n>> > > even insufficient on x86[-64], but it's definitely insufficient on\n>> arm.\n>> > >\n>> > In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier |\n>> Microsoft\n>> > Learn\n>> > <\n>> https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170\n>> >\n>>\n>> I'd just ignore that, that's just pushing towards more modern stuff that's\n>> more applicable to C++ than C.\n>>\n>>\n>> > I did try using atomic_thread_fence as per atomic_thread_fence -\n>> > cppreference.com\n>> > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n>>\n>> The semantics of atomic_thread_fence are, uh, very odd. I'd just use\n>> MemoryBarrier().\n>>\n>> #define S_UNLOCK(lock) \\\n> do { MemoryBarrier(); (*(lock)) = 0; } while (0)\n>\n> #endif\n>\n> Has no effect.\n>\n> I have no idea if that is what you meant that I should do ?\n>\n> Dave\n>\n\n\nRevisiting this:\n\nAndrew, can you explain the difference between ninja test (which passes)\nand what the build farm does. The buildfarm crashes.\n\nDave\n\nOn Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks> wrote:On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:Hi,\n\nOn 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\n> > I think I might have been on to something - if my human emulation of a\n> > preprocessor isn't wrong, we'd end up with\n> >\n> > #define S_UNLOCK(lock) \\\n> > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n> >\n> > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\n> > limits *compiler* level reordering, not CPU level reordering. I think it's\n> > even insufficient on x86[-64], but it's definitely insufficient on arm.\n> >\n> In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\n> Learn\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\n\nI'd just ignore that, that's just pushing towards more modern stuff that's\nmore applicable to C++ than C.\n\n\n> I did try using atomic_thread_fence as per atomic_thread_fence -\n> cppreference.com\n> <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n\nThe semantics of atomic_thread_fence are, uh, very odd. I'd just use\nMemoryBarrier().\n#define S_UNLOCK(lock) \\ do { MemoryBarrier(); (*(lock)) = 0; } while (0)#endif Has no effect.I have no idea if that is what you meant that I should do ?DaveRevisiting this:Andrew, can you explain the difference between ninja test (which passes) and what the build farm does. The buildfarm crashes.Dave","msg_date":"Tue, 24 Sep 2024 14:31:10 -0400","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, 24 Sept 2024 at 14:31, Dave Cramer <davecramer@postgres.rocks>\nwrote:\n\n>\n>\n> On Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks>\n> wrote:\n>\n>>\n>>\n>>\n>> On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:\n>>\n>>> Hi,\n>>>\n>>> On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\n>>> > > I think I might have been on to something - if my human emulation of\n>>> a\n>>> > > preprocessor isn't wrong, we'd end up with\n>>> > >\n>>> > > #define S_UNLOCK(lock) \\\n>>> > > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n>>> > >\n>>> > > on msvc + arm. And that's entirely insufficient -\n>>> _ReadWriteBarrier() just\n>>> > > limits *compiler* level reordering, not CPU level reordering. I\n>>> think it's\n>>> > > even insufficient on x86[-64], but it's definitely insufficient on\n>>> arm.\n>>> > >\n>>> > In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier |\n>>> Microsoft\n>>> > Learn\n>>> > <\n>>> https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170\n>>> >\n>>>\n>>> I'd just ignore that, that's just pushing towards more modern stuff\n>>> that's\n>>> more applicable to C++ than C.\n>>>\n>>>\n>>> > I did try using atomic_thread_fence as per atomic_thread_fence -\n>>> > cppreference.com\n>>> > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n>>>\n>>> The semantics of atomic_thread_fence are, uh, very odd. I'd just use\n>>> MemoryBarrier().\n>>>\n>>> #define S_UNLOCK(lock) \\\n>> do { MemoryBarrier(); (*(lock)) = 0; } while (0)\n>>\n>> #endif\n>>\n>> Has no effect.\n>>\n>> I have no idea if that is what you meant that I should do ?\n>>\n>> Dave\n>>\n>\n>\n> Revisiting this:\n>\n> Andrew, can you explain the difference between ninja test (which passes)\n> and what the build farm does. The buildfarm crashes.\n>\n\nI have a dmp file with a stack trace if anyone is interested\n\nDave\n\n>\n> Dave\n>\n\nOn Tue, 24 Sept 2024 at 14:31, Dave Cramer <davecramer@postgres.rocks> wrote:On Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks> wrote:On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:Hi,\n\nOn 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\n> > I think I might have been on to something - if my human emulation of a\n> > preprocessor isn't wrong, we'd end up with\n> >\n> > #define S_UNLOCK(lock) \\\n> > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n> >\n> > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\n> > limits *compiler* level reordering, not CPU level reordering. I think it's\n> > even insufficient on x86[-64], but it's definitely insufficient on arm.\n> >\n> In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\n> Learn\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\n\nI'd just ignore that, that's just pushing towards more modern stuff that's\nmore applicable to C++ than C.\n\n\n> I did try using atomic_thread_fence as per atomic_thread_fence -\n> cppreference.com\n> <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n\nThe semantics of atomic_thread_fence are, uh, very odd. I'd just use\nMemoryBarrier().\n#define S_UNLOCK(lock) \\ do { MemoryBarrier(); (*(lock)) = 0; } while (0)#endif Has no effect.I have no idea if that is what you meant that I should do ?DaveRevisiting this:Andrew, can you explain the difference between ninja test (which passes) and what the build farm does. The buildfarm crashes.I have a dmp file with a stack trace if anyone is interestedDave Dave","msg_date":"Tue, 24 Sep 2024 16:01:30 -0400","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-09-24 16:01:30 -0400, Dave Cramer wrote:\n> I have a dmp file with a stack trace if anyone is interested\n\n/me raises his hand\n\n\n","msg_date":"Tue, 24 Sep 2024 16:30:55 -0400","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On 2024-09-24 Tu 2:31 PM, Dave Cramer wrote:\n>\n>\n> On Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks> \n> wrote:\n>\n>\n>\n>\n> On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de>\n> wrote:\n>\n> Hi,\n>\n> On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\n> > > I think I might have been on to something - if my human\n> emulation of a\n> > > preprocessor isn't wrong, we'd end up with\n> > >\n> > > #define S_UNLOCK(lock) \\\n> > > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\n> > >\n> > > on msvc + arm. And that's entirely insufficient -\n> _ReadWriteBarrier() just\n> > > limits *compiler* level reordering, not CPU level\n> reordering. I think it's\n> > > even insufficient on x86[-64], but it's definitely\n> insufficient on arm.\n> > >\n> > In fact ReadWriteBarrier has been deprecated\n> _ReadWriteBarrier | Microsoft\n> > Learn\n> >\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\n>\n> I'd just ignore that, that's just pushing towards more modern\n> stuff that's\n> more applicable to C++ than C.\n>\n>\n> > I did try using atomic_thread_fence as per atomic_thread_fence -\n> > cppreference.com <http://cppreference.com>\n> > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n>\n> The semantics of atomic_thread_fence are, uh, very odd. I'd\n> just use\n> MemoryBarrier().\n>\n> #defineS_UNLOCK(lock) \\\n> do{ MemoryBarrier(); (*(lock)) =0; } while(0)\n> #endif\n> Has no effect.\n>\n> I have no idea if that is what you meant that I should do ?\n>\n> Dave\n>\n>\n>\n> Revisiting this:\n>\n> Andrew, can you explain the difference between ninja test (which \n> passes) and what the build farm does. The buildfarm crashes.\n\n\nThe buildfarm client performs these steps:\n\n\n meson test -C $pgsql --no-rebuild --suite setup\n meson test -t $meson_test_timeout $jflag -C $pgsql --logbase checklog --print-errorlogs --no-rebuild --suite regress --test-args=--no-locale\n meson test -t $meson_test_timeout $jflag -C $pgsql --print-errorlogs --no-rebuild --logbase checkworld --no-suite setup --no-suite regress\n foreach tested locale: meson test -t $meson_test_timeout $jflag -v -C $pgsql --no-rebuild --print-errorlogs --setup running --suite regress-running --logbase regress-installcheck-$locale\n\n\n$pgsql is the build root, $jflag is setting the number of jobs\n\n\nIOW, we do test suite setup, run the core regression tests, run all the \nremaining non-install tests, then run the install tests for each locale.\n\nWe don't call ninja directly, but I don't see why that should make a \ndifference.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2024-09-24 Tu 2:31 PM, Dave Cramer\n wrote:\n\n\n\n\n\n\n\n\nOn Tue, 13 Feb 2024 at\n 16:28, Dave Cramer <davecramer@postgres.rocks> wrote:\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nOn Tue, 13 Feb 2024\n at 12:52, Andres Freund <andres@anarazel.de>\n wrote:\n\nHi,\n\n On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\n > > I think I might have been on to something\n - if my human emulation of a\n > > preprocessor isn't wrong, we'd end up with\n > >\n > > #define S_UNLOCK(lock) \\\n > > do { _ReadWriteBarrier();\n (*(lock)) = 0; } while (0)\n > >\n > > on msvc + arm. And that's entirely\n insufficient - _ReadWriteBarrier() just\n > > limits *compiler* level reordering, not\n CPU level reordering. I think it's\n > > even insufficient on x86[-64], but it's\n definitely insufficient on arm.\n > >\n > In fact ReadWriteBarrier has been deprecated\n _ReadWriteBarrier | Microsoft\n > Learn\n > <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\n\n I'd just ignore that, that's just pushing towards\n more modern stuff that's\n more applicable to C++ than C.\n\n\n > I did try using atomic_thread_fence as per\n atomic_thread_fence -\n > cppreference.com\n > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\n\n The semantics of atomic_thread_fence are, uh, very\n odd. I'd just use\n MemoryBarrier().\n\n\n\n#define S_UNLOCK(lock) \\ do { MemoryBarrier(); (*(lock)) = 0; } while (0)\n#endif\n\n \nHas no effect.\n\n\nI have no idea if that is what you meant that I\n should do ?\n\n\nDave\n\n\n\n\n\n\n\n\nRevisiting this:\n\n\nAndrew, can you explain the difference between ninja test\n (which passes) and what the build farm does. The buildfarm\n crashes.\n\n\n\n\n\nThe buildfarm client performs these steps:\n\n\n\nmeson test -C $pgsql --no-rebuild --suite setup\nmeson test -t $meson_test_timeout $jflag -C $pgsql --logbase checklog --print-errorlogs --no-rebuild --suite regress --test-args=--no-locale\nmeson test -t $meson_test_timeout $jflag -C $pgsql --print-errorlogs --no-rebuild --logbase checkworld --no-suite setup --no-suite regress \nforeach tested locale: meson test -t $meson_test_timeout $jflag -v -C $pgsql --no-rebuild --print-errorlogs --setup running --suite regress-running --logbase regress-installcheck-$locale\n\n\n\n\n$pgsql is the build root, $jflag is setting the number of jobs\n\n\nIOW, we do test suite setup, run the core regression tests, run\n all the remaining non-install tests, then run the install tests\n for each locale.\nWe don't call ninja directly, but I don't see why that should\n make a difference.\n\n\n\ncheers\n\n\nandrew\n\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Sat, 28 Sep 2024 09:01:15 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Tue, Feb 13, 2024 at 10:01 AM Dave Cramer <davecramer@postgres.rocks> wrote:\n> > postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C\n> postgres.exe!resize(dshash_table * hash_table, unsigned __int64 new_size_log2) Line 879 C\n> postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const void * key, bool * found) Line 480 C\n> postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int dboid, unsigned int objoid, bool create, bool * created_entry) Line 455 C\n\nHmm, we can see that the spinlock stuff is too weak (assumes TSO), but\nthere aren't any spinlocks near that code so I think something else is\nwrong. What does pg_read_barrier() compile to? I mention that\nbecause it's used near there in a somewhat complicated way. But I'm a\nbit confused because by reading through the code it looks like it\nshould be too strong, not too weak. I think it should fall back to\npg_memory_barrier_impl() for lack of pg_read_barrier_impl(), which\nshould produce MemoryBarrier() and I guess DMB SY (ARM instruction for\nfull system data memory barrier). So maybe some other pg_atomic_XXX\noperations are falling back to code that doesn't work.\n\nJust for experimentation, you could see what happens if you redirect\nall that stuff to C11's <stdatomic.h>. Here's a quick-and-dirty patch\nfor that, which works on my ARM Mac, but I have no idea if it'll work\non MSVC (our CI uses MSVC 2019; they only added <stdatomic.h> in MSVC\n2022, but that's the same edition that added ARM so you must have it).\nIt's not a very finely tuned patch (for example I think several calls\nshould use the _explicit() variants and be weaker, but they err on the\nside of being too strong, so the code they generate is probably not as\nfast as it could be; I lost interested when the general idea was\nrejected for now, but it might help learn something about the MSVC+ARM\nsituation).\n\nCombined with the patch that redirects spinlocks to atomics[1], there\nwould be no 'hand rolled' code relating to memory order or atomics,\njust standard modern C. There are obviously other architecture tests\nhere and there, and some of them will be wrong (some of them are wrong\nalready for MSVC on x86), and that might be fixed by project\nstandardisation[2].\n\n[1] https://www.postgresql.org/message-id/CA%2BhUKGJ%2BoA%2B62iUZ-EQb5R2cAOW3Y942ZoOtzOD%3D1sQ05iNg6Q%40mail.gmail.com\n[2] https://www.postgresql.org/message-id/CA%2BhUKG%2BOu%2BC%2BpXioo_W1gKcwRCEN1WNrLBBPSMJyxjgc%2B1JEJA%40mail.gmail.com","msg_date":"Sun, 29 Sep 2024 13:02:21 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"},{"msg_contents":"On Sat, 28 Sept 2024 at 20:03, Thomas Munro <thomas.munro@gmail.com> wrote:\n\n> On Tue, Feb 13, 2024 at 10:01 AM Dave Cramer <davecramer@postgres.rocks>\n> wrote:\n> > > postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C\n> > postgres.exe!resize(dshash_table * hash_table, unsigned __int64\n> new_size_log2) Line 879 C\n> > postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const\n> void * key, bool * found) Line 480 C\n> > postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int\n> dboid, unsigned int objoid, bool create, bool * created_entry) Line 455 C\n>\n> Hmm, we can see that the spinlock stuff is too weak (assumes TSO), but\n> there aren't any spinlocks near that code so I think something else is\n> wrong. What does pg_read_barrier() compile to? I mention that\n> because it's used near there in a somewhat complicated way. But I'm a\n> bit confused because by reading through the code it looks like it\n> should be too strong, not too weak. I think it should fall back to\n> pg_memory_barrier_impl() for lack of pg_read_barrier_impl(), which\n> should produce MemoryBarrier() and I guess DMB SY (ARM instruction for\n> full system data memory barrier). So maybe some other pg_atomic_XXX\n> operations are falling back to code that doesn't work.\n>\n> Just for experimentation, you could see what happens if you redirect\n> all that stuff to C11's <stdatomic.h>. Here's a quick-and-dirty patch\n> for that, which works on my ARM Mac, but I have no idea if it'll work\n> on MSVC (our CI uses MSVC 2019; they only added <stdatomic.h> in MSVC\n> 2022, but that's the same edition that added ARM so you must have it).\n> It's not a very finely tuned patch (for example I think several calls\n> should use the _explicit() variants and be weaker, but they err on the\n> side of being too strong, so the code they generate is probably not as\n> fast as it could be; I lost interested when the general idea was\n> rejected for now, but it might help learn something about the MSVC+ARM\n> situation).\n>\n> Combined with the patch that redirects spinlocks to atomics[1], there\n> would be no 'hand rolled' code relating to memory order or atomics,\n> just standard modern C. There are obviously other architecture tests\n> here and there, and some of them will be wrong (some of them are wrong\n> already for MSVC on x86), and that might be fixed by project\n> standardisation[2].\n>\n> [1]\n> https://www.postgresql.org/message-id/CA%2BhUKGJ%2BoA%2B62iUZ-EQb5R2cAOW3Y942ZoOtzOD%3D1sQ05iNg6Q%40mail.gmail.com\n> [2]\n> https://www.postgresql.org/message-id/CA%2BhUKG%2BOu%2BC%2BpXioo_W1gKcwRCEN1WNrLBBPSMJyxjgc%2B1JEJA%40mail.gmail.com\n\n\n\nThomas,\n\nThanks, I will be able to try this when I get back from NY. So Thurs or so.\n\nDave\n\nOn Sat, 28 Sept 2024 at 20:03, Thomas Munro <thomas.munro@gmail.com> wrote:On Tue, Feb 13, 2024 at 10:01 AM Dave Cramer <davecramer@postgres.rocks> wrote:\n> > postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C\n> postgres.exe!resize(dshash_table * hash_table, unsigned __int64 new_size_log2) Line 879 C\n> postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const void * key, bool * found) Line 480 C\n> postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int dboid, unsigned int objoid, bool create, bool * created_entry) Line 455 C\n\nHmm, we can see that the spinlock stuff is too weak (assumes TSO), but\nthere aren't any spinlocks near that code so I think something else is\nwrong. What does pg_read_barrier() compile to? I mention that\nbecause it's used near there in a somewhat complicated way. But I'm a\nbit confused because by reading through the code it looks like it\nshould be too strong, not too weak. I think it should fall back to\npg_memory_barrier_impl() for lack of pg_read_barrier_impl(), which\nshould produce MemoryBarrier() and I guess DMB SY (ARM instruction for\nfull system data memory barrier). So maybe some other pg_atomic_XXX\noperations are falling back to code that doesn't work.\n\nJust for experimentation, you could see what happens if you redirect\nall that stuff to C11's <stdatomic.h>. Here's a quick-and-dirty patch\nfor that, which works on my ARM Mac, but I have no idea if it'll work\non MSVC (our CI uses MSVC 2019; they only added <stdatomic.h> in MSVC\n2022, but that's the same edition that added ARM so you must have it).\nIt's not a very finely tuned patch (for example I think several calls\nshould use the _explicit() variants and be weaker, but they err on the\nside of being too strong, so the code they generate is probably not as\nfast as it could be; I lost interested when the general idea was\nrejected for now, but it might help learn something about the MSVC+ARM\nsituation).\n\nCombined with the patch that redirects spinlocks to atomics[1], there\nwould be no 'hand rolled' code relating to memory order or atomics,\njust standard modern C. There are obviously other architecture tests\nhere and there, and some of them will be wrong (some of them are wrong\nalready for MSVC on x86), and that might be fixed by project\nstandardisation[2].\n\n[1] https://www.postgresql.org/message-id/CA%2BhUKGJ%2BoA%2B62iUZ-EQb5R2cAOW3Y942ZoOtzOD%3D1sQ05iNg6Q%40mail.gmail.com\n[2] https://www.postgresql.org/message-id/CA%2BhUKG%2BOu%2BC%2BpXioo_W1gKcwRCEN1WNrLBBPSMJyxjgc%2B1JEJA%40mail.gmail.comThomas,Thanks, I will be able to try this when I get back from NY. So Thurs or so.Dave","msg_date":"Sun, 29 Sep 2024 06:31:40 -0400","msg_from":"Dave Cramer <davecramer@postgres.rocks>","msg_from_op":false,"msg_subject":"Re: [PATCH] Add native windows on arm64 support"}],"string":"[\n {\n \"msg_contents\": \"Hello,\\n\\nI have created a patch that adds support for building Postgres for the\\nwindows/arm64 platform using the MSVC toolchain. Following changes have\\nbeen included\\n\\n1. Extend MSVC scripts to handle ARM64 platform.\\n2. Add arm64 definition of spin_delay function.\\n3. Exclude arm_acle.h import with MSVC compiler.\\n\\nCompilation steps are consistent and similar to other windows platforms.\\n\\nThe change has been tested on windows/x86_64 and windows/arm64 and all\\nregression tests passes on both platforms.\\n\\nThanks,\\nNiyas\",\n \"msg_date\": \"Wed, 23 Feb 2022 10:09:36 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 23, 2022 at 11:09 PM Niyas Sait <niyas.sait@linaro.org> wrote:\\n> I have created a patch that adds support for building Postgres for the windows/arm64 platform using the MSVC toolchain. Following changes have been included\\n>\\n> 1. Extend MSVC scripts to handle ARM64 platform.\\n> 2. Add arm64 definition of spin_delay function.\\n> 3. Exclude arm_acle.h import with MSVC compiler.\\n>\\n> Compilation steps are consistent and similar to other windows platforms.\\n>\\n> The change has been tested on windows/x86_64 and windows/arm64 and all regression tests passes on both platforms.\\n\\n+ # arm64 linker only supports dynamic base address\\n+ my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' : 'false';\\n...\\n+ <RandomizedBaseAddress>$cfgrandbaseaddress</RandomizedBaseAddress>\\n\\nDoes that mean that you can't turn off ASLR on this arch? Does it\\ncause random backend failures due to inability to map shared memory at\\nthe expected address? Our experience with EXEC_BACKEND on various\\nUnix systems tells us that you have to turn off ASLR if you want 100%\\nsuccess rate on starting backends, but I guess it depends on the\\ndetails of how the randomisation is done.\\n\\nAny interest in providing a build farm animal, so that we could know\\nthat this works?\\n\\n\\n\",\n \"msg_date\": \"Sat, 19 Mar 2022 09:50:06 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Thanks, Thomas for reviewing the patch!\\n\\nYes, we cannot turn off ASLR for Arm64 targets with MSVC. I haven't seen\\nany issues so far on win/arm64 with this option turned off. I guess it\\nshould be okay. If we really need ASLR turned off, then I guess might have\\nto use clang.\\n\\nYes, we could look into providing a build machine. Do you have any\\nreference to what the CI system looks like now for PostgresSQL and how to\\nadd new workers etc.?\\n\\nNiyas\\n\\nOn Fri, 18 Mar 2022 at 20:50, Thomas Munro <thomas.munro@gmail.com> wrote:\\n\\n> On Wed, Feb 23, 2022 at 11:09 PM Niyas Sait <niyas.sait@linaro.org> wrote:\\n> > I have created a patch that adds support for building Postgres for the\\n> windows/arm64 platform using the MSVC toolchain. Following changes have\\n> been included\\n> >\\n> > 1. Extend MSVC scripts to handle ARM64 platform.\\n> > 2. Add arm64 definition of spin_delay function.\\n> > 3. Exclude arm_acle.h import with MSVC compiler.\\n> >\\n> > Compilation steps are consistent and similar to other windows platforms.\\n> >\\n> > The change has been tested on windows/x86_64 and windows/arm64 and all\\n> regression tests passes on both platforms.\\n>\\n> + # arm64 linker only supports dynamic base address\\n> + my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' :\\n> 'false';\\n> ...\\n> + <RandomizedBaseAddress>$cfgrandbaseaddress</RandomizedBaseAddress>\\n>\\n> Does that mean that you can't turn off ASLR on this arch? Does it\\n> cause random backend failures due to inability to map shared memory at\\n> the expected address? Our experience with EXEC_BACKEND on various\\n> Unix systems tells us that you have to turn off ASLR if you want 100%\\n> success rate on starting backends, but I guess it depends on the\\n> details of how the randomisation is done.\\n>\\n> Any interest in providing a build farm animal, so that we could know\\n> that this works?\\n>\\n\\nThanks, Thomas for reviewing the patch!Yes, we cannot turn off ASLR for Arm64 targets with MSVC. I haven't seen any issues so far on win/arm64 with this option turned off. I guess it should be okay. If we really need ASLR turned off, then I guess might have to use clang.Yes, we could look into providing a build machine. Do you have any reference to what the CI system looks like now for PostgresSQL and how to add new workers etc.?NiyasOn Fri, 18 Mar 2022 at 20:50, Thomas Munro <thomas.munro@gmail.com> wrote:On Wed, Feb 23, 2022 at 11:09 PM Niyas Sait <niyas.sait@linaro.org> wrote:\\n> I have created a patch that adds support for building Postgres for the windows/arm64 platform using the MSVC toolchain. Following changes have been included\\n>\\n> 1. Extend MSVC scripts to handle ARM64 platform.\\n> 2. Add arm64 definition of spin_delay function.\\n> 3. Exclude arm_acle.h import with MSVC compiler.\\n>\\n> Compilation steps are consistent and similar to other windows platforms.\\n>\\n> The change has been tested on windows/x86_64 and windows/arm64 and all regression tests passes on both platforms.\\n\\n+ # arm64 linker only supports dynamic base address\\n+ my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' : 'false';\\n...\\n+ <RandomizedBaseAddress>$cfgrandbaseaddress</RandomizedBaseAddress>\\n\\nDoes that mean that you can't turn off ASLR on this arch? Does it\\ncause random backend failures due to inability to map shared memory at\\nthe expected address? Our experience with EXEC_BACKEND on various\\nUnix systems tells us that you have to turn off ASLR if you want 100%\\nsuccess rate on starting backends, but I guess it depends on the\\ndetails of how the randomisation is done.\\n\\nAny interest in providing a build farm animal, so that we could know\\nthat this works?\",\n \"msg_date\": \"Tue, 22 Mar 2022 09:37:46 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nPlease don't top-post here. See\\nhttps://wiki.postgresql.org/wiki/Mailing_Lists#Email_etiquette_mechanics.\\n\\nOn Tue, Mar 22, 2022 at 09:37:46AM +0000, Niyas Sait wrote:\\n>\\n> Yes, we could look into providing a build machine. Do you have any\\n> reference to what the CI system looks like now for PostgresSQL and how to\\n> add new workers etc.?\\n\\nIt's all documented at\\nhttps://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto.\\n\\n\\n\",\n \"msg_date\": \"Tue, 22 Mar 2022 18:30:11 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 22, 2022 at 11:30 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> On Tue, Mar 22, 2022 at 09:37:46AM +0000, Niyas Sait wrote:\\n> > Yes, we could look into providing a build machine. Do you have any\\n> > reference to what the CI system looks like now for PostgresSQL and how to\\n> > add new workers etc.?\\n>\\n> It's all documented at\\n> https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto.\\n\\nIt seems likely there will be more and more Windows/ARM users, so yeah\\nhaving a machine to test that combination would be great. I wonder if\\nASLR isn't breaking for you by good luck only...\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Mar 2022 16:30:58 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-23 16:30:58 +1300, Thomas Munro wrote:\\n> On Tue, Mar 22, 2022 at 11:30 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> > On Tue, Mar 22, 2022 at 09:37:46AM +0000, Niyas Sait wrote:\\n> > > Yes, we could look into providing a build machine. Do you have any\\n> > > reference to what the CI system looks like now for PostgresSQL and how to\\n> > > add new workers etc.?\\n> >\\n> > It's all documented at\\n> > https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto.\\n> \\n> It seems likely there will be more and more Windows/ARM users, so yeah\\n> having a machine to test that combination would be great. I wonder if\\n> ASLR isn't breaking for you by good luck only...\\n\\nI think we've generally seen the ASLR issue become less prominent on\\nwindows. Whether that's because of the silent retries we added, or because\\njust about everyone moved to 64bit windows / PG, I don't know. I'd guess both,\\nwith 64bit being the larger influence.\\n\\nWonder if it's worth adding some debug logging to the retry code and stop\\ndisabling ASLR on 64bit windows... It's imo pretty crazy that we loop up to\\n100 times in internal_forkexec() around CreateProcess() &&\\npgwin32_ReserveSharedMemoryRegion() without, as far as I can see, a single\\ndebug message.\\n\\nI don't think we can infer too much about the failure rate on windows from the\\n!windows EXEC_BACKEND rates. The two internal_forkexec() implementations\\nbehaves just too differently.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Mar 2022 18:15:31 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 24, 2022 at 2:15 PM Andres Freund <andres@anarazel.de> wrote:\\n> I think we've generally seen the ASLR issue become less prominent on\\n> windows. Whether that's because of the silent retries we added, or because\\n> just about everyone moved to 64bit windows / PG, I don't know. I'd guess both,\\n> with 64bit being the larger influence.\\n>\\n> Wonder if it's worth adding some debug logging to the retry code and stop\\n> disabling ASLR on 64bit windows... It's imo pretty crazy that we loop up to\\n> 100 times in internal_forkexec() around CreateProcess() &&\\n> pgwin32_ReserveSharedMemoryRegion() without, as far as I can see, a single\\n> debug message.\\n\\nYeah. I think we should commit this patch, but decree that\\nWindows/aarch64 support is experimental only for now. That allows a\\nbuild farm animal to be set up. Then we add a bit of extra logging\\nand see how it does running our test suite over time and learn more.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 11:38:44 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 11:38:44AM +1300, Thomas Munro wrote:\\n> Yeah. I think we should commit this patch, but decree that\\n> Windows/aarch64 support is experimental only for now. That allows a\\n> build farm animal to be set up. Then we add a bit of extra logging\\n> and see how it does running our test suite over time and learn more.\\n\\nI don't have such a setup so my testing capabilities are limited.\\nDoes anybody have one? I think that we could be flexible for this\\npatch, even after feature freeze as it introduces something entirely\\nnew without impacting the existing code. The patch has been moved to\\nthe next CF for now.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 7 Apr 2022 15:36:19 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Fri, Mar 25, 2022 at 11:38:44AM +1300, Thomas Munro wrote:\\n>> Yeah. I think we should commit this patch, but decree that\\n>> Windows/aarch64 support is experimental only for now. That allows a\\n>> build farm animal to be set up. Then we add a bit of extra logging\\n>> and see how it does running our test suite over time and learn more.\\n\\n> I don't have such a setup so my testing capabilities are limited.\\n> Does anybody have one? I think that we could be flexible for this\\n> patch, even after feature freeze as it introduces something entirely\\n> new without impacting the existing code. The patch has been moved to\\n> the next CF for now.\\n\\nI dunno, the lack of any in-house capability for this makes me very\\nnervous. If it causes problems down the road, how will we debug it?\\nSo it seems like the sort of patch to put in at the beginning of a\\ndevelopment cycle, not post-feature-freeze.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Apr 2022 09:40:43 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"> On 7 Apr 2022, at 15:40, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Michael Paquier <michael@paquier.xyz> writes:\\n>> On Fri, Mar 25, 2022 at 11:38:44AM +1300, Thomas Munro wrote:\\n>>> Yeah. I think we should commit this patch, but decree that\\n>>> Windows/aarch64 support is experimental only for now. That allows a\\n>>> build farm animal to be set up. Then we add a bit of extra logging\\n>>> and see how it does running our test suite over time and learn more.\\n> \\n>> I don't have such a setup so my testing capabilities are limited.\\n>> Does anybody have one? I think that we could be flexible for this\\n>> patch, even after feature freeze as it introduces something entirely\\n>> new without impacting the existing code. The patch has been moved to\\n>> the next CF for now.\\n> \\n> I dunno, the lack of any in-house capability for this makes me very\\n> nervous. If it causes problems down the road, how will we debug it?\\n\\nIf those with an interest in such platform support cannot spare the cycles to\\nat least run a buildfarm member, then it seems a stretch for us to maintain\\nsuch support with any confidence.\\n\\n> So it seems like the sort of patch to put in at the beginning of a\\n> development cycle, not post-feature-freeze.\\n\\n+1\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 16:55:27 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-07 09:40:43 -0400, Tom Lane wrote:\\n> If it causes problems down the road, how will we debug it?\\n\\nIf what causes problems down the road? Afaics the patch doesn't change\\nanything outside of windows-on-arm, so it shouldn't cause any breakage we care\\nabout until we get a buildfarm animal.\\n\\nWe've traditionally been somewhat relaxed about adding support for new\\nplatforms, on similar notions. That said:\\n\\n> So it seems like the sort of patch to put in at the beginning of a\\n> development cycle, not post-feature-freeze.\\n\\nThere doesn't seem to be a great urgency, and there's plenty stuff going on\\nright now. I can see us merging it post branching off, and then backpatching\\nit a bit later?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 10:22:58 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2022-04-07 09:40:43 -0400, Tom Lane wrote:\\n>> If it causes problems down the road, how will we debug it?\\n\\n> If what causes problems down the road? Afaics the patch doesn't change\\n> anything outside of windows-on-arm, so it shouldn't cause any breakage we care\\n> about until we get a buildfarm animal.\\n\\nDo we have a volunteer to run a buildfarm animal? I don't see much\\npoint in thinking about this till there is one ready to go.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 07 Apr 2022 13:42:49 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-07 13:42:49 -0400, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > On 2022-04-07 09:40:43 -0400, Tom Lane wrote:\\n> >> If it causes problems down the road, how will we debug it?\\n>\\n> > If what causes problems down the road? Afaics the patch doesn't change\\n> > anything outside of windows-on-arm, so it shouldn't cause any breakage we care\\n> > about until we get a buildfarm animal.\\n>\\n> Do we have a volunteer to run a buildfarm animal? I don't see much\\n> point in thinking about this till there is one ready to go.\\n\\nOP said that they might be able to:\\nhttps://www.postgresql.org/message-id/CAFPTBD_NZb%3Dq_6uE-QV%2BS%2Bpm%3DRc9sBKrg8CQ_Y3dc27Awrm7cQ%40mail.gmail.com\\n\\nNiyas, any updates on that?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 10:53:59 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"> Niyas, any updates on that?\\n\\nSorry for the delay! Configuring the scripts took some time. I have\\nsuccessfully run the builfarm animal script using my git repository [1]\\nwhich contains the proposed patch on a Windows Arm64 machine.\\n\\nI made a request to add a new machine to build farm through [2].\\n\\nI believe we should be able to enable the build farm machine once the\\nchange has been merged.\\n\\nThanks,\\nNiyas\\n\\n[1] https://github.com/nsait-linaro/postgres.git\\n[2] https://buildfarm.postgresql.org/cgi-bin/register-form.pl\\n\\n\\nOn Thu, 7 Apr 2022 at 18:54, Andres Freund <andres@anarazel.de> wrote:\\n\\n> Hi,\\n>\\n> On 2022-04-07 13:42:49 -0400, Tom Lane wrote:\\n> > Andres Freund <andres@anarazel.de> writes:\\n> > > On 2022-04-07 09:40:43 -0400, Tom Lane wrote:\\n> > >> If it causes problems down the road, how will we debug it?\\n> >\\n> > > If what causes problems down the road? Afaics the patch doesn't change\\n> > > anything outside of windows-on-arm, so it shouldn't cause any breakage\\n> we care\\n> > > about until we get a buildfarm animal.\\n> >\\n> > Do we have a volunteer to run a buildfarm animal? I don't see much\\n> > point in thinking about this till there is one ready to go.\\n>\\n> OP said that they might be able to:\\n>\\n> https://www.postgresql.org/message-id/CAFPTBD_NZb%3Dq_6uE-QV%2BS%2Bpm%3DRc9sBKrg8CQ_Y3dc27Awrm7cQ%40mail.gmail.com\\n>\\n> Niyas, any updates on that?\\n>\\n> Greetings,\\n>\\n> Andres Freund\\n>\\n\\n > Niyas, any updates on that?Sorry for the delay! Configuring the scripts took some time. I have successfully run the builfarm animal script using my git repository [1] which contains the proposed patch on a Windows Arm64 machine.I made a request to add a new machine to build farm through [2].I believe we should be able to enable the build farm machine once the change has been merged.Thanks,Niyas[1] https://github.com/nsait-linaro/postgres.git[2] https://buildfarm.postgresql.org/cgi-bin/register-form.plOn Thu, 7 Apr 2022 at 18:54, Andres Freund <andres@anarazel.de> wrote:Hi,\\n\\nOn 2022-04-07 13:42:49 -0400, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > On 2022-04-07 09:40:43 -0400, Tom Lane wrote:\\n> >> If it causes problems down the road, how will we debug it?\\n>\\n> > If what causes problems down the road? Afaics the patch doesn't change\\n> > anything outside of windows-on-arm, so it shouldn't cause any breakage we care\\n> > about until we get a buildfarm animal.\\n>\\n> Do we have a volunteer to run a buildfarm animal? I don't see much\\n> point in thinking about this till there is one ready to go.\\n\\nOP said that they might be able to:\\nhttps://www.postgresql.org/message-id/CAFPTBD_NZb%3Dq_6uE-QV%2BS%2Bpm%3DRc9sBKrg8CQ_Y3dc27Awrm7cQ%40mail.gmail.com\\n\\nNiyas, any updates on that?\\n\\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Tue, 19 Apr 2022 15:22:30 +0100\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 19, 2022 at 03:22:30PM +0100, Niyas Sait wrote:\\n> Sorry for the delay! Configuring the scripts took some time. I have\\n> successfully run the builfarm animal script using my git repository [1]\\n> which contains the proposed patch on a Windows Arm64 machine.\\n> \\n> I made a request to add a new machine to build farm through [2].\\n\\nHave you tested with the amount of coverage provided by vcregress.pl?\\n\\nAnother thing I was wondering about is if it would be possible to have\\nthis option in Travis, but that does not seem to be the case:\\nhttps://docs.travis-ci.com/user/reference/windows/#windows-version\\n\\n> I believe we should be able to enable the build farm machine once the\\n> change has been merged.\\n\\nBetter to wait for the beginning of the development cycle at this\\nstage, based on all the replies received. That would bring us to the\\nbeginning of July.\\n\\n+ if ($solution->{platform} eq 'ARM64') {\\n+ push(@pgportfiles, 'pg_crc32c_armv8_choose.c');\\n+ push(@pgportfiles, 'pg_crc32c_armv8.c');\\n+ } else {\\n+ push(@pgportfiles, 'pg_crc32c_sse42_choose.c');\\n+ push(@pgportfiles, 'pg_crc32c_sse42.c');\\n+ }\\n\\n+++ b/src/port/pg_crc32c_armv8.c\\n+#ifndef _MSC_VER\\n #include <arm_acle.h>\\n+#endif\\n[ ... ]\\n+#ifdef _M_ARM64\\n+ /*\\n+ * arm64 way of hinting processor for spin loops optimisations\\n+ * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\\n+ */\\n+ __isb(_ARM64_BARRIER_SY);\\n+#else\\nI think that such extra optimizations had better be in a separate\\npatch, and we should focus on getting the build done first.\\n\\n+ # arm64 linker only supports dynamic base address\\n+ my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' : 'false';\\nThis issue is still lying around, and you may have been lucky. Would\\nthere be any issues to remove this change to get a basic support in?\\nAs mentioned upthread, there is a long history of Postgres with ASLR.\\n\\nThis would mean that a basic patch could be done with just the changes\\nfor gendef.pl, and the first part of the changes inMSBuildProject.pm.\\nWould that not be enough?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 20 Apr 2022 09:28:09 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"> Have you tested with the amount of coverage provided by vcregress.pl?\\n\\nI built and ran the relevant tests with the help of run_build.pl.\\n\\nI think following tests are executed - check, contribcheck, ecpgcheck,\\ninstallcheck, isolationcheck, modulescheck, and upgradecheck.\\n\\n> Another thing I was wondering about is if it would be possible to have\\n> this option in Travis, but that does not seem to be the case:\\n> https://docs.travis-ci.com/user/reference/windows/#windows-version\\n\\nYes, I think Travis doesn't yet support Windows/Arm64 platform.\\n\\n> + if ($solution->{platform} eq 'ARM64') {\\n> + push(@pgportfiles, 'pg_crc32c_armv8_choose.c');\\n> + push(@pgportfiles, 'pg_crc32c_armv8.c');\\n> + } else {\\n> + push(@pgportfiles, 'pg_crc32c_sse42_choose.c');\\n> + push(@pgportfiles, 'pg_crc32c_sse42.c');\\n> + }\\n>\\n> +++ b/src/port/pg_crc32c_armv8.c\\n> +#ifndef _MSC_VER\\n> #include <arm_acle.h>\\n> +#endif\\n> [ ... ]\\n> +#ifdef _M_ARM64\\n> + /*\\n> + * arm64 way of hinting processor for spin loops optimisations\\n> + * ref:\\nhttps://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\\n> + */\\n> + __isb(_ARM64_BARRIER_SY);\\n> +#else\\n> I think that such extra optimizations had better be in a separate\\n> patch, and we should focus on getting the build done first.\\n\\n> This would mean that a basic patch could be done with just the changes\\n> for gendef.pl, and the first part of the changes inMSBuildProject.pm.\\n> Would that not be enough?\\n\\nI cannot build without any of the above changes. Nothing specific for\\noptimization\\nis added to this patch.\\n\\n> + # arm64 linker only supports dynamic base address\\n> + my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' :\\n'false';\\n> This issue is still lying around, and you may have been lucky. Would\\n> there be any issues to remove this change to get a basic support in?\\n> As mentioned upthread, there is a long history of Postgres with ASLR.\\n\\nMSVC linker doesn't allow non-random base addresses for Arm64 platforms.\\nIt is needed for basic support.\\n\\nNiyas\\n\\n> Have you tested with the amount of coverage provided by vcregress.pl?I built and ran the relevant tests with the help of run_build.pl.I think following tests are executed - check, contribcheck, ecpgcheck, installcheck, isolationcheck, modulescheck, and upgradecheck.> Another thing I was wondering about is if it would be possible to have> this option in Travis, but that does not seem to be the case:> https://docs.travis-ci.com/user/reference/windows/#windows-versionYes, I think Travis doesn't yet support Windows/Arm64 platform.> + if ($solution->{platform} eq 'ARM64') {> + push(@pgportfiles, 'pg_crc32c_armv8_choose.c');> + push(@pgportfiles, 'pg_crc32c_armv8.c');> + } else {> + push(@pgportfiles, 'pg_crc32c_sse42_choose.c');> + push(@pgportfiles, 'pg_crc32c_sse42.c');> + }> > +++ b/src/port/pg_crc32c_armv8.c> +#ifndef _MSC_VER> #include <arm_acle.h>> +#endif> [ ... ]> +#ifdef _M_ARM64> + /*> + * arm64 way of hinting processor for spin loops optimisations> + * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq> + */> + __isb(_ARM64_BARRIER_SY);> +#else> I think that such extra optimizations had better be in a separate> patch, and we should focus on getting the build done first.> This would mean that a basic patch could be done with just the changes> for gendef.pl, and the first part of the changes inMSBuildProject.pm.> Would that not be enough?I cannot build without any of the above changes. Nothing specific for optimizationis added to this patch.> + # arm64 linker only supports dynamic base address> + my $cfgrandbaseaddress = $self->{platform} eq 'ARM64' ? 'true' : 'false';> This issue is still lying around, and you may have been lucky. Would> there be any issues to remove this change to get a basic support in?> As mentioned upthread, there is a long history of Postgres with ASLR.MSVC linker doesn't allow non-random base addresses for Arm64 platforms.It is needed for basic support.Niyas\",\n \"msg_date\": \"Wed, 20 Apr 2022 10:43:06 +0100\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 20, 2022 at 10:43:06AM +0100, Niyas Sait wrote:\\n>> This issue is still lying around, and you may have been lucky. Would\\n>> there be any issues to remove this change to get a basic support in?\\n>> As mentioned upthread, there is a long history of Postgres with ASLR.\\n> \\n> MSVC linker doesn't allow non-random base addresses for Arm64 platforms.\\n> It is needed for basic support.\\n\\nWhy does it not allow that? Is that just a limitation of the\\ncompiler? If yes, what is the error happening? This question is not\\nexactly answered yet as of this thread. I may be missing a reference\\nabout that in the upstream docs, but I see nowhere an explanation\\nabout the reason and the why. That's also one of the first questions\\nfrom Thomas upthread.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 21 Apr 2022 13:07:18 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"> Why does it not allow that? Is that just a limitation of the\\n> compiler? If yes, what is the error happening? This question is not\\n> exactly answered yet as of this thread. I may be missing a reference\\n> about that in the upstream docs, but I see nowhere an explanation\\n> about the reason and the why. That's also one of the first questions\\n> from Thomas upthread.\\n\\nThe following error occurs:\\n\\nLINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64'\\ntarget machine; link without '/DYNAMICBASE:NO\\n\\nThis seems to be a deliberate restriction for Arm64 targets. However, no\\nreferences were provided. To clarify, I have posted a question [1] on the\\ncommunity channel of Visual Studio.\\n\\nNiyas\\n\\n[1]\\nhttps://developercommunity.visualstudio.com/t/LINK-:-fatal-error-LNK1246:-DYNAMICBAS/10020163\\n\\n\\n\\n\\n\\nOn Thu, 21 Apr 2022 at 05:07, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Wed, Apr 20, 2022 at 10:43:06AM +0100, Niyas Sait wrote:\\n> >> This issue is still lying around, and you may have been lucky. Would\\n> >> there be any issues to remove this change to get a basic support in?\\n> >> As mentioned upthread, there is a long history of Postgres with ASLR.\\n> >\\n> > MSVC linker doesn't allow non-random base addresses for Arm64 platforms.\\n> > It is needed for basic support.\\n>\\n> Why does it not allow that? Is that just a limitation of the\\n> compiler? If yes, what is the error happening? This question is not\\n> exactly answered yet as of this thread. I may be missing a reference\\n> about that in the upstream docs, but I see nowhere an explanation\\n> about the reason and the why. That's also one of the first questions\\n> from Thomas upthread.\\n> --\\n> Michael\\n>\\n\\n> Why does it not allow that? Is that just a limitation of the> compiler? If yes, what is the error happening? This question is not> exactly answered yet as of this thread. I may be missing a reference> about that in the upstream docs, but I see nowhere an explanation> about the reason and the why. That's also one of the first questions> from Thomas upthread.The following error occurs:LINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64' target machine; link without '/DYNAMICBASE:NOThis seems to be a deliberate restriction for Arm64 targets. However, no references were provided. To clarify, I have posted a question [1] on the community channel of Visual Studio.Niyas[1] https://developercommunity.visualstudio.com/t/LINK-:-fatal-error-LNK1246:-DYNAMICBAS/10020163On Thu, 21 Apr 2022 at 05:07, Michael Paquier <michael@paquier.xyz> wrote:On Wed, Apr 20, 2022 at 10:43:06AM +0100, Niyas Sait wrote:\\n>> This issue is still lying around, and you may have been lucky. Would\\n>> there be any issues to remove this change to get a basic support in?\\n>> As mentioned upthread, there is a long history of Postgres with ASLR.\\n> \\n> MSVC linker doesn't allow non-random base addresses for Arm64 platforms.\\n> It is needed for basic support.\\n\\nWhy does it not allow that? Is that just a limitation of the\\ncompiler? If yes, what is the error happening? This question is not\\nexactly answered yet as of this thread. I may be missing a reference\\nabout that in the upstream docs, but I see nowhere an explanation\\nabout the reason and the why. That's also one of the first questions\\nfrom Thomas upthread.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 21 Apr 2022 10:21:04 +0100\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 21, 2022 at 10:21:04AM +0100, Niyas Sait wrote:\\n> The following error occurs:\\n> \\n> LINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64'\\n> target machine; link without '/DYNAMICBASE:NO\\n\\nOkay, that's interesting. In light of things like 7f3e17b, that may\\nbe annoying. Perhaps newer Windows versions are able to handle that\\nbetter. I am wondering if it would be worth re-evaluating this\\nchange, and attempt to re-enable that in environments other than\\narm64. This could become interesting if we consider that there have\\nbeen talks to cut the support for a bunch of Windows versions to focus\\non the newer ones only.\\n\\n> This seems to be a deliberate restriction for Arm64 targets. However, no\\n> references were provided. To clarify, I have posted a question [1] on the\\n> community channel of Visual Studio.\\n\\nThanks for doing that! Your post is just a couple of days old, let's\\nsee if we get a reply on that.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 25 Apr 2022 10:17:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"> Thanks for doing that! Your post is just a couple of days old, let's\\n> see if we get a reply on that.\\n\\nMicrosoft updated documentation [1] and clarified that ASLR cannot be\\ndisabled for Arm64 targets.\\n\\nBecause ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures,\\n> /DYNAMICBASE:NO isn't supported for these targets.\\n\\n\\nThanks\\nNiyas\\n\\n[1]\\nhttps://docs.microsoft.com/en-us/cpp/build/reference/dynamicbase?view=msvc-170\\n\\n\\nOn Mon, 25 Apr 2022 at 02:17, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Thu, Apr 21, 2022 at 10:21:04AM +0100, Niyas Sait wrote:\\n> > The following error occurs:\\n> >\\n> > LINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64'\\n> > target machine; link without '/DYNAMICBASE:NO\\n>\\n> Okay, that's interesting. In light of things like 7f3e17b, that may\\n> be annoying. Perhaps newer Windows versions are able to handle that\\n> better. I am wondering if it would be worth re-evaluating this\\n> change, and attempt to re-enable that in environments other than\\n> arm64. This could become interesting if we consider that there have\\n> been talks to cut the support for a bunch of Windows versions to focus\\n> on the newer ones only.\\n>\\n> > This seems to be a deliberate restriction for Arm64 targets. However, no\\n> > references were provided. To clarify, I have posted a question [1] on the\\n> > community channel of Visual Studio.\\n>\\n> Thanks for doing that! Your post is just a couple of days old, let's\\n> see if we get a reply on that.\\n> --\\n> Michael\\n>\\n\\n> Thanks for doing that! Your post is just a couple of days old, let's> see if we get a reply on that.Microsoft updated documentation [1] and clarified that ASLR cannot be disabled for Arm64 targets.Because ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures, /DYNAMICBASE:NO isn't supported for these targets.ThanksNiyas[1] https://docs.microsoft.com/en-us/cpp/build/reference/dynamicbase?view=msvc-170On Mon, 25 Apr 2022 at 02:17, Michael Paquier <michael@paquier.xyz> wrote:On Thu, Apr 21, 2022 at 10:21:04AM +0100, Niyas Sait wrote:\\n> The following error occurs:\\n> \\n> LINK : fatal error LNK1246: '/DYNAMICBASE:NO' not compatible with 'ARM64'\\n> target machine; link without '/DYNAMICBASE:NO\\n\\nOkay, that's interesting. In light of things like 7f3e17b, that may\\nbe annoying. Perhaps newer Windows versions are able to handle that\\nbetter. I am wondering if it would be worth re-evaluating this\\nchange, and attempt to re-enable that in environments other than\\narm64. This could become interesting if we consider that there have\\nbeen talks to cut the support for a bunch of Windows versions to focus\\non the newer ones only.\\n\\n> This seems to be a deliberate restriction for Arm64 targets. However, no\\n> references were provided. To clarify, I have posted a question [1] on the\\n> community channel of Visual Studio.\\n\\nThanks for doing that! Your post is just a couple of days old, let's\\nsee if we get a reply on that.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 9 May 2022 12:44:22 +0100\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Mon, May 09, 2022 at 12:44:22PM +0100, Niyas Sait wrote:\\n> Microsoft updated documentation [1] and clarified that ASLR cannot be\\n> disabled for Arm64 targets.\\n> \\n> Because ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures,\\n> > /DYNAMICBASE:NO isn't supported for these targets.\\n\\nBetter than nothing, I guess, though this does not stand as a reason\\nexplaining why this choice has been made. And it looks like we will\\nnever know. We are going to need more advanced testing to check that\\nDYNAMICBASE is stable enough if we begin to enable it. Perhaps we\\nshould really look at that for all the build targets we currently\\nsupport rather than just ARM, for the most modern Win32 environments\\nif we are going to cut support for most of the oldest releases..\\n--\\nMichael\",\n \"msg_date\": \"Tue, 10 May 2022 10:01:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hello,\\n\\nPlease find a new patch (no further changes) rebased on top of the master.\\n\\nOn Tue, 10 May 2022 at 02:02, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Mon, May 09, 2022 at 12:44:22PM +0100, Niyas Sait wrote:\\n> > Microsoft updated documentation [1] and clarified that ASLR cannot be\\n> > disabled for Arm64 targets.\\n> >\\n> > Because ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures,\\n> > > /DYNAMICBASE:NO isn't supported for these targets.\\n>\\n> Better than nothing, I guess, though this does not stand as a reason\\n> explaining why this choice has been made. And it looks like we will\\n> never know. We are going to need more advanced testing to check that\\n> DYNAMICBASE is stable enough if we begin to enable it. Perhaps we\\n> should really look at that for all the build targets we currently\\n> support rather than just ARM, for the most modern Win32 environments\\n> if we are going to cut support for most of the oldest releases..\\n> --\\n> Michael\\n>\\n\\n\\n-- \\nNiyas Sait\\nLinaro Limited\",\n \"msg_date\": \"Thu, 14 Jul 2022 09:36:13 +0100\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-05-10 10:01:55 +0900, Michael Paquier wrote:\\n> On Mon, May 09, 2022 at 12:44:22PM +0100, Niyas Sait wrote:\\n> > Microsoft updated documentation [1] and clarified that ASLR cannot be\\n> > disabled for Arm64 targets.\\n> > \\n> > Because ASLR can't be disabled on ARM, ARM64, or ARM64EC architectures,\\n> > > /DYNAMICBASE:NO isn't supported for these targets.\\n> \\n> Better than nothing, I guess, though this does not stand as a reason\\n> explaining why this choice has been made. And it looks like we will\\n> never know. We are going to need more advanced testing to check that\\n> DYNAMICBASE is stable enough if we begin to enable it. Perhaps we\\n> should really look at that for all the build targets we currently\\n> support rather than just ARM, for the most modern Win32 environments\\n> if we are going to cut support for most of the oldest releases..\\n\\nI accidentally did a lot of testing with DYNAMICBASE - I accidentally\\nmistranslated MSBuildProject.pm's <RandomizedBaseAddress>false</> to adding\\n/DYNAMICBASE in the meson port. Survived several hundred CI cycles and dozens\\nof local runs without observing any problems related to that.\\n\\nWhich doesn't surprise me, given the way we reserve space in the new process.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 25 Aug 2022 18:29:07 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 25, 2022 at 06:29:07PM -0700, Andres Freund wrote:\\n> I accidentally did a lot of testing with DYNAMICBASE - I accidentally\\n> mistranslated MSBuildProject.pm's <RandomizedBaseAddress>false</> to adding\\n> /DYNAMICBASE in the meson port. Survived several hundred CI cycles and dozens\\n> of local runs without observing any problems related to that.\\n> \\n> Which doesn't surprise me, given the way we reserve space in the new process.\\n\\nStill, we had problems with that in the past and I don't recall huge\\nchanges in the way we allocate shmem on WIN32. Could there be some\\nstuff in Windows itself that would explain more stability? I would\\nnot mind switching back to DYNAMICBASE on HEAD seeing your results.\\nThat would simplify this thread's patch a bit as well.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 26 Aug 2022 11:09:41 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-08-26 11:09:41 +0900, Michael Paquier wrote:\\n> On Thu, Aug 25, 2022 at 06:29:07PM -0700, Andres Freund wrote:\\n> > I accidentally did a lot of testing with DYNAMICBASE - I accidentally\\n> > mistranslated MSBuildProject.pm's <RandomizedBaseAddress>false</> to adding\\n> > /DYNAMICBASE in the meson port. Survived several hundred CI cycles and dozens\\n> > of local runs without observing any problems related to that.\\n> > \\n> > Which doesn't surprise me, given the way we reserve space in the new process.\\n> \\n> Still, we had problems with that in the past and I don't recall huge\\n> changes in the way we allocate shmem on WIN32.\\n\\nIt's really not great that 7f3e17b4827 disabled randomization without even a\\ncomment as to why...\\n\\nThere actually seems to have been a lot of work in that area. See 617dc6d299c\\n/ bcbf2346d69 and as well as the retrying in 45e004fb4e39. Those weren't\\nprevented by us disabling ASLR.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 25 Aug 2022 20:29:43 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 25, 2022 at 08:29:43PM -0700, Andres Freund wrote:\\n> It's really not great that 7f3e17b4827 disabled randomization without even a\\n> comment as to why...\\n\\nThis story is on this thread, with some processes not able to start:\\nhttps://www.postgresql.org/message-id/BD0D89EC2438455C9DE0DC94D36912F4@maumau\\n\\n> There actually seems to have been a lot of work in that area. See 617dc6d299c\\n> / bcbf2346d69 and as well as the retrying in 45e004fb4e39. Those weren't\\n> prevented by us disabling ASLR.\\n\\nIndeed. 45e004f looks like the most interesting bit here. FWIW, I\\nwould not mind re-enabling that on HEAD, as of something like the\\nattached. I have done a dozen of runs without seeing a test failure,\\nand knowing that we don't support anything older than Win10 makes me\\nfeel safer about this change. Any objections?\\n--\\nMichael\",\n \"msg_date\": \"Sat, 27 Aug 2022 11:33:51 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> Indeed. 45e004f looks like the most interesting bit here. FWIW, I\\n> would not mind re-enabling that on HEAD, as of something like the\\n> attached. I have done a dozen of runs without seeing a test failure,\\n> and knowing that we don't support anything older than Win10 makes me\\n> feel safer about this change. Any objections?\\n\\nWe're early enough in the v16 cycle to have plenty of time to detect\\nany problems, so I see little reason not to try it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 26 Aug 2022 23:02:58 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-08-27 11:33:51 +0900, Michael Paquier wrote:\\n> FWIW, I would not mind re-enabling that on HEAD, as of something like the\\n> attached. I have done a dozen of runs without seeing a test failure, and\\n> knowing that we don't support anything older than Win10 makes me feel safer\\n> about this change. Any objections?\\n\\n+1. We don't have a choice about it on other architectures and it seems like a\\nbad idea to disable on its own.\\n\\nI checked, and it looks like we didn't add --disable-dynamicbase, so ASLR\\nwasn't disabled for mingw builds.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sat, 27 Aug 2022 12:27:57 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Sat, Aug 27, 2022 at 12:27:57PM -0700, Andres Freund wrote:\\n> I checked, and it looks like we didn't add --disable-dynamicbase, so ASLR\\n> wasn't disabled for mingw builds.\\n\\nTrue enough. I have applied the patch to re-enable that on HEAD.\\nLet's now see what happens in the next couple of days. Popcorn is\\nready here.\\n--\\nMichael\",\n \"msg_date\": \"Sun, 28 Aug 2022 16:23:08 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> True enough. I have applied the patch to re-enable that on HEAD.\\n> Let's now see what happens in the next couple of days. Popcorn is\\n> ready here.\\n\\nHmm:\\n\\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-28%2010%3A30%3A29\\n\\nMaybe that's just unrelated noise, but it sure looks weird: after\\npassing the core regression tests in \\\"make check\\\", it failed in\\npg_upgrade's run with\\n\\ndiff -w -U3 H:/prog/bf/root/HEAD/pgsql.build/src/test/regress/expected/xml.out H:/prog/bf/root/HEAD/pgsql.build/src/bin/pg_upgrade/tmp_check/results/xml.out\\n--- H:/prog/bf/root/HEAD/pgsql.build/src/test/regress/expected/xml.out\\tWed May 18 18:30:12 2022\\n+++ H:/prog/bf/root/HEAD/pgsql.build/src/bin/pg_upgrade/tmp_check/results/xml.out\\tSun Aug 28 06:55:14 2022\\n@@ -1557,7 +1557,7 @@\\n \\\\\\\\x\\n SELECT * FROM XMLTABLE('*' PASSING '<e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>' COLUMNS x xml PATH 'node()', y xml PATH '/');\\n -[ RECORD 1 ]-----------------------------------------------------------\\n-x | pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\\n+x | pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\\n y | <e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>+\\n | \\n \\n\\nI don't recall ever seeing a failure quite like this.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 28 Aug 2022 10:09:53 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-08-28 10:09:53 -0400, Tom Lane wrote:\\n> Michael Paquier <michael@paquier.xyz> writes:\\n> > True enough. I have applied the patch to re-enable that on HEAD.\\n> > Let's now see what happens in the next couple of days. Popcorn is\\n> > ready here.\\n>\\n> Hmm:\\n>\\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-28%2010%3A30%3A29\\n>\\n> Maybe that's just unrelated noise, but it sure looks weird: after\\n> passing the core regression tests in \\\"make check\\\", it failed in\\n> pg_upgrade's run with\\n>\\n> diff -w -U3 H:/prog/bf/root/HEAD/pgsql.build/src/test/regress/expected/xml.out H:/prog/bf/root/HEAD/pgsql.build/src/bin/pg_upgrade/tmp_check/results/xml.out\\n> --- H:/prog/bf/root/HEAD/pgsql.build/src/test/regress/expected/xml.out\\tWed May 18 18:30:12 2022\\n> +++ H:/prog/bf/root/HEAD/pgsql.build/src/bin/pg_upgrade/tmp_check/results/xml.out\\tSun Aug 28 06:55:14 2022\\n> @@ -1557,7 +1557,7 @@\\n> \\\\\\\\x\\n> SELECT * FROM XMLTABLE('*' PASSING '<e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>' COLUMNS x xml PATH 'node()', y xml PATH '/');\\n> -[ RECORD 1 ]-----------------------------------------------------------\\n> -x | pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\\n> +x | pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\\n> y | <e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>+\\n> |\\n\\nPretty weird, agreed. But hard to see how it could be caused by the\\nrandomization change, except that perhaps it could highlight a preexisting\\nbug?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sun, 28 Aug 2022 08:41:54 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2022-08-28 10:09:53 -0400, Tom Lane wrote:\\n>> -x | pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\\n>> +x | pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\\n\\n> Pretty weird, agreed. But hard to see how it could be caused by the\\n> randomization change, except that perhaps it could highlight a preexisting\\n> bug?\\n\\nI have no idea either. I agree there *shouldn't* be any connection,\\nso if ASLR is somehow triggering this then whatever is failing is\\nalmost certainly buggy on its own terms. But there's a lot of\\nmoving parts here (mumble libxml mumble). I'm going to wait to see\\nif it reproduces before spending much effort.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 28 Aug 2022 11:51:19 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Sun, Aug 28, 2022 at 11:51:19AM -0400, Tom Lane wrote:\\n> I have no idea either. I agree there *shouldn't* be any connection,\\n> so if ASLR is somehow triggering this then whatever is failing is\\n> almost certainly buggy on its own terms. But there's a lot of\\n> moving parts here (mumble libxml mumble). I'm going to wait to see\\n> if it reproduces before spending much effort.\\n\\nI have noticed that yesterday, but cannot think much about it. This\\nbasically changes the position of \\\"<!--c1-->\\\" for the first record,\\nleaving the second one untouched:\\n<!--c1--><?pi arg?><![CDATA[&ent1]]>\\n<?pi arg?><![CDATA[&ent1]]><!--c1-->\\n\\nI am not used to xmltable(), but I wonder if there is something in one\\nof these support functions in xml.c that gets influenced by the\\nrandomization. That sounds a bit hairy as make check passed in\\nbowerbird, and I have noticed at least two other Windows hosts running\\nTAP that passed. Or that's just something with libxml itself.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 29 Aug 2022 10:12:20 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 29, 2022 at 10:12:20AM +0900, Michael Paquier wrote:\\n> I have noticed that yesterday, but cannot think much about it. This\\n> basically changes the position of \\\"<!--c1-->\\\" for the first record,\\n> leaving the second one untouched:\\n> <!--c1--><?pi arg?><![CDATA[&ent1]]>\\n> <?pi arg?><![CDATA[&ent1]]><!--c1-->\\n> \\n> I am not used to xmltable(), but I wonder if there is something in one\\n> of these support functions in xml.c that gets influenced by the\\n> randomization. That sounds a bit hairy as make check passed in\\n> bowerbird, and I have noticed at least two other Windows hosts running\\n> TAP that passed. Or that's just something with libxml itself.\\n\\nThis is amazing. The issue has showed up a second time in a row in\\nbowerbird, as of:\\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-29%2013%3A30%3A32\\n\\nI don't know what to think about ASLR that manipulates the comment in\\nthis XML object under VS 2017 (perhaps a compiler issue?), but it\\nshould be possible to go back to green simply by removing \\\"<!--c1-->\\\"\\nfrom the input string. Creating an extra output pattern here would be\\nvery costly, as xml_2.out and xml.out have outputs for --with-libxml.\\n\\nWould people object if I do that for now?\\n--\\nMichael\",\n \"msg_date\": \"Tue, 30 Aug 2022 08:48:38 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> This is amazing. The issue has showed up a second time in a row in\\n> bowerbird, as of:\\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-29%2013%3A30%3A32\\n\\nThat is fascinating.\\n\\n> I don't know what to think about ASLR that manipulates the comment in\\n> this XML object under VS 2017 (perhaps a compiler issue?), but it\\n> should be possible to go back to green simply by removing \\\"<!--c1-->\\\"\\n> from the input string. Creating an extra output pattern here would be\\n> very costly, as xml_2.out and xml.out have outputs for --with-libxml.\\n\\n> Would people object if I do that for now?\\n\\nLet's let it go for a few more runs. I want to know whether it\\nreproduces 100% or only sometimes.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 19:58:31 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-08-30 08:48:38 +0900, Michael Paquier wrote:\\n> On Mon, Aug 29, 2022 at 10:12:20AM +0900, Michael Paquier wrote:\\n> > I have noticed that yesterday, but cannot think much about it. This\\n> > basically changes the position of \\\"<!--c1-->\\\" for the first record,\\n> > leaving the second one untouched:\\n> > <!--c1--><?pi arg?><![CDATA[&ent1]]>\\n> > <?pi arg?><![CDATA[&ent1]]><!--c1-->\\n> > \\n> > I am not used to xmltable(), but I wonder if there is something in one\\n> > of these support functions in xml.c that gets influenced by the\\n> > randomization. That sounds a bit hairy as make check passed in\\n> > bowerbird, and I have noticed at least two other Windows hosts running\\n> > TAP that passed. Or that's just something with libxml itself.\\n> \\n> This is amazing. The issue has showed up a second time in a row in\\n> bowerbird, as of:\\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-29%2013%3A30%3A32\\n> \\n> I don't know what to think about ASLR that manipulates the comment in\\n> this XML object under VS 2017 (perhaps a compiler issue?), but it\\n> should be possible to go back to green simply by removing \\\"<!--c1-->\\\"\\n> from the input string. Creating an extra output pattern here would be\\n> very costly, as xml_2.out and xml.out have outputs for --with-libxml.\\n> \\n> Would people object if I do that for now?\\n\\nI do. Clearly something is wrong, what do we gain by averting our eyes? An\\nalternative output file strikes me as an equally bad idea - it's not like this\\nis a case where there are legitimately different outputs?\\n\\nBowerbird is the only animal testing libxml with visual studio afaics. And CI\\ndoesn't have libxml either. So we have no idea if it's the msvc version that\\nmatters.\\n\\nAndrew, what version of libxml is installed? Afaict we don't see that anywhere\\nwith msvc.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 17:09:54 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-08-29 19:58:31 -0400, Tom Lane wrote:\\n> Michael Paquier <michael@paquier.xyz> writes:\\n> > This is amazing. The issue has showed up a second time in a row in\\n> > bowerbird, as of:\\n> > https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-29%2013%3A30%3A32\\n> \\n> That is fascinating.\\n\\n> > I don't know what to think about ASLR that manipulates the comment in\\n> > this XML object under VS 2017 (perhaps a compiler issue?), but it\\n> > should be possible to go back to green simply by removing \\\"<!--c1-->\\\"\\n> > from the input string. Creating an extra output pattern here would be\\n> > very costly, as xml_2.out and xml.out have outputs for --with-libxml.\\n> \\n> > Would people object if I do that for now?\\n> \\n> Let's let it go for a few more runs. I want to know whether it\\n> reproduces 100% or only sometimes.\\n\\nThe weirdest part is that it only happens as part of the the pg_upgrade test.\\n\\nI just tested it in my test windows vm (win 10, vs 2019), with a build of\\nlibxml I had around (2.9.7), and the regression tests passed both \\\"normally\\\"\\nand within pg_upgrade.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 17:33:56 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 29, 2022 at 05:33:56PM -0700, Andres Freund wrote:\\n> The weirdest part is that it only happens as part of the the pg_upgrade test.\\n\\nmake check has just failed:\\nhttps://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-30%2001%3A15%3A13\\n\\n> I just tested it in my test windows vm (win 10, vs 2019), with a build of\\n> libxml I had around (2.9.7), and the regression tests passed both \\\"normally\\\"\\n> and within pg_upgrade.\\n\\nbowerbird is feeling from c:\\\\\\\\prog\\\\\\\\3p64\\\\\\\\include\\\\\\\\libxml2 and\\nc:\\\\\\\\prog\\\\\\\\3p64\\\\\\\\lib\\\\\\\\libxml2.lib. I am not sure which version of\\nlibxml this is, and the other animals of MSVC don't use libxml so it\\nis not possible to correlate that only to VS 2017, either.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 30 Aug 2022 10:35:35 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Mon, Aug 29, 2022 at 05:33:56PM -0700, Andres Freund wrote:\\n>> The weirdest part is that it only happens as part of the the pg_upgrade test.\\n\\n> make check has just failed:\\n> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-30%2001%3A15%3A13\\n\\nSo it *is* probabilistic, which is pretty much what you'd expect\\nif ASLR triggers it. That brings us no closer to understanding\\nwhat the mechanism is, though.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 29 Aug 2022 22:00:10 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 29, 2022 at 10:00:10PM -0400, Tom Lane wrote:\\n> Michael Paquier <michael@paquier.xyz> writes:\\n> > On Mon, Aug 29, 2022 at 05:33:56PM -0700, Andres Freund wrote:\\n> >> The weirdest part is that it only happens as part of the the pg_upgrade test.\\n> \\n> > make check has just failed:\\n> > https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2022-08-30%2001%3A15%3A13\\n> \\n> So it *is* probabilistic, which is pretty much what you'd expect\\n> if ASLR triggers it. That brings us no closer to understanding\\n> what the mechanism is, though.\\n\\nThere have been more failures, always switching the input from\\n\\\"pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\\\"\\nto \\\"pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\\\".\\n\\nUsing a PATH of node() influences the output. I am not verse unto\\nXMLTABLE, but could it be an issue where each node is parsed and we\\nhave something like a qsort() applied on the pointer addresses for\\neach part or something like that, causing the output to become\\nunstable?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 31 Aug 2022 08:36:01 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> Using a PATH of node() influences the output. I am not verse unto\\n> XMLTABLE, but could it be an issue where each node is parsed and we\\n> have something like a qsort() applied on the pointer addresses for\\n> each part or something like that, causing the output to become\\n> unstable?\\n\\nI'm not sure. I dug into this enough to find that the output from\\nthis query is generated in xml.c lines 4635ff:\\n\\n /* Concatenate serialized values */\\n initStringInfo(&str);\\n for (int i = 0; i < count; i++)\\n {\\n textstr =\\n xml_xmlnodetoxmltype(xpathobj->nodesetval->nodeTab[i],\\n xtCxt->xmlerrcxt);\\n\\n appendStringInfoText(&str, textstr);\\n }\\n cstr = str.data;\\n\\nSo evidently, the problem occurs because the elements of the\\nnodesetval->nodeTab[] array are in a different order than we expect.\\n\\nI looked into <xpath.h> and found the definition of the \\\"nodesetval\\\"\\nstruct, and the comments are eye-opening to say the least:\\n\\n/*\\n * A node-set (an unordered collection of nodes without duplicates).\\n */\\ntypedef struct _xmlNodeSet xmlNodeSet;\\ntypedef xmlNodeSet *xmlNodeSetPtr;\\nstruct _xmlNodeSet {\\n int nodeNr;\\t\\t\\t/* number of nodes in the set */\\n int nodeMax;\\t\\t/* size of the array as allocated */\\n xmlNodePtr *nodeTab;\\t/* array of nodes in no particular order */\\n /* @@ with_ns to check wether namespace nodes should be looked at @@ */\\n};\\n\\nIt seems like maybe we're relying on an ordering we should not.\\nYet surely the ordering of the pieces of this output is meaningful?\\nAre we using the wrong libxml API to create this result?\\n\\nAnyway, I'm now suspicious that we've accidentally exposed a logic\\nbug in the XMLTABLE code, rather than anything wrong with the\\nASLR stuff.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 30 Aug 2022 20:07:27 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 31, 2022 at 08:36:01AM +0900, Michael Paquier wrote:\\n> There have been more failures, always switching the input from\\n> \\\"pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post\\\"\\n> to \\\"pre<?pi arg?><![CDATA[&ent1]]><!--c1--><n2>&deep</n2>post\\\".\\n> \\n> Using a PATH of node() influences the output. I am not verse unto\\n> XMLTABLE, but could it be an issue where each node is parsed and we\\n> have something like a qsort() applied on the pointer addresses for\\n> each part or something like that, causing the output to become\\n> unstable?\\n\\nHmm. I think that I may have an idea here after looking at our xml.c\\nand xpath.c in libxml2/. From what I understand, we process the PATH\\nthrough XmlTableGetValue() that builds a XML node path in\\nxmlXPathCompiledEval(). The interesting part comes from libxml2's\\nxmlXPathCompiledEvalInternal(), where I think we don't apply a sort on\\nthe contents generated. Hence, I am wondering if the solution here\\nwould be to do one xmlXPathNodeSetSort(xpathobj->nodesetval) after\\ncompiling the path with xmlXPathCompiledEval() in XmlTableGetValue().\\nThis should ensure that the items are ordered even if ASLR mixes if\\nthe pointer positions.\\n\\nA complete solution would involve more code paths, but we'd need only\\none change in XmlTableGetValue() for the current regression tests to\\nwork. I don't have an environment where I can reproduce that, so that\\nwould be up to the buildfarm to stress this solution..\\n\\nThoughts?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 31 Aug 2022 09:15:00 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 30, 2022 at 08:07:27PM -0400, Tom Lane wrote:\\n> It seems like maybe we're relying on an ordering we should not.\\n> Yet surely the ordering of the pieces of this output is meaningful?\\n> Are we using the wrong libxml API to create this result?\\n> \\n> Anyway, I'm now suspicious that we've accidentally exposed a logic\\n> bug in the XMLTABLE code, rather than anything wrong with the\\n> ASLR stuff.\\n\\nFunny. I have reached basically the same conclusion as you a couple\\nof minutes ago, but I also think that I have found what we need to do\\nhere to ensure the ordering of the nodes generated by the libxml\\ncode.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 31 Aug 2022 09:16:50 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> Hence, I am wondering if the solution here\\n> would be to do one xmlXPathNodeSetSort(xpathobj->nodesetval) after\\n> compiling the path with xmlXPathCompiledEval() in XmlTableGetValue().\\n\\nHmm, I see that function declared in <xpathInternals.h>, which\\nsure doesn't give me a warm feeling that we're supposed to call it\\ndirectly. But maybe there's an approved way to get the result.\\n\\nOr perhaps this test case is wrong, and instead of \\\"node()\\\" we\\nneed to write something that specifies a sorted result?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 30 Aug 2022 20:28:58 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 30, 2022 at 08:28:58PM -0400, Tom Lane wrote:\\n> Hmm, I see that function declared in <xpathInternals.h>, which\\n> sure doesn't give me a warm feeling that we're supposed to call it\\n> directly. But maybe there's an approved way to get the result.\\n> \\n> Or perhaps this test case is wrong, and instead of \\\"node()\\\" we\\n> need to write something that specifies a sorted result?\\n\\nI am no specialist on this matter.. Still, looking around, this looks\\nto be the kind of job that xsl:sort would do? We cannot use it here\\nfor an XML-only context, though, and it does not seem like there is a\\ndirect way to enforce the ordering per the nature of the XPath\\nlanguage, either. I may be missing something of course.\\n\\nPlease note that I am also getting a bit the cold feet with the idea\\nof enforcing a sorting of the elements in the generated output,\\nactually, as this could have a performance penalty for all users\\nparticularly on large blobs of data.\\n\\nAt the end, I'd like to think that it would be wiser to just remove\\nentirely the test for node() or reduce the contents of the string to\\nbe able to strictly control the output order (say a simple\\n'<e>pre<![CDATA[&ent1]]>post</e>' would do the trick). I cannot think\\nabout a better idea now. Note that removing the test case where we\\nhave node() has no impact on the coverage of xml.c.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 31 Aug 2022 15:09:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> At the end, I'd like to think that it would be wiser to just remove\\n> entirely the test for node() or reduce the contents of the string to\\n> be able to strictly control the output order (say a simple\\n> '<e>pre<![CDATA[&ent1]]>post</e>' would do the trick). I cannot think\\n> about a better idea now. Note that removing the test case where we\\n> have node() has no impact on the coverage of xml.c.\\n\\nYeah, I confirm that: local code-coverage testing shows no change\\nin the number of lines reached in xml.c when I remove that column:\\n\\n-SELECT * FROM XMLTABLE('*' PASSING '<e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>' COLUMNS x xml PATH 'node()', y xml PATH '/');\\n+SELECT * FROM XMLTABLE('*' PASSING '<e>pre<!--c1--><?pi arg?><![CDATA[&ent1]]><n2>&deep</n2>post</e>' COLUMNS y xml PATH '/');\\n\\nDropping the query altogether does result in a reduction in the\\nnumber of lines hit.\\n\\nI did not check the other XMLTABLE infrastructure, so what probably\\nis best to do is keep the two output columns but change 'node()'\\nto something with a more stable result; any preferences?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 31 Aug 2022 13:33:53 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 31, 2022 at 01:33:53PM -0400, Tom Lane wrote:\\n> I did not check the other XMLTABLE infrastructure, so what probably\\n> is best to do is keep the two output columns but change 'node()'\\n> to something with a more stable result; any preferences?\\n\\nThe input object could also be reworked so as we would not have any\\nordering issues, say \\\"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\\\".\\nChanging only the path, you could use \\\"/e/n2\\\" instead of \\\"node()\\\", so\\nas we'd always get the most internal member.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 1 Sep 2022 09:32:38 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> The input object could also be reworked so as we would not have any\\n> ordering issues, say \\\"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\\\".\\n> Changing only the path, you could use \\\"/e/n2\\\" instead of \\\"node()\\\", so\\n> as we'd always get the most internal member.\\n\\nDone that way.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 31 Aug 2022 22:22:06 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 31, 2022 at 10:22:06PM -0400, Tom Lane wrote:\\n> Michael Paquier <michael@paquier.xyz> writes:\\n> > The input object could also be reworked so as we would not have any\\n> > ordering issues, say \\\"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\\\".\\n> > Changing only the path, you could use \\\"/e/n2\\\" instead of \\\"node()\\\", so\\n> > as we'd always get the most internal member.\\n> \\n> Done that way.\\n\\nCool, thanks. bowerbird looks happy after its first run.\\n\\nAn argument in favor of backpatching is if one decides to build the\\ncode with MSVC and patch the scripts to enable ASLR. However, nobody\\nhas complained about that yet, so fine by me to leave this change only\\non HEAD for now.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 1 Sep 2022 13:41:59 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"2022年9月1日(木) 13:42 Michael Paquier <michael@paquier.xyz>:\\n>\\n> On Wed, Aug 31, 2022 at 10:22:06PM -0400, Tom Lane wrote:\\n> > Michael Paquier <michael@paquier.xyz> writes:\\n> > > The input object could also be reworked so as we would not have any\\n> > > ordering issues, say \\\"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\\\".\\n> > > Changing only the path, you could use \\\"/e/n2\\\" instead of \\\"node()\\\", so\\n> > > as we'd always get the most internal member.\\n> >\\n> > Done that way.\\n>\\n> Cool, thanks. bowerbird looks happy after its first run.\\n>\\n> An argument in favor of backpatching is if one decides to build the\\n> code with MSVC and patch the scripts to enable ASLR. However, nobody\\n> has complained about that yet, so fine by me to leave this change only\\n> on HEAD for now.\\n\\nApologies for the thread bump, but there is an entry for this thread\\nin the current CF\\npreviously marked \\\"Needs review\\\":\\n\\n https://commitfest.postgresql.org/40/3561/\\n\\nI've gone ahead and marked it as \\\"Committed\\\", as that appears to have happened\\nback in August as 36389a060c.\\n\\nIf for whatever reason that was the Wrong Thing To Do, please let me know.\\n\\nRegards\\n\\nIan Barwick\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Nov 2022 17:38:09 +0900\",\n \"msg_from\": \"Ian Lawrence Barwick <barwick@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"> I've gone ahead and marked it as \\\"Committed\\\", as that appears to have\\nhappened\\n> back in August as 36389a060c.\\n> If for whatever reason that was the Wrong Thing To Do, please let me know.\\n\\n36389a060c commit enables ASLR for windows but does not include other\\nchanges required for Arm64.\\n\\nI've attached a new version of the patch which excludes the already merged\\nASLR changes and add\\nsmall changes to handle latest changes in the build scripts.\\n\\nOn Thu, 3 Nov 2022 at 08:38, Ian Lawrence Barwick <barwick@gmail.com> wrote:\\n\\n> 2022年9月1日(木) 13:42 Michael Paquier <michael@paquier.xyz>:\\n> >\\n> > On Wed, Aug 31, 2022 at 10:22:06PM -0400, Tom Lane wrote:\\n> > > Michael Paquier <michael@paquier.xyz> writes:\\n> > > > The input object could also be reworked so as we would not have any\\n> > > > ordering issues, say\\n> \\\"<e>pre<f><g><n2>&deep</n2></g></f>post</e>\\\".\\n> > > > Changing only the path, you could use \\\"/e/n2\\\" instead of \\\"node()\\\", so\\n> > > > as we'd always get the most internal member.\\n> > >\\n> > > Done that way.\\n> >\\n> > Cool, thanks. bowerbird looks happy after its first run.\\n> >\\n> > An argument in favor of backpatching is if one decides to build the\\n> > code with MSVC and patch the scripts to enable ASLR. However, nobody\\n> > has complained about that yet, so fine by me to leave this change only\\n> > on HEAD for now.\\n>\\n> Apologies for the thread bump, but there is an entry for this thread\\n> in the current CF\\n> previously marked \\\"Needs review\\\":\\n>\\n> https://commitfest.postgresql.org/40/3561/\\n>\\n> I've gone ahead and marked it as \\\"Committed\\\", as that appears to have\\n> happened\\n> back in August as 36389a060c.\\n>\\n> If for whatever reason that was the Wrong Thing To Do, please let me know.\\n>\\n> Regards\\n>\\n> Ian Barwick\\n>\\n\\n\\n-- \\nNiyas Sait\\nLinaro Limited\",\n \"msg_date\": \"Thu, 3 Nov 2022 11:06:46 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 03, 2022 at 11:06:46AM +0000, Niyas Sait wrote:\\n> > I've gone ahead and marked it as \\\"Committed\\\", as that appears to have\\n> happened\\n> > back in August as 36389a060c.\\n> > If for whatever reason that was the Wrong Thing To Do, please let me know.\\n> \\n> 36389a060c commit enables ASLR for windows but does not include other\\n> changes required for Arm64.\\n\\nYes, marking this patch as committed is incorrect. That's just me\\nlagging behind for the review :)\\n\\nI have switched that back as \\\"Needs review\\\" for the moment.\\n\\n> I've attached a new version of the patch which excludes the already merged\\n> ASLR changes and add\\n> small changes to handle latest changes in the build scripts.\\n\\nThanks.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 3 Nov 2022 20:43:04 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-11-03 11:06:46 +0000, Niyas Sait wrote:\\n> I've attached a new version of the patch which excludes the already merged\\n> ASLR changes and add\\n> small changes to handle latest changes in the build scripts.\\n\\nNote that we're planning to remove the custom windows build scripts before the\\nnext release, relying on the meson build instead.\\n\\n\\n> diff --git a/src/include/storage/s_lock.h b/src/include/storage/s_lock.h\\n> index 8b19ab160f..bf6a6dba35 100644\\n> --- a/src/include/storage/s_lock.h\\n> +++ b/src/include/storage/s_lock.h\\n> @@ -708,13 +708,21 @@ typedef LONG slock_t;\\n> #define SPIN_DELAY() spin_delay()\\n> \\n> /* If using Visual C++ on Win64, inline assembly is unavailable.\\n> - * Use a _mm_pause intrinsic instead of rep nop.\\n> + * Use _mm_pause (x64) or __isb(arm64) intrinsic instead of rep nop.\\n> */\\n> #if defined(_WIN64)\\n> static __forceinline void\\n> spin_delay(void)\\n> {\\n> +#ifdef _M_ARM64\\n> +\\t/*\\n> +\\t * arm64 way of hinting processor for spin loops optimisations\\n> +\\t * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\\n> +\\t*/\\n> +\\t__isb(_ARM64_BARRIER_SY);\\n> +#else\\n> \\t_mm_pause();\\n> +#endif\\n> }\\n> #else\\n> static __forceinline void\\n\\nI think we should just apply this, there seems very little risk of making\\nanything worse, given the gating to _WIN64 && _M_ARM64.\\n\\n\\n> diff --git a/src/port/pg_crc32c_armv8.c b/src/port/pg_crc32c_armv8.c\\n> index 9e301f96f6..981718752f 100644\\n> --- a/src/port/pg_crc32c_armv8.c\\n> +++ b/src/port/pg_crc32c_armv8.c\\n> @@ -14,7 +14,9 @@\\n> */\\n> #include \\\"c.h\\\"\\n> \\n> +#ifndef _MSC_VER\\n> #include <arm_acle.h>\\n> +#endif\\n> \\n> #include \\\"port/pg_crc32c.h\\\"\\n\\nThis won't suffice with the meson build, since the relevant configure test\\nalso uses arm_acle.h:\\nelif host_cpu == 'arm' or host_cpu == 'aarch64'\\n\\n prog = '''\\n#include <arm_acle.h>\\n\\nint main(void)\\n{\\n unsigned int crc = 0;\\n crc = __crc32cb(crc, 0);\\n crc = __crc32ch(crc, 0);\\n crc = __crc32cw(crc, 0);\\n crc = __crc32cd(crc, 0);\\n\\n /* return computed value, to prevent the above being optimized away */\\n return crc == 0;\\n}\\n'''\\n\\n if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n args: test_c_args)\\n # Use ARM CRC Extension unconditionally\\n cdata.set('USE_ARMV8_CRC32C', 1)\\n have_optimized_crc = true\\n elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\\n args: test_c_args + ['-march=armv8-a+crc'])\\n # Use ARM CRC Extension, with runtime check\\n cflags_crc += '-march=armv8-a+crc'\\n cdata.set('USE_ARMV8_CRC32C', false)\\n cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n have_optimized_crc = true\\n endif\\nendif\\n\\nThe meson checking logic is used both for msvc and other compilers, so this\\nwill need to work with both.\\n\\n\\n> diff --git a/src/tools/msvc/gendef.pl b/src/tools/msvc/gendef.pl\\n> index d6bed1ce15..f1e0ff446b 100644\\n> --- a/src/tools/msvc/gendef.pl\\n> +++ b/src/tools/msvc/gendef.pl\\n> @@ -120,9 +120,9 @@ sub writedef\\n> \\t{\\n> \\t\\tmy $isdata = $def->{$f} eq 'data';\\n> \\n> -\\t\\t# Strip the leading underscore for win32, but not x64\\n> +\\t\\t# Strip the leading underscore for win32, but not x64 and arm64\\n> \\t\\t$f =~ s/^_//\\n> -\\t\\t unless ($arch eq \\\"x86_64\\\");\\n> +\\t\\t unless ($arch ne \\\"x86\\\");\\n> \\n> \\t\\t# Emit just the name if it's a function symbol, or emit the name\\n> \\t\\t# decorated with the DATA option for variables.\\n> @@ -143,7 +143,7 @@ sub writedef\\n> sub usage\\n> {\\n> \\tdie(\\\"Usage: gendef.pl --arch <arch> --deffile <deffile> --tempdir <tempdir> files-or-directories\\\\n\\\"\\n> -\\t\\t . \\\" arch: x86 | x86_64\\\\n\\\"\\n> +\\t\\t . \\\" arch: x86 | x86_64 | arm64 \\\\n\\\"\\n> \\t\\t . \\\" deffile: path of the generated file\\\\n\\\"\\n> \\t\\t . \\\" tempdir: directory for temporary files\\\\n\\\"\\n> \\t\\t . \\\" files or directories: object files or directory containing object files\\\\n\\\"\\n> @@ -160,7 +160,7 @@ GetOptions(\\n> \\t'tempdir:s' => \\\\$tempdir,) or usage();\\n> \\n> usage(\\\"arch: $arch\\\")\\n> - unless ($arch eq 'x86' || $arch eq 'x86_64');\\n> + unless ($arch eq 'x86' || $arch eq 'x86_64' || $arch eq 'arm64');\\n> \\n> my @files;\\n> \\n\\nSeems reasonable.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sat, 5 Nov 2022 11:31:36 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Sat, Nov 05, 2022 at 11:31:36AM -0700, Andres Freund wrote:\\n> On 2022-11-03 11:06:46 +0000, Niyas Sait wrote:\\n> Note that we're planning to remove the custom windows build scripts before the\\n> next release, relying on the meson build instead.\\n\\nYoupi.\\n\\n>> #if defined(_WIN64)\\n>> static __forceinline void\\n>> spin_delay(void)\\n>> {\\n>> +#ifdef _M_ARM64\\n>> +\\t/*\\n>> +\\t * arm64 way of hinting processor for spin loops optimisations\\n>> +\\t * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\\n>> +\\t*/\\n>> +\\t__isb(_ARM64_BARRIER_SY);\\n>> +#else\\n>> \\t_mm_pause();\\n>> +#endif\\n>> }\\n>> #else\\n>> static __forceinline void\\n> \\n> I think we should just apply this, there seems very little risk of making\\n> anything worse, given the gating to _WIN64 && _M_ARM64.\\n\\nSeems so. Hmm, where does _ARM64_BARRIER_SY come from? Perhaps it\\nwould be better to have a comment referring to it from a different\\nplace than the forums of arm, like some actual docs?\\n\\n> The meson checking logic is used both for msvc and other compilers, so this\\n> will need to work with both.\\n\\nThe patch has been switched as waiting on author for now.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 7 Nov 2022 15:58:13 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 07, 2022 at 03:58:13PM +0900, Michael Paquier wrote:\\n> The patch has been switched as waiting on author for now.\\n\\nSeeing no replies after three weeks, I have marked the patch as\\nreturned with feedback for now.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 1 Dec 2022 16:42:19 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"\\nOn 05/11/2022 18:31, Andres Freund wrote:\\n> On 2022-11-03 11:06:46 +0000, Niyas Sait wrote:\\n>> I've attached a new version of the patch which excludes the already merged\\n>> ASLR changes and add\\n>> small changes to handle latest changes in the build scripts.\\n> Note that we're planning to remove the custom windows build scripts before the\\n> next release, relying on the meson build instead.\\n> \\n\\nThanks. I will add changes to add meson build support.\\n\\n> This won't suffice with the meson build, since the relevant configure test\\n> also uses arm_acle.h:\\n> elif host_cpu == 'arm' or host_cpu == 'aarch64'\\n> \\n> prog = '''\\n> #include <arm_acle.h>\\n> \\n> int main(void)\\n> {\\n> unsigned int crc = 0;\\n> crc = __crc32cb(crc, 0);\\n> crc = __crc32ch(crc, 0);\\n> crc = __crc32cw(crc, 0);\\n> crc = __crc32cd(crc, 0);\\n> \\n> /* return computed value, to prevent the above being optimized away */\\n> return crc == 0;\\n> }\\n> '''\\n> \\n> if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n> args: test_c_args)\\n> # Use ARM CRC Extension unconditionally\\n> cdata.set('USE_ARMV8_CRC32C', 1)\\n> have_optimized_crc = true\\n> elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\\n> args: test_c_args + ['-march=armv8-a+crc'])\\n> # Use ARM CRC Extension, with runtime check\\n> cflags_crc += '-march=armv8-a+crc'\\n> cdata.set('USE_ARMV8_CRC32C', false)\\n> cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> have_optimized_crc = true\\n> endif\\n> endif\\n> \\n> The meson checking logic is used both for msvc and other compilers, so this\\n> will need to work with both.\\n\\nYes, will handle that.\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Thu, 1 Dec 2022 17:17:53 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"\\nOn 07/11/2022 06:58, Michael Paquier wrote:\\n>>> #if defined(_WIN64)\\n>>> static __forceinline void\\n>>> spin_delay(void)\\n>>> {\\n>>> +#ifdef _M_ARM64\\n>>> +\\t/*\\n>>> +\\t * arm64 way of hinting processor for spin loops optimisations\\n>>> +\\t * ref: https://community.arm.com/support-forums/f/infrastructure-solutions-forum/48654/ssetoneon-faq\\n>>> +\\t*/\\n>>> +\\t__isb(_ARM64_BARRIER_SY);\\n>>> +#else\\n>>> \\t_mm_pause();\\n>>> +#endif\\n>>> }\\n>>> #else\\n>>> static __forceinline void\\n>>\\n>> I think we should just apply this, there seems very little risk of making\\n>> anything worse, given the gating to _WIN64 && _M_ARM64.\\n> \\n> Seems so. Hmm, where does _ARM64_BARRIER_SY come from? Perhaps it\\n> would be better to have a comment referring to it from a different\\n> place than the forums of arm, like some actual docs?\\n\\n\\n_ARM64_BARRIER_SY is defined in Microsoft Arm64 intrinsic documentation \\n- \\nhttps://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-170#BarrierRestrictions\\n\\nI couldn't find something more official for the sse2neon library part.\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Thu, 1 Dec 2022 17:20:20 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hello,\\n\\nI've attached a new revision of the patch (v4) and includes following \\nchanges,\\n\\n1. Add support for meson build system\\n2. Extend MSVC scripts to handle ARM64 platform.\\n3. Add arm64 definition of spin_delay function.\\n4. Exclude arm_acle.h import with MSVC compiler.\\n\\nV3->V4: Add support for meson build system\\nV2->V3: Removed ASLR enablement and rebased on master.\\nV1->V2: Rebased on top of master\\n\\n-- \\nNiyas\",\n \"msg_date\": \"Thu, 1 Dec 2022 17:53:47 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 01, 2022 at 05:53:47PM +0000, Niyas Sait wrote:\\n> I've attached a new revision of the patch (v4) and includes following\\n> changes,\\n> \\n> 1. Add support for meson build system\\n> 2. Extend MSVC scripts to handle ARM64 platform.\\n> 3. Add arm64 definition of spin_delay function.\\n> 4. Exclude arm_acle.h import with MSVC compiler.\\n> \\n> V3->V4: Add support for meson build system\\n> V2->V3: Removed ASLR enablement and rebased on master.\\n> V1->V2: Rebased on top of master\\n\\nThanks for the updated version. I have been looking at it closely and\\nit looks like it should be able to do the job (no arm64 machine for\\nWindows here, sigh).\\n\\nI have one tiny comment about this part:\\n\\n- USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => undef,\\n+ USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => 1,\\n\\nShouldn't we only enable this flag when we are under aarch64?\\nSimilarly, I don't think that it is a good idea to switch on\\nUSE_SSE42_CRC32C_WITH_RUNTIME_CHECK all the time. We should only set\\nit when building under x86 and x86_64.\\n\\nI would also add your link [1] in s_lock.h.\\n\\n[1]: https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics#BarrierRestrictions\\n--\\nMichael\",\n \"msg_date\": \"Fri, 2 Dec 2022 14:02:54 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 2, 2022 at 12:20 AM Niyas Sait <niyas.sait@linaro.org> wrote:\\n>\\n>\\n> On 07/11/2022 06:58, Michael Paquier wrote:\\n> > Seems so. Hmm, where does _ARM64_BARRIER_SY come from? Perhaps it\\n> > would be better to have a comment referring to it from a different\\n> > place than the forums of arm, like some actual docs?\\n>\\n>\\n> _ARM64_BARRIER_SY is defined in Microsoft Arm64 intrinsic documentation\\n> -\\n>\\nhttps://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-170#BarrierRestrictions\\n\\nIn particular, at the bottom of that section:\\n\\n\\\"For the __isb intrinsic, the only restriction that is currently valid is\\n_ARM64_BARRIER_SY; all other values are reserved by the architecture.\\\"\\n\\nThis corresponds to\\n\\nhttps://developer.arm.com/documentation/ddi0596/2021-06/Base-Instructions/ISB--Instruction-Synchronization-Barrier-\\n\\nwhich says\\n\\n\\\"SY Full system barrier operation, encoded as CRm = 0b1111. Can be omitted.\\\"\\n\\n> I couldn't find something more official for the sse2neon library part.\\n\\nNot quite sure what this is referring to, but it seems we can just point to\\nthe __aarch64__ section in the same file, which uses the same instruction:\\n\\nspin_delay(void)\\n{\\n __asm__ __volatile__(\\n \\\" isb; \\\\n\\\");\\n}\\n\\n...and which already explains the choice with a comment.\\n\\nAbout v4:\\n\\n+ * Use _mm_pause (x64) or __isb(arm64) intrinsic instead of rep nop.\\n\\nNeed a space here after __isb.\\n\\n+ if cc.get_id() == 'msvc'\\n+ cdata.set('USE_ARMV8_CRC32C', false)\\n+ cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n+ have_optimized_crc = true\\n+ else\\n\\nThat seems like a heavy-handed way to force it. Could we just use the same\\ngating in the test program that the patch puts in the code of interest?\\nNamely:\\n\\n+#ifndef _MSC_VER\\n #include <arm_acle.h>\\n+#endif\\n\\n--\\nJohn Naylor\\nEDB: http://www.enterprisedb.com\\n\\nOn Fri, Dec 2, 2022 at 12:20 AM Niyas Sait <niyas.sait@linaro.org> wrote:>>> On 07/11/2022 06:58, Michael Paquier wrote:> > Seems so. Hmm, where does _ARM64_BARRIER_SY come from? Perhaps it> > would be better to have a comment referring to it from a different> > place than the forums of arm, like some actual docs?>>> _ARM64_BARRIER_SY is defined in Microsoft Arm64 intrinsic documentation> -> https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-170#BarrierRestrictionsIn particular, at the bottom of that section:\\\"For the __isb intrinsic, the only restriction that is currently valid is _ARM64_BARRIER_SY; all other values are reserved by the architecture.\\\"This corresponds tohttps://developer.arm.com/documentation/ddi0596/2021-06/Base-Instructions/ISB--Instruction-Synchronization-Barrier-which says\\\"SY Full system barrier operation, encoded as CRm = 0b1111. Can be omitted.\\\"> I couldn't find something more official for the sse2neon library part.Not quite sure what this is referring to, but it seems we can just point to the __aarch64__ section in the same file, which uses the same instruction:spin_delay(void){ __asm__ __volatile__( \\\" isb;\\t\\t\\t\\t\\\\n\\\");}...and which already explains the choice with a comment.About v4:+ * Use _mm_pause (x64) or __isb(arm64) intrinsic instead of rep nop.Need a space here after __isb.+ if cc.get_id() == 'msvc'+ cdata.set('USE_ARMV8_CRC32C', false)+ cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)+ have_optimized_crc = true+ elseThat seems like a heavy-handed way to force it. Could we just use the same gating in the test program that the patch puts in the code of interest? Namely:+#ifndef _MSC_VER #include <arm_acle.h>+#endif--John NaylorEDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Fri, 2 Dec 2022 12:41:06 +0700\",\n \"msg_from\": \"John Naylor <john.naylor@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"\\nOn 02/12/2022 05:02, Michael Paquier wrote:\\n> Thanks for the updated version. I have been looking at it closely and\\n> it looks like it should be able to do the job (no arm64 machine for\\n> Windows here, sigh).\\n> \\n> I have one tiny comment about this part:\\n> \\n> - USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => undef,\\n> + USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => 1,\\n> \\n> Shouldn't we only enable this flag when we are under aarch64?\\n> Similarly, I don't think that it is a good idea to switch on\\n> USE_SSE42_CRC32C_WITH_RUNTIME_CHECK all the time. We should only set\\n> it when building under x86 and x86_64.\\n> \\n\\nOk, I will fix ARM64 specific one in the next revision.\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Dec 2022 09:37:43 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 02/12/2022 05:41, John Naylor wrote:\\n>> I couldn't find something more official for the sse2neon library part.\\n> Not quite sure what this is referring to, but it seems we can just point to\\n> the __aarch64__ section in the same file, which uses the same instruction:\\n> \\n> spin_delay(void)\\n> {\\n> __asm__ __volatile__(\\n> \\\" isb; \\\\n\\\");\\n> }\\n> \\n> ...and which already explains the choice with a comment.\\n\\nGood point. Will add the comment.\\n\\n> + if cc.get_id() == 'msvc'\\n> + cdata.set('USE_ARMV8_CRC32C', false)\\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> + have_optimized_crc = true\\n> + else\\n> \\n> That seems like a heavy-handed way to force it. Could we just use the same\\n> gating in the test program that the patch puts in the code of interest?\\n> Namely:\\n> \\n> +#ifndef _MSC_VER\\n> #include <arm_acle.h>\\n> +#endif\\nI took a similar approach as x86 MSVC code. I don't think the test \\nprogram would work with MSVC. The compiler options are not MSVC friendly.\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Fri, 2 Dec 2022 09:50:20 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hello,\\n\\nI've attached a new revision of the patch (v5) and includes following \\nchanges,\\n\\n1. Add support for meson build system\\n2. Extend MSVC scripts to handle ARM64 platform.\\n3. Add arm64 definition of spin_delay function.\\n4. Exclude arm_acle.h import with MSVC compiler.\\n\\nV4->V5: * Added reference to Microsoft arm64 intrinsic documentation\\n * Conditionally enable USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK\\n * Fixed styling issue spotted by Ian Lawerence Barwick\\nV3->V4: Add support for meson build system\\nV2->V3: Removed ASLR enablement and rebased on master.\\nV1->V2: Rebased on top of master\\n\\n-- \\nNiyas\",\n \"msg_date\": \"Fri, 2 Dec 2022 11:09:15 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 02, 2022 at 11:09:15AM +0000, Niyas Sait wrote:\\n> I've attached a new revision of the patch (v5) and includes following\\n> changes,\\n> \\n> 1. Add support for meson build system\\n> 2. Extend MSVC scripts to handle ARM64 platform.\\n> 3. Add arm64 definition of spin_delay function.\\n> 4. Exclude arm_acle.h import with MSVC compiler.\\n\\nHmm. There are still a few things that need some attention here:\\n- USE_SSE42_CRC32C_WITH_RUNTIME_CHECK should not be set for aarch64.\\n- This is missing updates for ICU. Looking at the upstream code,\\nBuild.Windows.ProjectConfiguration.props uses libARM64 and binARM64\\nfor the output library and binary paths.\\n- This is missing updates for krb5. For this case, I am seeing no\\ntraces of packages for aarch64, so I guess that we could just fail\\nhard until someone cares enough to ping us about what to do here.\\n- There were zero changes in the docs, but we need to update at least\\nthe section about architectures supported for the 64-bit builds.\\n- Last comes OpenSSL, that supports amd64_arm64 as build target (see\\nNOTES-WINDOWS.md), and the library names are libssl.lib and\\nlibcrypto.lib. Looking at\\nhttps://slproweb.com/products/Win32OpenSSL.html, there are\\nexperimental builds for arm64 with OpenSSL 3.0. Niyas or somebody\\nelse, could you look at the contents of lib/VC/ and see what we could\\nrely on for the debug builds after installing this MSI? We should\\nrely on something like lib/VC/sslcrypto64MD.lib or\\nlib/VC/sslcrypto32MD.lib, but for arm64.\\n\\nWith meson gaining in maturity, perhaps that's not the most urgent\\nthing as we will likely remove src/tools/msvc/ soon but I'd rather do\\nthat right anyway as much as I can to avoid an incorrect state in the\\ntree at any time in its history.\\n\\n- USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => undef,\\n+ USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => $self->{platform} eq \\\"ARM64\\\" ? : 1 : undef,\\nDid you actually test this patch? This won't work at all with perl,\\nper se the double colon after the question mark.\\n\\nFor now, please find attached an updated patch with all the fixes I\\ncould come up with.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 5 Dec 2022 14:12:41 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"\\nOn 05/12/2022 05:12, Michael Paquier wrote:\\n> - Last comes OpenSSL, that supports amd64_arm64 as build target (see\\n> NOTES-WINDOWS.md), and the library names are libssl.lib and\\n> libcrypto.lib. Looking at\\n> https://slproweb.com/products/Win32OpenSSL.html, there are\\n> experimental builds for arm64 with OpenSSL 3.0. Niyas or somebody\\n> else, could you look at the contents of lib/VC/ and see what we could\\n> rely on for the debug builds after installing this MSI? We should\\n> rely on something like lib/VC/sslcrypto64MD.lib or\\n> lib/VC/sslcrypto32MD.lib, but for arm64.\\n\\nI tried that installer. And I can see following libraries installed in \\nlib/VC location.\\n\\nlibcryptoarm64MD.lib\\nlibcryptoarm64MDd.lib\\nlibcryptoarm64MT.lib\\nlibcryptoarm64MTd.lib\\nlibsslarm64MD.lib\\nlibsslarm64MDd.lib\\nlibsslarm64MT.lib\\nlibsslarm64MTd.lib\\n\\n> - USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => undef,\\n> + USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK => $self->{platform} eq \\\"ARM64\\\" ? : 1 : undef,\\n> Did you actually test this patch? This won't work at all with perl,\\n> per se the double colon after the question mark.\\n\\n\\nYes I did a full build. I am not sure why I didn't see any error with \\nthat. My perl skills are very limited and I started with something bit \\nmore naive like \\\"$self->{platform} == \\\"ARM64\\\" ? : 1 : undef\\\" But that \\ndidn't work and I changed to using eq and the compilation was fine. \\nThanks for fixing the patch.\\n\\n > For now, please find attached an updated patch with all the fixes I\\n > could come up with.\\n\\nThanks. I did a quick sanity build with your updated patch and looks fine.\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Mon, 5 Dec 2022 08:54:52 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-12-05 14:12:41 +0900, Michael Paquier wrote:\\n> With meson gaining in maturity, perhaps that's not the most urgent\\n> thing as we will likely remove src/tools/msvc/ soon but I'd rather do\\n> that right anyway as much as I can to avoid an incorrect state in the\\n> tree at any time in its history.\\n\\nI'd actually argue that we should just not add win32 support to\\nsrc/tools/msvc/.\\n\\n\\n> --- a/src/include/storage/s_lock.h\\n> +++ b/src/include/storage/s_lock.h\\n> @@ -708,13 +708,21 @@ typedef LONG slock_t;\\n> #define SPIN_DELAY() spin_delay()\\n> \\n> /* If using Visual C++ on Win64, inline assembly is unavailable.\\n> - * Use a _mm_pause intrinsic instead of rep nop.\\n> + * Use _mm_pause (x64) or __isb (arm64) intrinsic instead of rep nop.\\n> */\\n> #if defined(_WIN64)\\n> static __forceinline void\\n> spin_delay(void)\\n> {\\n> +#ifdef _M_ARM64\\n> +\\t/*\\n> +\\t * See spin_delay aarch64 inline assembly definition above for details\\n> +\\t * ref: https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics#BarrierRestrictions\\n> +\\t*/\\n> +\\t__isb(_ARM64_BARRIER_SY);\\n> +#else\\n> \\t_mm_pause();\\n> +#endif\\n> }\\n> #else\\n> static __forceinline void\\n\\nThis looks somewhat wrong to me. We end up with some ifdefs on the function\\ndefintion level, and some others inside the function body. I think it should\\nbe either or.\\n\\n\\n\\n> diff --git a/meson.build b/meson.build\\n> index 725e10d815..e354ad7650 100644\\n> --- a/meson.build\\n> +++ b/meson.build\\n> @@ -1944,7 +1944,13 @@ int main(void)\\n> \\n> elif host_cpu == 'arm' or host_cpu == 'aarch64'\\n> \\n> - prog = '''\\n> + if cc.get_id() == 'msvc'\\n> + cdata.set('USE_ARMV8_CRC32C', false)\\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> + have_optimized_crc = true\\n> + else\\n> +\\n> + prog = '''\\n> #include <arm_acle.h>\\n> \\n> int main(void)\\n\\nWhy does this need to be hardcoded? The compiler probe should just work for\\nmsvc.\\n\\n\\n> @@ -1960,18 +1966,19 @@ int main(void)\\n> }\\n> '''\\n> \\n> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n> - args: test_c_args)\\n> - # Use ARM CRC Extension unconditionally\\n> - cdata.set('USE_ARMV8_CRC32C', 1)\\n> - have_optimized_crc = true\\n> - elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\\n> - args: test_c_args + ['-march=armv8-a+crc'])\\n> - # Use ARM CRC Extension, with runtime check\\n> - cflags_crc += '-march=armv8-a+crc'\\n> - cdata.set('USE_ARMV8_CRC32C', false)\\n> - cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> - have_optimized_crc = true\\n> + if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n> + args: test_c_args)\\n> + # Use ARM CRC Extension unconditionally\\n> + cdata.set('USE_ARMV8_CRC32C', 1)\\n> + have_optimized_crc = true\\n> + elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\\n> + args: test_c_args + ['-march=armv8-a+crc'])\\n> + # Use ARM CRC Extension, with runtime check\\n> + cflags_crc += '-march=armv8-a+crc'\\n> + cdata.set('USE_ARMV8_CRC32C', false)\\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> + have_optimized_crc = true\\n> + endif\\n> endif\\n> endif\\n\\nAnd then this reindent wouldn't be needed.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 5 Dec 2022 10:14:49 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 05, 2022 at 10:14:49AM -0800, Andres Freund wrote:\\n> On 2022-12-05 14:12:41 +0900, Michael Paquier wrote:\\n>> With meson gaining in maturity, perhaps that's not the most urgent\\n>> thing as we will likely remove src/tools/msvc/ soon but I'd rather do\\n>> that right anyway as much as I can to avoid an incorrect state in the\\n>> tree at any time in its history.\\n> \\n> I'd actually argue that we should just not add win32 support to\\n> src/tools/msvc/.\\n\\nAre you arguing about ripping out the existing win32 or do nothing for\\narm64? I guess the later, but these words also sound like the former\\nto me.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 6 Dec 2022 08:31:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-12-06 08:31:16 +0900, Michael Paquier wrote:\\n> On Mon, Dec 05, 2022 at 10:14:49AM -0800, Andres Freund wrote:\\n> > On 2022-12-05 14:12:41 +0900, Michael Paquier wrote:\\n> >> With meson gaining in maturity, perhaps that's not the most urgent\\n> >> thing as we will likely remove src/tools/msvc/ soon but I'd rather do\\n> >> that right anyway as much as I can to avoid an incorrect state in the\\n> >> tree at any time in its history.\\n> > \\n> > I'd actually argue that we should just not add win32 support to\\n> > src/tools/msvc/.\\n> \\n> Are you arguing about ripping out the existing win32 or do nothing for\\n> arm64? I guess the later, but these words also sound like the former\\n> to me.\\n\\nYes, that's a typo. I meant that we shouldn't add arm64 support to\\nsrc/tools/msvc/.\\n\\nI do think we should rip out src/tools/msvc/ soon-ish, but we need\\nbuildfarm support first.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 5 Dec 2022 15:48:45 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 05, 2022 at 03:48:45PM -0800, Andres Freund wrote:\\n> Yes, that's a typo. I meant that we shouldn't add arm64 support to\\n> src/tools/msvc/.\\n\\nOkay, thanks.\\n\\n> I do think we should rip out src/tools/msvc/ soon-ish, but we need\\n> buildfarm support first.\\n\\nSure, let's see where it goes.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 6 Dec 2022 09:16:17 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 05/12/2022 18:14, Andres Freund wrote:\\n\\n>> With meson gaining in maturity, perhaps that's not the most urgent\\n>> thing as we will likely remove src/tools/msvc/ soon but I'd rather do\\n>> that right anyway as much as I can to avoid an incorrect state in the\\n>> tree at any time in its history.\\n> \\n> I'd actually argue that we should just not add win32 support to\\n> src/tools/msvc/.\\n> \\n> \\n\\nI think the old build system specific part is really minimal in the \\npatch. I can strip out those if that's preferred.\\n\\n\\n>> --- a/src/include/storage/s_lock.h\\n>> +++ b/src/include/storage/s_lock.h\\n>> @@ -708,13 +708,21 @@ typedef LONG slock_t;\\n>> #define SPIN_DELAY() spin_delay()\\n>> \\n>> /* If using Visual C++ on Win64, inline assembly is unavailable.\\n>> - * Use a _mm_pause intrinsic instead of rep nop.\\n>> + * Use _mm_pause (x64) or __isb (arm64) intrinsic instead of rep nop.\\n>> */\\n>> #if defined(_WIN64)\\n>> static __forceinline void\\n>> spin_delay(void)\\n>> {\\n>> +#ifdef _M_ARM64\\n>> +\\t/*\\n>> +\\t * See spin_delay aarch64 inline assembly definition above for details\\n>> +\\t * ref: https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics#BarrierRestrictions\\n>> +\\t*/\\n>> +\\t__isb(_ARM64_BARRIER_SY);\\n>> +#else\\n>> \\t_mm_pause();\\n>> +#endif\\n>> }\\n>> #else\\n>> static __forceinline void\\n> \\n> This looks somewhat wrong to me. We end up with some ifdefs on the function\\n> defintion level, and some others inside the function body. I think it should\\n> be either or.\\n> \\n\\nOk, I can add an MSVC/ARM64 specific function.\\n\\n>> diff --git a/meson.build b/meson.build\\n>> index 725e10d815..e354ad7650 100644\\n>> --- a/meson.build\\n>> +++ b/meson.build\\n>> @@ -1944,7 +1944,13 @@ int main(void)\\n>> \\n>> elif host_cpu == 'arm' or host_cpu == 'aarch64'\\n>> \\n>> - prog = '''\\n>> + if cc.get_id() == 'msvc'\\n>> + cdata.set('USE_ARMV8_CRC32C', false)\\n>> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n>> + have_optimized_crc = true\\n>> + else\\n>> +\\n>> + prog = '''\\n>> #include <arm_acle.h>\\n>> \\n>> int main(void)\\n> Why does this need to be hardcoded? The compiler probe should just work for\\n> msvc.\\n> \\n\\nThere are couple of minor issues in the code probe with MSVC such as \\narm_acle.h needs to be removed and requires an explicit import of \\nintrin.h. But even with those fixes, USE_ARMV8_CRC32C would be set and \\nno runtime CRC extension check will be performed. Since CRC extension is \\noptional in ARMv8, It would be better to use the CRC variant with \\nruntime check. So I end up following the x64 variant and hardcoded the \\nflags in case of ARM64 and MSVC.\\n\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Dec 2022 13:38:37 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 12, 2022 at 01:38:37PM +0000, Niyas Sait wrote:\\n> On 05/12/2022 18:14, Andres Freund wrote:\\n> I think the old build system specific part is really minimal in the patch. I\\n> can strip out those if that's preferred.\\n\\nRemoving all the changes from src/tools/msvc/ is an approach that\\nworks for me.\\n\\n>> Why does this need to be hardcoded? The compiler probe should just work for\\n>> msvc.\\n> \\n> There are couple of minor issues in the code probe with MSVC such as\\n> arm_acle.h needs to be removed and requires an explicit import of intrin.h.\\n> But even with those fixes, USE_ARMV8_CRC32C would be set and no runtime CRC\\n> extension check will be performed. Since CRC extension is optional in ARMv8,\\n> It would be better to use the CRC variant with runtime check. So I end up\\n> following the x64 variant and hardcoded the flags in case of ARM64 and MSVC.\\n\\nHm.. Andres, do you have access to Windows hosts with ARM64\\nunderneath to validate any of that? No way to blame Cirrus for not\\nhaving this option, they already do a lot :)\\n--\\nMichael\",\n \"msg_date\": \"Tue, 13 Dec 2022 08:56:33 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 12/12/2022 23:56, Michael Paquier wrote:\\n> On Mon, Dec 12, 2022 at 01:38:37PM +0000, Niyas Sait wrote:\\n>> On 05/12/2022 18:14, Andres Freund wrote:\\n>> I think the old build system specific part is really minimal in the patch. I\\n>> can strip out those if that's preferred.\\n> Removing all the changes from src/tools/msvc/ is an approach that\\n> works for me.\\n> \\n\\nI've attached a new version (v7) that removes changes to support old \\nMSVC build system.\\n\\n\\n-- \\nNiyas\",\n \"msg_date\": \"Fri, 16 Dec 2022 10:52:23 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"\\nOn 16/12/2022 10:52, Niyas Sait wrote:\\n> On 12/12/2022 23:56, Michael Paquier wrote:\\n>> On Mon, Dec 12, 2022 at 01:38:37PM +0000, Niyas Sait wrote:\\n>>> On 05/12/2022 18:14, Andres Freund wrote:\\n>>> I think the old build system specific part is really minimal in the \\n>>> patch. I\\n>>> can strip out those if that's preferred.\\n>> Removing all the changes from src/tools/msvc/ is an approach that\\n>> works for me.\\n>>\\n> \\n> I've attached a new version (v7) that removes changes to support old \\n> MSVC build system.\\n\\n\\nHello,\\n\\nCan I get review for the latest patch please ?\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Tue, 17 Jan 2023 09:19:58 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-12-16 10:52:23 +0000, Niyas Sait wrote:\\n> Subject: [PATCH v7] Enable postgres native build for windows-arm64 platform\\n\\n> elif host_cpu == 'arm' or host_cpu == 'aarch64'\\n> \\n> - prog = '''\\n> + if cc.get_id() == 'msvc'\\n> + cdata.set('USE_ARMV8_CRC32C', false)\\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> + have_optimized_crc = true\\n\\nI dimly recall that windows might actually require the relevant extension on\\narm?\\n\\n\\n> + else\\n> + prog = '''\\n> #include <arm_acle.h>\\n\\nI'd just make this include #ifdef _MSV_VER (or whatever it is).\\n\\n\\n> int main(void)\\n> @@ -1960,18 +1966,19 @@ int main(void)\\n> }\\n> '''\\n> \\n> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n> - args: test_c_args)\\n> - # Use ARM CRC Extension unconditionally\\n> - cdata.set('USE_ARMV8_CRC32C', 1)\\n> - have_optimized_crc = true\\n> - elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\\n> - args: test_c_args + ['-march=armv8-a+crc'])\\n> - # Use ARM CRC Extension, with runtime check\\n> - cflags_crc += '-march=armv8-a+crc'\\n> - cdata.set('USE_ARMV8_CRC32C', false)\\n> - cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> - have_optimized_crc = true\\n> + if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n> + args: test_c_args)\\n\\nSeems like it'd be easier to read if you don't re-indent this, but just have\\nthe cc.get_id() == 'msvc' part of this if/else-if.\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 17 Jan 2023 14:51:31 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 17/01/2023 22:51, Andres Freund wrote:\\n> Hi,\\n> \\n> On 2022-12-16 10:52:23 +0000, Niyas Sait wrote:\\n>> Subject: [PATCH v7] Enable postgres native build for windows-arm64 platform\\n> \\n>> elif host_cpu == 'arm' or host_cpu == 'aarch64'\\n>> \\n>> - prog = '''\\n>> + if cc.get_id() == 'msvc'\\n>> + cdata.set('USE_ARMV8_CRC32C', false)\\n>> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n>> + have_optimized_crc = true\\n> \\n> I dimly recall that windows might actually require the relevant extension on\\n> arm?\\n\\nDo you mean we don't need the runtime checks for CRC ?\\n\\nCRC is an optional extension for ARMv8 base architecture. I am not sure \\nif windows make it mandatory to have this implementation.\\n\\n> \\n>> + else\\n>> + prog = '''\\n>> #include <arm_acle.h>\\n> \\n> I'd just make this include #ifdef _MSV_VER (or whatever it is).\\n\\nThe code snippet is not used for MSVC part. I am not sure why we need to \\nadd the #ifdef _MSC_VER.\\n\\n>> int main(void)\\n>> @@ -1960,18 +1966,19 @@ int main(void)\\n>> }\\n>> '''\\n>> \\n>> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n>> - args: test_c_args)\\n>> - # Use ARM CRC Extension unconditionally\\n>> - cdata.set('USE_ARMV8_CRC32C', 1)\\n>> - have_optimized_crc = true\\n>> - elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd with -march=armv8-a+crc',\\n>> - args: test_c_args + ['-march=armv8-a+crc'])\\n>> - # Use ARM CRC Extension, with runtime check\\n>> - cflags_crc += '-march=armv8-a+crc'\\n>> - cdata.set('USE_ARMV8_CRC32C', false)\\n>> - cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n>> - have_optimized_crc = true\\n>> + if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n>> + args: test_c_args)\\n> \\n> Seems like it'd be easier to read if you don't re-indent this, but just have\\n> the cc.get_id() == 'msvc' part of this if/else-if.\\n> \\n\\nYes that looks better. will do it in next patch.\\n\\nThanks Andres for the review.\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Thu, 19 Jan 2023 10:09:01 +0000\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 19/01/2023 10:09, Niyas Sait wrote:\\n> \\n> \\n> On 17/01/2023 22:51, Andres Freund wrote:\\n\\n>>> int main(void)\\n>>> @@ -1960,18 +1966,19 @@ int main(void)\\n>>> }\\n>>> '''\\n>>> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and \\n>>> __crc32cd without -march=armv8-a+crc',\\n>>> - args: test_c_args)\\n>>> - # Use ARM CRC Extension unconditionally\\n>>> - cdata.set('USE_ARMV8_CRC32C', 1)\\n>>> - have_optimized_crc = true\\n>>> - elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and \\n>>> __crc32cd with -march=armv8-a+crc',\\n>>> - args: test_c_args + ['-march=armv8-a+crc'])\\n>>> - # Use ARM CRC Extension, with runtime check\\n>>> - cflags_crc += '-march=armv8-a+crc'\\n>>> - cdata.set('USE_ARMV8_CRC32C', false)\\n>>> - cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n>>> - have_optimized_crc = true\\n>>> + if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and \\n>>> __crc32cd without -march=armv8-a+crc',\\n>>> + args: test_c_args)\\n>>\\n>> Seems like it'd be easier to read if you don't re-indent this, but \\n>> just have\\n>> the cc.get_id() == 'msvc' part of this if/else-if.\\n>>\\n> \\n\\nI've attached a new version (v8) to fix the above indentation issue.\\n\\nCould someone please help with the review ?\\n\\n-- \\nNiyas\",\n \"msg_date\": \"Tue, 2 May 2023 08:51:09 +0100\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, May 02, 2023 at 08:51:09AM +0100, Niyas Sait wrote:\\n> I've attached a new version (v8) to fix the above indentation issue.\\n> \\n> Could someone please help with the review ?\\n\\nSeeing how simple this has become, I would be really tempted to get\\nthat applied, especially if there's a buildfarm machine able to follow\\nup.. On the one hand, we are in a stability period for v16, so it may\\nnot be the best moment ever. On the other hand, waiting one more year\\nsounds like a waste for a patch that just adds new code paths.\\n\\nRMT members, any thoughts?\\n--\\nMichael\",\n \"msg_date\": \"Sat, 6 May 2023 12:09:22 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> Seeing how simple this has become, I would be really tempted to get\\n> that applied, especially if there's a buildfarm machine able to follow\\n> up.. On the one hand, we are in a stability period for v16, so it may\\n> not be the best moment ever. On the other hand, waiting one more year\\n> sounds like a waste for a patch that just adds new code paths.\\n\\nIndeed, there's not much in this patch ... but it's impossible to see\\nhow \\\"run on an entirely new platform\\\" isn't a new feature. Moreover,\\nit's a platform that very few of us will have any ability to support\\nor debug for. I can't see how it's a good idea to shove this in\\npost-feature-freeze. Let's plan to push it in when v17 opens.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 06 May 2023 00:35:40 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Sat, May 06, 2023 at 12:35:40AM -0400, Tom Lane wrote:\\n> Indeed, there's not much in this patch ... but it's impossible to see\\n> how \\\"run on an entirely new platform\\\" isn't a new feature. Moreover,\\n> it's a platform that very few of us will have any ability to support\\n> or debug for. I can't see how it's a good idea to shove this in\\n> post-feature-freeze. Let's plan to push it in when v17 opens.\\n\\nFine by me. I have added an entry in the CF app to remember about\\nthis stuff as there was nothing present yet:\\nhttps://commitfest.postgresql.org/43/4309/\\n--\\nMichael\",\n \"msg_date\": \"Sun, 7 May 2023 09:37:50 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2023-05-06 00:35:40 -0400, Tom Lane wrote:\\n> Michael Paquier <michael@paquier.xyz> writes:\\n> > Seeing how simple this has become, I would be really tempted to get\\n> > that applied, especially if there's a buildfarm machine able to follow\\n> > up.. On the one hand, we are in a stability period for v16, so it may\\n> > not be the best moment ever. On the other hand, waiting one more year\\n> > sounds like a waste for a patch that just adds new code paths.\\n> \\n> Indeed, there's not much in this patch ... but it's impossible to see\\n> how \\\"run on an entirely new platform\\\" isn't a new feature. Moreover,\\n> it's a platform that very few of us will have any ability to support\\n> or debug for. I can't see how it's a good idea to shove this in\\n> post-feature-freeze. Let's plan to push it in when v17 opens.\\n\\nI don't really have feelings either way - but haven't we gone further and even\\nbackpatched things like spinlock support for new arches in the past?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 May 2023 12:17:31 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> I don't really have feelings either way - but haven't we gone further and even\\n> backpatched things like spinlock support for new arches in the past?\\n\\nMmmm ... don't really think those cases were comparable. We weren't\\nadding support for a whole new OS. Now, you might argue that Windows\\non arm64 will be just like Windows on x86_64, but I think the jury\\nis still out on that. Microsoft was so Intel-only for so many years\\nthat I bet they've had to change quite a bit to make it go on ARM.\\n\\nAlso, the cases of back-patched spinlock support that I can find\\nin the last few years were pretty low-risk. I'll grant that\\nc32fcac56 was a bit blue-sky because hardly anybody had RISC-V\\nat that point, but by the same token anybody relying on it at the\\ntime would be dealing with a beta-grade OS too. On the other hand,\\n1c72d82c2 was immediately testable in the buildfarm, and f3bd00c01\\nwas importing code already verified by our OpenBSD packagers.\\n\\nAs I said upthread, this seems like something to put in at the\\nbeginning of a dev cycle, not post-feature-freeze.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 08 May 2023 15:58:41 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2023-05-08 Mo 15:58, Tom Lane wrote:\\n> Andres Freund<andres@anarazel.de> writes:\\n>> I don't really have feelings either way - but haven't we gone further and even\\n>> backpatched things like spinlock support for new arches in the past?\\n> Mmmm ... don't really think those cases were comparable. We weren't\\n> adding support for a whole new OS. Now, you might argue that Windows\\n> on arm64 will be just like Windows on x86_64, but I think the jury\\n> is still out on that. Microsoft was so Intel-only for so many years\\n> that I bet they've had to change quite a bit to make it go on ARM.\\n>\\n> Also, the cases of back-patched spinlock support that I can find\\n> in the last few years were pretty low-risk. I'll grant that\\n> c32fcac56 was a bit blue-sky because hardly anybody had RISC-V\\n> at that point, but by the same token anybody relying on it at the\\n> time would be dealing with a beta-grade OS too. On the other hand,\\n> 1c72d82c2 was immediately testable in the buildfarm, and f3bd00c01\\n> was importing code already verified by our OpenBSD packagers.\\n>\\n> As I said upthread, this seems like something to put in at the\\n> beginning of a dev cycle, not post-feature-freeze.\\n> \\t\\t\\t\\n\\n\\nWe will definitely want buildfarm support. I don't have such a machine \\nand am not likely to have one any time soon. I do run drongo and \\nfairywren on an EC2 instance, but AWS doesn't seem to have support for \\nWindows on ARM. Maybe it's available on Azure (maybe then some of our \\ncolleagues working at Microsoft could arrange such a beast for me to set \\nup potential buildfarm animal, or else run one themselves.)\\n\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2023-05-08 Mo 15:58, Tom Lane wrote:\\n\\n\\nAndres Freund <andres@anarazel.de> writes:\\n\\n\\nI don't really have feelings either way - but haven't we gone further and even\\nbackpatched things like spinlock support for new arches in the past?\\n\\n\\n\\nMmmm ... don't really think those cases were comparable. We weren't\\nadding support for a whole new OS. Now, you might argue that Windows\\non arm64 will be just like Windows on x86_64, but I think the jury\\nis still out on that. Microsoft was so Intel-only for so many years\\nthat I bet they've had to change quite a bit to make it go on ARM.\\n\\nAlso, the cases of back-patched spinlock support that I can find\\nin the last few years were pretty low-risk. I'll grant that\\nc32fcac56 was a bit blue-sky because hardly anybody had RISC-V\\nat that point, but by the same token anybody relying on it at the\\ntime would be dealing with a beta-grade OS too. On the other hand,\\n1c72d82c2 was immediately testable in the buildfarm, and f3bd00c01\\nwas importing code already verified by our OpenBSD packagers.\\n\\nAs I said upthread, this seems like something to put in at the\\nbeginning of a dev cycle, not post-feature-freeze.\\n\\t\\t\\t\\n\\n\\n\\nWe will definitely want buildfarm support. I don't have such a\\n machine and am not likely to have one any time soon. I do run\\n drongo and fairywren on an EC2 instance, but AWS doesn't seem to\\n have support for Windows on ARM. Maybe it's available on Azure\\n (maybe then some of our colleagues working at Microsoft could\\n arrange such a beast for me to set up potential buildfarm animal,\\n or else run one themselves.)\\n\\n\\n\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Wed, 10 May 2023 09:24:39 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, May 10, 2023 at 09:24:39AM -0400, Andrew Dunstan wrote:\\n> We will definitely want buildfarm support. I don't have such a machine and\\n> am not likely to have one any time soon. I do run drongo and fairywren on an\\n> EC2 instance, but AWS doesn't seem to have support for Windows on ARM. Maybe\\n> it's available on Azure (maybe then some of our colleagues working at\\n> Microsoft could arrange such a beast for me to set up potential buildfarm\\n> animal, or else run one themselves.)\\n\\nUnfortunately AWS does not offer ARM on Windows with an EC2, that's\\nsomething I have also looked at setting up :/\\n\\nAnyway, Niyat has mentioned me that he was working on that, but it\\nseems that it was off-list. Niyat, could you confirm this part?\\n--\\nMichael\",\n \"msg_date\": \"Thu, 11 May 2023 08:34:29 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, May 11, 2023 at 11:34 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> On Wed, May 10, 2023 at 09:24:39AM -0400, Andrew Dunstan wrote:\\n> > We will definitely want buildfarm support. I don't have such a machine and\\n> > am not likely to have one any time soon. I do run drongo and fairywren on an\\n> > EC2 instance, but AWS doesn't seem to have support for Windows on ARM. Maybe\\n> > it's available on Azure (maybe then some of our colleagues working at\\n> > Microsoft could arrange such a beast for me to set up potential buildfarm\\n> > animal, or else run one themselves.)\\n>\\n> Unfortunately AWS does not offer ARM on Windows with an EC2, that's\\n> something I have also looked at setting up :/\\n\\nAzure does have an image \\\"Microsoft Windows 11 Preview arm64\\\" to run\\non Ampere CPUs, which says it's a pre-release version intended for\\nvalidation, which sounds approximately like what we want. I will try\\nto find out more.\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 May 2023 12:17:25 +1200\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, May 11, 2023 at 12:17:25PM +1200, Thomas Munro wrote:\\n> Azure does have an image \\\"Microsoft Windows 11 Preview arm64\\\" to run\\n> on Ampere CPUs, which says it's a pre-release version intended for\\n> validation, which sounds approximately like what we want. I will try\\n> to find out more.\\n\\nInteresting. Thanks for mentioning it.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 11 May 2023 13:19:54 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 11/05/2023 00:34, Michael Paquier wrote:\\n> On Wed, May 10, 2023 at 09:24:39AM -0400, Andrew Dunstan wrote:\\n>> We will definitely want buildfarm support. I don't have such a machine and\\n>> am not likely to have one any time soon. I do run drongo and fairywren on an\\n>> EC2 instance, but AWS doesn't seem to have support for Windows on ARM. Maybe\\n>> it's available on Azure (maybe then some of our colleagues working at\\n>> Microsoft could arrange such a beast for me to set up potential buildfarm\\n>> animal, or else run one themselves.)\\n> Unfortunately AWS does not offer ARM on Windows with an EC2, that's\\n> something I have also looked at setting up :/\\n> \\n> Anyway, Niyat has mentioned me that he was working on that, but it\\n> seems that it was off-list. Niyat, could you confirm this part?\\n> --\\n> Michael\\n\\nIf we have Azure VM credits, we could use that to deploy buildfarm machines.\\n\\nOtherwise, I can try to allocate a machine from Linaro Windows Lab for \\nthe buildbot.\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 May 2023 08:57:57 +0100\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hello,\\n\\nJust wanted to give a heads up that I am moving to a new role outside \\nLinaro and as a result wouldn't be able to continue contributing.\\n\\nAnthony Roberts from Linaro is going to support the enablement work.\\n\\n-- \\nNiyas\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Jun 2023 13:25:17 +0100\",\n \"msg_from\": \"Niyas Sait <niyas.sait@linaro.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, May 11, 2023 at 01:19:54PM +0900, Michael Paquier wrote:\\n> On Thu, May 11, 2023 at 12:17:25PM +1200, Thomas Munro wrote:\\n>> Azure does have an image \\\"Microsoft Windows 11 Preview arm64\\\" to run\\n>> on Ampere CPUs, which says it's a pre-release version intended for\\n>> validation, which sounds approximately like what we want. I will try\\n>> to find out more.\\n> \\n> Interesting. Thanks for mentioning it.\\n\\nNow that v17 is open, I was looking at v8 posted at [1] and I don't\\nhave much more to add about it. The problem about the lack of\\nbuildfarm machine is still annoying, but I don't see a reason not to\\nmove forward and let people play with this stuff on HEAD. At least\\nthat would be progress. Any thoughts?\\n\\nThomas, what's the state of ARM support for Windows on Azure? Is that\\nstill in preview?\\n\\n[1]: https://www.postgresql.org/message-id/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56%40linaro.org\\n--\\nMichael\",\n \"msg_date\": \"Tue, 4 Jul 2023 16:34:36 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi All,\\n\\nJust following up on this, has there been any movement?\\n\\nI did see another message go into the thread here with no reply: \\nhttps://www.postgresql.org/message-id/ZKPLjLjIjwN2lxkg%40paquier.xyz\\n\\nThanks,\\nAnthony\\n\\nOn 14/06/2023 13:25, Niyas Sait wrote:\\n> Hello,\\n>\\n> Just wanted to give a heads up that I am moving to a new role outside \\n> Linaro and as a result wouldn't be able to continue contributing.\\n>\\n> Anthony Roberts from Linaro is going to support the enablement work.\\n>\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Aug 2023 09:41:44 +0100\",\n \"msg_from\": \"Anthony Roberts <anthony.roberts@linaro.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 17, 2023 at 09:41:44AM +0100, Anthony Roberts wrote:\\n> Just following up on this, has there been any movement?\\n> \\n> I did see another message go into the thread here with no reply:\\n> https://www.postgresql.org/message-id/ZKPLjLjIjwN2lxkg%40paquier.xyz\\n\\nI don't have an environment to test the patch, but I don't object to\\nit per se. However, I don't really want to move forward completely\\nblindly as well in the long-term.\\n\\nAs mentioned to Niyas, could it be possible to provide to the\\ncommunity a buildfarm machine that would be able to test this\\nenvironment? I am not sure what's your status on that. Perhaps this\\nis already set up and you are just waiting for the patch to be merged?\\n--\\nMichael\",\n \"msg_date\": \"Fri, 18 Aug 2023 07:28:04 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nWhich of these are you looking for us to provide:\\n\\n* A machine for you to directly access (via a VPN)\\n* A machine we just run a worker script on that automatically picks up \\nthe builds as required\\n* Us to do downstream CI\\n\\nAll are possible, but preferably not option 1, as it would mean straight \\nup pulling out a machine from our build farm, and it has to go through \\nall sorts of approvals internally. If it's the only way forward I can \\nkick it up the chain though.\\n\\nOption 2 and 3 are ones we do for various other projects (ie. 2 - CMake, \\n3 - OpenSSL)\\n\\nThanks,\\nAnthony\\n\\nOn 17/08/2023 23:28, Michael Paquier wrote:\\n> On Thu, Aug 17, 2023 at 09:41:44AM +0100, Anthony Roberts wrote:\\n>> Just following up on this, has there been any movement?\\n>>\\n>> I did see another message go into the thread here with no reply:\\n>> https://www.postgresql.org/message-id/ZKPLjLjIjwN2lxkg%40paquier.xyz\\n> I don't have an environment to test the patch, but I don't object to\\n> it per se. However, I don't really want to move forward completely\\n> blindly as well in the long-term.\\n>\\n> As mentioned to Niyas, could it be possible to provide to the\\n> community a buildfarm machine that would be able to test this\\n> environment? I am not sure what's your status on that. Perhaps this\\n> is already set up and you are just waiting for the patch to be merged?\\n> --\\n> Michael\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Aug 2023 10:40:39 +0100\",\n \"msg_from\": \"Anthony Roberts <anthony.roberts@linaro.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 25, 2023 at 10:40:39AM +0100, Anthony Roberts wrote:\\n> Which of these are you looking for us to provide:\\n> \\n> * A machine for you to directly access (via a VPN)\\n> * A machine we just run a worker script on that automatically picks up the\\n> builds as required\\n> * Us to do downstream CI\\n> \\n> All are possible, but preferably not option 1, as it would mean straight up\\n> pulling out a machine from our build farm, and it has to go through all\\n> sorts of approvals internally. If it's the only way forward I can kick it up\\n> the chain though.\\n> \\n> Option 2 and 3 are ones we do for various other projects (ie. 2 - CMake, 3 -\\n> OpenSSL)\\n\\nThe community has its own CI facility. Here is a link of how to set\\nup a machine:\\nhttps://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto\\n\\nAnd all the results are published by the maintainers of the machines\\non a periodic basis where the buildfarm client code is set:\\nhttps://buildfarm.postgresql.org/cgi-bin/show_status.pl\\n\\nThe final result would be for you to maintain the machine so as we are\\nable to see if anything breaks with it. Adding extra dependencies to\\nthe build like OpenSSL would be nice, but based on the contents of the\\npatch to add this port that does not seem mandatory to me, either.\\n\\nOnce the machine is ready, you will need to request a buildfarm\\nmachine name and a key to be able to begin publishing the reports of\\nthe builds to the community buildfarm. Once the machine is ready to\\ngo, just let me know and I'd be OK to merge the patch (I still want to\\ndo a final review of it in case I've missed something, but I can move\\non with that before the buildfarm machine is set up).\\n--\\nMichael\",\n \"msg_date\": \"Fri, 25 Aug 2023 19:17:27 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Thanks for the link - that looks basically the same as what we do for \\nCMake (option 2), so should be relatively easy.\\n\\nI will have a chat to relevant people about setting a machine up \\nproperly for it.\\n\\nThanks,\\nAnthony\\n\\nOn 25/08/2023 11:17, Michael Paquier wrote:\\n> On Fri, Aug 25, 2023 at 10:40:39AM +0100, Anthony Roberts wrote:\\n>> Which of these are you looking for us to provide:\\n>>\\n>> * A machine for you to directly access (via a VPN)\\n>> * A machine we just run a worker script on that automatically picks up the\\n>> builds as required\\n>> * Us to do downstream CI\\n>>\\n>> All are possible, but preferably not option 1, as it would mean straight up\\n>> pulling out a machine from our build farm, and it has to go through all\\n>> sorts of approvals internally. If it's the only way forward I can kick it up\\n>> the chain though.\\n>>\\n>> Option 2 and 3 are ones we do for various other projects (ie. 2 - CMake, 3 -\\n>> OpenSSL)\\n> The community has its own CI facility. Here is a link of how to set\\n> up a machine:\\n> https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto\\n>\\n> And all the results are published by the maintainers of the machines\\n> on a periodic basis where the buildfarm client code is set:\\n> https://buildfarm.postgresql.org/cgi-bin/show_status.pl\\n>\\n> The final result would be for you to maintain the machine so as we are\\n> able to see if anything breaks with it. Adding extra dependencies to\\n> the build like OpenSSL would be nice, but based on the contents of the\\n> patch to add this port that does not seem mandatory to me, either.\\n>\\n> Once the machine is ready, you will need to request a buildfarm\\n> machine name and a key to be able to begin publishing the reports of\\n> the builds to the community buildfarm. Once the machine is ready to\\n> go, just let me know and I'd be OK to merge the patch (I still want to\\n> do a final review of it in case I've missed something, but I can move\\n> on with that before the buildfarm machine is set up).\\n> --\\n> Michael\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Aug 2023 11:34:18 +0100\",\n \"msg_from\": \"Anthony Roberts <anthony.roberts@linaro.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"> On 25 Aug 2023, at 11:40, Anthony Roberts <anthony.roberts@linaro.org> wrote:\\n\\n> * A machine for you to directly access (via a VPN)\\n\\nThat could be very helpful to provide to developers who are trying to debug\\nvery wicked bugs which requires access for analysis. Such things are on a case\\nby case basis of course at your discretion, but it would most likely be very\\nappreciated when/if the need exists.\\n\\n> * A machine we just run a worker script on that automatically picks up the builds as required\\n\\nThis sounds like what we have for the buildfarm, where machines independently\\nbuilds the branches from the Git repo by using the buildfarm client as worker.\\n\\n> * Us to do downstream CI\\n\\nThis sounds like the CFBot, where we (currently) use Cirrus with our own\\nworkers for building and testing the patches which are being worked on.\\n\\n\\thttp://cfbot.cputube.org/\\n\\nBoth the two latter options are very helpful for the community, and serve two\\ndistincly different purposes. Option 2 is required for a platform to be\\nconsidered supported whereas option 3 assists developers in being proactive\\nrather than just reactive.\\n\\n--\\nDaniel Gustafsson\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Aug 2023 13:48:21 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nJust a follow-up: having spoken to the relevant people, sadly, right now,\\nwe do not have the capacity to be pulling machines out of our day-to-day CI\\ntasks to dedicate to specific projects.\\n\\nI can, however, offer a few alternatives:\\n\\n1. If you have capacity on your linux x64 machines, it is possible to\\nemulate WoA, and run tests that way:\\nhttps://www.linaro.org/blog/emulate-windows-on-arm/\\nThis is something that works well for projects that we have contributed it\\nto (ie, sse2neon uses it via github actions) - if you run into any issues\\nwith this, we are able to try and fix them on our side, as likely more than\\npostgres would benefit.\\n\\n2. If there is any scope for it on your side at all, a Samsung Galaxy Book\\nGo 14\\\" can be had from ebay for ~£140.\\nUnderstandably, this might not be an option for you.\\n\\nIf it helps, if this is merged, we do have the possibility of running our\\nown downstream CI builds nightly for the tip of your main branch here:\\nhttps://gitlab.com/Linaro/windowsonarm/nightly\\nThis is something we do for a few projects that do not have full upstream\\nCI.\\n\\nThanks,\\nAnthony\\n\\n\\nOn Fri, 25 Aug 2023 at 11:34, Anthony Roberts <anthony.roberts@linaro.org>\\nwrote:\\n\\n> Thanks for the link - that looks basically the same as what we do for\\n> CMake (option 2), so should be relatively easy.\\n>\\n> I will have a chat to relevant people about setting a machine up\\n> properly for it.\\n>\\n> Thanks,\\n> Anthony\\n>\\n> On 25/08/2023 11:17, Michael Paquier wrote:\\n> > On Fri, Aug 25, 2023 at 10:40:39AM +0100, Anthony Roberts wrote:\\n> >> Which of these are you looking for us to provide:\\n> >>\\n> >> * A machine for you to directly access (via a VPN)\\n> >> * A machine we just run a worker script on that automatically picks up\\n> the\\n> >> builds as required\\n> >> * Us to do downstream CI\\n> >>\\n> >> All are possible, but preferably not option 1, as it would mean\\n> straight up\\n> >> pulling out a machine from our build farm, and it has to go through all\\n> >> sorts of approvals internally. If it's the only way forward I can kick\\n> it up\\n> >> the chain though.\\n> >>\\n> >> Option 2 and 3 are ones we do for various other projects (ie. 2 -\\n> CMake, 3 -\\n> >> OpenSSL)\\n> > The community has its own CI facility. Here is a link of how to set\\n> > up a machine:\\n> > https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto\\n> >\\n> > And all the results are published by the maintainers of the machines\\n> > on a periodic basis where the buildfarm client code is set:\\n> > https://buildfarm.postgresql.org/cgi-bin/show_status.pl\\n> >\\n> > The final result would be for you to maintain the machine so as we are\\n> > able to see if anything breaks with it. Adding extra dependencies to\\n> > the build like OpenSSL would be nice, but based on the contents of the\\n> > patch to add this port that does not seem mandatory to me, either.\\n> >\\n> > Once the machine is ready, you will need to request a buildfarm\\n> > machine name and a key to be able to begin publishing the reports of\\n> > the builds to the community buildfarm. Once the machine is ready to\\n> > go, just let me know and I'd be OK to merge the patch (I still want to\\n> > do a final review of it in case I've missed something, but I can move\\n> > on with that before the buildfarm machine is set up).\\n> > --\\n> > Michael\\n>\\n\\nHi,Just a follow-up: having spoken to the relevant people, sadly, right now, we do not have the capacity to be pulling machines out of our day-to-day CI tasks to dedicate to specific projects.I can, however, offer a few alternatives:1. If you have capacity on your linux x64 machines, it is possible to emulate WoA, and run tests that way: https://www.linaro.org/blog/emulate-windows-on-arm/This is something that works well for projects that we have contributed it to (ie, sse2neon uses it via github actions) - if you run into any issues with this, we are able to try and fix them on our side, as likely more than postgres would benefit.2. If there is any scope for it on your side at all, a Samsung Galaxy Book Go 14\\\" can be had from ebay for ~£140.Understandably, this might not be an option for you.If it helps, if this is merged, we do have the possibility of running our own downstream CI builds nightly for the tip of your main branch here: https://gitlab.com/Linaro/windowsonarm/nightlyThis is something we do for a few projects that do not have full upstream CI.Thanks,AnthonyOn Fri, 25 Aug 2023 at 11:34, Anthony Roberts <anthony.roberts@linaro.org> wrote:Thanks for the link - that looks basically the same as what we do for \\nCMake (option 2), so should be relatively easy.\\n\\nI will have a chat to relevant people about setting a machine up \\nproperly for it.\\n\\nThanks,\\nAnthony\\n\\nOn 25/08/2023 11:17, Michael Paquier wrote:\\n> On Fri, Aug 25, 2023 at 10:40:39AM +0100, Anthony Roberts wrote:\\n>> Which of these are you looking for us to provide:\\n>>\\n>> * A machine for you to directly access (via a VPN)\\n>> * A machine we just run a worker script on that automatically picks up the\\n>> builds as required\\n>> * Us to do downstream CI\\n>>\\n>> All are possible, but preferably not option 1, as it would mean straight up\\n>> pulling out a machine from our build farm, and it has to go through all\\n>> sorts of approvals internally. If it's the only way forward I can kick it up\\n>> the chain though.\\n>>\\n>> Option 2 and 3 are ones we do for various other projects (ie. 2 - CMake, 3 -\\n>> OpenSSL)\\n> The community has its own CI facility. Here is a link of how to set\\n> up a machine:\\n> https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto\\n>\\n> And all the results are published by the maintainers of the machines\\n> on a periodic basis where the buildfarm client code is set:\\n> https://buildfarm.postgresql.org/cgi-bin/show_status.pl\\n>\\n> The final result would be for you to maintain the machine so as we are\\n> able to see if anything breaks with it. Adding extra dependencies to\\n> the build like OpenSSL would be nice, but based on the contents of the\\n> patch to add this port that does not seem mandatory to me, either.\\n>\\n> Once the machine is ready, you will need to request a buildfarm\\n> machine name and a key to be able to begin publishing the reports of\\n> the builds to the community buildfarm. Once the machine is ready to\\n> go, just let me know and I'd be OK to merge the patch (I still want to\\n> do a final review of it in case I've missed something, but I can move\\n> on with that before the buildfarm machine is set up).\\n> --\\n> Michael\",\n \"msg_date\": \"Wed, 30 Aug 2023 16:12:16 +0100\",\n \"msg_from\": \"Anthony Roberts <anthony.roberts@linaro.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 30, 2023 at 04:12:16PM +0100, Anthony Roberts wrote:\\n> Just a follow-up: having spoken to the relevant people, sadly, right now,\\n> we do not have the capacity to be pulling machines out of our day-to-day CI\\n> tasks to dedicate to specific projects.\\n\\nOkay, thanks for letting us know.\\n\\n> I can, however, offer a few alternatives:\\n> \\n> 1. If you have capacity on your linux x64 machines, it is possible to\\n> emulate WoA, and run tests that way:\\n> https://www.linaro.org/blog/emulate-windows-on-arm/\\n> This is something that works well for projects that we have contributed it\\n> to (ie, sse2neon uses it via github actions) - if you run into any issues\\n> with this, we are able to try and fix them on our side, as likely more than\\n> postgres would benefit.\\n> \\n> 2. If there is any scope for it on your side at all, a Samsung Galaxy Book\\n> Go 14\\\" can be had from ebay for ~£140.\\n> Understandably, this might not be an option for you.\\n> \\n> If it helps, if this is merged, we do have the possibility of running our\\n> own downstream CI builds nightly for the tip of your main branch here:\\n> https://gitlab.com/Linaro/windowsonarm/nightly\\n> This is something we do for a few projects that do not have full upstream\\n> CI.\\n\\nHonestly, I am not sure how to proceed here. Having something in the\\nbuildfarm is necessary IMO, because this is the central place that\\ncommunity members look at when it comes to monitoring the stability of\\na patch committed.\\n\\nAny opinions from the others?\\n--\\nMichael\",\n \"msg_date\": \"Thu, 31 Aug 2023 13:22:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Wed, Aug 30, 2023 at 04:12:16PM +0100, Anthony Roberts wrote:\\n>> Just a follow-up: having spoken to the relevant people, sadly, right now,\\n>> we do not have the capacity to be pulling machines out of our day-to-day CI\\n>> tasks to dedicate to specific projects.\\n\\n> Okay, thanks for letting us know.\\n\\nToo bad.\\n\\n> Honestly, I am not sure how to proceed here. Having something in the\\n> buildfarm is necessary IMO, because this is the central place that\\n> community members look at when it comes to monitoring the stability of\\n> a patch committed.\\n> Any opinions from the others?\\n\\nI agree. I'm really uncomfortable with claiming support for\\nWindows-on-ARM if we don't have a buildfarm member testing it.\\nFor other platforms that have a track record of multiple\\nhardware support, it might not be a stretch ... but Windows was\\nso resolutely Intel-only for so long that \\\"it works on ARM\\\" is\\na proposition that I won't trust without hard evidence. There\\nare too many bits of that system that might not have gotten the\\nword yet, or at least not gotten sufficient testing.\\n\\nMy vote for this is we don't commit without a buildfarm member.\\n\\n(As a comparison point, I'd still be unsure about our support\\nfor macOS-on-ARM if we didn't have buildfarm support for that.\\nThat exists, and is being paid for out of my own pocket.\\nI do not have much sympathy for a corporation that claims\\nit can't afford to support one measly buildfarm animal.)\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Aug 2023 00:44:07 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 31, 2023 at 4:22 PM Michael Paquier <michael@paquier.xyz> wrote:\\n> Honestly, I am not sure how to proceed here. Having something in the\\n> buildfarm is necessary IMO, because this is the central place that\\n> community members look at when it comes to monitoring the stability of\\n> a patch committed.\\n\\nIt's tricky, because usually the org that wants the thing to work\\nsupplies the thing, and I personally think that provides a meaningful\\n\\\"demand\\\" signal too. For example, we have external Solaris\\nrepresentation, but no users of AIX represented at this point, and I\\nconsider that to be quite interesting information. Plenty of people\\nwant PG to work well on their OS or architecture.\\n\\nBut, supposing Azure credits were suddenly available (just between\\nyou, me and the mailing list, I heard that this should be imminent as\\nsome kind of grant to PG US, as a result of the Cirrus thing), do we\\nhave a Windows-savvy volunteer to feed and water an animal? I feel\\nlike this case is inevitable anyway because people in our own\\ncommunity are probably going to be using Windows ARM laptops sooner or\\nlater...\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Aug 2023 16:57:05 +1200\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Thomas Munro <thomas.munro@gmail.com> writes:\\n> But, supposing Azure credits were suddenly available (just between\\n> you, me and the mailing list, I heard that this should be imminent as\\n> some kind of grant to PG US, as a result of the Cirrus thing), do we\\n> have a Windows-savvy volunteer to feed and water an animal? I feel\\n> like this case is inevitable anyway because people in our own\\n> community are probably going to be using Windows ARM laptops sooner or\\n> later...\\n\\nThat's a good point. The hardware resources to support a buildfarm\\nanimal are really pretty trivial these days. What is important is\\nan animal owner who is willing to help chase down problems when they\\narise.\\n\\nMaybe we just need to wait for some community members to acquire\\nthose laptops.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Aug 2023 01:06:53 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 31, 2023 at 01:06:53AM -0400, Tom Lane wrote:\\n> That's a good point. The hardware resources to support a buildfarm\\n> animal are really pretty trivial these days. What is important is\\n> an animal owner who is willing to help chase down problems when they\\n> arise.\\n\\nUnfortunately I don't see myself running that at home. It is too hot\\nand humid in Summer so the machine would not last. Something in the\\ncloud would be OK for me, but AWS has no option for a Windows host\\nwith ARM yet.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 31 Aug 2023 14:33:56 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 02.05.23 09:51, Niyas Sait wrote:\\n> diff --git a/doc/src/sgml/install-windows.sgml b/doc/src/sgml/install-windows.sgml\\n> index cbc70a039c..2ecd5fcf38 100644\\n> --- a/doc/src/sgml/install-windows.sgml\\n> +++ b/doc/src/sgml/install-windows.sgml\\n> @@ -352,7 +352,8 @@ $ENV{MSBFLAGS}=\\\"/m\\\";\\n> <title>Special Considerations for 64-Bit Windows</title>\\n> \\n> <para>\\n> - PostgreSQL will only build for the x64 architecture on 64-bit Windows.\\n> + PostgreSQL will only build for the x64 and ARM64 architectures on 64-bit\\n> + Windows.\\n> </para>\\n\\nAre there other possible architectures besides x64 and ARM64? Itanium? \\nOr should we just delete this sentence?\\n\\n> diff --git a/meson.build b/meson.build\\n> index 096044628c..6cca212fae 100644\\n> --- a/meson.build\\n> +++ b/meson.build\\n> @@ -2045,8 +2045,11 @@ int main(void)\\n> elif host_cpu == 'arm' or host_cpu == 'aarch64'\\n> \\n> prog = '''\\n> +#ifdef _MSC_VER\\n> +#include <intrin.h>\\n> +#else\\n> #include <arm_acle.h>\\n> -\\n> +#endif\\n\\nMaybe keep the whitespace here.\\n\\n> - if cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n> + if cc.get_id() == 'msvc'\\n> + cdata.set('USE_ARMV8_CRC32C', false)\\n> + cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> + have_optimized_crc = true\\n> + elif cc.links(prog, name: '__crc32cb, __crc32ch, __crc32cw, and __crc32cd without -march=armv8-a+crc',\\n\\nAdd a comment here.\\n\\n> diff --git a/src/port/pg_crc32c_armv8.c b/src/port/pg_crc32c_armv8.c\\n> index d8fae510cf..3d7eb748ff 100644\\n> --- a/src/port/pg_crc32c_armv8.c\\n> +++ b/src/port/pg_crc32c_armv8.c\\n> @@ -14,7 +14,9 @@\\n> */\\n> #include \\\"c.h\\\"\\n> \\n> +#ifndef _MSC_VER\\n> #include <arm_acle.h>\\n> +#endif\\n\\nAlso add a comment here.\\n\\n> \\n> #include \\\"port/pg_crc32c.h\\\"\\n> \\n> diff --git a/src/tools/msvc/gendef.pl b/src/tools/msvc/gendef.pl\\n> index e7cbefcbc3..934dc17b96 100644\\n> --- a/src/tools/msvc/gendef.pl\\n> +++ b/src/tools/msvc/gendef.pl\\n> @@ -120,9 +120,9 @@ sub writedef\\n> \\t{\\n> \\t\\tmy $isdata = $def->{$f} eq 'data';\\n> \\n> -\\t\\t# Strip the leading underscore for win32, but not x64\\n> +\\t\\t# Strip the leading underscore for win32, but not x64 and aarch64\\n\\nIs x64 the opposite of win32? Does this make sense? Should we reverse \\nthe logic here and single out the one variant where the stripping is \\nnecessary?\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Sep 2023 21:09:12 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter@eisentraut.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 31.08.23 06:44, Tom Lane wrote:\\n> I agree. I'm really uncomfortable with claiming support for\\n> Windows-on-ARM if we don't have a buildfarm member testing it.\\n> For other platforms that have a track record of multiple\\n> hardware support, it might not be a stretch ... but Windows was\\n> so resolutely Intel-only for so long that \\\"it works on ARM\\\" is\\n> a proposition that I won't trust without hard evidence. There\\n> are too many bits of that system that might not have gotten the\\n> word yet, or at least not gotten sufficient testing.\\n> \\n> My vote for this is we don't commit without a buildfarm member.\\n\\nI think we can have a multi-tiered approach, where we can commit support \\nbut consider it experimental until we have buildfarm coverage.\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Sep 2023 21:12:46 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter@eisentraut.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2023-08-31 14:33:56 +0900, Michael Paquier wrote:\\n> On Thu, Aug 31, 2023 at 01:06:53AM -0400, Tom Lane wrote:\\n> > That's a good point. The hardware resources to support a buildfarm\\n> > animal are really pretty trivial these days. What is important is\\n> > an animal owner who is willing to help chase down problems when they\\n> > arise.\\n> \\n> Unfortunately I don't see myself running that at home. It is too hot\\n> and humid in Summer so the machine would not last. Something in the\\n> cloud would be OK for me, but AWS has no option for a Windows host\\n> with ARM yet.\\n\\nFWIW, we could run ARM windows as part of CI. It'd be a bit of scripting work,\\nas google doesn't yet have readymade VM images with windows for ARM. Right now\\nthe CI VM images for windows are based on top of the google images. Scripting\\ngeneration of ARM images would be some one-off work, but after that...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Sep 2023 12:32:46 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"> On 13 Sep 2023, at 21:12, Peter Eisentraut <peter@eisentraut.org> wrote:\\n> \\n> On 31.08.23 06:44, Tom Lane wrote:\\n>> I agree. I'm really uncomfortable with claiming support for\\n>> Windows-on-ARM if we don't have a buildfarm member testing it.\\n>> For other platforms that have a track record of multiple\\n>> hardware support, it might not be a stretch ... but Windows was\\n>> so resolutely Intel-only for so long that \\\"it works on ARM\\\" is\\n>> a proposition that I won't trust without hard evidence. There\\n>> are too many bits of that system that might not have gotten the\\n>> word yet, or at least not gotten sufficient testing.\\n>> My vote for this is we don't commit without a buildfarm member.\\n> \\n> I think we can have a multi-tiered approach, where we can commit support but consider it experimental until we have buildfarm coverage.\\n\\nIf it's experimental it should probably be behind an opt-in flag in\\nautoconf/meson, or be reverted by the time REL_17_STABLE branches unless\\ncoverage has materialized by then.\\n\\n--\\nDaniel Gustafsson\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 14 Sep 2023 11:39:20 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 14.09.23 11:39, Daniel Gustafsson wrote:\\n>> On 13 Sep 2023, at 21:12, Peter Eisentraut <peter@eisentraut.org> wrote:\\n>>\\n>> On 31.08.23 06:44, Tom Lane wrote:\\n>>> I agree. I'm really uncomfortable with claiming support for\\n>>> Windows-on-ARM if we don't have a buildfarm member testing it.\\n>>> For other platforms that have a track record of multiple\\n>>> hardware support, it might not be a stretch ... but Windows was\\n>>> so resolutely Intel-only for so long that \\\"it works on ARM\\\" is\\n>>> a proposition that I won't trust without hard evidence. There\\n>>> are too many bits of that system that might not have gotten the\\n>>> word yet, or at least not gotten sufficient testing.\\n>>> My vote for this is we don't commit without a buildfarm member.\\n>>\\n>> I think we can have a multi-tiered approach, where we can commit support but consider it experimental until we have buildfarm coverage.\\n> \\n> If it's experimental it should probably be behind an opt-in flag in\\n> autoconf/meson, or be reverted by the time REL_17_STABLE branches unless\\n> coverage has materialized by then.\\n\\nThe author's email is bouncing now, due to job change, so it's unlikely \\nwe will see any progress on this anymore. I am setting it to returned \\nwith feedback.\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 19 Sep 2023 10:33:48 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter@eisentraut.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nThis was covered earlier in the thread - I have taken this on in Niyas' \\nstead.\\n\\nWas there an explicit request for something there? I was under the \\nimpression that this was all just suggestion/theory at the moment.\\n\\nThanks,\\nAnthony\\n\\nOn 19/09/2023 09:33, Peter Eisentraut wrote:\\n> On 14.09.23 11:39, Daniel Gustafsson wrote:\\n>>> On 13 Sep 2023, at 21:12, Peter Eisentraut <peter@eisentraut.org> \\n>>> wrote:\\n>>>\\n>>> On 31.08.23 06:44, Tom Lane wrote:\\n>>>> I agree. I'm really uncomfortable with claiming support for\\n>>>> Windows-on-ARM if we don't have a buildfarm member testing it.\\n>>>> For other platforms that have a track record of multiple\\n>>>> hardware support, it might not be a stretch ... but Windows was\\n>>>> so resolutely Intel-only for so long that \\\"it works on ARM\\\" is\\n>>>> a proposition that I won't trust without hard evidence. There\\n>>>> are too many bits of that system that might not have gotten the\\n>>>> word yet, or at least not gotten sufficient testing.\\n>>>> My vote for this is we don't commit without a buildfarm member.\\n>>>\\n>>> I think we can have a multi-tiered approach, where we can commit \\n>>> support but consider it experimental until we have buildfarm coverage.\\n>>\\n>> If it's experimental it should probably be behind an opt-in flag in\\n>> autoconf/meson, or be reverted by the time REL_17_STABLE branches unless\\n>> coverage has materialized by then.\\n>\\n> The author's email is bouncing now, due to job change, so it's \\n> unlikely we will see any progress on this anymore. I am setting it to \\n> returned with feedback.\\n>\\n\\n\\n\",\n \"msg_date\": \"Tue, 19 Sep 2023 13:35:17 +0100\",\n \"msg_from\": \"Anthony Roberts <anthony.roberts@linaro.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, Sep 19, 2023 at 01:35:17PM +0100, Anthony Roberts wrote:\\n> Was there an explicit request for something there? I was under the\\n> impression that this was all just suggestion/theory at the moment.\\n\\nYes. The addition of a buildfarm member registered into the community\\nfacilities, so as it is possible to provide back to developers dynamic\\nfeedback to the new configuration setup you would like to see\\nsupported in PostgreSQL. This has been mentioned a few times on this\\nthread, around these places:\\nhttps://www.postgresql.org/message-id/20220322103011.i6z2tuj4hle23wgx@jrouhaud\\nhttps://www.postgresql.org/message-id/DBD80715-E56B-40EB-84AA-ACE70198E7AD@yesql.se\\nhttps://www.postgresql.org/message-id/6D1E2719-FA49-42A5-A6FF-0B0072BF631D@yesql.se\\n--\\nMichael\",\n \"msg_date\": \"Wed, 20 Sep 2023 07:59:53 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, 19 Sept 2023 at 23:48, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Tue, Sep 19, 2023 at 01:35:17PM +0100, Anthony Roberts wrote:\\n> > Was there an explicit request for something there? I was under the\\n> > impression that this was all just suggestion/theory at the moment.\\n>\\n> Yes. The addition of a buildfarm member registered into the community\\n> facilities, so as it is possible to provide back to developers dynamic\\n> feedback to the new configuration setup you would like to see\\n> supported in PostgreSQL. This has been mentioned a few times on this\\n> thread, around these places:\\n>\\n> https://www.postgresql.org/message-id/20220322103011.i6z2tuj4hle23wgx@jrouhaud\\n>\\n> https://www.postgresql.org/message-id/DBD80715-E56B-40EB-84AA-ACE70198E7AD@yesql.se\\n>\\n> https://www.postgresql.org/message-id/6D1E2719-FA49-42A5-A6FF-0B0072BF631D@yesql.se\\n> --\\n> Michael\\n>\\n\\nThe attached patch works with v17. I will work on getting a buildfarm\\nanimal up shortly\",\n \"msg_date\": \"Tue, 23 Jan 2024 08:46:37 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks> wrote:\\n\\n>\\n>\\n> On Tue, 19 Sept 2023 at 23:48, Michael Paquier <michael@paquier.xyz>\\n> wrote:\\n>\\n>> On Tue, Sep 19, 2023 at 01:35:17PM +0100, Anthony Roberts wrote:\\n>> > Was there an explicit request for something there? I was under the\\n>> > impression that this was all just suggestion/theory at the moment.\\n>>\\n>> Yes. The addition of a buildfarm member registered into the community\\n>> facilities, so as it is possible to provide back to developers dynamic\\n>> feedback to the new configuration setup you would like to see\\n>> supported in PostgreSQL. This has been mentioned a few times on this\\n>> thread, around these places:\\n>>\\n>> https://www.postgresql.org/message-id/20220322103011.i6z2tuj4hle23wgx@jrouhaud\\n>>\\n>> https://www.postgresql.org/message-id/DBD80715-E56B-40EB-84AA-ACE70198E7AD@yesql.se\\n>>\\n>> https://www.postgresql.org/message-id/6D1E2719-FA49-42A5-A6FF-0B0072BF631D@yesql.se\\n>> --\\n>> Michael\\n>>\\n>\\n> The attached patch works with v17. I will work on getting a buildfarm\\n> animal up shortly\\n>\\n>\\nI can build using meson manually, but for some reason building with the\\nbuildfarm fails.\\n\\nI'm not sure which files to attach to this to help. Andrew, what files do\\nyou need ?\\n\\nDave\\n\\n>\\n>\\n\\nOn Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks> wrote:On Tue, 19 Sept 2023 at 23:48, Michael Paquier <michael@paquier.xyz> wrote:On Tue, Sep 19, 2023 at 01:35:17PM +0100, Anthony Roberts wrote:\\n> Was there an explicit request for something there? I was under the\\n> impression that this was all just suggestion/theory at the moment.\\n\\nYes. The addition of a buildfarm member registered into the community\\nfacilities, so as it is possible to provide back to developers dynamic\\nfeedback to the new configuration setup you would like to see\\nsupported in PostgreSQL. This has been mentioned a few times on this\\nthread, around these places:\\nhttps://www.postgresql.org/message-id/20220322103011.i6z2tuj4hle23wgx@jrouhaud\\nhttps://www.postgresql.org/message-id/DBD80715-E56B-40EB-84AA-ACE70198E7AD@yesql.se\\nhttps://www.postgresql.org/message-id/6D1E2719-FA49-42A5-A6FF-0B0072BF631D@yesql.se\\n--\\nMichaelThe attached patch works with v17. I will work on getting a buildfarm animal up shortlyI can build using meson manually, but for some reason building with the buildfarm fails.I'm not sure which files to attach to this to help. Andrew, what files do you need ?Dave\",\n \"msg_date\": \"Tue, 23 Jan 2024 16:13:05 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 23, 2024 at 04:13:05PM -0500, Dave Cramer wrote:\\n> On Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks> wrote:\\n>> The attached patch works with v17. I will work on getting a buildfarm\\n>> animal up shortly\\n\\nThanks for doing that!\\n\\n> I can build using meson manually, but for some reason building with the\\n> buildfarm fails.\\n> \\n> I'm not sure which files to attach to this to help. Andrew, what files do\\n> you need ?\\n\\nProbably build.ninja and the contents of meson-logs/ would offer\\nenough information?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 24 Jan 2024 08:32:08 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, 23 Jan 2024 at 18:32, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Tue, Jan 23, 2024 at 04:13:05PM -0500, Dave Cramer wrote:\\n> > On Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks>\\n> wrote:\\n> >> The attached patch works with v17. I will work on getting a buildfarm\\n> >> animal up shortly\\n>\\n> Thanks for doing that!\\n>\\n> > I can build using meson manually, but for some reason building with the\\n> > buildfarm fails.\\n> >\\n> > I'm not sure which files to attach to this to help. Andrew, what files do\\n> > you need ?\\n>\\n> Probably build.ninja and the contents of meson-logs/ would offer\\n> enough information?\\n> --\\n> Michael\\n>\\n\\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\\nadd some options, but the patch above needs to be applied to build it.\\n\\nDave\\n\\nOn Tue, 23 Jan 2024 at 18:32, Michael Paquier <michael@paquier.xyz> wrote:On Tue, Jan 23, 2024 at 04:13:05PM -0500, Dave Cramer wrote:\\n> On Tue, 23 Jan 2024 at 08:46, Dave Cramer <davecramer@postgres.rocks> wrote:\\n>> The attached patch works with v17. I will work on getting a buildfarm\\n>> animal up shortly\\n\\nThanks for doing that!\\n\\n> I can build using meson manually, but for some reason building with the\\n> buildfarm fails.\\n> \\n> I'm not sure which files to attach to this to help. Andrew, what files do\\n> you need ?\\n\\nProbably build.ninja and the contents of meson-logs/ would offer\\nenough information?\\n--\\nMichaelI managed to get it to build the vcvarsall arch needs to be x64. I need to add some options, but the patch above needs to be applied to build it.Dave\",\n \"msg_date\": \"Wed, 24 Jan 2024 06:45:21 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\\n> add some options, but the patch above needs to be applied to build it.\\n\\nNice. If I may ask, what kind of host and/or configuration have you\\nused to reach a state where the code can be compiled and run tests\\nwith meson? If you have found specific steps, it may be a good thing\\nto document that on the wiki, say around [1].\\n\\nPerhaps you have not included TAP? It may be fine in terms of runtime\\nchecks and coverage.\\n\\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\\n--\\nMichael\",\n \"msg_date\": \"Thu, 25 Jan 2024 09:02:30 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, 24 Jan 2024 at 19:03, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n> > I managed to get it to build the vcvarsall arch needs to be x64. I need\\n> to\\n> > add some options, but the patch above needs to be applied to build it.\\n>\\n> Nice. If I may ask, what kind of host and/or configuration have you\\n> used to reach a state where the code can be compiled and run tests\\n> with meson? If you have found specific steps, it may be a good thing\\n> to document that on the wiki, say around [1].\\n>\\n\\nI am running a mac-mini M2 with parallels running windows pro 11\\nThe only thing required is this patch. Running in a native developer\\nprompt\\n\\nmeson setup build --prefix=c:\\\\postgres\\ncd build\\nninja\\nninja install\\nninja test\\n\\nall run without errors\\nI also have buildfarm running and will run that today\\n\\nDave\\n\\n>\\n> Perhaps you have not included TAP? It may be fine in terms of runtime\\n> checks and coverage.\\n>\\n\\nDoes TAP have to be explicitly added to the meson build ?\\n\\nDave\\n\\nOn Wed, 24 Jan 2024 at 19:03, Michael Paquier <michael@paquier.xyz> wrote:On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\\n> add some options, but the patch above needs to be applied to build it.\\n\\nNice. If I may ask, what kind of host and/or configuration have you\\nused to reach a state where the code can be compiled and run tests\\nwith meson? If you have found specific steps, it may be a good thing\\nto document that on the wiki, say around [1].I am running a mac-mini M2 with parallels running windows pro 11The only thing required is this patch. Running in a native developer prompt meson setup build --prefix=c:\\\\postgrescd buildninja ninja installninja testall run without errorsI also have buildfarm running and will run that todayDave\\n\\nPerhaps you have not included TAP? It may be fine in terms of runtime\\nchecks and coverage.Does TAP have to be explicitly added to the meson build ?Dave\",\n \"msg_date\": \"Thu, 25 Jan 2024 08:31:07 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, 25 Jan 2024 at 08:31, Dave Cramer <davecramer@postgres.rocks> wrote:\\n\\n>\\n>\\n> On Wed, 24 Jan 2024 at 19:03, Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n>> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n>> > I managed to get it to build the vcvarsall arch needs to be x64. I need\\n>> to\\n>> > add some options, but the patch above needs to be applied to build it.\\n>>\\n>> Nice. If I may ask, what kind of host and/or configuration have you\\n>> used to reach a state where the code can be compiled and run tests\\n>> with meson? If you have found specific steps, it may be a good thing\\n>> to document that on the wiki, say around [1].\\n>>\\n>\\n> I am running a mac-mini M2 with parallels running windows pro 11\\n> The only thing required is this patch. Running in a native developer\\n> prompt\\n>\\n> meson setup build --prefix=c:\\\\postgres\\n> cd build\\n> ninja\\n> ninja install\\n> ninja test\\n>\\n> all run without errors\\n> I also have buildfarm running and will run that today\\n>\\n> Dave\\n>\\n>>\\n>> Perhaps you have not included TAP? It may be fine in terms of runtime\\n>> checks and coverage.\\n>>\\n>\\n> Does TAP have to be explicitly added to the meson build ?\\n>\\n> Dave\\n>\\n\\n\\nI tried running my buildfarm using my git repo and my branch, but get the\\nfollowing error\\nStatus Line: 492 bad branch parameter\\nContent:\\nbad branch parameter fix_arm\\n\\nWeb txn failed with status: 1\\n\\nOn Thu, 25 Jan 2024 at 08:31, Dave Cramer <davecramer@postgres.rocks> wrote:On Wed, 24 Jan 2024 at 19:03, Michael Paquier <michael@paquier.xyz> wrote:On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\\n> add some options, but the patch above needs to be applied to build it.\\n\\nNice. If I may ask, what kind of host and/or configuration have you\\nused to reach a state where the code can be compiled and run tests\\nwith meson? If you have found specific steps, it may be a good thing\\nto document that on the wiki, say around [1].I am running a mac-mini M2 with parallels running windows pro 11The only thing required is this patch. Running in a native developer prompt meson setup build --prefix=c:\\\\postgrescd buildninja ninja installninja testall run without errorsI also have buildfarm running and will run that todayDave\\n\\nPerhaps you have not included TAP? It may be fine in terms of runtime\\nchecks and coverage.Does TAP have to be explicitly added to the meson build ?DaveI tried running my buildfarm using my git repo and my branch, but get the following errorStatus Line: 492 bad branch parameterContent:bad branch parameter fix_armWeb txn failed with status: 1\",\n \"msg_date\": \"Thu, 25 Jan 2024 08:45:52 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-01-24 We 19:02, Michael Paquier wrote:\\n> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n>> I managed to get it to build the vcvarsall arch needs to be x64. I need to\\n>> add some options, but the patch above needs to be applied to build it.\\n> Nice. If I may ask, what kind of host and/or configuration have you\\n> used to reach a state where the code can be compiled and run tests\\n> with meson? If you have found specific steps, it may be a good thing\\n> to document that on the wiki, say around [1].\\n>\\n> Perhaps you have not included TAP? It may be fine in terms of runtime\\n> checks and coverage.\\n>\\n> [1]:https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\\n\\n\\n\\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think we \\nreally want to build with x64_arm64, i.e. to generate native arm64 \\nbinaries. Setting just x64 will not do that, AIUI.\\n\\nI tried that with the buidfarm, setting that in the config file's call \\nto PGBuild::VSenv::getenv().\\n\\nThat upset msvc_gendef.pl, so I added this there to keep it happy:\\n\\n $arch = 'x86_64' if $arch eq 'aarch64';\\n\\nAfter that things went ok until I got this:\\n\\n[1453/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\nFAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n\\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library src\\\\backend\\\\postgres.exe.lib\\nstorage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol \\n_mm_pause referenced in function perform_spin_delay\\nsrc\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n\\n\\nI haven't made further progress, but I will return to it in the next day \\nor so.\\n\\nWhile this will be nice to have, I think it won't really matter until \\nthere is ARM64 support in released versions of Windows Server. AFAICT \\nthey still only sell versions for x86_64\\n\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-01-24 We 19:02, Michael Paquier\\n wrote:\\n\\n\\nOn Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n\\n\\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\\nadd some options, but the patch above needs to be applied to build it.\\n\\n\\n\\nNice. If I may ask, what kind of host and/or configuration have you\\nused to reach a state where the code can be compiled and run tests\\nwith meson? If you have found specific steps, it may be a good thing\\nto document that on the wiki, say around [1].\\n\\nPerhaps you have not included TAP? It may be fine in terms of runtime\\nchecks and coverage.\\n\\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\\n\\n\\n\\n\\n\\n\\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think\\n we really want to build with x64_arm64, i.e. to generate native\\n arm64 binaries. Setting just x64 will not do that, AIUI.\\nI tried that with the buidfarm, setting that in the config file's\\n call to PGBuild::VSenv::getenv().\\nThat upset msvc_gendef.pl, so I added this there to keep it\\n happy:\\n\\n$arch = 'x86_64' if $arch eq 'aarch64';\\n\\nAfter that things went ok until I got this:\\n[1453/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library src\\\\backend\\\\postgres.exe.lib\\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\\n symbol _mm_pause referenced in function perform_spin_delay\\n src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved\\n externals \\n\\n\\n\\nI haven't made further progress, but I will return to it in the\\n next day or so.\\nWhile this will be nice to have, I think it won't really matter\\n until there is ARM64 support in released versions of Windows\\n Server. AFAICT they still only sell versions for x86_64\\n\\n\\n\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 25 Jan 2024 12:30:55 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-01-24 We 19:02, Michael Paquier wrote:\\n>\\n> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n>\\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\\n> add some options, but the patch above needs to be applied to build it.\\n>\\n> Nice. If I may ask, what kind of host and/or configuration have you\\n> used to reach a state where the code can be compiled and run tests\\n> with meson? If you have found specific steps, it may be a good thing\\n> to document that on the wiki, say around [1].\\n>\\n> Perhaps you have not included TAP? It may be fine in terms of runtime\\n> checks and coverage.\\n>\\n> [1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\\n>\\n>\\n>\\n> I now have an ARM64 machine, so I set up a W11 ARM64 VM. I think we really\\n> want to build with x64_arm64, i.e. to generate native arm64 binaries.\\n> Setting just x64 will not do that, AIUI.\\n>\\n> I tried that with the buidfarm, setting that in the config file's call to\\n> PGBuild::VSenv::getenv().\\n>\\n> That upset msvc_gendef.pl, so I added this there to keep it happy:\\n>\\n> $arch = 'x86_64' if $arch eq 'aarch64';\\n>\\n> After that things went ok until I got this:\\n>\\n> [1453/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n> \\\"link\\\" @src/backend/postgres.exe.rsp\\n> Creating library src\\\\backend\\\\postgres.exe.lib\\n> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\\n> _mm_pause referenced in function perform_spin_delay\\n> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n>\\n>\\n> I haven't made further progress, but I will return to it in the next day\\n> or so.\\n>\\n> While this will be nice to have, I think it won't really matter until\\n> there is ARM64 support in released versions of Windows Server. AFAICT they\\n> still only sell versions for x86_64\\n>\\n\\nI need it to build clients. The clients need arm64 libraries to link against\\n\\nDave\\n\\nOn Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-01-24 We 19:02, Michael Paquier\\n wrote:\\n\\n\\nOn Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n\\n\\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\\nadd some options, but the patch above needs to be applied to build it.\\n\\n\\nNice. If I may ask, what kind of host and/or configuration have you\\nused to reach a state where the code can be compiled and run tests\\nwith meson? If you have found specific steps, it may be a good thing\\nto document that on the wiki, say around [1].\\n\\nPerhaps you have not included TAP? It may be fine in terms of runtime\\nchecks and coverage.\\n\\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\\n\\n\\n\\n\\n\\n\\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think\\n we really want to build with x64_arm64, i.e. to generate native\\n arm64 binaries. Setting just x64 will not do that, AIUI.\\nI tried that with the buidfarm, setting that in the config file's\\n call to PGBuild::VSenv::getenv().\\nThat upset msvc_gendef.pl, so I added this there to keep it\\n happy:\\n\\n$arch = 'x86_64' if $arch eq 'aarch64';\\n\\nAfter that things went ok until I got this:\\n[1453/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library src\\\\backend\\\\postgres.exe.lib\\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\\n symbol _mm_pause referenced in function perform_spin_delay\\n src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved\\n externals \\n\\n\\n\\nI haven't made further progress, but I will return to it in the\\n next day or so.\\nWhile this will be nice to have, I think it won't really matter\\n until there is ARM64 support in released versions of Windows\\n Server. AFAICT they still only sell versions for x86_64I need it to build clients. The clients need arm64 libraries to link againstDave\",\n \"msg_date\": \"Thu, 25 Jan 2024 14:07:11 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-01-25 Th 08:45, Dave Cramer wrote:\\n>\\n>\\n>\\n>\\n> I tried running my buildfarm using my git repo and my branch, but get \\n> the following error\\n> Status Line: 492 bad branch parameter\\n> Content:\\n> bad branch parameter fix_arm\\n>\\n> Web txn failed with status: 1\\n\\n\\n\\nYou can't use your own branch with the buildfarm, you need to let it set \\nup its own branches.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-01-25 Th 08:45, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nI tried running my buildfarm using my git repo and my\\n branch, but get the following error\\nStatus Line: 492 bad branch parameter\\nContent:\\nbad branch parameter fix_arm\\n\\n\\nWeb txn failed with status: 1 \\n\\n\\n\\n\\n\\n\\n\\n\\n\\nYou can't use your own branch with the buildfarm, you need to let\\n it set up its own branches.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 25 Jan 2024 14:31:43 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, 25 Jan 2024 at 14:31, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-01-25 Th 08:45, Dave Cramer wrote:\\n>\\n>\\n>\\n>\\n>\\n> I tried running my buildfarm using my git repo and my branch, but get the\\n> following error\\n> Status Line: 492 bad branch parameter\\n> Content:\\n> bad branch parameter fix_arm\\n>\\n> Web txn failed with status: 1\\n>\\n>\\n>\\n> You can't use your own branch with the buildfarm, you need to let it set\\n> up its own branches.\\n>\\n\\nSo I guess I have to wait until this patch is merged in ?\\n\\nDave\\n\\nOn Thu, 25 Jan 2024 at 14:31, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-01-25 Th 08:45, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nI tried running my buildfarm using my git repo and my\\n branch, but get the following error\\nStatus Line: 492 bad branch parameter\\nContent:\\nbad branch parameter fix_arm\\n\\n\\nWeb txn failed with status: 1 \\n\\n\\n\\n\\n\\n\\n\\n\\n\\nYou can't use your own branch with the buildfarm, you need to let\\n it set up its own branches.So I guess I have to wait until this patch is merged in ?Dave\",\n \"msg_date\": \"Thu, 25 Jan 2024 15:56:20 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-01-24 We 19:02, Michael Paquier wrote:\\n>\\n> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n>\\n> I managed to get it to build the vcvarsall arch needs to be x64. I need to\\n> add some options, but the patch above needs to be applied to build it.\\n>\\n> Nice. If I may ask, what kind of host and/or configuration have you\\n> used to reach a state where the code can be compiled and run tests\\n> with meson? If you have found specific steps, it may be a good thing\\n> to document that on the wiki, say around [1].\\n>\\n> Perhaps you have not included TAP? It may be fine in terms of runtime\\n> checks and coverage.\\n>\\n> [1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\\n>\\n>\\n>\\n> I now have an ARM64 machine, so I set up a W11 ARM64 VM. I think we really\\n> want to build with x64_arm64, i.e. to generate native arm64 binaries.\\n> Setting just x64 will not do that, AIUI.\\n>\\n> I tried that with the buidfarm, setting that in the config file's call to\\n> PGBuild::VSenv::getenv().\\n>\\n> That upset msvc_gendef.pl, so I added this there to keep it happy:\\n>\\n> $arch = 'x86_64' if $arch eq 'aarch64';\\n>\\n> After that things went ok until I got this:\\n>\\n> [1453/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n> \\\"link\\\" @src/backend/postgres.exe.rsp\\n> Creating library src\\\\backend\\\\postgres.exe.lib\\n> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\\n> _mm_pause referenced in function perform_spin_delay\\n> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n>\\n>\\n> I haven't made further progress, but I will return to it in the next day\\n> or so.\\n>\\n> While this will be nice to have, I think it won't really matter until\\n> there is ARM64 support in released versions of Windows Server. AFAICT they\\n> still only sell versions for x86_64\\n>\\n\\nI've tried it with my patch attached previously and x64_arm64 and it works\\nfine. builds using the buildfarm as well.\\n\\nIs there a definitive way to figure out if the binaries are x64_arm64\\n\\nDave\\n\\nOn Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-01-24 We 19:02, Michael Paquier\\n wrote:\\n\\n\\nOn Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n\\n\\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\\nadd some options, but the patch above needs to be applied to build it.\\n\\n\\nNice. If I may ask, what kind of host and/or configuration have you\\nused to reach a state where the code can be compiled and run tests\\nwith meson? If you have found specific steps, it may be a good thing\\nto document that on the wiki, say around [1].\\n\\nPerhaps you have not included TAP? It may be fine in terms of runtime\\nchecks and coverage.\\n\\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\\n\\n\\n\\n\\n\\n\\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think\\n we really want to build with x64_arm64, i.e. to generate native\\n arm64 binaries. Setting just x64 will not do that, AIUI.\\nI tried that with the buidfarm, setting that in the config file's\\n call to PGBuild::VSenv::getenv().\\nThat upset msvc_gendef.pl, so I added this there to keep it\\n happy:\\n\\n$arch = 'x86_64' if $arch eq 'aarch64';\\n\\nAfter that things went ok until I got this:\\n[1453/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library src\\\\backend\\\\postgres.exe.lib\\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\\n symbol _mm_pause referenced in function perform_spin_delay\\n src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved\\n externals \\n\\n\\n\\nI haven't made further progress, but I will return to it in the\\n next day or so.\\nWhile this will be nice to have, I think it won't really matter\\n until there is ARM64 support in released versions of Windows\\n Server. AFAICT they still only sell versions for x86_64I've tried it with my patch attached previously and x64_arm64 and it works fine. builds using the buildfarm as well.Is there a definitive way to figure out if the binaries are x64_arm64Dave\",\n \"msg_date\": \"Thu, 25 Jan 2024 16:01:28 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi David,\\n\\nUnix \\\"file\\\" or \\\"dumpbin /headers\\\" in vcvarsall are your best bets.\\n\\nThanks,\\nAnthony\\n\\nOn Thu, 25 Jan 2024, 21:01 Dave Cramer, <davecramer@postgres.rocks> wrote:\\n\\n>\\n>\\n> On Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>>\\n>> On 2024-01-24 We 19:02, Michael Paquier wrote:\\n>>\\n>> On Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n>>\\n>> I managed to get it to build the vcvarsall arch needs to be x64. I need to\\n>> add some options, but the patch above needs to be applied to build it.\\n>>\\n>> Nice. If I may ask, what kind of host and/or configuration have you\\n>> used to reach a state where the code can be compiled and run tests\\n>> with meson? If you have found specific steps, it may be a good thing\\n>> to document that on the wiki, say around [1].\\n>>\\n>> Perhaps you have not included TAP? It may be fine in terms of runtime\\n>> checks and coverage.\\n>>\\n>> [1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\\n>>\\n>>\\n>>\\n>> I now have an ARM64 machine, so I set up a W11 ARM64 VM. I think we\\n>> really want to build with x64_arm64, i.e. to generate native arm64\\n>> binaries. Setting just x64 will not do that, AIUI.\\n>>\\n>> I tried that with the buidfarm, setting that in the config file's call to\\n>> PGBuild::VSenv::getenv().\\n>>\\n>> That upset msvc_gendef.pl, so I added this there to keep it happy:\\n>>\\n>> $arch = 'x86_64' if $arch eq 'aarch64';\\n>>\\n>> After that things went ok until I got this:\\n>>\\n>> [1453/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n>> \\\"link\\\" @src/backend/postgres.exe.rsp\\n>> Creating library src\\\\backend\\\\postgres.exe.lib\\n>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\\n>> _mm_pause referenced in function perform_spin_delay\\n>> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n>>\\n>>\\n>> I haven't made further progress, but I will return to it in the next day\\n>> or so.\\n>>\\n>> While this will be nice to have, I think it won't really matter until\\n>> there is ARM64 support in released versions of Windows Server. AFAICT they\\n>> still only sell versions for x86_64\\n>>\\n>\\n> I've tried it with my patch attached previously and x64_arm64 and it works\\n> fine. builds using the buildfarm as well.\\n>\\n> Is there a definitive way to figure out if the binaries are x64_arm64\\n>\\n> Dave\\n>\\n\\nHi David,Unix \\\"file\\\" or \\\"dumpbin /headers\\\" in vcvarsall are your best bets.Thanks,AnthonyOn Thu, 25 Jan 2024, 21:01 Dave Cramer, <davecramer@postgres.rocks> wrote:On Thu, 25 Jan 2024 at 12:30, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-01-24 We 19:02, Michael Paquier\\n wrote:\\n\\n\\nOn Wed, Jan 24, 2024 at 06:45:21AM -0500, Dave Cramer wrote:\\n\\n\\nI managed to get it to build the vcvarsall arch needs to be x64. I need to\\nadd some options, but the patch above needs to be applied to build it.\\n\\n\\nNice. If I may ask, what kind of host and/or configuration have you\\nused to reach a state where the code can be compiled and run tests\\nwith meson? If you have found specific steps, it may be a good thing\\nto document that on the wiki, say around [1].\\n\\nPerhaps you have not included TAP? It may be fine in terms of runtime\\nchecks and coverage.\\n\\n[1]: https://wiki.postgresql.org/wiki/PostgreSQL_Buildfarm_Howto#Running_on_Windows\\n\\n\\n\\n\\n\\n\\nI now have an ARM64 machine, so I set up a W11 ARM64 VM. I think\\n we really want to build with x64_arm64, i.e. to generate native\\n arm64 binaries. Setting just x64 will not do that, AIUI.\\nI tried that with the buidfarm, setting that in the config file's\\n call to PGBuild::VSenv::getenv().\\nThat upset msvc_gendef.pl, so I added this there to keep it\\n happy:\\n\\n$arch = 'x86_64' if $arch eq 'aarch64';\\n\\nAfter that things went ok until I got this:\\n[1453/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library src\\\\backend\\\\postgres.exe.lib\\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\\n symbol _mm_pause referenced in function perform_spin_delay\\n src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved\\n externals \\n\\n\\n\\nI haven't made further progress, but I will return to it in the\\n next day or so.\\nWhile this will be nice to have, I think it won't really matter\\n until there is ARM64 support in released versions of Windows\\n Server. AFAICT they still only sell versions for x86_64I've tried it with my patch attached previously and x64_arm64 and it works fine. builds using the buildfarm as well.Is there a definitive way to figure out if the binaries are x64_arm64Dave\",\n \"msg_date\": \"Thu, 25 Jan 2024 21:03:51 +0000\",\n \"msg_from\": \"Anthony Roberts <anthony.roberts@linaro.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-01-25 Th 15:56, Dave Cramer wrote:\\n>\\n>\\n> On Thu, 25 Jan 2024 at 14:31, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>\\n> On 2024-01-25 Th 08:45, Dave Cramer wrote:\\n>>\\n>>\\n>>\\n>>\\n>> I tried running my buildfarm using my git repo and my branch, but\\n>> get the following error\\n>> Status Line: 492 bad branch parameter\\n>> Content:\\n>> bad branch parameter fix_arm\\n>>\\n>> Web txn failed with status: 1\\n>\\n>\\n>\\n> You can't use your own branch with the buildfarm, you need to let\\n> it set up its own branches.\\n>\\n>\\n> So I guess I have to wait until this patch is merged in ?\\n>\\n>\\n\\nNot quite.\\n\\nThere's a switch that lets you test your own code. If you say \\n--from-source /path/to/repo it will run in whatever state the repo is \\nin. It won't care what branch you are on, and it won't try to update the \\nrepo. It won't report the results to the server, but it will build and \\ntest everything just like in a regular run. (On the \\\"eat your own dog \\nfood\\\" principle I use this mode a lot.) If you use that mode you \\nprobably also want to specify --verbose as well.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-01-25 Th 15:56, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Thu, 25 Jan 2024 at\\n 14:31, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-01-25 Th 08:45, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nI tried running my buildfarm using my git\\n repo and my branch, but get the following error\\nStatus Line: 492 bad branch parameter\\nContent:\\nbad branch parameter fix_arm\\n\\n\\nWeb txn failed with status: 1 \\n\\n\\n\\n\\n\\n\\n\\n\\n\\nYou can't use your own branch with the buildfarm, you\\n need to let it set up its own branches.\\n\\n\\n\\n\\nSo I guess I have to wait until this patch is merged in ?\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nNot quite. \\n\\nThere's a switch that lets you test your own code. If you say\\n --from-source /path/to/repo it will run in whatever state the\\n repo is in. It won't care what branch you are on, and it won't try\\n to update the repo. It won't report the results to the server, but\\n it will build and test everything just like in a regular run. (On\\n the \\\"eat your own dog food\\\" principle I use this mode a lot.) If\\n you use that mode you probably also want to specify --verbose as\\n well.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 25 Jan 2024 16:14:25 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, 25 Jan 2024 at 16:04, Anthony Roberts <anthony.roberts@linaro.org>\\nwrote:\\n\\n> Hi David,\\n>\\n> Unix \\\"file\\\" or \\\"dumpbin /headers\\\" in vcvarsall are your best bets.\\n>\\n> Thanks,\\n> Anthony\\n>\\n\\n\\nSo there is another way, select the file in Windows Explorer and right\\nclick, in the compatibility tab if the \\\"Windows on ARM\\\" is greyed out it is\\nan arm binary.\\n\\nSo far mine are not :(\\n\\nThanks,\\n\\nDave\\n\\n>\\n\\nOn Thu, 25 Jan 2024 at 16:04, Anthony Roberts <anthony.roberts@linaro.org> wrote:Hi David,Unix \\\"file\\\" or \\\"dumpbin /headers\\\" in vcvarsall are your best bets.Thanks,AnthonySo there is another way, select the file in Windows Explorer and right click, in the compatibility tab if the \\\"Windows on ARM\\\" is greyed out it is an arm binary.So far mine are not :(Thanks,Dave\",\n \"msg_date\": \"Thu, 25 Jan 2024 16:17:15 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>\\n>\\n> On Thu, 25 Jan 2024 at 16:04, Anthony Roberts \\n> <anthony.roberts@linaro.org> wrote:\\n>\\n> Hi David,\\n>\\n> Unix \\\"file\\\" or \\\"dumpbin /headers\\\" in vcvarsall are your best bets.\\n>\\n> Thanks,\\n> Anthony\\n>\\n>\\n>\\n> So there is another way, select the file in Windows Explorer and right \\n> click, in the compatibility tab if the \\\"Windows on ARM\\\" is greyed out \\n> it is an arm binary.\\n>\\n> So far mine are not :(\\n>\\n\\nYeah, I think the default Developer Command Prompt for VS2022 is set up \\nfor x86 builds. AIUI you should start by executing \\\"vcvarsall x64_arm64\\\".\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-01-25 Th 16:17, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Thu, 25 Jan 2024 at\\n 16:04, Anthony Roberts <anthony.roberts@linaro.org>\\n wrote:\\n\\n\\nHi David,\\n \\n\\nUnix \\\"file\\\" or \\\"dumpbin /headers\\\" in\\n vcvarsall are your best bets.\\n\\n\\nThanks,\\nAnthony\\n\\n\\n\\n\\n\\n\\nSo there is another way, select the file in Windows\\n Explorer and right click, in the compatibility tab if the\\n \\\"Windows on ARM\\\" is greyed out it is an arm binary.\\n\\n\\nSo far mine are not :(\\n\\n\\n\\n\\n\\n\\n\\nYeah, I think the default Developer Command Prompt for VS2022 is\\n set up for x86 builds. AIUI you should start by executing\\n \\\"vcvarsall x64_arm64\\\".\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 25 Jan 2024 16:32:39 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>\\n>\\n>\\n> On Thu, 25 Jan 2024 at 16:04, Anthony Roberts <anthony.roberts@linaro.org>\\n> wrote:\\n>\\n>> Hi David,\\n>>\\n>> Unix \\\"file\\\" or \\\"dumpbin /headers\\\" in vcvarsall are your best bets.\\n>>\\n>> Thanks,\\n>> Anthony\\n>>\\n>\\n>\\n> So there is another way, select the file in Windows Explorer and right\\n> click, in the compatibility tab if the \\\"Windows on ARM\\\" is greyed out it is\\n> an arm binary.\\n>\\n> So far mine are not :(\\n>\\n>\\n> Yeah, I think the default Developer Command Prompt for VS2022 is set up\\n> for x86 builds. AIUI you should start by executing \\\"vcvarsall x64_arm64\\\".\\n>\\n\\nYup, now I'm in the same state you are\\n\\nDave\\n\\nOn Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-01-25 Th 16:17, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\nOn Thu, 25 Jan 2024 at\\n 16:04, Anthony Roberts <anthony.roberts@linaro.org>\\n wrote:\\n\\n\\nHi David,\\n \\n\\nUnix \\\"file\\\" or \\\"dumpbin /headers\\\" in\\n vcvarsall are your best bets.\\n\\n\\nThanks,\\nAnthony\\n\\n\\n\\n\\n\\n\\nSo there is another way, select the file in Windows\\n Explorer and right click, in the compatibility tab if the\\n \\\"Windows on ARM\\\" is greyed out it is an arm binary.\\n\\n\\nSo far mine are not :(\\n\\n\\n\\n\\n\\n\\n\\nYeah, I think the default Developer Command Prompt for VS2022 is\\n set up for x86 builds. AIUI you should start by executing\\n \\\"vcvarsall x64_arm64\\\".Yup, now I'm in the same state you are Dave\",\n \"msg_date\": \"Thu, 25 Jan 2024 16:52:30 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>> Yeah, I think the default Developer Command Prompt for VS2022 is set up\\n>> for x86 builds. AIUI you should start by executing \\\"vcvarsall x64_arm64\\\".\\n> \\n> Yup, now I'm in the same state you are\\n\\nWait a minute here. Based on [1], x64_arm64 means you can use a x64\\nhost and you'll be able to produce ARM64 builds, still these will not\\nbe able to run on the host where they were built. How much of the\\npatch posted upthread is required to produce such builds? Basically \\neverything from it, I guess, so as build dependencies can be\\nsatisfied?\\n\\n[1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n--\\nMichael\",\n \"msg_date\": \"Fri, 26 Jan 2024 10:32:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"\\nOn 2024-01-25 Th 20:32, Michael Paquier wrote:\\n> On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n>> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>>> Yeah, I think the default Developer Command Prompt for VS2022 is set up\\n>>> for x86 builds. AIUI you should start by executing \\\"vcvarsall x64_arm64\\\".\\n>> Yup, now I'm in the same state you are\\n> Wait a minute here. Based on [1], x64_arm64 means you can use a x64\\n> host and you'll be able to produce ARM64 builds, still these will not\\n> be able to run on the host where they were built. How much of the\\n> patch posted upthread is required to produce such builds? Basically\\n> everything from it, I guess, so as build dependencies can be\\n> satisfied?\\n>\\n> [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n\\n\\nIf you look at the table here x86 and x64 are the only supported host \\narchitectures. But that's OK, the x64 binaries will run on arm64 (W11 \\nARM64 has x64 emulation builtin). If that didn't work Dave and I would \\nnot have got as far as we have. But you want the x64_arm64 argument to \\nvcvarsall so you will get ARM64 output.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 26 Jan 2024 07:36:25 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\\n> wrote:\\n> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n> >>> Yeah, I think the default Developer Command Prompt for VS2022 is set up\\n> >>> for x86 builds. AIUI you should start by executing \\\"vcvarsall\\n> x64_arm64\\\".\\n> >> Yup, now I'm in the same state you are\\n> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\\n> > host and you'll be able to produce ARM64 builds, still these will not\\n> > be able to run on the host where they were built. How much of the\\n> > patch posted upthread is required to produce such builds? Basically\\n> > everything from it, I guess, so as build dependencies can be\\n> > satisfied?\\n> >\\n> > [1]:\\n> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>\\n>\\n> If you look at the table here x86 and x64 are the only supported host\\n> architectures. But that's OK, the x64 binaries will run on arm64 (W11\\n> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\\n> not have got as far as we have. But you want the x64_arm64 argument to\\n> vcvarsall so you will get ARM64 output.\\n>\\n\\nI've rebuilt it using x64_arm64 and with the attached (very naive patch)\\nand I still get an x64 binary :(\\n\\n>\\n>\",\n \"msg_date\": \"Fri, 26 Jan 2024 09:18:50 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n>\\n>\\n> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>\\n> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\\n> <andrew@dunslane.net> wrote:\\n> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n> >>> Yeah, I think the default Developer Command Prompt for VS2022\\n> is set up\\n> >>> for x86 builds. AIUI you should start by executing \\\"vcvarsall\\n> x64_arm64\\\".\\n> >> Yup, now I'm in the same state you are\\n> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\\n> > host and you'll be able to produce ARM64 builds, still these\\n> will not\\n> > be able to run on the host where they were built. How much of the\\n> > patch posted upthread is required to produce such builds? Basically\\n> > everything from it, I guess, so as build dependencies can be\\n> > satisfied?\\n> >\\n> > [1]:\\n> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>\\n>\\n> If you look at the table here x86 and x64 are the only supported host\\n> architectures. But that's OK, the x64 binaries will run on arm64 (W11\\n> ARM64 has x64 emulation builtin). If that didn't work Dave and I\\n> would\\n> not have got as far as we have. But you want the x64_arm64\\n> argument to\\n> vcvarsall so you will get ARM64 output.\\n>\\n>\\n> I've rebuilt it using x64_arm64 and with the attached (very naive \\n> patch) and I still get an x64 binary :(\\n\\n\\nI am definitely getting ARM64 binaries (e.g. for initdb.exe the Windows \\non ARM compatibility setting is greyed out)\\n\\n\\nI'll try your patch.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-01-26 Fr 09:18, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 26 Jan 2024 at\\n 07:36, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer\\n wrote:\\n >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n >>> Yeah, I think the default Developer Command\\n Prompt for VS2022 is set up\\n >>> for x86 builds. AIUI you should start by\\n executing \\\"vcvarsall x64_arm64\\\".\\n >> Yup, now I'm in the same state you are\\n > Wait a minute here. Based on [1], x64_arm64 means\\n you can use a x64\\n > host and you'll be able to produce ARM64 builds,\\n still these will not\\n > be able to run on the host where they were built. \\n How much of the\\n > patch posted upthread is required to produce such\\n builds? Basically\\n > everything from it, I guess, so as build dependencies\\n can be\\n > satisfied?\\n >\\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n\\n\\n If you look at the table here x86 and x64 are the only\\n supported host \\n architectures. But that's OK, the x64 binaries will run on\\n arm64 (W11 \\n ARM64 has x64 emulation builtin). If that didn't work Dave\\n and I would \\n not have got as far as we have. But you want the x64_arm64\\n argument to \\n vcvarsall so you will get ARM64 output.\\n\\n\\n\\nI've rebuilt it using x64_arm64 and with the attached\\n (very naive patch) and I still get an x64 binary :(\\n\\n\\n\\n\\n\\n\\nI am definitely getting ARM64 binaries (e.g. for initdb.exe the\\n Windows on ARM compatibility setting is greyed out)\\n\\n\\nI'll try your patch.\\n\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 29 Jan 2024 10:35:51 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n>\\n>\\n> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>\\n> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\\n> <andrew@dunslane.net> wrote:\\n> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n> >>> Yeah, I think the default Developer Command Prompt for VS2022\\n> is set up\\n> >>> for x86 builds. AIUI you should start by executing \\\"vcvarsall\\n> x64_arm64\\\".\\n> >> Yup, now I'm in the same state you are\\n> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\\n> > host and you'll be able to produce ARM64 builds, still these\\n> will not\\n> > be able to run on the host where they were built. How much of the\\n> > patch posted upthread is required to produce such builds? Basically\\n> > everything from it, I guess, so as build dependencies can be\\n> > satisfied?\\n> >\\n> > [1]:\\n> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>\\n>\\n> If you look at the table here x86 and x64 are the only supported host\\n> architectures. But that's OK, the x64 binaries will run on arm64 (W11\\n> ARM64 has x64 emulation builtin). If that didn't work Dave and I\\n> would\\n> not have got as far as we have. But you want the x64_arm64\\n> argument to\\n> vcvarsall so you will get ARM64 output.\\n>\\n>\\n> I've rebuilt it using x64_arm64 and with the attached (very naive \\n> patch) and I still get an x64 binary :(\\n\\n\\nWith this patch I still get a build error, but it's different :-)\\n\\n\\n[1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\nFAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n\\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library src\\\\backend\\\\postgres.exe.lib\\n\\nstorage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol \\nspin_delay referenced in function perform_spin_delay\\n\\nsrc\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-01-26 Fr 09:18, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 26 Jan 2024 at\\n 07:36, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer\\n wrote:\\n >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n >>> Yeah, I think the default Developer Command\\n Prompt for VS2022 is set up\\n >>> for x86 builds. AIUI you should start by\\n executing \\\"vcvarsall x64_arm64\\\".\\n >> Yup, now I'm in the same state you are\\n > Wait a minute here. Based on [1], x64_arm64 means\\n you can use a x64\\n > host and you'll be able to produce ARM64 builds,\\n still these will not\\n > be able to run on the host where they were built. \\n How much of the\\n > patch posted upthread is required to produce such\\n builds? Basically\\n > everything from it, I guess, so as build dependencies\\n can be\\n > satisfied?\\n >\\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n\\n\\n If you look at the table here x86 and x64 are the only\\n supported host \\n architectures. But that's OK, the x64 binaries will run on\\n arm64 (W11 \\n ARM64 has x64 emulation builtin). If that didn't work Dave\\n and I would \\n not have got as far as we have. But you want the x64_arm64\\n argument to \\n vcvarsall so you will get ARM64 output.\\n\\n\\n\\nI've rebuilt it using x64_arm64 and with the attached\\n (very naive patch) and I still get an x64 binary :(\\n\\n\\n\\n\\n\\n\\nWith this patch I still get a build error, but it's different :-)\\n\\n\\n[1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library src\\\\backend\\\\postgres.exe.lib\\n\\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\\n symbol spin_delay referenced in function perform_spin_delay\\n\\n src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved\\n externals\\n\\n\\n\\ncheers\\n\\n\\nandrew\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 29 Jan 2024 11:16:41 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Dave Cramer\\nwww.postgres.rocks\\n\\n\\nOn Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n>\\n>\\n>\\n> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>>\\n>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\\n>> wrote:\\n>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>> >>> Yeah, I think the default Developer Command Prompt for VS2022 is set\\n>> up\\n>> >>> for x86 builds. AIUI you should start by executing \\\"vcvarsall\\n>> x64_arm64\\\".\\n>> >> Yup, now I'm in the same state you are\\n>> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\\n>> > host and you'll be able to produce ARM64 builds, still these will not\\n>> > be able to run on the host where they were built. How much of the\\n>> > patch posted upthread is required to produce such builds? Basically\\n>> > everything from it, I guess, so as build dependencies can be\\n>> > satisfied?\\n>> >\\n>> > [1]:\\n>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>>\\n>>\\n>> If you look at the table here x86 and x64 are the only supported host\\n>> architectures. But that's OK, the x64 binaries will run on arm64 (W11\\n>> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\\n>> not have got as far as we have. But you want the x64_arm64 argument to\\n>> vcvarsall so you will get ARM64 output.\\n>>\\n>\\n> I've rebuilt it using x64_arm64 and with the attached (very naive patch)\\n> and I still get an x64 binary :(\\n>\\n>\\n> With this patch I still get a build error, but it's different :-)\\n>\\n>\\n> [1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n> \\\"link\\\" @src/backend/postgres.exe.rsp\\n> Creating library src\\\\backend\\\\postgres.exe.lib\\n>\\n> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\\n> spin_delay referenced in function perform_spin_delay\\n>\\n> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n>\\n\\nDid you add the latest lock.patch ?\\n\\nDave\\n\\nDave Cramerwww.postgres.rocksOn Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-01-26 Fr 09:18, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 26 Jan 2024 at\\n 07:36, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer\\n wrote:\\n >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n >>> Yeah, I think the default Developer Command\\n Prompt for VS2022 is set up\\n >>> for x86 builds. AIUI you should start by\\n executing \\\"vcvarsall x64_arm64\\\".\\n >> Yup, now I'm in the same state you are\\n > Wait a minute here. Based on [1], x64_arm64 means\\n you can use a x64\\n > host and you'll be able to produce ARM64 builds,\\n still these will not\\n > be able to run on the host where they were built. \\n How much of the\\n > patch posted upthread is required to produce such\\n builds? Basically\\n > everything from it, I guess, so as build dependencies\\n can be\\n > satisfied?\\n >\\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n\\n\\n If you look at the table here x86 and x64 are the only\\n supported host \\n architectures. But that's OK, the x64 binaries will run on\\n arm64 (W11 \\n ARM64 has x64 emulation builtin). If that didn't work Dave\\n and I would \\n not have got as far as we have. But you want the x64_arm64\\n argument to \\n vcvarsall so you will get ARM64 output.\\n\\n\\n\\nI've rebuilt it using x64_arm64 and with the attached\\n (very naive patch) and I still get an x64 binary :(\\n\\n\\n\\n\\n\\n\\nWith this patch I still get a build error, but it's different :-)\\n\\n\\n[1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library src\\\\backend\\\\postgres.exe.lib\\n\\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\\n symbol spin_delay referenced in function perform_spin_delay\\n\\n src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved\\n externalsDid you add the latest lock.patch ?Dave\",\n \"msg_date\": \"Mon, 29 Jan 2024 11:20:18 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-01-29 Mo 11:20, Dave Cramer wrote:\\n>\\n> Dave Cramer\\n> www.postgres.rocks\\n>\\n>\\n> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>\\n> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n>>\\n>>\\n>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan\\n>> <andrew@dunslane.net> wrote:\\n>>\\n>>\\n>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\\n>> <andrew@dunslane.net> wrote:\\n>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>> >>> Yeah, I think the default Developer Command Prompt for\\n>> VS2022 is set up\\n>> >>> for x86 builds. AIUI you should start by executing\\n>> \\\"vcvarsall x64_arm64\\\".\\n>> >> Yup, now I'm in the same state you are\\n>> > Wait a minute here. Based on [1], x64_arm64 means you can\\n>> use a x64\\n>> > host and you'll be able to produce ARM64 builds, still\\n>> these will not\\n>> > be able to run on the host where they were built. How much\\n>> of the\\n>> > patch posted upthread is required to produce such builds? \\n>> Basically\\n>> > everything from it, I guess, so as build dependencies can be\\n>> > satisfied?\\n>> >\\n>> > [1]:\\n>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>>\\n>>\\n>> If you look at the table here x86 and x64 are the only\\n>> supported host\\n>> architectures. But that's OK, the x64 binaries will run on\\n>> arm64 (W11\\n>> ARM64 has x64 emulation builtin). If that didn't work Dave\\n>> and I would\\n>> not have got as far as we have. But you want the x64_arm64\\n>> argument to\\n>> vcvarsall so you will get ARM64 output.\\n>>\\n>>\\n>> I've rebuilt it using x64_arm64 and with the attached (very\\n>> naive patch) and I still get an x64 binary :(\\n>\\n>\\n> With this patch I still get a build error, but it's different :-)\\n>\\n>\\n> [1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n> \\\"link\\\" @src/backend/postgres.exe.rsp\\n> Creating library src\\\\backend\\\\postgres.exe.lib\\n>\\n> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external\\n> symbol spin_delay referenced in function perform_spin_delay\\n>\\n> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n>\\n>\\n> Did you add the latest lock.patch ?\\n>\\n>\\n\\n\\nI'm a bit confused about exactly what needs to be applied. Can you \\nsupply a complete patch to be applied to a pristine checkout that will \\nlet me build?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-01-29 Mo 11:20, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Mon, 29 Jan 2024 at\\n 11:16, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 26 Jan\\n 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n > On Thu, Jan 25, 2024 at 04:52:30PM -0500,\\n Dave Cramer wrote:\\n >> On Thu, 25 Jan 2024 at 16:32, Andrew\\n Dunstan <andrew@dunslane.net>\\n wrote:\\n >>> On 2024-01-25 Th 16:17, Dave Cramer\\n wrote:\\n >>> Yeah, I think the default Developer\\n Command Prompt for VS2022 is set up\\n >>> for x86 builds. AIUI you should\\n start by executing \\\"vcvarsall x64_arm64\\\".\\n >> Yup, now I'm in the same state you are\\n > Wait a minute here. Based on [1],\\n x64_arm64 means you can use a x64\\n > host and you'll be able to produce ARM64\\n builds, still these will not\\n > be able to run on the host where they were\\n built. How much of the\\n > patch posted upthread is required to\\n produce such builds? Basically\\n > everything from it, I guess, so as build\\n dependencies can be\\n > satisfied?\\n >\\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n\\n\\n If you look at the table here x86 and x64 are\\n the only supported host \\n architectures. But that's OK, the x64 binaries\\n will run on arm64 (W11 \\n ARM64 has x64 emulation builtin). If that didn't\\n work Dave and I would \\n not have got as far as we have. But you want the\\n x64_arm64 argument to \\n vcvarsall so you will get ARM64 output.\\n\\n\\n\\nI've rebuilt it using x64_arm64 and with the\\n attached (very naive patch) and I still get an\\n x64 binary :(\\n\\n\\n\\n\\n\\n\\nWith this patch I still get a build error, but it's\\n different :-)\\n\\n\\n[1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe\\n src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library src\\\\backend\\\\postgres.exe.lib\\n\\n storage_lmgr_s_lock.c.obj : error LNK2019: unresolved\\n external symbol spin_delay referenced in function\\n perform_spin_delay\\n\\n src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1\\n unresolved externals\\n\\n\\n\\n\\n\\nDid you add the latest lock.patch ?\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nI'm a bit confused about exactly what needs to be applied. Can\\n you supply a complete patch to be applied to a pristine checkout\\n that will let me build?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 30 Jan 2024 08:38:40 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\\n>\\n>\\n> Dave Cramer\\n> www.postgres.rocks\\n>\\n>\\n> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>>\\n>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n>>\\n>>\\n>>\\n>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>\\n>>>\\n>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\\n>>> wrote:\\n>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>>> >>> Yeah, I think the default Developer Command Prompt for VS2022 is set\\n>>> up\\n>>> >>> for x86 builds. AIUI you should start by executing \\\"vcvarsall\\n>>> x64_arm64\\\".\\n>>> >> Yup, now I'm in the same state you are\\n>>> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\\n>>> > host and you'll be able to produce ARM64 builds, still these will not\\n>>> > be able to run on the host where they were built. How much of the\\n>>> > patch posted upthread is required to produce such builds? Basically\\n>>> > everything from it, I guess, so as build dependencies can be\\n>>> > satisfied?\\n>>> >\\n>>> > [1]:\\n>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>>>\\n>>>\\n>>> If you look at the table here x86 and x64 are the only supported host\\n>>> architectures. But that's OK, the x64 binaries will run on arm64 (W11\\n>>> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\\n>>> not have got as far as we have. But you want the x64_arm64 argument to\\n>>> vcvarsall so you will get ARM64 output.\\n>>>\\n>>\\n>> I've rebuilt it using x64_arm64 and with the attached (very naive patch)\\n>> and I still get an x64 binary :(\\n>>\\n>>\\n>> With this patch I still get a build error, but it's different :-)\\n>>\\n>>\\n>> [1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n>> \\\"link\\\" @src/backend/postgres.exe.rsp\\n>> Creating library src\\\\backend\\\\postgres.exe.lib\\n>>\\n>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\\n>> spin_delay referenced in function perform_spin_delay\\n>>\\n>> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n>>\\n>\\n> Did you add the latest lock.patch ?\\n>\\n>\\n>\\n>\\n> I'm a bit confused about exactly what needs to be applied. Can you supply\\n> a complete patch to be applied to a pristine checkout that will let me\\n> build?\\n>\\n>\\n> cheers\\n>\\n\\nSee attached.\\n\\nDave\",\n \"msg_date\": \"Tue, 30 Jan 2024 09:50:10 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-01-30 Tu 09:50, Dave Cramer wrote:\\n>\\n>\\n> On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>\\n> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\\n>>\\n>> Dave Cramer\\n>> www.postgres.rocks <http://www.postgres.rocks>\\n>>\\n>>\\n>> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan\\n>> <andrew@dunslane.net> wrote:\\n>>\\n>>\\n>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n>>>\\n>>>\\n>>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan\\n>>> <andrew@dunslane.net> wrote:\\n>>>\\n>>>\\n>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer\\n>>> wrote:\\n>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\\n>>> <andrew@dunslane.net> wrote:\\n>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>>> >>> Yeah, I think the default Developer Command Prompt\\n>>> for VS2022 is set up\\n>>> >>> for x86 builds. AIUI you should start by executing\\n>>> \\\"vcvarsall x64_arm64\\\".\\n>>> >> Yup, now I'm in the same state you are\\n>>> > Wait a minute here. Based on [1], x64_arm64 means you\\n>>> can use a x64\\n>>> > host and you'll be able to produce ARM64 builds, still\\n>>> these will not\\n>>> > be able to run on the host where they were built. How\\n>>> much of the\\n>>> > patch posted upthread is required to produce such\\n>>> builds? Basically\\n>>> > everything from it, I guess, so as build dependencies\\n>>> can be\\n>>> > satisfied?\\n>>> >\\n>>> > [1]:\\n>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>>>\\n>>>\\n>>> If you look at the table here x86 and x64 are the only\\n>>> supported host\\n>>> architectures. But that's OK, the x64 binaries will run\\n>>> on arm64 (W11\\n>>> ARM64 has x64 emulation builtin). If that didn't work\\n>>> Dave and I would\\n>>> not have got as far as we have. But you want the\\n>>> x64_arm64 argument to\\n>>> vcvarsall so you will get ARM64 output.\\n>>>\\n>>>\\n>>> I've rebuilt it using x64_arm64 and with the attached (very\\n>>> naive patch) and I still get an x64 binary :(\\n>>\\n>>\\n>> With this patch I still get a build error, but it's different :-)\\n>>\\n>>\\n>> [1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n>> \\\"link\\\" @src/backend/postgres.exe.rsp\\n>> Creating library src\\\\backend\\\\postgres.exe.lib\\n>>\\n>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved\\n>> external symbol spin_delay referenced in function\\n>> perform_spin_delay\\n>>\\n>> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved\\n>> externals\\n>>\\n>>\\n>> Did you add the latest lock.patch ?\\n>>\\n>>\\n>\\n>\\n> I'm a bit confused about exactly what needs to be applied. Can you\\n> supply a complete patch to be applied to a pristine checkout that\\n> will let me build?\\n>\\n>\\n> cheers\\n>\\n>\\n> See attached.\\n>\\n>\\n\\nNo, that is what is giving me the error shown above (just tried again to \\nbe certain). And it's not surprising, as patch 2 #ifdef's out the \\ndefinition of spin_delay().\\n\\nIf you can get a complete build with these patches then I suspect you're \\nnot doing a proper ARM64 build.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-01-30 Tu 09:50, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Tue, 30 Jan 2024 at\\n 08:38, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-01-29 Mo 11:20, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Mon, 29 Jan\\n 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn\\n Fri, 26 Jan 2024 at 07:36, Andrew\\n Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n On 2024-01-25 Th 20:32, Michael\\n Paquier wrote:\\n > On Thu, Jan 25, 2024 at\\n 04:52:30PM -0500, Dave Cramer wrote:\\n >> On Thu, 25 Jan 2024 at 16:32,\\n Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n >>> On 2024-01-25 Th 16:17,\\n Dave Cramer wrote:\\n >>> Yeah, I think the default\\n Developer Command Prompt for VS2022 is\\n set up\\n >>> for x86 builds. AIUI you\\n should start by executing \\\"vcvarsall\\n x64_arm64\\\".\\n >> Yup, now I'm in the same\\n state you are\\n > Wait a minute here. Based on\\n [1], x64_arm64 means you can use a x64\\n > host and you'll be able to\\n produce ARM64 builds, still these will\\n not\\n > be able to run on the host where\\n they were built. How much of the\\n > patch posted upthread is required\\n to produce such builds? Basically\\n > everything from it, I guess, so\\n as build dependencies can be\\n > satisfied?\\n >\\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n\\n\\n If you look at the table here x86 and\\n x64 are the only supported host \\n architectures. But that's OK, the x64\\n binaries will run on arm64 (W11 \\n ARM64 has x64 emulation builtin). If\\n that didn't work Dave and I would \\n not have got as far as we have. But\\n you want the x64_arm64 argument to \\n vcvarsall so you will get ARM64\\n output.\\n\\n\\n\\nI've rebuilt it using x64_arm64\\n and with the attached (very naive\\n patch) and I still get an x64 binary\\n :(\\n\\n\\n\\n\\n\\n\\nWith this patch I still get a build error,\\n but it's different :-)\\n\\n\\n[1406/2088] \\\"link\\\"\\n @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe\\n src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library\\n src\\\\backend\\\\postgres.exe.lib\\n\\n storage_lmgr_s_lock.c.obj : error LNK2019:\\n unresolved external symbol spin_delay\\n referenced in function perform_spin_delay\\n\\n src\\\\backend\\\\postgres.exe : fatal error\\n LNK1120: 1 unresolved externals\\n\\n\\n\\n\\n\\nDid you add the latest lock.patch ?\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nI'm a bit confused about exactly what needs to be\\n applied. Can you supply a complete patch to be applied\\n to a pristine checkout that will let me build?\\n\\n\\ncheers\\n\\n\\n\\n\\nSee attached.\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nNo, that is what is giving me the error shown above (just tried\\n again to be certain). And it's not surprising, as patch 2 #ifdef's\\n out the definition of spin_delay().\\nIf you can get a complete build with these patches then I suspect\\n you're not doing a proper ARM64 build.\\n\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 30 Jan 2024 16:56:19 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 30, 2024 at 4:56 PM Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-01-30 Tu 09:50, Dave Cramer wrote:\\n>\\n>\\n>\\n> On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>>\\n>> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\\n>>\\n>>\\n>> Dave Cramer\\n>> www.postgres.rocks\\n>>\\n>>\\n>> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>\\n>>>\\n>>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n>>>\\n>>>\\n>>>\\n>>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net>\\n>>> wrote:\\n>>>\\n>>>>\\n>>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n>>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n>>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\\n>>>> wrote:\\n>>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>>>> >>> Yeah, I think the default Developer Command Prompt for VS2022 is\\n>>>> set up\\n>>>> >>> for x86 builds. AIUI you should start by executing \\\"vcvarsall\\n>>>> x64_arm64\\\".\\n>>>> >> Yup, now I'm in the same state you are\\n>>>> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\\n>>>> > host and you'll be able to produce ARM64 builds, still these will not\\n>>>> > be able to run on the host where they were built. How much of the\\n>>>> > patch posted upthread is required to produce such builds? Basically\\n>>>> > everything from it, I guess, so as build dependencies can be\\n>>>> > satisfied?\\n>>>> >\\n>>>> > [1]:\\n>>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>>>>\\n>>>>\\n>>>> If you look at the table here x86 and x64 are the only supported host\\n>>>> architectures. But that's OK, the x64 binaries will run on arm64 (W11\\n>>>> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\\n>>>> not have got as far as we have. But you want the x64_arm64 argument to\\n>>>> vcvarsall so you will get ARM64 output.\\n>>>>\\n>>>\\n>>> I've rebuilt it using x64_arm64 and with the attached (very naive\\n>>> patch) and I still get an x64 binary :(\\n>>>\\n>>>\\n>>> With this patch I still get a build error, but it's different :-)\\n>>>\\n>>>\\n>>> [1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n>>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n>>> \\\"link\\\" @src/backend/postgres.exe.rsp\\n>>> Creating library src\\\\backend\\\\postgres.exe.lib\\n>>>\\n>>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\\n>>> spin_delay referenced in function perform_spin_delay\\n>>>\\n>>> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n>>>\\n>>\\n>> Did you add the latest lock.patch ?\\n>>\\n>>\\n>>\\n>>\\n>> I'm a bit confused about exactly what needs to be applied. Can you supply\\n>> a complete patch to be applied to a pristine checkout that will let me\\n>> build?\\n>>\\n>>\\n>> cheers\\n>>\\n>\\n> See attached.\\n>\\n>\\n>\\n> No, that is what is giving me the error shown above (just tried again to\\n> be certain). And it's not surprising, as patch 2 #ifdef's out the\\n> definition of spin_delay().\\n>\\n> If you can get a complete build with these patches then I suspect you're\\n> not doing a proper ARM64 build.\\n>\\n>\\nOkay I will look when I get home in a week\\n\\nDave\\n\\n>\\n> cheers\\n>\\n>\\n> andrew\\n>\\n> --\\n> Andrew Dunstan\\n> EDB: https://www.enterprisedb.com\\n>\\n>\\n\\nOn Tue, Jan 30, 2024 at 4:56 PM Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-01-30 Tu 09:50, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\nOn Tue, 30 Jan 2024 at\\n 08:38, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-01-29 Mo 11:20, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Mon, 29 Jan\\n 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn\\n Fri, 26 Jan 2024 at 07:36, Andrew\\n Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n On 2024-01-25 Th 20:32, Michael\\n Paquier wrote:\\n > On Thu, Jan 25, 2024 at\\n 04:52:30PM -0500, Dave Cramer wrote:\\n >> On Thu, 25 Jan 2024 at 16:32,\\n Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n >>> On 2024-01-25 Th 16:17,\\n Dave Cramer wrote:\\n >>> Yeah, I think the default\\n Developer Command Prompt for VS2022 is\\n set up\\n >>> for x86 builds. AIUI you\\n should start by executing \\\"vcvarsall\\n x64_arm64\\\".\\n >> Yup, now I'm in the same\\n state you are\\n > Wait a minute here. Based on\\n [1], x64_arm64 means you can use a x64\\n > host and you'll be able to\\n produce ARM64 builds, still these will\\n not\\n > be able to run on the host where\\n they were built. How much of the\\n > patch posted upthread is required\\n to produce such builds? Basically\\n > everything from it, I guess, so\\n as build dependencies can be\\n > satisfied?\\n >\\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n\\n\\n If you look at the table here x86 and\\n x64 are the only supported host \\n architectures. But that's OK, the x64\\n binaries will run on arm64 (W11 \\n ARM64 has x64 emulation builtin). If\\n that didn't work Dave and I would \\n not have got as far as we have. But\\n you want the x64_arm64 argument to \\n vcvarsall so you will get ARM64\\n output.\\n\\n\\n\\nI've rebuilt it using x64_arm64\\n and with the attached (very naive\\n patch) and I still get an x64 binary\\n :(\\n\\n\\n\\n\\n\\n\\nWith this patch I still get a build error,\\n but it's different :-)\\n\\n\\n[1406/2088] \\\"link\\\"\\n @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe\\n src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library\\n src\\\\backend\\\\postgres.exe.lib\\n\\n storage_lmgr_s_lock.c.obj : error LNK2019:\\n unresolved external symbol spin_delay\\n referenced in function perform_spin_delay\\n\\n src\\\\backend\\\\postgres.exe : fatal error\\n LNK1120: 1 unresolved externals\\n\\n\\n\\n\\n\\nDid you add the latest lock.patch ?\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nI'm a bit confused about exactly what needs to be\\n applied. Can you supply a complete patch to be applied\\n to a pristine checkout that will let me build?\\n\\n\\ncheers\\n\\n\\n\\n\\nSee attached.\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nNo, that is what is giving me the error shown above (just tried\\n again to be certain). And it's not surprising, as patch 2 #ifdef's\\n out the definition of spin_delay().\\nIf you can get a complete build with these patches then I suspect\\n you're not doing a proper ARM64 build.\\n\\nOkay I will look when I get home in a weekDave \\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Tue, 30 Jan 2024 17:54:01 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-01-30 Tu 17:54, Dave Cramer wrote:\\n>\\n>\\n>\\n> On Tue, Jan 30, 2024 at 4:56 PM Andrew Dunstan <andrew@dunslane.net> \\n> wrote:\\n>\\n>\\n> On 2024-01-30 Tu 09:50, Dave Cramer wrote:\\n>>\\n>>\\n>> On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan\\n>> <andrew@dunslane.net> wrote:\\n>>\\n>>\\n>> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\\n>>>\\n>>> Dave Cramer\\n>>> www.postgres.rocks <http://www.postgres.rocks>\\n>>>\\n>>>\\n>>> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan\\n>>> <andrew@dunslane.net> wrote:\\n>>>\\n>>>\\n>>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n>>>>\\n>>>>\\n>>>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan\\n>>>> <andrew@dunslane.net> wrote:\\n>>>>\\n>>>>\\n>>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n>>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave\\n>>>> Cramer wrote:\\n>>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan\\n>>>> <andrew@dunslane.net> wrote:\\n>>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>>>> >>> Yeah, I think the default Developer Command\\n>>>> Prompt for VS2022 is set up\\n>>>> >>> for x86 builds. AIUI you should start by\\n>>>> executing \\\"vcvarsall x64_arm64\\\".\\n>>>> >> Yup, now I'm in the same state you are\\n>>>> > Wait a minute here. Based on [1], x64_arm64 means\\n>>>> you can use a x64\\n>>>> > host and you'll be able to produce ARM64 builds,\\n>>>> still these will not\\n>>>> > be able to run on the host where they were built.\\n>>>> How much of the\\n>>>> > patch posted upthread is required to produce such\\n>>>> builds? Basically\\n>>>> > everything from it, I guess, so as build\\n>>>> dependencies can be\\n>>>> > satisfied?\\n>>>> >\\n>>>> > [1]:\\n>>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>>>>\\n>>>>\\n>>>> If you look at the table here x86 and x64 are the\\n>>>> only supported host\\n>>>> architectures. But that's OK, the x64 binaries will\\n>>>> run on arm64 (W11\\n>>>> ARM64 has x64 emulation builtin). If that didn't\\n>>>> work Dave and I would\\n>>>> not have got as far as we have. But you want the\\n>>>> x64_arm64 argument to\\n>>>> vcvarsall so you will get ARM64 output.\\n>>>>\\n>>>>\\n>>>> I've rebuilt it using x64_arm64 and with the attached\\n>>>> (very naive patch) and I still get an x64 binary :(\\n>>>\\n>>>\\n>>> With this patch I still get a build error, but it's\\n>>> different :-)\\n>>>\\n>>>\\n>>> [1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n>>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n>>> \\\"link\\\" @src/backend/postgres.exe.rsp\\n>>> Creating library src\\\\backend\\\\postgres.exe.lib\\n>>>\\n>>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved\\n>>> external symbol spin_delay referenced in function\\n>>> perform_spin_delay\\n>>>\\n>>> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1\\n>>> unresolved externals\\n>>>\\n>>>\\n>>> Did you add the latest lock.patch ?\\n>>>\\n>>>\\n>>\\n>>\\n>> I'm a bit confused about exactly what needs to be applied.\\n>> Can you supply a complete patch to be applied to a pristine\\n>> checkout that will let me build?\\n>>\\n>>\\n>> cheers\\n>>\\n>>\\n>> See attached.\\n>>\\n>>\\n>\\n> No, that is what is giving me the error shown above (just tried\\n> again to be certain). And it's not surprising, as patch 2 #ifdef's\\n> out the definition of spin_delay().\\n>\\n> If you can get a complete build with these patches then I suspect\\n> you're not doing a proper ARM64 build.\\n>\\n>\\n> Okay I will look when I get home in a week\\n\\n\\nI made some progress. The attached is mostly taken from \\n<https://postgr.es/m/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56@linaro.org>\\n\\nWith it applied I was able to get a successful build using the buildfarm \\nclient. However, there are access violations when running some tests, so \\nthere is still some work to do, apparently.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\",\n \"msg_date\": \"Wed, 31 Jan 2024 10:20:54 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 31.01.24 16:20, Andrew Dunstan wrote:\\n> - PostgreSQL will only build for the x64 architecture on 64-bit Windows. \\n> + PostgreSQL will only build for the x64 and ARM64 architecture on \\n> 64-bit Windows.\\n\\nAre there any other 64-bit architectures for Windows?\\n\\nPossibly, the original sentence was meant to communicate that ARM was \\nnot supported, in which case it could now be removed?\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 31 Jan 2024 16:34:13 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter@eisentraut.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"\\nOn 2024-01-31 We 10:34, Peter Eisentraut wrote:\\n> On 31.01.24 16:20, Andrew Dunstan wrote:\\n>> - PostgreSQL will only build for the x64 architecture on 64-bit \\n>> Windows. + PostgreSQL will only build for the x64 and ARM64 \\n>> architecture on 64-bit Windows.\\n>\\n> Are there any other 64-bit architectures for Windows?\\n>\\n> Possibly, the original sentence was meant to communicate that ARM was \\n> not supported, in which case it could now be removed?\\n\\n\\nx86? That is in fact the default setting for VS even on ARM64.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 31 Jan 2024 10:41:57 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Wed, 31 Jan 2024 at 10:21, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-01-30 Tu 17:54, Dave Cramer wrote:\\n>\\n>\\n>\\n>\\n> On Tue, Jan 30, 2024 at 4:56 PM Andrew Dunstan <andrew@dunslane.net>\\n> wrote:\\n>\\n>>\\n>> On 2024-01-30 Tu 09:50, Dave Cramer wrote:\\n>>\\n>>\\n>>\\n>> On Tue, 30 Jan 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>\\n>>>\\n>>> On 2024-01-29 Mo 11:20, Dave Cramer wrote:\\n>>>\\n>>>\\n>>> Dave Cramer\\n>>> www.postgres.rocks\\n>>>\\n>>>\\n>>> On Mon, 29 Jan 2024 at 11:16, Andrew Dunstan <andrew@dunslane.net>\\n>>> wrote:\\n>>>\\n>>>>\\n>>>> On 2024-01-26 Fr 09:18, Dave Cramer wrote:\\n>>>>\\n>>>>\\n>>>>\\n>>>> On Fri, 26 Jan 2024 at 07:36, Andrew Dunstan <andrew@dunslane.net>\\n>>>> wrote:\\n>>>>\\n>>>>>\\n>>>>> On 2024-01-25 Th 20:32, Michael Paquier wrote:\\n>>>>> > On Thu, Jan 25, 2024 at 04:52:30PM -0500, Dave Cramer wrote:\\n>>>>> >> On Thu, 25 Jan 2024 at 16:32, Andrew Dunstan <andrew@dunslane.net>\\n>>>>> wrote:\\n>>>>> >>> On 2024-01-25 Th 16:17, Dave Cramer wrote:\\n>>>>> >>> Yeah, I think the default Developer Command Prompt for VS2022 is\\n>>>>> set up\\n>>>>> >>> for x86 builds. AIUI you should start by executing \\\"vcvarsall\\n>>>>> x64_arm64\\\".\\n>>>>> >> Yup, now I'm in the same state you are\\n>>>>> > Wait a minute here. Based on [1], x64_arm64 means you can use a x64\\n>>>>> > host and you'll be able to produce ARM64 builds, still these will not\\n>>>>> > be able to run on the host where they were built. How much of the\\n>>>>> > patch posted upthread is required to produce such builds? Basically\\n>>>>> > everything from it, I guess, so as build dependencies can be\\n>>>>> > satisfied?\\n>>>>> >\\n>>>>> > [1]:\\n>>>>> https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n>>>>>\\n>>>>>\\n>>>>> If you look at the table here x86 and x64 are the only supported host\\n>>>>> architectures. But that's OK, the x64 binaries will run on arm64 (W11\\n>>>>> ARM64 has x64 emulation builtin). If that didn't work Dave and I would\\n>>>>> not have got as far as we have. But you want the x64_arm64 argument to\\n>>>>> vcvarsall so you will get ARM64 output.\\n>>>>>\\n>>>>\\n>>>> I've rebuilt it using x64_arm64 and with the attached (very naive\\n>>>> patch) and I still get an x64 binary :(\\n>>>>\\n>>>>\\n>>>> With this patch I still get a build error, but it's different :-)\\n>>>>\\n>>>>\\n>>>> [1406/2088] \\\"link\\\" @src/backend/postgres.exe.rsp\\n>>>> FAILED: src/backend/postgres.exe src/backend/postgres.pdb\\n>>>> \\\"link\\\" @src/backend/postgres.exe.rsp\\n>>>> Creating library src\\\\backend\\\\postgres.exe.lib\\n>>>>\\n>>>> storage_lmgr_s_lock.c.obj : error LNK2019: unresolved external symbol\\n>>>> spin_delay referenced in function perform_spin_delay\\n>>>>\\n>>>> src\\\\backend\\\\postgres.exe : fatal error LNK1120: 1 unresolved externals\\n>>>>\\n>>>\\n>>> Did you add the latest lock.patch ?\\n>>>\\n>>>\\n>>>\\n>>>\\n>>> I'm a bit confused about exactly what needs to be applied. Can you\\n>>> supply a complete patch to be applied to a pristine checkout that will let\\n>>> me build?\\n>>>\\n>>>\\n>>> cheers\\n>>>\\n>>\\n>> See attached.\\n>>\\n>>\\n>>\\n>> No, that is what is giving me the error shown above (just tried again to\\n>> be certain). And it's not surprising, as patch 2 #ifdef's out the\\n>> definition of spin_delay().\\n>>\\n>> If you can get a complete build with these patches then I suspect you're\\n>> not doing a proper ARM64 build.\\n>>\\n>\\n> Okay I will look when I get home in a week\\n>\\n>\\n> I made some progress. The attached is mostly taken from\\n> <https://postgr.es/m/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56@linaro.org>\\n> <https://postgr.es/m/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56@linaro.org>\\n>\\n> With it applied I was able to get a successful build using the buildfarm\\n> client. However, there are access violations when running some tests, so\\n> there is still some work to do, apparently.\\n>\\n>\\n> cheers\\n>\\n>\\n> andrew\\n>\\n> --\\n> Andrew Dunstan\\n> EDB: https://www.enterprisedb.com\\n>\\n>\\n\\nThanks, this patch works and\\ntesting with meon passes.\\n\\nI'll try the buildfarm next.\\n\\nDave\\n\\nOn Wed, 31 Jan 2024 at 10:21, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-01-30 Tu 17:54, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Tue, Jan 30, 2024 at\\n 4:56 PM Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-01-30 Tu 09:50, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\nOn Tue, 30 Jan\\n 2024 at 08:38, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-01-29 Mo 11:20, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Mon,\\n 29 Jan 2024 at 11:16, Andrew Dunstan\\n <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-01-26 Fr 09:18, Dave\\n Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri,\\n 26 Jan 2024 at 07:36, Andrew\\n Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n On 2024-01-25 Th 20:32,\\n Michael Paquier wrote:\\n > On Thu, Jan 25, 2024 at\\n 04:52:30PM -0500, Dave\\n Cramer wrote:\\n >> On Thu, 25 Jan 2024\\n at 16:32, Andrew Dunstan\\n <andrew@dunslane.net>\\n wrote:\\n >>> On 2024-01-25\\n Th 16:17, Dave Cramer wrote:\\n >>> Yeah, I think\\n the default Developer\\n Command Prompt for VS2022 is\\n set up\\n >>> for x86 builds.\\n AIUI you should start by\\n executing \\\"vcvarsall\\n x64_arm64\\\".\\n >> Yup, now I'm in the\\n same state you are\\n > Wait a minute here. \\n Based on [1], x64_arm64\\n means you can use a x64\\n > host and you'll be able\\n to produce ARM64 builds,\\n still these will not\\n > be able to run on the\\n host where they were built. \\n How much of the\\n > patch posted upthread\\n is required to produce such\\n builds? Basically\\n > everything from it, I\\n guess, so as build\\n dependencies can be\\n > satisfied?\\n >\\n > [1]: https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170\\n\\n\\n If you look at the table\\n here x86 and x64 are the\\n only supported host \\n architectures. But that's\\n OK, the x64 binaries will\\n run on arm64 (W11 \\n ARM64 has x64 emulation\\n builtin). If that didn't\\n work Dave and I would \\n not have got as far as we\\n have. But you want the\\n x64_arm64 argument to \\n vcvarsall so you will get\\n ARM64 output.\\n\\n\\n\\nI've rebuilt it using \\n x64_arm64 and with the\\n attached (very naive patch)\\n and I still get an x64\\n binary :(\\n\\n\\n\\n\\n\\n\\nWith this patch I still get a build\\n error, but it's different :-)\\n\\n\\n[1406/2088] \\\"link\\\"\\n @src/backend/postgres.exe.rsp\\n FAILED: src/backend/postgres.exe\\n src/backend/postgres.pdb\\n \\\"link\\\" @src/backend/postgres.exe.rsp\\n Creating library\\n src\\\\backend\\\\postgres.exe.lib\\n\\n storage_lmgr_s_lock.c.obj : error\\n LNK2019: unresolved external symbol\\n spin_delay referenced in function\\n perform_spin_delay\\n\\n src\\\\backend\\\\postgres.exe : fatal\\n error LNK1120: 1 unresolved\\n externals\\n\\n\\n\\n\\n\\nDid you add the latest lock.patch ?\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nI'm a bit confused about exactly what needs\\n to be applied. Can you supply a complete patch\\n to be applied to a pristine checkout that will\\n let me build?\\n\\n\\ncheers\\n\\n\\n\\n\\nSee attached.\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nNo, that is what is giving me the error shown above\\n (just tried again to be certain). And it's not\\n surprising, as patch 2 #ifdef's out the definition of\\n spin_delay().\\nIf you can get a complete build with these patches then\\n I suspect you're not doing a proper ARM64 build.\\n\\n\\n \\n\\n\\n\\nOkay I will look when I get home in a week\\n\\n\\n\\n\\n\\nI made some progress. The attached is mostly taken from \\n<https://postgr.es/m/dbee741f-b9b7-a0d5-1b1b-f9b532bb6f56@linaro.org>\\nWith it applied I was able to get a successful build using the\\n buildfarm client. However, there are access violations when\\n running some tests, so there is still some work to do, apparently.\\n\\n\\ncheers\\n\\n\\nandrew\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.comThanks, this patch works and testing with meon passes. I'll try the buildfarm next.Dave\",\n \"msg_date\": \"Tue, 6 Feb 2024 07:01:49 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n> Thanks, this patch works and\\n> testing with meson passes.\\n\\nOnly with the version posted at [1]? Interesting, that's the same\\ncontents as v8 posted upthread, minus src/tools/ because we don't need\\nto care about them anymore.\\n\\nAndrew, what's happening on the test side? It does not seem you've\\nmentioned any details about what is going wrong, or I have just missed\\nthem.\\n\\n> I'll try the buildfarm next.\\n\\n[1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n--\\nMichael\",\n \"msg_date\": \"Fri, 9 Feb 2024 14:26:34 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n> > Thanks, this patch works and\\n> > testing with meson passes.\\n>\\n> Only with the version posted at [1]? Interesting, that's the same\\n> contents as v8 posted upthread, minus src/tools/ because we don't need\\n> to care about them anymore.\\n>\\n> Andrew, what's happening on the test side? It does not seem you've\\n> mentioned any details about what is going wrong, or I have just missed\\n> them.\\n>\\n> > I'll try the buildfarm next.\\n>\\n> [1]:\\n> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n\\n\\ninterestingly meson test does not produce any error\\nThe buildfarm produces the following error for me:\\n\\n-SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n- FROM check_columns\\n- WHERE get_columns_length(coltypes) % 8 != 0 OR\\n- 'name'::regtype::oid = ANY(coltypes);\\n- relname | attname | coltypes | get_columns_length\\n----------+---------+----------+--------------------\\n-(0 rows)\\n-\\n+server closed the connection unexpectedly\\n+ This probably means the server terminated abnormally\\n+ before or while processing the request.\\n+connection to server was lost\\n\\nDave\\n\\nOn Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n> Thanks, this patch works and\\n> testing with meson passes.\\n\\nOnly with the version posted at [1]? Interesting, that's the same\\ncontents as v8 posted upthread, minus src/tools/ because we don't need\\nto care about them anymore.\\n\\nAndrew, what's happening on the test side? It does not seem you've\\nmentioned any details about what is going wrong, or I have just missed\\nthem.\\n\\n> I'll try the buildfarm next.\\n\\n[1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.netinterestingly meson test does not produce any errorThe buildfarm produces the following error for me:-SELECT relname, attname, coltypes, get_columns_length(coltypes)- FROM check_columns- WHERE get_columns_length(coltypes) % 8 != 0 OR- 'name'::regtype::oid = ANY(coltypes);- relname | attname | coltypes | get_columns_length ----------+---------+----------+---------------------(0 rows)-+server closed the connection unexpectedly+ This probably means the server terminated abnormally+ before or while processing the request.+connection to server was lost Dave\",\n \"msg_date\": \"Fri, 9 Feb 2024 07:18:13 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Dave Cramer\\nwww.postgres.rocks\\n\\n\\nOn Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks> wrote:\\n\\n>\\n>\\n>\\n>\\n> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n>> > Thanks, this patch works and\\n>> > testing with meson passes.\\n>>\\n>> Only with the version posted at [1]? Interesting, that's the same\\n>> contents as v8 posted upthread, minus src/tools/ because we don't need\\n>> to care about them anymore.\\n>>\\n>> Andrew, what's happening on the test side? It does not seem you've\\n>> mentioned any details about what is going wrong, or I have just missed\\n>> them.\\n>>\\n>> > I'll try the buildfarm next.\\n>>\\n>> [1]:\\n>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n>\\n>\\n> interestingly meson test does not produce any error\\n> The buildfarm produces the following error for me:\\n>\\n> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n> - FROM check_columns\\n> - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n> - 'name'::regtype::oid = ANY(coltypes);\\n> - relname | attname | coltypes | get_columns_length\\n> ----------+---------+----------+--------------------\\n> -(0 rows)\\n> -\\n> +server closed the connection unexpectedly\\n> + This probably means the server terminated abnormally\\n> + before or while processing the request.\\n> +connection to server was lost\\n>\\n\\nActually digging some more, here is the actual error\\n\\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\\n11204) was terminated by exception 0xC0000005\\n2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\\nrunning: VACUUM;\\n2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\\n\\\"ntstatus.h\\\" for a description of the hexadecimal value.\\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other\\nactive server processes\\n2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes\\nterminated; reinitializing\\n2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was\\ninterrupted; last known up at 2024-02-09 13:31:01 -05\\n\\n\\nDave\\n\\n>\\n> Dave\\n>\\n\\nDave Cramerwww.postgres.rocksOn Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks> wrote:On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n> Thanks, this patch works and\\n> testing with meson passes.\\n\\nOnly with the version posted at [1]? Interesting, that's the same\\ncontents as v8 posted upthread, minus src/tools/ because we don't need\\nto care about them anymore.\\n\\nAndrew, what's happening on the test side? It does not seem you've\\nmentioned any details about what is going wrong, or I have just missed\\nthem.\\n\\n> I'll try the buildfarm next.\\n\\n[1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.netinterestingly meson test does not produce any errorThe buildfarm produces the following error for me:-SELECT relname, attname, coltypes, get_columns_length(coltypes)- FROM check_columns- WHERE get_columns_length(coltypes) % 8 != 0 OR- 'name'::regtype::oid = ANY(coltypes);- relname | attname | coltypes | get_columns_length ----------+---------+----------+---------------------(0 rows)-+server closed the connection unexpectedly+ This probably means the server terminated abnormally+ before or while processing the request.+connection to server was lost Actually digging some more, here is the actual error2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \\\"ntstatus.h\\\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05Dave Dave\",\n \"msg_date\": \"Fri, 9 Feb 2024 14:23:46 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2024-02-09 14:23:46 -0500, Dave Cramer wrote:\\n> > interestingly meson test does not produce any error\\n> > The buildfarm produces the following error for me:\\n> >\\n> > -SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n> > - FROM check_columns\\n> > - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n> > - 'name'::regtype::oid = ANY(coltypes);\\n> > - relname | attname | coltypes | get_columns_length\\n> > ----------+---------+----------+--------------------\\n> > -(0 rows)\\n> > -\\n> > +server closed the connection unexpectedly\\n> > + This probably means the server terminated abnormally\\n> > + before or while processing the request.\\n> > +connection to server was lost\\n> >\\n> \\n> Actually digging some more, here is the actual error\\n> \\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\\n> 11204) was terminated by exception 0xC0000005\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\\n> running: VACUUM;\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\\n> \\\"ntstatus.h\\\" for a description of the hexadecimal value.\\n\\nThat's something like a segfault.\\n\\nOne suspicion I have is that src/port/pg_crc32c_armv8_choose.c possibly\\ndoesn't properly support msvc. It seems to assume that SIGILL can be trapped,\\nbut that IIRC doesn't work on windows.\\n\\nI'd check if the problem persists if you change\\ncdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\nto\\ncdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 0)\\n\\n\\nAlso, yikes, that's an ugly way of doing hardware detection. Jumping out of a\\nsignal handler into normal code. Brrr.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 9 Feb 2024 11:36:18 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Fri, 9 Feb 2024 at 14:36, Andres Freund <andres@anarazel.de> wrote:\\n\\n> Hi,\\n>\\n> On 2024-02-09 14:23:46 -0500, Dave Cramer wrote:\\n> > > interestingly meson test does not produce any error\\n> > > The buildfarm produces the following error for me:\\n> > >\\n> > > -SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n> > > - FROM check_columns\\n> > > - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n> > > - 'name'::regtype::oid = ANY(coltypes);\\n> > > - relname | attname | coltypes | get_columns_length\\n> > > ----------+---------+----------+--------------------\\n> > > -(0 rows)\\n> > > -\\n> > > +server closed the connection unexpectedly\\n> > > + This probably means the server terminated abnormally\\n> > > + before or while processing the request.\\n> > > +connection to server was lost\\n> > >\\n> >\\n> > Actually digging some more, here is the actual error\\n> >\\n> > 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\\n> > 11204) was terminated by exception 0xC0000005\\n> > 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\\n> > running: VACUUM;\\n> > 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\\n> > \\\"ntstatus.h\\\" for a description of the hexadecimal value.\\n>\\n> That's something like a segfault.\\n>\\n> One suspicion I have is that src/port/pg_crc32c_armv8_choose.c possibly\\n> doesn't properly support msvc. It seems to assume that SIGILL can be\\n> trapped,\\n> but that IIRC doesn't work on windows.\\n>\\n> I'd check if the problem persists if you change\\n> cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> to\\n> cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 0)\\n>\\n\\nThis results in\\n\\nFAILED: src/bin/pg_checksums/pg_checksums.exe\\nsrc/bin/pg_checksums/pg_checksums.pdb\\n\\\"link\\\" /MACHINE:ARM64 /OUT:src/bin/pg_checksums/pg_checksums.exe\\nsrc/bin/pg_checksums/pg_checksums.exe.p/win32ver.res\\nsrc/bin/pg_checksums/pg_checksums.exe.p/pg_checksums.c.obj \\\"/release\\\"\\n\\\"/nologo\\\" \\\"/DEBUG\\\" \\\"/PDB:src\\\\bin\\\\pg_checksums\\\\pg_checksums.pdb\\\"\\n\\\"/INCREMENTAL:NO\\\" \\\"/STACK:4194304\\\" \\\"/NOEXP\\\" \\\"src/fe_utils/libpgfeutils.a\\\"\\n\\\"src/common/libpgcommon.a\\\" \\\"src/port/libpgport.a\\\" \\\"ws2_32.lib\\\" \\\"ws2_32.lib\\\"\\n\\\"ws2_32.lib\\\" \\\"ws2_32.lib\\\" \\\"/SUBSYSTEM:CONSOLE\\\" \\\"kernel32.lib\\\" \\\"user32.lib\\\"\\n\\\"gdi32.lib\\\" \\\"winspool.lib\\\" \\\"shell32.lib\\\" \\\"ole32.lib\\\" \\\"oleaut32.lib\\\"\\n\\\"uuid.lib\\\" \\\"comdlg32.lib\\\" \\\"advapi32.lib\\\"\\nlibpgcommon.a(controldata_utils.c.obj) : error LNK2001: unresolved external\\nsymbol pg_comp_crc32c\\n\\n\\nDave\\n\\n\\n>\\n>\\n> Also, yikes, that's an ugly way of doing hardware detection. Jumping out\\n> of a\\n> signal handler into normal code. Brrr.\\n>\\n> Greetings,\\n>\\n> Andres Freund\\n>\\n\\nOn Fri, 9 Feb 2024 at 14:36, Andres Freund <andres@anarazel.de> wrote:Hi,\\n\\nOn 2024-02-09 14:23:46 -0500, Dave Cramer wrote:\\n> > interestingly meson test does not produce any error\\n> > The buildfarm produces the following error for me:\\n> >\\n> > -SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n> > - FROM check_columns\\n> > - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n> > - 'name'::regtype::oid = ANY(coltypes);\\n> > - relname | attname | coltypes | get_columns_length\\n> > ----------+---------+----------+--------------------\\n> > -(0 rows)\\n> > -\\n> > +server closed the connection unexpectedly\\n> > + This probably means the server terminated abnormally\\n> > + before or while processing the request.\\n> > +connection to server was lost\\n> >\\n> \\n> Actually digging some more, here is the actual error\\n> \\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\\n> 11204) was terminated by exception 0xC0000005\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\\n> running: VACUUM;\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\\n> \\\"ntstatus.h\\\" for a description of the hexadecimal value.\\n\\nThat's something like a segfault.\\n\\nOne suspicion I have is that src/port/pg_crc32c_armv8_choose.c possibly\\ndoesn't properly support msvc. It seems to assume that SIGILL can be trapped,\\nbut that IIRC doesn't work on windows.\\n\\nI'd check if the problem persists if you change\\ncdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\nto\\ncdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 0)This results in FAILED: src/bin/pg_checksums/pg_checksums.exe src/bin/pg_checksums/pg_checksums.pdb \\\"link\\\" /MACHINE:ARM64 /OUT:src/bin/pg_checksums/pg_checksums.exe src/bin/pg_checksums/pg_checksums.exe.p/win32ver.res src/bin/pg_checksums/pg_checksums.exe.p/pg_checksums.c.obj \\\"/release\\\" \\\"/nologo\\\" \\\"/DEBUG\\\" \\\"/PDB:src\\\\bin\\\\pg_checksums\\\\pg_checksums.pdb\\\" \\\"/INCREMENTAL:NO\\\" \\\"/STACK:4194304\\\" \\\"/NOEXP\\\" \\\"src/fe_utils/libpgfeutils.a\\\" \\\"src/common/libpgcommon.a\\\" \\\"src/port/libpgport.a\\\" \\\"ws2_32.lib\\\" \\\"ws2_32.lib\\\" \\\"ws2_32.lib\\\" \\\"ws2_32.lib\\\" \\\"/SUBSYSTEM:CONSOLE\\\" \\\"kernel32.lib\\\" \\\"user32.lib\\\" \\\"gdi32.lib\\\" \\\"winspool.lib\\\" \\\"shell32.lib\\\" \\\"ole32.lib\\\" \\\"oleaut32.lib\\\" \\\"uuid.lib\\\" \\\"comdlg32.lib\\\" \\\"advapi32.lib\\\"libpgcommon.a(controldata_utils.c.obj) : error LNK2001: unresolved external symbol pg_comp_crc32cDave \\n\\n\\nAlso, yikes, that's an ugly way of doing hardware detection. Jumping out of a\\nsignal handler into normal code. Brrr.\\n\\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Fri, 9 Feb 2024 15:32:10 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n>\\n> Dave Cramer\\n> www.postgres.rocks\\n>\\n>\\n> On Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks> \\n> wrote:\\n>\\n>\\n>\\n>\\n>\\n> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz>\\n> wrote:\\n>\\n> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n> > Thanks, this patch works and\\n> > testing with meson passes.\\n>\\n> Only with the version posted at [1]? Interesting, that's the same\\n> contents as v8 posted upthread, minus src/tools/ because we\\n> don't need\\n> to care about them anymore.\\n>\\n> Andrew, what's happening on the test side? It does not seem\\n> you've\\n> mentioned any details about what is going wrong, or I have\\n> just missed\\n> them.\\n>\\n> > I'll try the buildfarm next.\\n>\\n> [1]:\\n> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n>\\n>\\n> interestingly meson test does not produce any error\\n> The buildfarm produces the following error for me:\\n>\\n> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n> - FROM check_columns\\n> - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n> - 'name'::regtype::oid = ANY(coltypes);\\n> - relname | attname | coltypes | get_columns_length\\n> ----------+---------+----------+--------------------\\n> -(0 rows)\\n> -\\n> +server closed the connection unexpectedly\\n> +This probably means the server terminated abnormally\\n> +before or while processing the request.\\n> +connection to server was lost\\n>\\n>\\n> Actually digging some more, here is the actual error\\n>\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process \\n> (PID 11204) was terminated by exception 0xC0000005\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process \\n> was running: VACUUM;\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include \\n> file \\\"ntstatus.h\\\" for a description of the hexadecimal value.\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any \\n> other active server processes\\n> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server \\n> processes terminated; reinitializing\\n> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was \\n> interrupted; last known up at 2024-02-09 13:31:01 -05\\n>\\n>\\n>\\n\\nYes, this is pretty much what I saw.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-02-09 Fr 14:23, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 9 Feb 2024 at\\n 07:18, Dave Cramer <davecramer@postgres.rocks>\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 9 Feb 2024\\n at 00:26, Michael Paquier <michael@paquier.xyz>\\n wrote:\\n\\nOn Tue, Feb 06,\\n 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n > Thanks, this patch works and\\n > testing with meson passes.\\n\\n Only with the version posted at [1]? Interesting,\\n that's the same\\n contents as v8 posted upthread, minus src/tools/\\n because we don't need\\n to care about them anymore.\\n\\n Andrew, what's happening on the test side? It\\n does not seem you've\\n mentioned any details about what is going wrong,\\n or I have just missed\\n them.\\n\\n > I'll try the buildfarm next.\\n\\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n\\n\\ninterestingly meson test does not produce any\\n error\\nThe buildfarm produces the following error for\\n me:\\n\\n\\n-SELECT relname, attname, coltypes,\\n get_columns_length(coltypes)\\n- FROM check_columns\\n- WHERE get_columns_length(coltypes) % 8 != 0\\n OR\\n- 'name'::regtype::oid = ANY(coltypes);\\n- relname | attname | coltypes |\\n get_columns_length \\n----------+---------+----------+--------------------\\n-(0 rows)\\n-\\n+server closed the connection unexpectedly\\n+ This\\n probably means the server terminated abnormally\\n+ before\\n or while processing the request.\\n+connection to server was lost \\n\\n\\n\\n\\n\\n\\nActually digging some more, here is the actual error\\n\\n\\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \\\"ntstatus.h\\\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nYes, this is pretty much what I saw.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Sat, 10 Feb 2024 11:19:21 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n>\\n>\\n> Dave Cramer\\n> www.postgres.rocks\\n>\\n>\\n> On Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\\n> <davecramer@postgres.rocks> wrote:\\n>\\n>>\\n>>\\n>>\\n>>\\n>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz> wrote:\\n>>\\n>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n>>> > Thanks, this patch works and\\n>>> > testing with meson passes.\\n>>>\\n>>> Only with the version posted at [1]? Interesting, that's the same\\n>>> contents as v8 posted upthread, minus src/tools/ because we don't need\\n>>> to care about them anymore.\\n>>>\\n>>> Andrew, what's happening on the test side? It does not seem you've\\n>>> mentioned any details about what is going wrong, or I have just missed\\n>>> them.\\n>>>\\n>>> > I'll try the buildfarm next.\\n>>>\\n>>> [1]:\\n>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n>>\\n>>\\n>> interestingly meson test does not produce any error\\n>> The buildfarm produces the following error for me:\\n>>\\n>> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n>> - FROM check_columns\\n>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n>> - 'name'::regtype::oid = ANY(coltypes);\\n>> - relname | attname | coltypes | get_columns_length\\n>> ----------+---------+----------+--------------------\\n>> -(0 rows)\\n>> -\\n>> +server closed the connection unexpectedly\\n>> + This probably means the server terminated abnormally\\n>> + before or while processing the request.\\n>> +connection to server was lost\\n>>\\n>\\n> Actually digging some more, here is the actual error\\n>\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\\n> 11204) was terminated by exception 0xC0000005\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was\\n> running: VACUUM;\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\\n> \\\"ntstatus.h\\\" for a description of the hexadecimal value.\\n> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other\\n> active server processes\\n> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes\\n> terminated; reinitializing\\n> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was\\n> interrupted; last known up at 2024-02-09 13:31:01 -05\\n>\\n>\\n>\\n>\\n>\\nSo how does one debug this ?\\n\\nAlso if I `run meson test` I don't see this error. What does the buildfarm\\ndo differently?\\n\\nDave\\n\\n> Yes, this is pretty much what I saw.\\n>\\n>\\n> cheers\\n>\\n>\\n> andrew\\n>\\n> --\\n> Andrew Dunstan\\n> EDB: https://www.enterprisedb.com\\n>\\n>\\n\\nOn Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-02-09 Fr 14:23, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 9 Feb 2024 at\\n 07:18, Dave Cramer <davecramer@postgres.rocks>\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 9 Feb 2024\\n at 00:26, Michael Paquier <michael@paquier.xyz>\\n wrote:\\n\\nOn Tue, Feb 06,\\n 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n > Thanks, this patch works and\\n > testing with meson passes.\\n\\n Only with the version posted at [1]? Interesting,\\n that's the same\\n contents as v8 posted upthread, minus src/tools/\\n because we don't need\\n to care about them anymore.\\n\\n Andrew, what's happening on the test side? It\\n does not seem you've\\n mentioned any details about what is going wrong,\\n or I have just missed\\n them.\\n\\n > I'll try the buildfarm next.\\n\\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n\\n\\ninterestingly meson test does not produce any\\n error\\nThe buildfarm produces the following error for\\n me:\\n\\n\\n-SELECT relname, attname, coltypes,\\n get_columns_length(coltypes)\\n- FROM check_columns\\n- WHERE get_columns_length(coltypes) % 8 != 0\\n OR\\n- 'name'::regtype::oid = ANY(coltypes);\\n- relname | attname | coltypes |\\n get_columns_length \\n----------+---------+----------+--------------------\\n-(0 rows)\\n-\\n+server closed the connection unexpectedly\\n+ This\\n probably means the server terminated abnormally\\n+ before\\n or while processing the request.\\n+connection to server was lost \\n\\n\\n\\n\\n\\n\\nActually digging some more, here is the actual error\\n\\n\\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \\\"ntstatus.h\\\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nSo how does one debug this ?Also if I `run meson test` I don't see this error. What does the buildfarm do differently?Dave\\n\\nYes, this is pretty much what I saw.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Sat, 10 Feb 2024 12:20:12 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-02-10 Sa 12:20, Dave Cramer wrote:\\n>\\n>\\n> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>\\n> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n>>\\n>> Dave Cramer\\n>> www.postgres.rocks <http://www.postgres.rocks>\\n>>\\n>>\\n>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer\\n>> <davecramer@postgres.rocks> <mailto:davecramer@postgres.rocks> wrote:\\n>>\\n>>\\n>>\\n>>\\n>>\\n>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier\\n>> <michael@paquier.xyz> wrote:\\n>>\\n>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n>> > Thanks, this patch works and\\n>> > testing with meson passes.\\n>>\\n>> Only with the version posted at [1]? Interesting, that's\\n>> the same\\n>> contents as v8 posted upthread, minus src/tools/ because\\n>> we don't need\\n>> to care about them anymore.\\n>>\\n>> Andrew, what's happening on the test side? It does not\\n>> seem you've\\n>> mentioned any details about what is going wrong, or I\\n>> have just missed\\n>> them.\\n>>\\n>> > I'll try the buildfarm next.\\n>>\\n>> [1]:\\n>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n>>\\n>>\\n>> interestingly meson test does not produce any error\\n>> The buildfarm produces the following error for me:\\n>>\\n>> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n>> - FROM check_columns\\n>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n>> - 'name'::regtype::oid = ANY(coltypes);\\n>> - relname | attname | coltypes | get_columns_length\\n>> ----------+---------+----------+--------------------\\n>> -(0 rows)\\n>> -\\n>> +server closed the connection unexpectedly\\n>> +This probably means the server terminated abnormally\\n>> +before or while processing the request.\\n>> +connection to server was lost\\n>>\\n>>\\n>> Actually digging some more, here is the actual error\\n>>\\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server\\n>> process (PID 11204) was terminated by exception 0xC0000005\\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed\\n>> process was running: VACUUM;\\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C\\n>> include file \\\"ntstatus.h\\\" for a description of the hexadecimal value.\\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating\\n>> any other active server processes\\n>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server\\n>> processes terminated; reinitializing\\n>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system\\n>> was interrupted; last known up at 2024-02-09 13:31:01 -05\\n>>\\n>>\\n>>\\n>\\n>\\n> So how does one debug this ?\\n>\\n> Also if I `run meson test` I don't see this error. What does the \\n> buildfarm do differently?\\n\\n\\nFirst it does this:\\n\\n\\n meson test -C $pgsql --no-rebuild --suite setup\\n\\n\\nThen it does this (jflag is for the number of jobs):\\n\\n\\n meson test -t $meson_test_timeout $jflag -C $pgsql --logbase\\n checklog --print-errorlogs --no-rebuild --suite regress\\n --test-args=--no-locale\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-02-10 Sa 12:20, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Sat, 10 Feb 2024 at\\n 11:19, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 9 Feb\\n 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri,\\n 9 Feb 2024 at 00:26, Michael Paquier\\n <michael@paquier.xyz>\\n wrote:\\n\\nOn\\n Tue, Feb 06, 2024 at 07:01:49AM -0500,\\n Dave Cramer wrote:\\n > Thanks, this patch works and\\n > testing with meson passes.\\n\\n Only with the version posted at [1]? \\n Interesting, that's the same\\n contents as v8 posted upthread, minus\\n src/tools/ because we don't need\\n to care about them anymore.\\n\\n Andrew, what's happening on the test\\n side? It does not seem you've\\n mentioned any details about what is\\n going wrong, or I have just missed\\n them.\\n\\n > I'll try the buildfarm next.\\n\\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n\\n\\ninterestingly meson test does not\\n produce any error\\nThe buildfarm produces the following\\n error for me:\\n\\n\\n-SELECT relname, attname, coltypes,\\n get_columns_length(coltypes)\\n- FROM check_columns\\n- WHERE get_columns_length(coltypes)\\n % 8 != 0 OR\\n- 'name'::regtype::oid =\\n ANY(coltypes);\\n- relname | attname | coltypes |\\n get_columns_length \\n----------+---------+----------+--------------------\\n-(0 rows)\\n-\\n+server closed the connection\\n unexpectedly\\n+ This\\n probably means the server terminated\\n abnormally\\n+ before\\n or while processing the request.\\n+connection to server was lost \\n\\n\\n\\n\\n\\n\\nActually digging some more, here is the\\n actual error\\n\\n\\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \\\"ntstatus.h\\\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nSo how does one debug this ?\\n\\n\\nAlso if I `run meson test` I don't see this error. What\\n does the buildfarm do differently?\\n\\n\\n\\n\\n\\nFirst it does this:\\n\\n\\n\\nmeson test -C $pgsql --no-rebuild --suite setup\\n\\n\\n\\nThen it does this (jflag is for the number of jobs):\\n\\n\\n\\nmeson test -t $meson_test_timeout $jflag -C $pgsql --logbase\\n checklog --print-errorlogs --no-rebuild --suite regress\\n --test-args=--no-locale\\n\\n\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Sat, 10 Feb 2024 13:28:29 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Sat, 10 Feb 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-02-10 Sa 12:20, Dave Cramer wrote:\\n>\\n>\\n>\\n> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>>\\n>> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n>>\\n>>\\n>> Dave Cramer\\n>> www.postgres.rocks\\n>>\\n>>\\n>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\\n>> <davecramer@postgres.rocks> wrote:\\n>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz>\\n>>> wrote:\\n>>>\\n>>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n>>>> > Thanks, this patch works and\\n>>>> > testing with meson passes.\\n>>>>\\n>>>> Only with the version posted at [1]? Interesting, that's the same\\n>>>> contents as v8 posted upthread, minus src/tools/ because we don't need\\n>>>> to care about them anymore.\\n>>>>\\n>>>> Andrew, what's happening on the test side? It does not seem you've\\n>>>> mentioned any details about what is going wrong, or I have just missed\\n>>>> them.\\n>>>>\\n>>>> > I'll try the buildfarm next.\\n>>>>\\n>>>> [1]:\\n>>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n>>>\\n>>>\\n>>> interestingly meson test does not produce any error\\n>>> The buildfarm produces the following error for me:\\n>>>\\n>>> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n>>> - FROM check_columns\\n>>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n>>> - 'name'::regtype::oid = ANY(coltypes);\\n>>> - relname | attname | coltypes | get_columns_length\\n>>> ----------+---------+----------+--------------------\\n>>> -(0 rows)\\n>>> -\\n>>> +server closed the connection unexpectedly\\n>>> + This probably means the server terminated abnormally\\n>>> + before or while processing the request.\\n>>> +connection to server was lost\\n>>>\\n>>\\n>> Actually digging some more, here is the actual error\\n>>\\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\\n>> 11204) was terminated by exception 0xC0000005\\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process\\n>> was running: VACUUM;\\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\\n>> \\\"ntstatus.h\\\" for a description of the hexadecimal value.\\n>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any\\n>> other active server processes\\n>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes\\n>> terminated; reinitializing\\n>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was\\n>> interrupted; last known up at 2024-02-09 13:31:01 -05\\n>>\\n>>\\n>>\\n>>\\n>>\\n> So how does one debug this ?\\n>\\n> Also if I `run meson test` I don't see this error. What does the buildfarm\\n> do differently?\\n>\\n>\\n> First it does this:\\n>\\n>\\n> meson test -C $pgsql --no-rebuild --suite setup\\n>\\n>\\n> Then it does this (jflag is for the number of jobs):\\n>\\n>\\n> meson test -t $meson_test_timeout $jflag -C $pgsql --logbase checklog\\n> --print-errorlogs --no-rebuild --suite regress --test-args=--no-locale\\n>\\n>\\n\\nrunning the above manually produces no errors ??\\n\\nDave\\n\\nOn Sat, 10 Feb 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-02-10 Sa 12:20, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Sat, 10 Feb 2024 at\\n 11:19, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 9 Feb\\n 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri,\\n 9 Feb 2024 at 00:26, Michael Paquier\\n <michael@paquier.xyz>\\n wrote:\\n\\nOn\\n Tue, Feb 06, 2024 at 07:01:49AM -0500,\\n Dave Cramer wrote:\\n > Thanks, this patch works and\\n > testing with meson passes.\\n\\n Only with the version posted at [1]? \\n Interesting, that's the same\\n contents as v8 posted upthread, minus\\n src/tools/ because we don't need\\n to care about them anymore.\\n\\n Andrew, what's happening on the test\\n side? It does not seem you've\\n mentioned any details about what is\\n going wrong, or I have just missed\\n them.\\n\\n > I'll try the buildfarm next.\\n\\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n\\n\\ninterestingly meson test does not\\n produce any error\\nThe buildfarm produces the following\\n error for me:\\n\\n\\n-SELECT relname, attname, coltypes,\\n get_columns_length(coltypes)\\n- FROM check_columns\\n- WHERE get_columns_length(coltypes)\\n % 8 != 0 OR\\n- 'name'::regtype::oid =\\n ANY(coltypes);\\n- relname | attname | coltypes |\\n get_columns_length \\n----------+---------+----------+--------------------\\n-(0 rows)\\n-\\n+server closed the connection\\n unexpectedly\\n+ This\\n probably means the server terminated\\n abnormally\\n+ before\\n or while processing the request.\\n+connection to server was lost \\n\\n\\n\\n\\n\\n\\nActually digging some more, here is the\\n actual error\\n\\n\\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \\\"ntstatus.h\\\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nSo how does one debug this ?\\n\\n\\nAlso if I `run meson test` I don't see this error. What\\n does the buildfarm do differently?\\n\\n\\n\\n\\n\\nFirst it does this:\\n\\n\\n\\nmeson test -C $pgsql --no-rebuild --suite setup\\n\\n\\n\\nThen it does this (jflag is for the number of jobs):\\n\\n\\n\\nmeson test -t $meson_test_timeout $jflag -C $pgsql --logbase\\n checklog --print-errorlogs --no-rebuild --suite regress\\n --test-args=--no-localerunning the above manually produces no errors ??Dave\",\n \"msg_date\": \"Mon, 12 Feb 2024 08:51:45 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-02-12 Mo 08:51, Dave Cramer wrote:\\n>\\n>\\n> On Sat, 10 Feb 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>\\n> On 2024-02-10 Sa 12:20, Dave Cramer wrote:\\n>>\\n>>\\n>> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan\\n>> <andrew@dunslane.net> wrote:\\n>>\\n>>\\n>> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n>>>\\n>>> Dave Cramer\\n>>> www.postgres.rocks <http://www.postgres.rocks>\\n>>>\\n>>>\\n>>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer\\n>>> <davecramer@postgres.rocks>\\n>>> <mailto:davecramer@postgres.rocks> wrote:\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier\\n>>> <michael@paquier.xyz> wrote:\\n>>>\\n>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave\\n>>> Cramer wrote:\\n>>> > Thanks, this patch works and\\n>>> > testing with meson passes.\\n>>>\\n>>> Only with the version posted at [1]? Interesting,\\n>>> that's the same\\n>>> contents as v8 posted upthread, minus src/tools/\\n>>> because we don't need\\n>>> to care about them anymore.\\n>>>\\n>>> Andrew, what's happening on the test side? It does\\n>>> not seem you've\\n>>> mentioned any details about what is going wrong, or\\n>>> I have just missed\\n>>> them.\\n>>>\\n>>> > I'll try the buildfarm next.\\n>>>\\n>>> [1]:\\n>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n>>>\\n>>>\\n>>> interestingly meson test does not produce any error\\n>>> The buildfarm produces the following error for me:\\n>>>\\n>>> -SELECT relname, attname, coltypes,\\n>>> get_columns_length(coltypes)\\n>>> - FROM check_columns\\n>>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n>>> - 'name'::regtype::oid = ANY(coltypes);\\n>>> - relname | attname | coltypes | get_columns_length\\n>>> ----------+---------+----------+--------------------\\n>>> -(0 rows)\\n>>> -\\n>>> +server closed the connection unexpectedly\\n>>> +This probably means the server terminated abnormally\\n>>> +before or while processing the request.\\n>>> +connection to server was lost\\n>>>\\n>>>\\n>>> Actually digging some more, here is the actual error\\n>>>\\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server\\n>>> process (PID 11204) was terminated by exception 0xC0000005\\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL:\\n>>> Failed process was running: VACUUM;\\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C\\n>>> include file \\\"ntstatus.h\\\" for a description of the\\n>>> hexadecimal value.\\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG:\\n>>> terminating any other active server processes\\n>>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all\\n>>> server processes terminated; reinitializing\\n>>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database\\n>>> system was interrupted; last known up at 2024-02-09 13:31:01 -05\\n>>>\\n>>>\\n>>>\\n>>\\n>>\\n>> So how does one debug this ?\\n>>\\n>> Also if I `run meson test` I don't see this error. What does the\\n>> buildfarm do differently?\\n>\\n>\\n> First it does this:\\n>\\n>\\n> meson test -C $pgsql --no-rebuild --suite setup\\n>\\n>\\n> Then it does this (jflag is for the number of jobs):\\n>\\n>\\n> meson test -t $meson_test_timeout $jflag -C $pgsql --logbase\\n> checklog --print-errorlogs --no-rebuild --suite regress\\n> --test-args=--no-locale\\n>\\n>\\n>\\n> running the above manually produces no errors ??\\n>\\n>\\n\\nNot for me. I get the error I previously reported, It's an access \\nviolation error.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-02-12 Mo 08:51, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Sat, 10 Feb 2024 at\\n 13:28, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-02-10 Sa 12:20, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Sat, 10 Feb\\n 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn\\n Fri, 9 Feb 2024 at 07:18, Dave Cramer\\n <davecramer@postgres.rocks>\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 9\\n Feb 2024 at 00:26, Michael\\n Paquier <michael@paquier.xyz>\\n wrote:\\n\\nOn\\n Tue, Feb 06, 2024 at\\n 07:01:49AM -0500, Dave Cramer\\n wrote:\\n > Thanks, this patch works\\n and\\n > testing with meson\\n passes.\\n\\n Only with the version posted\\n at [1]? Interesting, that's\\n the same\\n contents as v8 posted\\n upthread, minus src/tools/\\n because we don't need\\n to care about them anymore.\\n\\n Andrew, what's happening on\\n the test side? It does not\\n seem you've\\n mentioned any details about\\n what is going wrong, or I have\\n just missed\\n them.\\n\\n > I'll try the buildfarm\\n next.\\n\\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n\\n\\ninterestingly meson test\\n does not produce any error\\nThe buildfarm produces the\\n following error for me:\\n\\n\\n-SELECT relname, attname,\\n coltypes,\\n get_columns_length(coltypes)\\n- FROM check_columns\\n- WHERE\\n get_columns_length(coltypes) %\\n 8 != 0 OR\\n- \\n 'name'::regtype::oid =\\n ANY(coltypes);\\n- relname | attname |\\n coltypes | get_columns_length \\n----------+---------+----------+--------------------\\n-(0 rows)\\n-\\n+server closed the\\n connection unexpectedly\\n+ This\\n probably means the server\\n terminated abnormally\\n+ before\\n or while processing the\\n request.\\n+connection to server was\\n lost \\n\\n\\n\\n\\n\\n\\nActually digging some more, here is\\n the actual error\\n\\n\\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \\\"ntstatus.h\\\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nSo how does one debug this ?\\n\\n\\nAlso if I `run meson test` I don't see this\\n error. What does the buildfarm do differently?\\n\\n\\n\\n\\n\\nFirst it does this:\\n\\n\\n\\nmeson test -C $pgsql --no-rebuild --suite setup\\n\\n\\n\\nThen it does this (jflag is for the number of jobs):\\n\\n\\n\\nmeson test -t $meson_test_timeout $jflag -C $pgsql\\n --logbase checklog --print-errorlogs --no-rebuild\\n --suite regress --test-args=--no-locale\\n\\n\\n\\n\\n\\n\\n\\n\\nrunning the above manually produces no errors ??\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nNot for me. I get the error I previously reported, It's an access\\n violation error.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 12 Feb 2024 09:19:26 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Dave Cramer\\nwww.postgres.rocks\\n\\n\\nOn Mon, 12 Feb 2024 at 09:19, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n>\\n> On 2024-02-12 Mo 08:51, Dave Cramer wrote:\\n>\\n>\\n>\\n> On Sat, 10 Feb 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>>\\n>> On 2024-02-10 Sa 12:20, Dave Cramer wrote:\\n>>\\n>>\\n>>\\n>> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>>\\n>>>\\n>>> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n>>>\\n>>>\\n>>> Dave Cramer\\n>>> www.postgres.rocks\\n>>>\\n>>>\\n>>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer <davecramer@postgres.rocks>\\n>>> <davecramer@postgres.rocks> wrote:\\n>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier <michael@paquier.xyz>\\n>>>> wrote:\\n>>>>\\n>>>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave Cramer wrote:\\n>>>>> > Thanks, this patch works and\\n>>>>> > testing with meson passes.\\n>>>>>\\n>>>>> Only with the version posted at [1]? Interesting, that's the same\\n>>>>> contents as v8 posted upthread, minus src/tools/ because we don't need\\n>>>>> to care about them anymore.\\n>>>>>\\n>>>>> Andrew, what's happening on the test side? It does not seem you've\\n>>>>> mentioned any details about what is going wrong, or I have just missed\\n>>>>> them.\\n>>>>>\\n>>>>> > I'll try the buildfarm next.\\n>>>>>\\n>>>>> [1]:\\n>>>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n>>>>\\n>>>>\\n>>>> interestingly meson test does not produce any error\\n>>>> The buildfarm produces the following error for me:\\n>>>>\\n>>>> -SELECT relname, attname, coltypes, get_columns_length(coltypes)\\n>>>> - FROM check_columns\\n>>>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n>>>> - 'name'::regtype::oid = ANY(coltypes);\\n>>>> - relname | attname | coltypes | get_columns_length\\n>>>> ----------+---------+----------+--------------------\\n>>>> -(0 rows)\\n>>>> -\\n>>>> +server closed the connection unexpectedly\\n>>>> + This probably means the server terminated abnormally\\n>>>> + before or while processing the request.\\n>>>> +connection to server was lost\\n>>>>\\n>>>\\n>>> Actually digging some more, here is the actual error\\n>>>\\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID\\n>>> 11204) was terminated by exception 0xC0000005\\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process\\n>>> was running: VACUUM;\\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file\\n>>> \\\"ntstatus.h\\\" for a description of the hexadecimal value.\\n>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any\\n>>> other active server processes\\n>>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server\\n>>> processes terminated; reinitializing\\n>>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was\\n>>> interrupted; last known up at 2024-02-09 13:31:01 -05\\n>>>\\n>>>\\n>>>\\n>>>\\n>>>\\n>> So how does one debug this ?\\n>>\\n>> Also if I `run meson test` I don't see this error. What does the\\n>> buildfarm do differently?\\n>>\\n>>\\n>> First it does this:\\n>>\\n>>\\n>> meson test -C $pgsql --no-rebuild --suite setup\\n>>\\n>>\\n>> Then it does this (jflag is for the number of jobs):\\n>>\\n>>\\n>> meson test -t $meson_test_timeout $jflag -C $pgsql --logbase checklog\\n>> --print-errorlogs --no-rebuild --suite regress --test-args=--no-locale\\n>>\\n>>\\n>\\n> running the above manually produces no errors ??\\n>\\n>\\n>\\n> Not for me. I get the error I previously reported, It's an access\\n> violation error.\\n>\\n>\\n> cheers\\n>\\n>\\n> andrew\\n>\\n\\nOK, so I have managed to get a debugger attached to postgres.exe when it\\nfaults and the fault occurs at\\nhttps://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\\nspan is pointing to 0x0\\n\\nDave\\n\\nDave Cramerwww.postgres.rocksOn Mon, 12 Feb 2024 at 09:19, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n\\n\\nOn 2024-02-12 Mo 08:51, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\nOn Sat, 10 Feb 2024 at\\n 13:28, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-02-10 Sa 12:20, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Sat, 10 Feb\\n 2024 at 11:19, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn\\n Fri, 9 Feb 2024 at 07:18, Dave Cramer\\n <davecramer@postgres.rocks>\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Fri, 9\\n Feb 2024 at 00:26, Michael\\n Paquier <michael@paquier.xyz>\\n wrote:\\n\\nOn\\n Tue, Feb 06, 2024 at\\n 07:01:49AM -0500, Dave Cramer\\n wrote:\\n > Thanks, this patch works\\n and\\n > testing with meson\\n passes.\\n\\n Only with the version posted\\n at [1]? Interesting, that's\\n the same\\n contents as v8 posted\\n upthread, minus src/tools/\\n because we don't need\\n to care about them anymore.\\n\\n Andrew, what's happening on\\n the test side? It does not\\n seem you've\\n mentioned any details about\\n what is going wrong, or I have\\n just missed\\n them.\\n\\n > I'll try the buildfarm\\n next.\\n\\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n\\n\\ninterestingly meson test\\n does not produce any error\\nThe buildfarm produces the\\n following error for me:\\n\\n\\n-SELECT relname, attname,\\n coltypes,\\n get_columns_length(coltypes)\\n- FROM check_columns\\n- WHERE\\n get_columns_length(coltypes) %\\n 8 != 0 OR\\n- \\n 'name'::regtype::oid =\\n ANY(coltypes);\\n- relname | attname |\\n coltypes | get_columns_length \\n----------+---------+----------+--------------------\\n-(0 rows)\\n-\\n+server closed the\\n connection unexpectedly\\n+ This\\n probably means the server\\n terminated abnormally\\n+ before\\n or while processing the\\n request.\\n+connection to server was\\n lost \\n\\n\\n\\n\\n\\n\\nActually digging some more, here is\\n the actual error\\n\\n\\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \\\"ntstatus.h\\\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nSo how does one debug this ?\\n\\n\\nAlso if I `run meson test` I don't see this\\n error. What does the buildfarm do differently?\\n\\n\\n\\n\\n\\nFirst it does this:\\n\\n\\n\\nmeson test -C $pgsql --no-rebuild --suite setup\\n\\n\\n\\nThen it does this (jflag is for the number of jobs):\\n\\n\\n\\nmeson test -t $meson_test_timeout $jflag -C $pgsql\\n --logbase checklog --print-errorlogs --no-rebuild\\n --suite regress --test-args=--no-locale\\n\\n\\n\\n\\n\\n\\n\\n\\nrunning the above manually produces no errors ??\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nNot for me. I get the error I previously reported, It's an access\\n violation error.\\n\\n\\ncheers\\n\\n\\nandrewOK, so I have managed to get a debugger attached to postgres.exe when it faults and the fault occurs athttps://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869span is pointing to 0x0Dave\",\n \"msg_date\": \"Mon, 12 Feb 2024 11:44:23 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-02-12 Mo 11:44, Dave Cramer wrote:\\n>\\n> Dave Cramer\\n> www.postgres.rocks\\n>\\n>\\n> On Mon, 12 Feb 2024 at 09:19, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>\\n> On 2024-02-12 Mo 08:51, Dave Cramer wrote:\\n>>\\n>>\\n>> On Sat, 10 Feb 2024 at 13:28, Andrew Dunstan\\n>> <andrew@dunslane.net> wrote:\\n>>\\n>>\\n>> On 2024-02-10 Sa 12:20, Dave Cramer wrote:\\n>>>\\n>>>\\n>>> On Sat, 10 Feb 2024 at 11:19, Andrew Dunstan\\n>>> <andrew@dunslane.net> wrote:\\n>>>\\n>>>\\n>>> On 2024-02-09 Fr 14:23, Dave Cramer wrote:\\n>>>>\\n>>>> Dave Cramer\\n>>>> www.postgres.rocks <http://www.postgres.rocks>\\n>>>>\\n>>>>\\n>>>> On Fri, 9 Feb 2024 at 07:18, Dave Cramer\\n>>>> <davecramer@postgres.rocks>\\n>>>> <mailto:davecramer@postgres.rocks> wrote:\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>>\\n>>>> On Fri, 9 Feb 2024 at 00:26, Michael Paquier\\n>>>> <michael@paquier.xyz> wrote:\\n>>>>\\n>>>> On Tue, Feb 06, 2024 at 07:01:49AM -0500, Dave\\n>>>> Cramer wrote:\\n>>>> > Thanks, this patch works and\\n>>>> > testing with meson passes.\\n>>>>\\n>>>> Only with the version posted at [1]? \\n>>>> Interesting, that's the same\\n>>>> contents as v8 posted upthread, minus\\n>>>> src/tools/ because we don't need\\n>>>> to care about them anymore.\\n>>>>\\n>>>> Andrew, what's happening on the test side? It\\n>>>> does not seem you've\\n>>>> mentioned any details about what is going\\n>>>> wrong, or I have just missed\\n>>>> them.\\n>>>>\\n>>>> > I'll try the buildfarm next.\\n>>>>\\n>>>> [1]:\\n>>>> https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n>>>>\\n>>>>\\n>>>> interestingly meson test does not produce any error\\n>>>> The buildfarm produces the following error for me:\\n>>>>\\n>>>> -SELECT relname, attname, coltypes,\\n>>>> get_columns_length(coltypes)\\n>>>> - FROM check_columns\\n>>>> - WHERE get_columns_length(coltypes) % 8 != 0 OR\\n>>>> - 'name'::regtype::oid = ANY(coltypes);\\n>>>> - relname | attname | coltypes | get_columns_length\\n>>>> ----------+---------+----------+--------------------\\n>>>> -(0 rows)\\n>>>> -\\n>>>> +server closed the connection unexpectedly\\n>>>> +This probably means the server terminated abnormally\\n>>>> +before or while processing the request.\\n>>>> +connection to server was lost\\n>>>>\\n>>>>\\n>>>> Actually digging some more, here is the actual error\\n>>>>\\n>>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG:\\n>>>> server process (PID 11204) was terminated by exception\\n>>>> 0xC0000005\\n>>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL:\\n>>>> Failed process was running: VACUUM;\\n>>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] HINT:\\n>>>> See C include file \\\"ntstatus.h\\\" for a description of\\n>>>> the hexadecimal value.\\n>>>> 2024-02-09 13:31:11.008 -05 postmaster[10672] LOG:\\n>>>> terminating any other active server processes\\n>>>> 2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all\\n>>>> server processes terminated; reinitializing\\n>>>> 2024-02-09 13:31:11.034 -05 startup[6152] LOG:\\n>>>> database system was interrupted; last known up at\\n>>>> 2024-02-09 13:31:01 -05\\n>>>>\\n>>>>\\n>>>>\\n>>>\\n>>>\\n>>> So how does one debug this ?\\n>>>\\n>>> Also if I `run meson test` I don't see this error. What does\\n>>> the buildfarm do differently?\\n>>\\n>>\\n>> First it does this:\\n>>\\n>>\\n>> meson test -C $pgsql --no-rebuild --suite setup\\n>>\\n>>\\n>> Then it does this (jflag is for the number of jobs):\\n>>\\n>>\\n>> meson test -t $meson_test_timeout $jflag -C $pgsql\\n>> --logbase checklog --print-errorlogs --no-rebuild --suite\\n>> regress --test-args=--no-locale\\n>>\\n>>\\n>>\\n>> running the above manually produces no errors ??\\n>>\\n>>\\n>\\n> Not for me. I get the error I previously reported, It's an access\\n> violation error.\\n>\\n>\\n>\\n>\\n> OK, so I have managed to get a debugger attached to postgres.exe when \\n> it faults and the fault occurs at\\n> https://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\\n> span is pointing to 0x0\\n>\\n>\\n\\nFurther data point. If I select a build type of 'debug' instead of \\n'debugoptimized' the error disappears.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-02-12 Mo 11:44, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nDave Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Mon, 12 Feb 2024 at\\n 09:19, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-02-12 Mo 08:51, Dave Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\nOn Sat, 10 Feb\\n 2024 at 13:28, Andrew Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-02-10 Sa 12:20, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn\\n Sat, 10 Feb 2024 at 11:19, Andrew\\n Dunstan <andrew@dunslane.net>\\n wrote:\\n\\n\\n\\n\\n\\nOn 2024-02-09 Fr 14:23, Dave\\n Cramer wrote:\\n\\n\\n\\n\\n\\n\\n\\nDave\\n Cramer\\n www.postgres.rocks\\n\\n\\n\\n\\n\\n\\n\\nOn Fri,\\n 9 Feb 2024 at 07:18, Dave\\n Cramer <davecramer@postgres.rocks>\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn\\n Fri, 9 Feb 2024 at\\n 00:26, Michael\\n Paquier <michael@paquier.xyz>\\n wrote:\\n\\nOn\\n Tue, Feb 06, 2024\\n at 07:01:49AM\\n -0500, Dave Cramer\\n wrote:\\n > Thanks, this\\n patch works and\\n > testing with\\n meson passes.\\n\\n Only with the\\n version posted at\\n [1]? Interesting,\\n that's the same\\n contents as v8\\n posted upthread,\\n minus src/tools/\\n because we don't\\n need\\n to care about them\\n anymore.\\n\\n Andrew, what's\\n happening on the\\n test side? It\\n does not seem\\n you've\\n mentioned any\\n details about what\\n is going wrong, or\\n I have just missed\\n them.\\n\\n > I'll try the\\n buildfarm next.\\n\\n [1]: https://www.postgresql.org/message-id/ea42654a-3dc4-98b0-335b-56b7ec5e5687@dunslane.net\\n\\n\\ninterestingly\\n meson test does\\n not produce any\\n error\\nThe buildfarm\\n produces the\\n following error\\n for me:\\n\\n\\n-SELECT\\n relname, attname,\\n coltypes,\\n get_columns_length(coltypes)\\n- FROM\\n check_columns\\n- WHERE\\n get_columns_length(coltypes)\\n % 8 != 0 OR\\n- \\n 'name'::regtype::oid\\n = ANY(coltypes);\\n- relname |\\n attname | coltypes\\n |\\n get_columns_length \\n----------+---------+----------+--------------------\\n-(0 rows)\\n-\\n+server closed\\n the connection\\n unexpectedly\\n+ This\\n probably means the\\n server terminated\\n abnormally\\n+ before\\n or while\\n processing the\\n request.\\n+connection to\\n server was lost \\n\\n\\n\\n\\n\\n\\nActually digging some\\n more, here is the actual\\n error\\n\\n\\n2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: server process (PID 11204) was terminated by exception 0xC00000052024-02-09 13:31:11.008 -05 postmaster[10672] DETAIL: Failed process was running: VACUUM;2024-02-09 13:31:11.008 -05 postmaster[10672] HINT: See C include file \\\"ntstatus.h\\\" for a description of the hexadecimal value.2024-02-09 13:31:11.008 -05 postmaster[10672] LOG: terminating any other active server processes2024-02-09 13:31:11.013 -05 postmaster[10672] LOG: all server processes terminated; reinitializing2024-02-09 13:31:11.034 -05 startup[6152] LOG: database system was interrupted; last known up at 2024-02-09 13:31:01 -05\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nSo how does one debug this ?\\n\\n\\nAlso if I `run meson test` I don't\\n see this error. What does the\\n buildfarm do differently?\\n\\n\\n\\n\\n\\nFirst it does this:\\n\\n\\n\\nmeson test -C $pgsql --no-rebuild --suite\\n setup\\n\\n\\n\\nThen it does this (jflag is for the number\\n of jobs):\\n\\n\\n\\nmeson test -t $meson_test_timeout $jflag\\n -C $pgsql --logbase checklog\\n --print-errorlogs --no-rebuild --suite\\n regress --test-args=--no-locale\\n\\n\\n\\n\\n\\n\\n\\n\\nrunning the above manually produces no errors\\n ??\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nNot for me. I get the error I previously reported,\\n It's an access violation error.\\n\\n\\n\\n\\n\\n\\n\\nOK, so I have managed to get a debugger attached to\\n postgres.exe when it faults and the fault occurs at\\nhttps://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\\nspan is pointing to 0x0\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nFurther data point. If I select a build type of 'debug' instead\\n of 'debugoptimized' the error disappears.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 12 Feb 2024 13:28:40 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\\n> On 2024-02-12 Mo 11:44, Dave Cramer wrote:\\n> > OK, so I have managed to get a debugger attached to postgres.exe when it\\n> > faults and the fault occurs at\\n> > https://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\\n> > span is pointing to 0x0\\n> \\n> Further data point. If I select a build type of 'debug' instead of\\n> 'debugoptimized' the error disappears.\\n\\nOh, this is quite interesting. Dave, could you post the backtrace?\\n\\nI wonder if this indicates that we are either missing memory barriers\\nsomewhere or that the memory barriers we end up with on msvc + arm aren't\\ncorrect? Either could explain why the problem doesn't occur when building\\nwith optimizations.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Feb 2024 12:50:12 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2024-02-09 15:32:10 -0500, Dave Cramer wrote:\\n> On Fri, 9 Feb 2024 at 14:36, Andres Freund <andres@anarazel.de> wrote:\\n> > That's something like a segfault.\\n> >\\n> > One suspicion I have is that src/port/pg_crc32c_armv8_choose.c possibly\\n> > doesn't properly support msvc. It seems to assume that SIGILL can be\\n> > trapped,\\n> > but that IIRC doesn't work on windows.\\n> >\\n> > I'd check if the problem persists if you change\\n> > cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 1)\\n> > to\\n> > cdata.set('USE_ARMV8_CRC32C_WITH_RUNTIME_CHECK', 0)\\n> >\\n> \\n> This results in\\n> \\n> FAILED: src/bin/pg_checksums/pg_checksums.exe\\n> src/bin/pg_checksums/pg_checksums.pdb\\n> \\\"link\\\" /MACHINE:ARM64 /OUT:src/bin/pg_checksums/pg_checksums.exe\\n> src/bin/pg_checksums/pg_checksums.exe.p/win32ver.res\\n> src/bin/pg_checksums/pg_checksums.exe.p/pg_checksums.c.obj \\\"/release\\\"\\n> \\\"/nologo\\\" \\\"/DEBUG\\\" \\\"/PDB:src\\\\bin\\\\pg_checksums\\\\pg_checksums.pdb\\\"\\n> \\\"/INCREMENTAL:NO\\\" \\\"/STACK:4194304\\\" \\\"/NOEXP\\\" \\\"src/fe_utils/libpgfeutils.a\\\"\\n> \\\"src/common/libpgcommon.a\\\" \\\"src/port/libpgport.a\\\" \\\"ws2_32.lib\\\" \\\"ws2_32.lib\\\"\\n> \\\"ws2_32.lib\\\" \\\"ws2_32.lib\\\" \\\"/SUBSYSTEM:CONSOLE\\\" \\\"kernel32.lib\\\" \\\"user32.lib\\\"\\n> \\\"gdi32.lib\\\" \\\"winspool.lib\\\" \\\"shell32.lib\\\" \\\"ole32.lib\\\" \\\"oleaut32.lib\\\"\\n> \\\"uuid.lib\\\" \\\"comdlg32.lib\\\" \\\"advapi32.lib\\\"\\n> libpgcommon.a(controldata_utils.c.obj) : error LNK2001: unresolved external\\n> symbol pg_comp_crc32c\\n\\nThat's because have_optimized_crc ends up being set.\\n\\n\\nI'm somewhat unhappy about the msvc specific path in meson.build. Shouldn't at\\nthe very least the \\\"Use ARM CRC Extension unconditionally\\\" path be first. I\\nguess it's not really this patch's fault that the runtime detection for armv8\\ncrc is implemented this shoddily :(.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Feb 2024 12:53:52 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Mon, 12 Feb 2024 at 15:50, Andres Freund <andres@anarazel.de> wrote:\\n\\n> Hi,\\n>\\n> On 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\\n> > On 2024-02-12 Mo 11:44, Dave Cramer wrote:\\n> > > OK, so I have managed to get a debugger attached to postgres.exe when\\n> it\\n> > > faults and the fault occurs at\\n> > >\\n> https://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\\n> > > span is pointing to 0x0\\n> >\\n> > Further data point. If I select a build type of 'debug' instead of\\n> > 'debugoptimized' the error disappears.\\n>\\n> Oh, this is quite interesting. Dave, could you post the backtrace?\\n>\\n\\nAndres,\\n\\nI am using Visual Studio as the debugger. Here is what I have.\\n\\n> postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C\\n postgres.exe!resize(dshash_table * hash_table, unsigned __int64\\nnew_size_log2) Line 879 C\\n postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const void\\n* key, bool * found) Line 480 C\\n postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int dboid,\\nunsigned int objoid, bool create, bool * created_entry) Line 455 C\\n postgres.exe!pgstat_prep_pending_entry(PgStat_Kind kind, unsigned int\\ndboid, unsigned int objoid, bool * created_entry) Line 1123 C\\n [Inline Frame] postgres.exe!pgstat_prep_relation_pending(unsigned int)\\nLine 904 C\\n postgres.exe!pgstat_assoc_relation(RelationData * rel) Line 139 C\\n [Inline Frame] postgres.exe!ReadBufferExtended(RelationData *) Line 802 C\\n postgres.exe!ReadBuffer(RelationData * reln, unsigned int blockNum) Line\\n737 C\\n postgres.exe!read_seq_tuple(RelationData * rel, int * buf, HeapTupleData\\n* seqdatatuple) Line 1196 C\\n postgres.exe!AlterSequence(ParseState * pstate, AlterSeqStmt * stmt) Line\\n481 C\\n postgres.exe!ProcessUtilitySlow(ParseState * pstate, PlannedStmt * pstmt,\\nconst char * queryString, ProcessUtilityContext context, ParamListInfoData\\n* params, QueryEnvironment * queryEnv, _DestReceiver * dest,\\nQueryCompletion * qc) Line 1679 C\\n postgres.exe!standard_ProcessUtility(PlannedStmt * pstmt, const char *\\nqueryString, bool readOnlyTree, ProcessUtilityContext context,\\nParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver *\\ndest, QueryCompletion * qc) Line 1080 C\\n postgres.exe!ProcessUtility(PlannedStmt * pstmt, const char *\\nqueryString, bool readOnlyTree, ProcessUtilityContext context,\\nParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver *\\ndest, QueryCompletion * qc) Line 530 C\\n postgres.exe!ProcessUtilitySlow(ParseState * pstate, PlannedStmt * pstmt,\\nconst char * queryString, ProcessUtilityContext context, ParamListInfoData\\n* params, QueryEnvironment * queryEnv, _DestReceiver * dest,\\nQueryCompletion * qc) Line 1263 C\\n postgres.exe!standard_ProcessUtility(PlannedStmt * pstmt, const char *\\nqueryString, bool readOnlyTree, ProcessUtilityContext context,\\nParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver *\\ndest, QueryCompletion * qc) Line 1080 C\\n\\n\\nif there is a better debugger to use, please let me know\\n\\nDave\\n\\n\\n\\n>\\n> I wonder if this indicates that we are either missing memory barriers\\n> somewhere or that the memory barriers we end up with on msvc + arm aren't\\n> correct? Either could explain why the problem doesn't occur when building\\n> with optimizations.\\n>\\n> Greetings,\\n>\\n> Andres Freund\\n>\\n\\nOn Mon, 12 Feb 2024 at 15:50, Andres Freund <andres@anarazel.de> wrote:Hi,\\n\\nOn 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\\n> On 2024-02-12 Mo 11:44, Dave Cramer wrote:\\n> > OK, so I have managed to get a debugger attached to postgres.exe when it\\n> > faults and the fault occurs at\\n> > https://github.com/postgres/postgres/blob/09eb633e1baa3b7cd7929f3cc77f9c46f63c20b1/src/backend/utils/mmgr/dsa.c#L869\\n> > span is pointing to 0x0\\n> \\n> Further data point. If I select a build type of 'debug' instead of\\n> 'debugoptimized' the error disappears.\\n\\nOh, this is quite interesting. Dave, could you post the backtrace?Andres,I am using Visual Studio as the debugger. Here is what I have.> postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C postgres.exe!resize(dshash_table * hash_table, unsigned __int64 new_size_log2) Line 879 C postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const void * key, bool * found) Line 480 C postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int dboid, unsigned int objoid, bool create, bool * created_entry) Line 455 C postgres.exe!pgstat_prep_pending_entry(PgStat_Kind kind, unsigned int dboid, unsigned int objoid, bool * created_entry) Line 1123 C [Inline Frame] postgres.exe!pgstat_prep_relation_pending(unsigned int) Line 904 C postgres.exe!pgstat_assoc_relation(RelationData * rel) Line 139 C [Inline Frame] postgres.exe!ReadBufferExtended(RelationData *) Line 802 C postgres.exe!ReadBuffer(RelationData * reln, unsigned int blockNum) Line 737 C postgres.exe!read_seq_tuple(RelationData * rel, int * buf, HeapTupleData * seqdatatuple) Line 1196 C postgres.exe!AlterSequence(ParseState * pstate, AlterSeqStmt * stmt) Line 481 C postgres.exe!ProcessUtilitySlow(ParseState * pstate, PlannedStmt * pstmt, const char * queryString, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 1679 C postgres.exe!standard_ProcessUtility(PlannedStmt * pstmt, const char * queryString, bool readOnlyTree, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 1080 C postgres.exe!ProcessUtility(PlannedStmt * pstmt, const char * queryString, bool readOnlyTree, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 530 C postgres.exe!ProcessUtilitySlow(ParseState * pstate, PlannedStmt * pstmt, const char * queryString, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 1263 C postgres.exe!standard_ProcessUtility(PlannedStmt * pstmt, const char * queryString, bool readOnlyTree, ProcessUtilityContext context, ParamListInfoData * params, QueryEnvironment * queryEnv, _DestReceiver * dest, QueryCompletion * qc) Line 1080 Cif there is a better debugger to use, please let me knowDave \\n\\nI wonder if this indicates that we are either missing memory barriers\\nsomewhere or that the memory barriers we end up with on msvc + arm aren't\\ncorrect? Either could explain why the problem doesn't occur when building\\nwith optimizations.\\n\\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Mon, 12 Feb 2024 16:01:01 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2024-02-12 12:50:12 -0800, Andres Freund wrote:\\n> On 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\\n> I wonder if this indicates that we are either missing memory barriers\\n> somewhere or that the memory barriers we end up with on msvc + arm aren't\\n> correct? Either could explain why the problem doesn't occur when building\\n> with optimizations.\\n\\nI think I might have been on to something - if my human emulation of a\\npreprocessor isn't wrong, we'd end up with\\n\\n#define S_UNLOCK(lock)\\t\\\\\\n\\tdo { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n\\non msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\\nlimits *compiler* level reordering, not CPU level reordering. I think it's\\neven insufficient on x86[-64], but it's definitely insufficient on arm.\\n\\n\\n\\nAnother issue I see is that we have a bunch of code that's dependant on\\n__aarch64__ being set - but it doesn't look to me that it is set on msvc. One\\nobvious consequence of that is that 64bit atomics won't be used (see\\nsrc/include/port/atomics/arch-arm.h) - but that shouldn't cause this issue.\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 12 Feb 2024 13:19:15 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 10, 2024 at 8:36 AM Andres Freund <andres@anarazel.de> wrote:\\n> Also, yikes, that's an ugly way of doing hardware detection. Jumping out of a\\n> signal handler into normal code. Brrr.\\n\\nMaybe it's a little baroque but what's actually wrong with it?\\nOpenSSL does something similar during initialisation as a fallback\\n(see armcap.c), and I borrowed that particular idea because I didn't\\nwant the rat's nest of unportable strategies required to detect the\\nfeature otherwise. As people who ran OpenSSL-linked stuff under gdb\\non 32 bit ARM discover (on e.g. early 32 bit RPis and early iOS etc\\nsystems the instruction was missing, and the signal dropped you into\\nthe debugger, so there are lots of reports of that if you google this\\ntopic). FWIW psql also long-jumps out of signal handlers, arguably\\nmore questionably because it's an async signal and the program counter\\nmay be inside readline...\\n\\nAs for Windows, let's see... SIGILL doesn't even work (the identifier\\nis defined for compatibility, but illegal instruction exception is not\\nconnected up to it), not to mention that in the backend we redirect\\nsigaction() to our own fake co-operative signal system so this\\ntechnique wouldn't work even if SIGILL did. But I would be surprised\\nif anyone actually has a platform that can run Windows that doesn't\\nhave this instruction (if they did I think we'd see a crash with\\nuntrapped illegal instruction exception 0xC000001D, let me put that\\nhere to help google to find this message if it ever happens...).\\nCRC32C is required by ARMv8.1[1], and was optional but AFAIK present\\nin general purpose ARMv8.0 cores including lowly RPis. In other words\\nanything < 8 years old[2] has it certainly, and older stuff probably\\ndoes too if it is a 64 bit server/PC/laptop/RPi? If you look for\\nreports of systems without it they seem to be pre-ARMv8, or early\\nphones/tablets (?)?\\n\\nIt looks like Windows 11 requires ARMv8.1[3] (\\\"the Windows kernel\\ndropped ARMv8.0 support shortly after builds 22621 [= 2022], now\\nrequiring and enforcing a minimum of ARMv8.1 and thus the new atomics\\ninstructions\\\", that atomic/lock stuff being a pretty solid reason to\\nwant to do so as discussed on this list before...). What if we just\\nassumed that the same must be effectively true in practice for Windows\\n10 until we hear otherwise? If a reasonable non-hypothetical system\\nshows up that rejects this instruction in the short time before we\\ndisclaim support for 10 (whose EOL is next year?), we can decide\\nwhether to just turn off the CRC32 fast path for that OS version, or\\nfigure out the runtime feature-probe patch. Clearly we need to\\nminimise Windows magic in this project due to lack of hackers, so\\nwriting code proactively for imaginary users seems like a bad plan...\\n\\n[1] https://developer.arm.com/documentation/ddi0597/2023-12/Base-Instructions/CRC32C--CRC32C-\\n[2] https://en.wikipedia.org/wiki/Comparison_of_ARM_processors#ARMv8-A\\n[3] http://www.emulators.com/docs/abc_history_of_woa.htm\\n\\n\\n\",\n \"msg_date\": \"Tue, 13 Feb 2024 15:07:17 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Dave Cramer\\nwww.postgres.rocks\\n\\n\\nOn Mon, 12 Feb 2024 at 16:19, Andres Freund <andres@anarazel.de> wrote:\\n\\n> Hi,\\n>\\n> On 2024-02-12 12:50:12 -0800, Andres Freund wrote:\\n> > On 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\\n> > I wonder if this indicates that we are either missing memory barriers\\n> > somewhere or that the memory barriers we end up with on msvc + arm aren't\\n> > correct? Either could explain why the problem doesn't occur when\\n> building\\n> > with optimizations.\\n>\\n> I think I might have been on to something - if my human emulation of a\\n> preprocessor isn't wrong, we'd end up with\\n>\\n> #define S_UNLOCK(lock) \\\\\\n> do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n>\\n> on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\\n> limits *compiler* level reordering, not CPU level reordering. I think it's\\n> even insufficient on x86[-64], but it's definitely insufficient on arm.\\n>\\nIn fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\\nLearn\\n<https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\\n\\nI did try using atomic_thread_fence as per atomic_thread_fence -\\ncppreference.com\\n<https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n\\n\\nHowever for some reason including #include <stdatomic.h>\\ncauses a bunch of compiler errors.\\n\\nc:\\\\Program Files\\\\Microsoft Visual\\nStudio\\\\2022\\\\Community\\\\VC\\\\Tools\\\\MSVC\\\\14.38.33130\\\\include\\\\vcruntime_c11_stdatomic.h(36):\\nerror C2061: syntax error: identifier 'atomic_bool'\\n\\nDave\\n\\nDave Cramerwww.postgres.rocksOn Mon, 12 Feb 2024 at 16:19, Andres Freund <andres@anarazel.de> wrote:Hi,\\n\\nOn 2024-02-12 12:50:12 -0800, Andres Freund wrote:\\n> On 2024-02-12 13:28:40 -0500, Andrew Dunstan wrote:\\n> I wonder if this indicates that we are either missing memory barriers\\n> somewhere or that the memory barriers we end up with on msvc + arm aren't\\n> correct? Either could explain why the problem doesn't occur when building\\n> with optimizations.\\n\\nI think I might have been on to something - if my human emulation of a\\npreprocessor isn't wrong, we'd end up with\\n\\n#define S_UNLOCK(lock) \\\\\\n do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n\\non msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\\nlimits *compiler* level reordering, not CPU level reordering. I think it's\\neven insufficient on x86[-64], but it's definitely insufficient on arm.In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft LearnI did try using atomic_thread_fence as per atomic_thread_fence - cppreference.comHowever for some reason including #include <stdatomic.h>causes a bunch of compiler errors.c:\\\\Program Files\\\\Microsoft Visual Studio\\\\2022\\\\Community\\\\VC\\\\Tools\\\\MSVC\\\\14.38.33130\\\\include\\\\vcruntime_c11_stdatomic.h(36): error C2061: syntax error: identifier 'atomic_bool'Dave\",\n \"msg_date\": \"Tue, 13 Feb 2024 12:49:33 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\\n> > I think I might have been on to something - if my human emulation of a\\n> > preprocessor isn't wrong, we'd end up with\\n> >\\n> > #define S_UNLOCK(lock) \\\\\\n> > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n> >\\n> > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\\n> > limits *compiler* level reordering, not CPU level reordering. I think it's\\n> > even insufficient on x86[-64], but it's definitely insufficient on arm.\\n> >\\n> In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\\n> Learn\\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\\n\\nI'd just ignore that, that's just pushing towards more modern stuff that's\\nmore applicable to C++ than C.\\n\\n\\n> I did try using atomic_thread_fence as per atomic_thread_fence -\\n> cppreference.com\\n> <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n\\nThe semantics of atomic_thread_fence are, uh, very odd. I'd just use\\nMemoryBarrier().\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 13 Feb 2024 09:52:17 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:\\n\\n> Hi,\\n>\\n> On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\\n> > > I think I might have been on to something - if my human emulation of a\\n> > > preprocessor isn't wrong, we'd end up with\\n> > >\\n> > > #define S_UNLOCK(lock) \\\\\\n> > > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n> > >\\n> > > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier()\\n> just\\n> > > limits *compiler* level reordering, not CPU level reordering. I think\\n> it's\\n> > > even insufficient on x86[-64], but it's definitely insufficient on arm.\\n> > >\\n> > In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier |\\n> Microsoft\\n> > Learn\\n> > <\\n> https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170\\n> >\\n>\\n> I'd just ignore that, that's just pushing towards more modern stuff that's\\n> more applicable to C++ than C.\\n>\\n>\\n> > I did try using atomic_thread_fence as per atomic_thread_fence -\\n> > cppreference.com\\n> > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n>\\n> The semantics of atomic_thread_fence are, uh, very odd. I'd just use\\n> MemoryBarrier().\\n>\\n> #define S_UNLOCK(lock) \\\\\\n do { MemoryBarrier(); (*(lock)) = 0; } while (0)\\n\\n#endif\\n\\nHas no effect.\\n\\nI have no idea if that is what you meant that I should do ?\\n\\nDave\\n\\nOn Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:Hi,\\n\\nOn 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\\n> > I think I might have been on to something - if my human emulation of a\\n> > preprocessor isn't wrong, we'd end up with\\n> >\\n> > #define S_UNLOCK(lock) \\\\\\n> > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n> >\\n> > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\\n> > limits *compiler* level reordering, not CPU level reordering. I think it's\\n> > even insufficient on x86[-64], but it's definitely insufficient on arm.\\n> >\\n> In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\\n> Learn\\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\\n\\nI'd just ignore that, that's just pushing towards more modern stuff that's\\nmore applicable to C++ than C.\\n\\n\\n> I did try using atomic_thread_fence as per atomic_thread_fence -\\n> cppreference.com\\n> <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n\\nThe semantics of atomic_thread_fence are, uh, very odd. I'd just use\\nMemoryBarrier().\\n#define S_UNLOCK(lock) \\\\ do { MemoryBarrier(); (*(lock)) = 0; } while (0)#endif Has no effect.I have no idea if that is what you meant that I should do ?Dave\",\n \"msg_date\": \"Tue, 13 Feb 2024 16:28:49 -0500\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks> wrote:\\n\\n>\\n>\\n>\\n> On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:\\n>\\n>> Hi,\\n>>\\n>> On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\\n>> > > I think I might have been on to something - if my human emulation of a\\n>> > > preprocessor isn't wrong, we'd end up with\\n>> > >\\n>> > > #define S_UNLOCK(lock) \\\\\\n>> > > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n>> > >\\n>> > > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier()\\n>> just\\n>> > > limits *compiler* level reordering, not CPU level reordering. I\\n>> think it's\\n>> > > even insufficient on x86[-64], but it's definitely insufficient on\\n>> arm.\\n>> > >\\n>> > In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier |\\n>> Microsoft\\n>> > Learn\\n>> > <\\n>> https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170\\n>> >\\n>>\\n>> I'd just ignore that, that's just pushing towards more modern stuff that's\\n>> more applicable to C++ than C.\\n>>\\n>>\\n>> > I did try using atomic_thread_fence as per atomic_thread_fence -\\n>> > cppreference.com\\n>> > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n>>\\n>> The semantics of atomic_thread_fence are, uh, very odd. I'd just use\\n>> MemoryBarrier().\\n>>\\n>> #define S_UNLOCK(lock) \\\\\\n> do { MemoryBarrier(); (*(lock)) = 0; } while (0)\\n>\\n> #endif\\n>\\n> Has no effect.\\n>\\n> I have no idea if that is what you meant that I should do ?\\n>\\n> Dave\\n>\\n\\n\\nRevisiting this:\\n\\nAndrew, can you explain the difference between ninja test (which passes)\\nand what the build farm does. The buildfarm crashes.\\n\\nDave\\n\\nOn Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks> wrote:On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:Hi,\\n\\nOn 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\\n> > I think I might have been on to something - if my human emulation of a\\n> > preprocessor isn't wrong, we'd end up with\\n> >\\n> > #define S_UNLOCK(lock) \\\\\\n> > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n> >\\n> > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\\n> > limits *compiler* level reordering, not CPU level reordering. I think it's\\n> > even insufficient on x86[-64], but it's definitely insufficient on arm.\\n> >\\n> In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\\n> Learn\\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\\n\\nI'd just ignore that, that's just pushing towards more modern stuff that's\\nmore applicable to C++ than C.\\n\\n\\n> I did try using atomic_thread_fence as per atomic_thread_fence -\\n> cppreference.com\\n> <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n\\nThe semantics of atomic_thread_fence are, uh, very odd. I'd just use\\nMemoryBarrier().\\n#define S_UNLOCK(lock) \\\\ do { MemoryBarrier(); (*(lock)) = 0; } while (0)#endif Has no effect.I have no idea if that is what you meant that I should do ?DaveRevisiting this:Andrew, can you explain the difference between ninja test (which passes) and what the build farm does. The buildfarm crashes.Dave\",\n \"msg_date\": \"Tue, 24 Sep 2024 14:31:10 -0400\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, 24 Sept 2024 at 14:31, Dave Cramer <davecramer@postgres.rocks>\\nwrote:\\n\\n>\\n>\\n> On Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks>\\n> wrote:\\n>\\n>>\\n>>\\n>>\\n>> On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:\\n>>\\n>>> Hi,\\n>>>\\n>>> On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\\n>>> > > I think I might have been on to something - if my human emulation of\\n>>> a\\n>>> > > preprocessor isn't wrong, we'd end up with\\n>>> > >\\n>>> > > #define S_UNLOCK(lock) \\\\\\n>>> > > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n>>> > >\\n>>> > > on msvc + arm. And that's entirely insufficient -\\n>>> _ReadWriteBarrier() just\\n>>> > > limits *compiler* level reordering, not CPU level reordering. I\\n>>> think it's\\n>>> > > even insufficient on x86[-64], but it's definitely insufficient on\\n>>> arm.\\n>>> > >\\n>>> > In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier |\\n>>> Microsoft\\n>>> > Learn\\n>>> > <\\n>>> https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170\\n>>> >\\n>>>\\n>>> I'd just ignore that, that's just pushing towards more modern stuff\\n>>> that's\\n>>> more applicable to C++ than C.\\n>>>\\n>>>\\n>>> > I did try using atomic_thread_fence as per atomic_thread_fence -\\n>>> > cppreference.com\\n>>> > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n>>>\\n>>> The semantics of atomic_thread_fence are, uh, very odd. I'd just use\\n>>> MemoryBarrier().\\n>>>\\n>>> #define S_UNLOCK(lock) \\\\\\n>> do { MemoryBarrier(); (*(lock)) = 0; } while (0)\\n>>\\n>> #endif\\n>>\\n>> Has no effect.\\n>>\\n>> I have no idea if that is what you meant that I should do ?\\n>>\\n>> Dave\\n>>\\n>\\n>\\n> Revisiting this:\\n>\\n> Andrew, can you explain the difference between ninja test (which passes)\\n> and what the build farm does. The buildfarm crashes.\\n>\\n\\nI have a dmp file with a stack trace if anyone is interested\\n\\nDave\\n\\n>\\n> Dave\\n>\\n\\nOn Tue, 24 Sept 2024 at 14:31, Dave Cramer <davecramer@postgres.rocks> wrote:On Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks> wrote:On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de> wrote:Hi,\\n\\nOn 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\\n> > I think I might have been on to something - if my human emulation of a\\n> > preprocessor isn't wrong, we'd end up with\\n> >\\n> > #define S_UNLOCK(lock) \\\\\\n> > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n> >\\n> > on msvc + arm. And that's entirely insufficient - _ReadWriteBarrier() just\\n> > limits *compiler* level reordering, not CPU level reordering. I think it's\\n> > even insufficient on x86[-64], but it's definitely insufficient on arm.\\n> >\\n> In fact ReadWriteBarrier has been deprecated _ReadWriteBarrier | Microsoft\\n> Learn\\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\\n\\nI'd just ignore that, that's just pushing towards more modern stuff that's\\nmore applicable to C++ than C.\\n\\n\\n> I did try using atomic_thread_fence as per atomic_thread_fence -\\n> cppreference.com\\n> <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n\\nThe semantics of atomic_thread_fence are, uh, very odd. I'd just use\\nMemoryBarrier().\\n#define S_UNLOCK(lock) \\\\ do { MemoryBarrier(); (*(lock)) = 0; } while (0)#endif Has no effect.I have no idea if that is what you meant that I should do ?DaveRevisiting this:Andrew, can you explain the difference between ninja test (which passes) and what the build farm does. The buildfarm crashes.I have a dmp file with a stack trace if anyone is interestedDave Dave\",\n \"msg_date\": \"Tue, 24 Sep 2024 16:01:30 -0400\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-09-24 16:01:30 -0400, Dave Cramer wrote:\\n> I have a dmp file with a stack trace if anyone is interested\\n\\n/me raises his hand\\n\\n\\n\",\n \"msg_date\": \"Tue, 24 Sep 2024 16:30:55 -0400\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On 2024-09-24 Tu 2:31 PM, Dave Cramer wrote:\\n>\\n>\\n> On Tue, 13 Feb 2024 at 16:28, Dave Cramer <davecramer@postgres.rocks> \\n> wrote:\\n>\\n>\\n>\\n>\\n> On Tue, 13 Feb 2024 at 12:52, Andres Freund <andres@anarazel.de>\\n> wrote:\\n>\\n> Hi,\\n>\\n> On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\\n> > > I think I might have been on to something - if my human\\n> emulation of a\\n> > > preprocessor isn't wrong, we'd end up with\\n> > >\\n> > > #define S_UNLOCK(lock) \\\\\\n> > > do { _ReadWriteBarrier(); (*(lock)) = 0; } while (0)\\n> > >\\n> > > on msvc + arm. And that's entirely insufficient -\\n> _ReadWriteBarrier() just\\n> > > limits *compiler* level reordering, not CPU level\\n> reordering. I think it's\\n> > > even insufficient on x86[-64], but it's definitely\\n> insufficient on arm.\\n> > >\\n> > In fact ReadWriteBarrier has been deprecated\\n> _ReadWriteBarrier | Microsoft\\n> > Learn\\n> >\\n> <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\\n>\\n> I'd just ignore that, that's just pushing towards more modern\\n> stuff that's\\n> more applicable to C++ than C.\\n>\\n>\\n> > I did try using atomic_thread_fence as per atomic_thread_fence -\\n> > cppreference.com <http://cppreference.com>\\n> > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n>\\n> The semantics of atomic_thread_fence are, uh, very odd. I'd\\n> just use\\n> MemoryBarrier().\\n>\\n> #defineS_UNLOCK(lock) \\\\\\n> do{ MemoryBarrier(); (*(lock)) =0; } while(0)\\n> #endif\\n> Has no effect.\\n>\\n> I have no idea if that is what you meant that I should do ?\\n>\\n> Dave\\n>\\n>\\n>\\n> Revisiting this:\\n>\\n> Andrew, can you explain the difference between ninja test (which \\n> passes) and what the build farm does. The buildfarm crashes.\\n\\n\\nThe buildfarm client performs these steps:\\n\\n\\n meson test -C $pgsql --no-rebuild --suite setup\\n meson test -t $meson_test_timeout $jflag -C $pgsql --logbase checklog --print-errorlogs --no-rebuild --suite regress --test-args=--no-locale\\n meson test -t $meson_test_timeout $jflag -C $pgsql --print-errorlogs --no-rebuild --logbase checkworld --no-suite setup --no-suite regress\\n foreach tested locale: meson test -t $meson_test_timeout $jflag -v -C $pgsql --no-rebuild --print-errorlogs --setup running --suite regress-running --logbase regress-installcheck-$locale\\n\\n\\n$pgsql is the build root, $jflag is setting the number of jobs\\n\\n\\nIOW, we do test suite setup, run the core regression tests, run all the \\nremaining non-install tests, then run the install tests for each locale.\\n\\nWe don't call ninja directly, but I don't see why that should make a \\ndifference.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2024-09-24 Tu 2:31 PM, Dave Cramer\\n wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Tue, 13 Feb 2024 at\\n 16:28, Dave Cramer <davecramer@postgres.rocks> wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\nOn Tue, 13 Feb 2024\\n at 12:52, Andres Freund <andres@anarazel.de>\\n wrote:\\n\\nHi,\\n\\n On 2024-02-13 12:49:33 -0500, Dave Cramer wrote:\\n > > I think I might have been on to something\\n - if my human emulation of a\\n > > preprocessor isn't wrong, we'd end up with\\n > >\\n > > #define S_UNLOCK(lock) \\\\\\n > > do { _ReadWriteBarrier();\\n (*(lock)) = 0; } while (0)\\n > >\\n > > on msvc + arm. And that's entirely\\n insufficient - _ReadWriteBarrier() just\\n > > limits *compiler* level reordering, not\\n CPU level reordering. I think it's\\n > > even insufficient on x86[-64], but it's\\n definitely insufficient on arm.\\n > >\\n > In fact ReadWriteBarrier has been deprecated\\n _ReadWriteBarrier | Microsoft\\n > Learn\\n > <https://learn.microsoft.com/en-us/cpp/intrinsics/readwritebarrier?view=msvc-170>\\n\\n I'd just ignore that, that's just pushing towards\\n more modern stuff that's\\n more applicable to C++ than C.\\n\\n\\n > I did try using atomic_thread_fence as per\\n atomic_thread_fence -\\n > cppreference.com\\n > <https://en.cppreference.com/w/c/atomic/atomic_thread_fence>\\n\\n The semantics of atomic_thread_fence are, uh, very\\n odd. I'd just use\\n MemoryBarrier().\\n\\n\\n\\n#define S_UNLOCK(lock) \\\\ do { MemoryBarrier(); (*(lock)) = 0; } while (0)\\n#endif\\n\\n \\nHas no effect.\\n\\n\\nI have no idea if that is what you meant that I\\n should do ?\\n\\n\\nDave\\n\\n\\n\\n\\n\\n\\n\\n\\nRevisiting this:\\n\\n\\nAndrew, can you explain the difference between ninja test\\n (which passes) and what the build farm does. The buildfarm\\n crashes.\\n\\n\\n\\n\\n\\nThe buildfarm client performs these steps:\\n\\n\\n\\nmeson test -C $pgsql --no-rebuild --suite setup\\nmeson test -t $meson_test_timeout $jflag -C $pgsql --logbase checklog --print-errorlogs --no-rebuild --suite regress --test-args=--no-locale\\nmeson test -t $meson_test_timeout $jflag -C $pgsql --print-errorlogs --no-rebuild --logbase checkworld --no-suite setup --no-suite regress \\nforeach tested locale: meson test -t $meson_test_timeout $jflag -v -C $pgsql --no-rebuild --print-errorlogs --setup running --suite regress-running --logbase regress-installcheck-$locale\\n\\n\\n\\n\\n$pgsql is the build root, $jflag is setting the number of jobs\\n\\n\\nIOW, we do test suite setup, run the core regression tests, run\\n all the remaining non-install tests, then run the install tests\\n for each locale.\\nWe don't call ninja directly, but I don't see why that should\\n make a difference.\\n\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Sat, 28 Sep 2024 09:01:15 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Tue, Feb 13, 2024 at 10:01 AM Dave Cramer <davecramer@postgres.rocks> wrote:\\n> > postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C\\n> postgres.exe!resize(dshash_table * hash_table, unsigned __int64 new_size_log2) Line 879 C\\n> postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const void * key, bool * found) Line 480 C\\n> postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int dboid, unsigned int objoid, bool create, bool * created_entry) Line 455 C\\n\\nHmm, we can see that the spinlock stuff is too weak (assumes TSO), but\\nthere aren't any spinlocks near that code so I think something else is\\nwrong. What does pg_read_barrier() compile to? I mention that\\nbecause it's used near there in a somewhat complicated way. But I'm a\\nbit confused because by reading through the code it looks like it\\nshould be too strong, not too weak. I think it should fall back to\\npg_memory_barrier_impl() for lack of pg_read_barrier_impl(), which\\nshould produce MemoryBarrier() and I guess DMB SY (ARM instruction for\\nfull system data memory barrier). So maybe some other pg_atomic_XXX\\noperations are falling back to code that doesn't work.\\n\\nJust for experimentation, you could see what happens if you redirect\\nall that stuff to C11's <stdatomic.h>. Here's a quick-and-dirty patch\\nfor that, which works on my ARM Mac, but I have no idea if it'll work\\non MSVC (our CI uses MSVC 2019; they only added <stdatomic.h> in MSVC\\n2022, but that's the same edition that added ARM so you must have it).\\nIt's not a very finely tuned patch (for example I think several calls\\nshould use the _explicit() variants and be weaker, but they err on the\\nside of being too strong, so the code they generate is probably not as\\nfast as it could be; I lost interested when the general idea was\\nrejected for now, but it might help learn something about the MSVC+ARM\\nsituation).\\n\\nCombined with the patch that redirects spinlocks to atomics[1], there\\nwould be no 'hand rolled' code relating to memory order or atomics,\\njust standard modern C. There are obviously other architecture tests\\nhere and there, and some of them will be wrong (some of them are wrong\\nalready for MSVC on x86), and that might be fixed by project\\nstandardisation[2].\\n\\n[1] https://www.postgresql.org/message-id/CA%2BhUKGJ%2BoA%2B62iUZ-EQb5R2cAOW3Y942ZoOtzOD%3D1sQ05iNg6Q%40mail.gmail.com\\n[2] https://www.postgresql.org/message-id/CA%2BhUKG%2BOu%2BC%2BpXioo_W1gKcwRCEN1WNrLBBPSMJyxjgc%2B1JEJA%40mail.gmail.com\",\n \"msg_date\": \"Sun, 29 Sep 2024 13:02:21 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n },\n {\n \"msg_contents\": \"On Sat, 28 Sept 2024 at 20:03, Thomas Munro <thomas.munro@gmail.com> wrote:\\n\\n> On Tue, Feb 13, 2024 at 10:01 AM Dave Cramer <davecramer@postgres.rocks>\\n> wrote:\\n> > > postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C\\n> > postgres.exe!resize(dshash_table * hash_table, unsigned __int64\\n> new_size_log2) Line 879 C\\n> > postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const\\n> void * key, bool * found) Line 480 C\\n> > postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int\\n> dboid, unsigned int objoid, bool create, bool * created_entry) Line 455 C\\n>\\n> Hmm, we can see that the spinlock stuff is too weak (assumes TSO), but\\n> there aren't any spinlocks near that code so I think something else is\\n> wrong. What does pg_read_barrier() compile to? I mention that\\n> because it's used near there in a somewhat complicated way. But I'm a\\n> bit confused because by reading through the code it looks like it\\n> should be too strong, not too weak. I think it should fall back to\\n> pg_memory_barrier_impl() for lack of pg_read_barrier_impl(), which\\n> should produce MemoryBarrier() and I guess DMB SY (ARM instruction for\\n> full system data memory barrier). So maybe some other pg_atomic_XXX\\n> operations are falling back to code that doesn't work.\\n>\\n> Just for experimentation, you could see what happens if you redirect\\n> all that stuff to C11's <stdatomic.h>. Here's a quick-and-dirty patch\\n> for that, which works on my ARM Mac, but I have no idea if it'll work\\n> on MSVC (our CI uses MSVC 2019; they only added <stdatomic.h> in MSVC\\n> 2022, but that's the same edition that added ARM so you must have it).\\n> It's not a very finely tuned patch (for example I think several calls\\n> should use the _explicit() variants and be weaker, but they err on the\\n> side of being too strong, so the code they generate is probably not as\\n> fast as it could be; I lost interested when the general idea was\\n> rejected for now, but it might help learn something about the MSVC+ARM\\n> situation).\\n>\\n> Combined with the patch that redirects spinlocks to atomics[1], there\\n> would be no 'hand rolled' code relating to memory order or atomics,\\n> just standard modern C. There are obviously other architecture tests\\n> here and there, and some of them will be wrong (some of them are wrong\\n> already for MSVC on x86), and that might be fixed by project\\n> standardisation[2].\\n>\\n> [1]\\n> https://www.postgresql.org/message-id/CA%2BhUKGJ%2BoA%2B62iUZ-EQb5R2cAOW3Y942ZoOtzOD%3D1sQ05iNg6Q%40mail.gmail.com\\n> [2]\\n> https://www.postgresql.org/message-id/CA%2BhUKG%2BOu%2BC%2BpXioo_W1gKcwRCEN1WNrLBBPSMJyxjgc%2B1JEJA%40mail.gmail.com\\n\\n\\n\\nThomas,\\n\\nThanks, I will be able to try this when I get back from NY. So Thurs or so.\\n\\nDave\\n\\nOn Sat, 28 Sept 2024 at 20:03, Thomas Munro <thomas.munro@gmail.com> wrote:On Tue, Feb 13, 2024 at 10:01 AM Dave Cramer <davecramer@postgres.rocks> wrote:\\n> > postgres.exe!dsa_free(dsa_area * area, unsigned __int64 dp) Line 869 C\\n> postgres.exe!resize(dshash_table * hash_table, unsigned __int64 new_size_log2) Line 879 C\\n> postgres.exe!dshash_find_or_insert(dshash_table * hash_table, const void * key, bool * found) Line 480 C\\n> postgres.exe!pgstat_get_entry_ref(PgStat_Kind kind, unsigned int dboid, unsigned int objoid, bool create, bool * created_entry) Line 455 C\\n\\nHmm, we can see that the spinlock stuff is too weak (assumes TSO), but\\nthere aren't any spinlocks near that code so I think something else is\\nwrong. What does pg_read_barrier() compile to? I mention that\\nbecause it's used near there in a somewhat complicated way. But I'm a\\nbit confused because by reading through the code it looks like it\\nshould be too strong, not too weak. I think it should fall back to\\npg_memory_barrier_impl() for lack of pg_read_barrier_impl(), which\\nshould produce MemoryBarrier() and I guess DMB SY (ARM instruction for\\nfull system data memory barrier). So maybe some other pg_atomic_XXX\\noperations are falling back to code that doesn't work.\\n\\nJust for experimentation, you could see what happens if you redirect\\nall that stuff to C11's <stdatomic.h>. Here's a quick-and-dirty patch\\nfor that, which works on my ARM Mac, but I have no idea if it'll work\\non MSVC (our CI uses MSVC 2019; they only added <stdatomic.h> in MSVC\\n2022, but that's the same edition that added ARM so you must have it).\\nIt's not a very finely tuned patch (for example I think several calls\\nshould use the _explicit() variants and be weaker, but they err on the\\nside of being too strong, so the code they generate is probably not as\\nfast as it could be; I lost interested when the general idea was\\nrejected for now, but it might help learn something about the MSVC+ARM\\nsituation).\\n\\nCombined with the patch that redirects spinlocks to atomics[1], there\\nwould be no 'hand rolled' code relating to memory order or atomics,\\njust standard modern C. There are obviously other architecture tests\\nhere and there, and some of them will be wrong (some of them are wrong\\nalready for MSVC on x86), and that might be fixed by project\\nstandardisation[2].\\n\\n[1] https://www.postgresql.org/message-id/CA%2BhUKGJ%2BoA%2B62iUZ-EQb5R2cAOW3Y942ZoOtzOD%3D1sQ05iNg6Q%40mail.gmail.com\\n[2] https://www.postgresql.org/message-id/CA%2BhUKG%2BOu%2BC%2BpXioo_W1gKcwRCEN1WNrLBBPSMJyxjgc%2B1JEJA%40mail.gmail.comThomas,Thanks, I will be able to try this when I get back from NY. So Thurs or so.Dave\",\n \"msg_date\": \"Sun, 29 Sep 2024 06:31:40 -0400\",\n \"msg_from\": \"Dave Cramer <davecramer@postgres.rocks>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Add native windows on arm64 support\"\n }\n]","isConversational":false}}},{"rowIdx":53564,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Attached are 3 small patches that improve the performance of numeric\ndivision. Patch 0002 seems to have the biggest impact, but I think\nthey're all worth including, since they're quite simple changes, with\nnoticeable performance gains.\n\n\nPatch 0001 vectorises the inner loop of div_var_fast(). This loop is\nalmost identical to the inner loop of mul_var(), which was vectorised\nby commit 8870917623. The thing preventing the div_var_fast() loop\nfrom being vectorised is the assignment to div[qi + i], and simply\nreplacing that with div_qi[i] where div_qi = &div[qi], in the same\nstyle as mul_var(), fixes that.\n\nOne difference between this and the mul_var() code is that it is also\nnecessary to add an explicit cast to get the compiler to generate\nmatching output, and give the best results. This is because in\nmul_var() the compiler recognises that var1digit is actually 16-bit,\nrather than 32-bit, and so it doesn't need to multiply the high part,\nbut in div_var_fast() it's not obvious to the compiler that qdigit\nalso fits in 16 bits, hence the cast.\n\nThe actual performance gain is typically quite modest, since\ndiv_var_fast() is always only a part of some more complex numeric\ncomputation, but it can be seen in cases like raising numbers to\nnegative integer powers, e.g.:\n\nCREATE TEMP TABLE num_test(x numeric);\nSELECT setseed(0);\nINSERT INTO num_test\n SELECT (SELECT ('1.'||string_agg((random()*9)::int::text, '')||x)::numeric\n FROM generate_series(1,100))\n FROM generate_series(1,100000) g(x);\n\nSELECT sum(x^(-2)) FROM num_test;\n\nTime: 112.949 ms (HEAD)\nTime: 98.537 ms (with patch)\n\n\nPatch 0002 simplifies the inner loop of div_var() (the guts of the\npublic-facing numeric division operator) by more closely combining the\nmultiplication and subtraction operations and folding the separate\n\"carry\" and \"borrow\" variables into a single \"borrow\", as suggested by\nthe old code comment.\n\nIMO, this makes the code easier to understand, as well as giving more\nsignificant performance gains:\n\nCREATE TEMP TABLE div_test(x numeric);\nSELECT setseed(0);\nINSERT INTO div_test\n SELECT (SELECT ('1.'||string_agg((random()*9)::int::text, ''))::numeric + x\n FROM generate_series(1,50))\n FROM generate_series(1,1000) g(x);\n\nSELECT sum(x/y) FROM div_test t1(x), div_test t2(y);\n\nTime: 1477.340 ms (HEAD)\nTime: 826.748 ms (with patch)\n\n\nPatch 0003 replaces some uses of div_var_fast() with div_var().\nSpecifically, when the divisor has just one or two base-NBASE digits,\ndiv_var() is faster. This is especially true for 1-digit divisors, due\nto the \"fast path\" code in div_var() to handle that. It's also just\nabout true for 2-digit divisors, and it occurs to me that that case\ncould potentially be optimised further with similar fast path code in\ndiv_var(), but I haven't tried that yet.\n\nOne-digit divisors are quite common. For example, in the Taylor series\nexpansions in exp_var() and ln_var(), since the number of Taylor\nseries terms never exceeds a few hundred in practice. Also,\nexp_var()'s argument reduction step divides by 2^ndiv2, which is\nroughly 100 times the input, rounded up to a power of two, and valid\ninputs are less than 6000, so this will always be one or two digits.\n\nTesting this with a bunch of random exp() and ln() computations I saw\na speed-up of 15-20%, and it reduced the run time of the numeric-big\nregression test by around 10%, which seems worth having.\n\nRegards,\nDean","msg_date":"Wed, 23 Feb 2022 13:34:59 +0000","msg_from":"Dean Rasheed <dean.a.rasheed@gmail.com>","msg_from_op":true,"msg_subject":"Some optimisations for numeric division"},{"msg_contents":"Dean Rasheed <dean.a.rasheed@gmail.com> writes:\n> Attached are 3 small patches that improve the performance of numeric\n> division. Patch 0002 seems to have the biggest impact, but I think\n> they're all worth including, since they're quite simple changes, with\n> noticeable performance gains.\n\nI took a quick look through these (just eyeball, didn't try to verify\nyour performance statements). I'm +1 on 0001 and 0002, but 0003 feels\na bit ad-hoc. It certainly *looks* weird for the allegedly faster\nfunction to be handing off to the allegedly slower one. I also wonder\nif we're leaving anything on the table by not exploiting\ndiv_var_fast's weaker roundoff guarantees in this case. Should we\nthink about a more thoroughgoing redesign of these functions' APIs?\nAnother idea is to only worry about the single-divisor-digit\noptimization, and just copy div_var's (very small) inner loop for that\ncase into div_var_fast.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 23 Feb 2022 15:55:12 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Some optimisations for numeric division"},{"msg_contents":"On Wed, 23 Feb 2022 at 20:55, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> I took a quick look through these (just eyeball, didn't try to verify\n> your performance statements).\n\nThanks for looking!\n\n> I'm +1 on 0001 and 0002, but 0003 feels\n> a bit ad-hoc. It certainly *looks* weird for the allegedly faster\n> function to be handing off to the allegedly slower one. I also wonder\n> if we're leaving anything on the table by not exploiting\n> div_var_fast's weaker roundoff guarantees in this case. Should we\n> think about a more thoroughgoing redesign of these functions' APIs?\n\nHmm, I'm not sure what kind of thing you had in mind.\n\nOne thought that occurred to me was that it's a bit silly that\nexp_var() and ln_var() have to use a NumericVar for what could just be\nan int, if we had a div_var_int() function that could divide by an\nint. Then both div_var() and div_var_fast() could hand off to it for\none and two digit divisors.\n\nRegards,\nDean\n\n\n","msg_date":"Wed, 23 Feb 2022 22:46:50 +0000","msg_from":"Dean Rasheed <dean.a.rasheed@gmail.com>","msg_from_op":true,"msg_subject":"Re: Some optimisations for numeric division"},{"msg_contents":"Dean Rasheed <dean.a.rasheed@gmail.com> writes:\n> On Wed, 23 Feb 2022 at 20:55, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>> I'm +1 on 0001 and 0002, but 0003 feels\n>> a bit ad-hoc. It certainly *looks* weird for the allegedly faster\n>> function to be handing off to the allegedly slower one. I also wonder\n>> if we're leaving anything on the table by not exploiting\n>> div_var_fast's weaker roundoff guarantees in this case. Should we\n>> think about a more thoroughgoing redesign of these functions' APIs?\n\n> Hmm, I'm not sure what kind of thing you had in mind.\n\nI'm not either, tbh. Just seems like this needs more than some\nhacking around the margins.\n\n> One thought that occurred to me was that it's a bit silly that\n> exp_var() and ln_var() have to use a NumericVar for what could just be\n> an int, if we had a div_var_int() function that could divide by an\n> int. Then both div_var() and div_var_fast() could hand off to it for\n> one and two digit divisors.\n\nOooh, that seems like a good idea.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 23 Feb 2022 17:55:02 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Some optimisations for numeric division"},{"msg_contents":"On Wed, 23 Feb 2022 at 22:55, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Dean Rasheed <dean.a.rasheed@gmail.com> writes:\n>\n> > One thought that occurred to me was that it's a bit silly that\n> > exp_var() and ln_var() have to use a NumericVar for what could just be\n> > an int, if we had a div_var_int() function that could divide by an\n> > int. Then both div_var() and div_var_fast() could hand off to it for\n> > one and two digit divisors.\n>\n> Oooh, that seems like a good idea.\n>\n\nOK, I've replaced the 0003 patch with one that does that instead. The\ndiv_var_int() API is slightly different in that it also accepts a\ndivisor weight argument, but the alternative would have been for the\ncallers to have to adjust both the result weight and rscale, which\nwould have been uglier.\n\nThere's a large block of code in div_var() that needs re-indenting,\nbut I think it would be better to leave that to a later pgindent run.\n\nThe performance results are quite pleasing. It's slightly faster than\nthe old one-digit div_var() code because it manages to avoid some\ndigit array copying, and for two digit divisors it's much faster:\n\nCREATE TEMP TABLE div_test(x numeric, y numeric);\nSELECT setseed(0);\nINSERT INTO div_test\n SELECT (SELECT (((x%9)+1)||string_agg((random()*9)::int::text, ''))::numeric\n FROM generate_series(1,50)),\n (SELECT ('1.'||string_agg((random()*9)::int::text,\n'')||(x%10)||'e3')::numeric\n FROM generate_series(1,6))\n FROM generate_series(1,5000) g(x);\nselect * from div_test limit 10;\n\nSELECT sum(t1.x/t2.y) FROM div_test t1, div_test t2;\n\nTime: 11600.034 ms (HEAD)\nTime: 9890.788 ms (with 0002)\nTime: 6202.851 ms (with 0003)\n\nAnd obviously it'll be a larger relative gain for div_var_fast(),\nsince that was slower to begin with in such cases.\n\nThis makes me think that it might also be worthwhile to follow this\nwith a similar div_var_int64() function on platforms with 128-bit\nintegers, which could then be used to handle 3- and 4-digit divisors,\nwhich are probably quite common in practice.\n\nAttached is the updated patch series (0001 and 0002 unchanged).\n\nRegards,\nDean","msg_date":"Fri, 25 Feb 2022 10:45:31 +0000","msg_from":"Dean Rasheed <dean.a.rasheed@gmail.com>","msg_from_op":true,"msg_subject":"Re: Some optimisations for numeric division"},{"msg_contents":"On Fri, 25 Feb 2022 at 10:45, Dean Rasheed <dean.a.rasheed@gmail.com> wrote:\n>\n> Attached is the updated patch series (0001 and 0002 unchanged).\n>\n\nAnd another update following feedback from the cfbot.\n\nRegards,\nDean","msg_date":"Fri, 25 Feb 2022 12:43:48 +0000","msg_from":"Dean Rasheed <dean.a.rasheed@gmail.com>","msg_from_op":true,"msg_subject":"Re: Some optimisations for numeric division"},{"msg_contents":"Dean Rasheed <dean.a.rasheed@gmail.com> writes:\n> And another update following feedback from the cfbot.\n\nThis patchset LGTM. On my machine there doesn't seem to be any\nmeasurable performance change for the numeric regression test,\nbut numeric_big gets about 15% faster.\n\nThe only additional thought I have, based on your comments about\n0001, is that maybe in mul_var we should declare var1digit as\nbeing NumericDigit not int. I tried that here and didn't see\nany material change in the generated assembly code (using gcc\n8.5.0), so you're right that modern gcc doesn't need any help\nthere; but I wonder if it could help on other compiler versions.\n\nI'll mark this RFC.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 25 Feb 2022 13:30:33 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Some optimisations for numeric division"},{"msg_contents":"On Fri, 25 Feb 2022 at 18:30, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Dean Rasheed <dean.a.rasheed@gmail.com> writes:\n> > And another update following feedback from the cfbot.\n>\n> This patchset LGTM. On my machine there doesn't seem to be any\n> measurable performance change for the numeric regression test,\n> but numeric_big gets about 15% faster.\n>\n\nYes, that matches my observations. Thanks for reviewing and testing.\n\n> The only additional thought I have, based on your comments about\n> 0001, is that maybe in mul_var we should declare var1digit as\n> being NumericDigit not int. I tried that here and didn't see\n> any material change in the generated assembly code (using gcc\n> 8.5.0), so you're right that modern gcc doesn't need any help\n> there; but I wonder if it could help on other compiler versions.\n>\n\nYes, that makes sense. Done that way.\n\nRegards,\nDean\n\n\n","msg_date":"Sun, 27 Feb 2022 11:23:27 +0000","msg_from":"Dean Rasheed <dean.a.rasheed@gmail.com>","msg_from_op":true,"msg_subject":"Re: Some optimisations for numeric division"}],"string":"[\n {\n \"msg_contents\": \"Attached are 3 small patches that improve the performance of numeric\\ndivision. Patch 0002 seems to have the biggest impact, but I think\\nthey're all worth including, since they're quite simple changes, with\\nnoticeable performance gains.\\n\\n\\nPatch 0001 vectorises the inner loop of div_var_fast(). This loop is\\nalmost identical to the inner loop of mul_var(), which was vectorised\\nby commit 8870917623. The thing preventing the div_var_fast() loop\\nfrom being vectorised is the assignment to div[qi + i], and simply\\nreplacing that with div_qi[i] where div_qi = &div[qi], in the same\\nstyle as mul_var(), fixes that.\\n\\nOne difference between this and the mul_var() code is that it is also\\nnecessary to add an explicit cast to get the compiler to generate\\nmatching output, and give the best results. This is because in\\nmul_var() the compiler recognises that var1digit is actually 16-bit,\\nrather than 32-bit, and so it doesn't need to multiply the high part,\\nbut in div_var_fast() it's not obvious to the compiler that qdigit\\nalso fits in 16 bits, hence the cast.\\n\\nThe actual performance gain is typically quite modest, since\\ndiv_var_fast() is always only a part of some more complex numeric\\ncomputation, but it can be seen in cases like raising numbers to\\nnegative integer powers, e.g.:\\n\\nCREATE TEMP TABLE num_test(x numeric);\\nSELECT setseed(0);\\nINSERT INTO num_test\\n SELECT (SELECT ('1.'||string_agg((random()*9)::int::text, '')||x)::numeric\\n FROM generate_series(1,100))\\n FROM generate_series(1,100000) g(x);\\n\\nSELECT sum(x^(-2)) FROM num_test;\\n\\nTime: 112.949 ms (HEAD)\\nTime: 98.537 ms (with patch)\\n\\n\\nPatch 0002 simplifies the inner loop of div_var() (the guts of the\\npublic-facing numeric division operator) by more closely combining the\\nmultiplication and subtraction operations and folding the separate\\n\\\"carry\\\" and \\\"borrow\\\" variables into a single \\\"borrow\\\", as suggested by\\nthe old code comment.\\n\\nIMO, this makes the code easier to understand, as well as giving more\\nsignificant performance gains:\\n\\nCREATE TEMP TABLE div_test(x numeric);\\nSELECT setseed(0);\\nINSERT INTO div_test\\n SELECT (SELECT ('1.'||string_agg((random()*9)::int::text, ''))::numeric + x\\n FROM generate_series(1,50))\\n FROM generate_series(1,1000) g(x);\\n\\nSELECT sum(x/y) FROM div_test t1(x), div_test t2(y);\\n\\nTime: 1477.340 ms (HEAD)\\nTime: 826.748 ms (with patch)\\n\\n\\nPatch 0003 replaces some uses of div_var_fast() with div_var().\\nSpecifically, when the divisor has just one or two base-NBASE digits,\\ndiv_var() is faster. This is especially true for 1-digit divisors, due\\nto the \\\"fast path\\\" code in div_var() to handle that. It's also just\\nabout true for 2-digit divisors, and it occurs to me that that case\\ncould potentially be optimised further with similar fast path code in\\ndiv_var(), but I haven't tried that yet.\\n\\nOne-digit divisors are quite common. For example, in the Taylor series\\nexpansions in exp_var() and ln_var(), since the number of Taylor\\nseries terms never exceeds a few hundred in practice. Also,\\nexp_var()'s argument reduction step divides by 2^ndiv2, which is\\nroughly 100 times the input, rounded up to a power of two, and valid\\ninputs are less than 6000, so this will always be one or two digits.\\n\\nTesting this with a bunch of random exp() and ln() computations I saw\\na speed-up of 15-20%, and it reduced the run time of the numeric-big\\nregression test by around 10%, which seems worth having.\\n\\nRegards,\\nDean\",\n \"msg_date\": \"Wed, 23 Feb 2022 13:34:59 +0000\",\n \"msg_from\": \"Dean Rasheed <dean.a.rasheed@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Some optimisations for numeric division\"\n },\n {\n \"msg_contents\": \"Dean Rasheed <dean.a.rasheed@gmail.com> writes:\\n> Attached are 3 small patches that improve the performance of numeric\\n> division. Patch 0002 seems to have the biggest impact, but I think\\n> they're all worth including, since they're quite simple changes, with\\n> noticeable performance gains.\\n\\nI took a quick look through these (just eyeball, didn't try to verify\\nyour performance statements). I'm +1 on 0001 and 0002, but 0003 feels\\na bit ad-hoc. It certainly *looks* weird for the allegedly faster\\nfunction to be handing off to the allegedly slower one. I also wonder\\nif we're leaving anything on the table by not exploiting\\ndiv_var_fast's weaker roundoff guarantees in this case. Should we\\nthink about a more thoroughgoing redesign of these functions' APIs?\\nAnother idea is to only worry about the single-divisor-digit\\noptimization, and just copy div_var's (very small) inner loop for that\\ncase into div_var_fast.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 15:55:12 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Some optimisations for numeric division\"\n },\n {\n \"msg_contents\": \"On Wed, 23 Feb 2022 at 20:55, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> I took a quick look through these (just eyeball, didn't try to verify\\n> your performance statements).\\n\\nThanks for looking!\\n\\n> I'm +1 on 0001 and 0002, but 0003 feels\\n> a bit ad-hoc. It certainly *looks* weird for the allegedly faster\\n> function to be handing off to the allegedly slower one. I also wonder\\n> if we're leaving anything on the table by not exploiting\\n> div_var_fast's weaker roundoff guarantees in this case. Should we\\n> think about a more thoroughgoing redesign of these functions' APIs?\\n\\nHmm, I'm not sure what kind of thing you had in mind.\\n\\nOne thought that occurred to me was that it's a bit silly that\\nexp_var() and ln_var() have to use a NumericVar for what could just be\\nan int, if we had a div_var_int() function that could divide by an\\nint. Then both div_var() and div_var_fast() could hand off to it for\\none and two digit divisors.\\n\\nRegards,\\nDean\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 22:46:50 +0000\",\n \"msg_from\": \"Dean Rasheed <dean.a.rasheed@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Some optimisations for numeric division\"\n },\n {\n \"msg_contents\": \"Dean Rasheed <dean.a.rasheed@gmail.com> writes:\\n> On Wed, 23 Feb 2022 at 20:55, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>> I'm +1 on 0001 and 0002, but 0003 feels\\n>> a bit ad-hoc. It certainly *looks* weird for the allegedly faster\\n>> function to be handing off to the allegedly slower one. I also wonder\\n>> if we're leaving anything on the table by not exploiting\\n>> div_var_fast's weaker roundoff guarantees in this case. Should we\\n>> think about a more thoroughgoing redesign of these functions' APIs?\\n\\n> Hmm, I'm not sure what kind of thing you had in mind.\\n\\nI'm not either, tbh. Just seems like this needs more than some\\nhacking around the margins.\\n\\n> One thought that occurred to me was that it's a bit silly that\\n> exp_var() and ln_var() have to use a NumericVar for what could just be\\n> an int, if we had a div_var_int() function that could divide by an\\n> int. Then both div_var() and div_var_fast() could hand off to it for\\n> one and two digit divisors.\\n\\nOooh, that seems like a good idea.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 17:55:02 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Some optimisations for numeric division\"\n },\n {\n \"msg_contents\": \"On Wed, 23 Feb 2022 at 22:55, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Dean Rasheed <dean.a.rasheed@gmail.com> writes:\\n>\\n> > One thought that occurred to me was that it's a bit silly that\\n> > exp_var() and ln_var() have to use a NumericVar for what could just be\\n> > an int, if we had a div_var_int() function that could divide by an\\n> > int. Then both div_var() and div_var_fast() could hand off to it for\\n> > one and two digit divisors.\\n>\\n> Oooh, that seems like a good idea.\\n>\\n\\nOK, I've replaced the 0003 patch with one that does that instead. The\\ndiv_var_int() API is slightly different in that it also accepts a\\ndivisor weight argument, but the alternative would have been for the\\ncallers to have to adjust both the result weight and rscale, which\\nwould have been uglier.\\n\\nThere's a large block of code in div_var() that needs re-indenting,\\nbut I think it would be better to leave that to a later pgindent run.\\n\\nThe performance results are quite pleasing. It's slightly faster than\\nthe old one-digit div_var() code because it manages to avoid some\\ndigit array copying, and for two digit divisors it's much faster:\\n\\nCREATE TEMP TABLE div_test(x numeric, y numeric);\\nSELECT setseed(0);\\nINSERT INTO div_test\\n SELECT (SELECT (((x%9)+1)||string_agg((random()*9)::int::text, ''))::numeric\\n FROM generate_series(1,50)),\\n (SELECT ('1.'||string_agg((random()*9)::int::text,\\n'')||(x%10)||'e3')::numeric\\n FROM generate_series(1,6))\\n FROM generate_series(1,5000) g(x);\\nselect * from div_test limit 10;\\n\\nSELECT sum(t1.x/t2.y) FROM div_test t1, div_test t2;\\n\\nTime: 11600.034 ms (HEAD)\\nTime: 9890.788 ms (with 0002)\\nTime: 6202.851 ms (with 0003)\\n\\nAnd obviously it'll be a larger relative gain for div_var_fast(),\\nsince that was slower to begin with in such cases.\\n\\nThis makes me think that it might also be worthwhile to follow this\\nwith a similar div_var_int64() function on platforms with 128-bit\\nintegers, which could then be used to handle 3- and 4-digit divisors,\\nwhich are probably quite common in practice.\\n\\nAttached is the updated patch series (0001 and 0002 unchanged).\\n\\nRegards,\\nDean\",\n \"msg_date\": \"Fri, 25 Feb 2022 10:45:31 +0000\",\n \"msg_from\": \"Dean Rasheed <dean.a.rasheed@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Some optimisations for numeric division\"\n },\n {\n \"msg_contents\": \"On Fri, 25 Feb 2022 at 10:45, Dean Rasheed <dean.a.rasheed@gmail.com> wrote:\\n>\\n> Attached is the updated patch series (0001 and 0002 unchanged).\\n>\\n\\nAnd another update following feedback from the cfbot.\\n\\nRegards,\\nDean\",\n \"msg_date\": \"Fri, 25 Feb 2022 12:43:48 +0000\",\n \"msg_from\": \"Dean Rasheed <dean.a.rasheed@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Some optimisations for numeric division\"\n },\n {\n \"msg_contents\": \"Dean Rasheed <dean.a.rasheed@gmail.com> writes:\\n> And another update following feedback from the cfbot.\\n\\nThis patchset LGTM. On my machine there doesn't seem to be any\\nmeasurable performance change for the numeric regression test,\\nbut numeric_big gets about 15% faster.\\n\\nThe only additional thought I have, based on your comments about\\n0001, is that maybe in mul_var we should declare var1digit as\\nbeing NumericDigit not int. I tried that here and didn't see\\nany material change in the generated assembly code (using gcc\\n8.5.0), so you're right that modern gcc doesn't need any help\\nthere; but I wonder if it could help on other compiler versions.\\n\\nI'll mark this RFC.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 13:30:33 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Some optimisations for numeric division\"\n },\n {\n \"msg_contents\": \"On Fri, 25 Feb 2022 at 18:30, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Dean Rasheed <dean.a.rasheed@gmail.com> writes:\\n> > And another update following feedback from the cfbot.\\n>\\n> This patchset LGTM. On my machine there doesn't seem to be any\\n> measurable performance change for the numeric regression test,\\n> but numeric_big gets about 15% faster.\\n>\\n\\nYes, that matches my observations. Thanks for reviewing and testing.\\n\\n> The only additional thought I have, based on your comments about\\n> 0001, is that maybe in mul_var we should declare var1digit as\\n> being NumericDigit not int. I tried that here and didn't see\\n> any material change in the generated assembly code (using gcc\\n> 8.5.0), so you're right that modern gcc doesn't need any help\\n> there; but I wonder if it could help on other compiler versions.\\n>\\n\\nYes, that makes sense. Done that way.\\n\\nRegards,\\nDean\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Feb 2022 11:23:27 +0000\",\n \"msg_from\": \"Dean Rasheed <dean.a.rasheed@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Some optimisations for numeric division\"\n }\n]","isConversational":false}}},{"rowIdx":53565,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi!\n\nDuring a Patroni PR discussion \n(https://github.com/zalando/patroni/pull/2225), we realised that if one \nwants to use the \"-c\" option on a typical Debian/Ubuntu installation \n(where the config resides below /etc/postgresql/), pg_rewind needs a way \nto be told where the postgresql.conf actually is.\n\nThe attached patch implements this as an option.\n\nI'm extremely unhappy that I called it \"--config-file\" here, while \n\"postgres\" itself wants \"--config_file\". But as the other dashed options \nof pg_rewind are, well, dashed and not underscored, it seemed to be \nbetter that the other way...\n\nAs the \"-c\" feature appeared in 12 and existing Debian installations are \nunable to use it right now, I suggest to even backport the patch.\n\nOh, and I'm still not subscribed to -hackers, so the usual \"please CC \nme\" applies!\n\nBest regards,\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato","msg_date":"Wed, 23 Feb 2022 16:28:09 +0100","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"On Wed, 2022-02-23 at 16:28 +0100, Gunnar \"Nick\" Bluth wrote:\r\n> I'm extremely unhappy that I called it \"--config-file\" here, while \r\n> \"postgres\" itself wants \"--config_file\". But as the other dashed options \r\n> of pg_rewind are, well, dashed and not underscored, it seemed to be \r\n> better that the other way...\r\n\r\nThe \"--\"-style options to postgres take either underscores or dashes in\r\nthe names (just double-checked with PG14), so this shouldn't be a\r\nproblem.\r\n\r\n--Jacob\r\n","msg_date":"Wed, 23 Feb 2022 17:21:15 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Hi Gunnar,\n\n> During a Patroni PR discussion\n> (https://github.com/zalando/patroni/pull/2225), we realised that if one\n> wants to use the \"-c\" option on a typical Debian/Ubuntu installation\n> (where the config resides below /etc/postgresql/), pg_rewind needs a way\n> to be told where the postgresql.conf actually is.\n>\n> The attached patch implements this as an option.\n>\n> [...]\n\nGood catch!\n\nCould you also implement a TAP test for the new code?\n\n-- \nBest regards,\nAleksander Alekseev\n\n\n","msg_date":"Thu, 24 Feb 2022 12:02:54 +0300","msg_from":"Aleksander Alekseev <aleksander@timescale.com>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Hello Gunnar,\n\nOn Thu, 24 Feb 2022 at 10:03, Aleksander Alekseev <aleksander@timescale.com>\nwrote:\n\n\n> > wants to use the \"-c\" option on a typical Debian/Ubuntu installation\n> > (where the config resides below /etc/postgresql/), pg_rewind needs a way\n> > to be told where the postgresql.conf actually is.\n> >\n> > The attached patch implements this as an option.\n> >\n> > [...]\n>\n> Good catch!\n>\n\nYeah, it is a known problem, and there was already another patch trying to\naddress it [1].\nHonestly, I like your approach much better, because, the previous make was\nassuming that the data_directory GUC is specified in the postgresql.conf,\nwhich is not very flexible.\n\n\n>\n> Could you also implement a TAP test for the new code?\n>\n\n+1\n\n\n+ <para>\n+ In case the <filename>postgresql.conf</filename> of your target\ncluster is not in the\n+ target pgdata and you want to use the <option>-c</option> option,\n+ provide a (relative or absolute) path to the\n<filename>postgresql.conf</filename> using this option.\n+ </para>\n\nIt took me a while to understand the meaning of <option>-c</option>. Maybe\nchanging it to <option>--restore-target-wal</option> will make it easier to\nunderstand.\n\n[1]\nhttps://www.postgresql.org/message-id/flat/16982-f12294cccd221480%40postgresql.org\n\nRegards,\n--\nAlexander Kukushkin\n\nHello Gunnar,On Thu, 24 Feb 2022 at 10:03, Aleksander Alekseev <aleksander@timescale.com> wrote: > wants to use the \"-c\" option on a typical Debian/Ubuntu installation\n> (where the config resides below /etc/postgresql/), pg_rewind needs a way\n> to be told where the postgresql.conf actually is.\n>\n> The attached patch implements this as an option.\n>\n> [...]\n\nGood catch!Yeah, it is a known problem, and there was already another patch trying to address it [1].Honestly, I like your approach much better, because, the previous make was assuming that the data_directory GUC is specified in the postgresql.conf, which is not very flexible. \n\nCould you also implement a TAP test for the new code?+1 + <para>+ In case the <filename>postgresql.conf</filename> of your target cluster is not in the + target pgdata and you want to use the <option>-c</option> option,+ provide a (relative or absolute) path to the <filename>postgresql.conf</filename> using this option.+ </para>It took me a while to understand the meaning of <option>-c</option>. Maybe changing it to <option>--restore-target-wal</option> will make it easier to understand.[1] https://www.postgresql.org/message-id/flat/16982-f12294cccd221480%40postgresql.orgRegards,--Alexander Kukushkin","msg_date":"Thu, 24 Feb 2022 10:27:30 +0100","msg_from":"Alexander Kukushkin <cyberdemn@gmail.com>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Hi hackers,\n\n> It took me a while to understand the meaning of <option>-c</option>. Maybe changing it to <option>--restore-target-wal</option> will make it easier to understand.\n\n+1, I \"stumbled\" while reading this too at first.\n\n-- \nBest regards,\nAleksander Alekseev\n\n\n","msg_date":"Thu, 24 Feb 2022 12:58:33 +0300","msg_from":"Aleksander Alekseev <aleksander@timescale.com>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"> On 24 Feb 2022, at 10:27, Alexander Kukushkin <cyberdemn@gmail.com> wrote:\n> \n> Hello Gunnar,\n> \n> On Thu, 24 Feb 2022 at 10:03, Aleksander Alekseev <aleksander@timescale.com> wrote:\n> \n> > wants to use the \"-c\" option on a typical Debian/Ubuntu installation\n> > (where the config resides below /etc/postgresql/), pg_rewind needs a way\n> > to be told where the postgresql.conf actually is.\n> >\n> > The attached patch implements this as an option.\n> >\n> > [...]\n> \n> Good catch!\n> \n> Yeah, it is a known problem, and there was already another patch trying to address it [1].\n\nThere is yet another related patch which provides a parameter to pass in\nrestore_command in cases where postgresq.conf isn't reachable:\n\n https://commitfest.postgresql.org/37/3213/\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Thu, 24 Feb 2022 14:08:40 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Am 24.02.2022 um 14:08 schrieb Daniel Gustafsson:\n>> On 24 Feb 2022, at 10:27, Alexander Kukushkin <cyberdemn@gmail.com> wrote:\n>>\n>> Hello Gunnar,\n>>\n>> On Thu, 24 Feb 2022 at 10:03, Aleksander Alekseev <aleksander@timescale.com> wrote:\n>> \n>>> wants to use the \"-c\" option on a typical Debian/Ubuntu installation\n>>> (where the config resides below /etc/postgresql/), pg_rewind needs a way\n>>> to be told where the postgresql.conf actually is.\n>>>\n>>> The attached patch implements this as an option.\n>>>\n>>> [...]\n>>\n>> Good catch!\n>>\n>> Yeah, it is a known problem, and there was already another patch trying to address it [1].\n> \n> There is yet another related patch which provides a parameter to pass in\n> restore_command in cases where postgresq.conf isn't reachable:\n> \n> https://commitfest.postgresql.org/37/3213/\n\nThat looks just as good to me, and it already has tests, so I'll happily \nstep down!\n\n(Note to myself: check the CF first next time ;-)\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato\n\n\n","msg_date":"Thu, 24 Feb 2022 14:43:17 +0100","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"> On 24 Feb 2022, at 14:43, Gunnar Nick Bluth <gunnar.bluth@pro-open.de> wrote:\n\n> That looks just as good to me, and it already has tests, so I'll happily step down!\n\nActually, I think this looks like a saner approach. Putting a config setting\nin two place (postgresql.conf and on the commandline for pg_rewind) is a recipe\nfor them diverging.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Thu, 24 Feb 2022 14:46:49 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Am 24.02.22 um 14:46 schrieb Daniel Gustafsson:\n>> On 24 Feb 2022, at 14:43, Gunnar Nick Bluth <gunnar.bluth@pro-open.de> wrote:\n> \n>> That looks just as good to me, and it already has tests, so I'll happily step down!\n> \n> Actually, I think this looks like a saner approach. Putting a config setting\n> in two place (postgresql.conf and on the commandline for pg_rewind) is a recipe\n> for them diverging.\n\nWell, well,\n\nreverted https://commitfest.postgresql.org/37/3562/ back to \"needs \nreview\" and moved it to \"Replication & Recovery\".\n\nPatch v2 includes:\n* doc improvements as suggested\n* tests shamelessly copied & adapted from \nhttps://commitfest.postgresql.org/37/3213/\n\nSince the need is pretty obvious (I *think* the Debian-ish userbase is \nquite significant alone), I'd love to see one of the approaches making \nit into 15!\n\nBest regards,\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato","msg_date":"Fri, 25 Feb 2022 10:35:49 +0100","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"On Fri, Feb 25, 2022 at 10:35:49AM +0100, Gunnar \"Nick\" Bluth wrote:\n> Am 24.02.22 um 14:46 schrieb Daniel Gustafsson:\n>> Actually, I think this looks like a saner approach. Putting a config setting\n>> in two place (postgresql.conf and on the commandline for pg_rewind) is a recipe\n>> for them diverging.\n\nFWIW, I have a bad feeling about passing down directly a command\nthrough an option itself part of a command, so what's discussed on\nthis thread is refreshing.\n\n+ } else {\n+ snprintf(postgres_cmd, sizeof(postgres_cmd),\n+ \"\\\"%s\\\" -D \\\"%s\\\" --config_file=\\\"%s\\\" -C restore_command\",\n+ postgres_exec_path, datadir_target, config_file);\n+ }\n\nShouldn't this one use appendShellString() on config_file?\n--\nMichael","msg_date":"Sat, 26 Feb 2022 14:51:14 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Am 26.02.22 um 06:51 schrieb Michael Paquier:\n> On Fri, Feb 25, 2022 at 10:35:49AM +0100, Gunnar \"Nick\" Bluth wrote:\n>> Am 24.02.22 um 14:46 schrieb Daniel Gustafsson:\n>>> Actually, I think this looks like a saner approach. Putting a config setting\n>>> in two place (postgresql.conf and on the commandline for pg_rewind) is a recipe\n>>> for them diverging.\n> \n> FWIW, I have a bad feeling about passing down directly a command\n> through an option itself part of a command, so what's discussed on\n> this thread is refreshing.\n\nTa.\n\n> \n> + } else {\n> + snprintf(postgres_cmd, sizeof(postgres_cmd),\n> + \"\\\"%s\\\" -D \\\"%s\\\" --config_file=\\\"%s\\\" -C restore_command\",\n> + postgres_exec_path, datadir_target, config_file);\n> + }\n> \n> Shouldn't this one use appendShellString() on config_file?\n\nIt probably should, yes. I don't fancy this repetitive code myself.\nBut then, pg_rewind as a whole could use an overhaul. I don't see any \nuse of PQExpBuffer in it, but a lot of char* ...\n\nI myself am not a C hacker (as you can probably tell already) and don't \nreally feel fit (enough) to rewrite pg_rewind.c to use \nfe_utils/string_utils.h :/\n\nI might give it a try anyhow, but that may be a bit... heavy,,, just to \nadd one option?\n\nWhile we're at it (I'm really good at opening cans of worms):\n09:47 $ grep -r -l --include \\*.c fe_utils/string_utils.h src/bin/ | wc -l\n28\n09:48 $ grep -r -L --include \\*.c fe_utils/string_utils.h src/bin/ | wc -l\n66\n\nGSOC project? ;-)\n\nBest,\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato\n\n\n","msg_date":"Sat, 26 Feb 2022 09:55:20 +0100","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"On Sat, Feb 26, 2022 at 09:55:20AM +0100, Gunnar \"Nick\" Bluth wrote:\n> Am 26.02.22 um 06:51 schrieb Michael Paquier:\n>> Shouldn't this one use appendShellString() on config_file?\n> \n> It probably should, yes. I don't fancy this repetitive code myself.\n> But then, pg_rewind as a whole could use an overhaul. I don't see any use of\n> PQExpBuffer in it, but a lot of char* ...\n\nHaving a lot of char* does not necessarily mean that all of them have\nto be changed to accomodate with PQExpBuffer. My guess that this is a\ncase-by-case, where we should apply that in places where it makes the\ncode cleaner to follow. In the case of your patch though, the two\nareas changed would make the logic correct, instead, because we have\nto apply correct quoting rules to any commands executed.\n\n> GSOC project? ;-)\n\nIt does not seem so, though there are surely more areas that could\ngain in clarity.\n--\nMichael","msg_date":"Sun, 27 Feb 2022 21:06:59 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Am 27.02.22 um 13:06 schrieb Michael Paquier:\n> On Sat, Feb 26, 2022 at 09:55:20AM +0100, Gunnar \"Nick\" Bluth wrote:\n>> Am 26.02.22 um 06:51 schrieb Michael Paquier:\n>>> Shouldn't this one use appendShellString() on config_file?\n>>\n>> It probably should, yes. I don't fancy this repetitive code myself.\n>> But then, pg_rewind as a whole could use an overhaul. I don't see any use of\n>> PQExpBuffer in it, but a lot of char* ...\n> \n> Having a lot of char* does not necessarily mean that all of them have\n> to be changed to accomodate with PQExpBuffer. My guess that this is a\n> case-by-case, where we should apply that in places where it makes the\n> code cleaner to follow. In the case of your patch though, the two\n> areas changed would make the logic correct, instead, because we have\n> to apply correct quoting rules to any commands executed.\n\nAlas! v3 attached.\n\n\n>> GSOC project? ;-)\n> \n> It does not seem so, though there are surely more areas that could\n> gain in clarity.\n\nThat's universally true ;-)\n\n\nBest regards,\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato","msg_date":"Sun, 27 Feb 2022 14:30:33 +0100","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"On Sun, Feb 27, 2022 at 02:30:33PM +0100, Gunnar \"Nick\" Bluth wrote:\n> That's universally true ;-)\n\n-# Internal routine to enable archive recovery command on a standby node\n+# Internal routine to enable archive recovery command on a standby node.\n+# Returns generated restore_command.\n sub enable_restoring\n {\n \tmy ($self, $root_node, $standby) = @_;\n@@ -1103,7 +1104,7 @@ restore_command = '$copy_command'\n \t{\n \t\t$self->set_recovery_mode();\n \t}\n-\treturn;\n+\treturn $copy_command;\n\nI don't like this change. This makes an existing routine do some\nextra work for something it is not designed for. You could get this\ninformation with a postgres -C command, for example. Now, you cannot\nuse postgres -C because of.. Reasons (1a9d802). But you could use a\npg_ctl command for the same purpose. I guess that we'd better\nrefactor this into a new routine of Cluster.pm where we execute a\npg_ctl command and return both stdout and stderr back to the caller?\n\nThe TAP part could be refactored on its own.\n\n+ In case the <filename>postgresql.conf</filename> of your\ntarget cluster is not in the\n+ target pgdata and you want to use the <option>-c /\n--restore-target-wal</option> option,\n+ provide a (relative or absolute) path to the\n<filename>postgresql.conf</filename> using this option.\n\nThis could do a better job to explain in which cases this option can\nbe used for. You could say something like that:\n\"This option specifies a path to a configuration file to be used when\nstarting the target cluster. This makes easier the use of\n-c/--restore-target-wal by setting restore_command in the extra\nconfiguration file specified via this option.\"\n\nHmm. What about the target cluster started in --single mode as of\nensureCleanShutdown()? Should we try to apply this option also in\nthis case?\n--\nMichael","msg_date":"Mon, 28 Feb 2022 20:56:24 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Am 28.02.22 um 12:56 schrieb Michael Paquier:\n> On Sun, Feb 27, 2022 at 02:30:33PM +0100, Gunnar \"Nick\" Bluth wrote:\n>> That's universally true ;-)\n> \n> -# Internal routine to enable archive recovery command on a standby node\n> +# Internal routine to enable archive recovery command on a standby node.\n> +# Returns generated restore_command.\n> sub enable_restoring\n> {\n> \tmy ($self, $root_node, $standby) = @_;\n> @@ -1103,7 +1104,7 @@ restore_command = '$copy_command'\n> \t{\n> \t\t$self->set_recovery_mode();\n> \t}\n> -\treturn;\n> +\treturn $copy_command;\n> \n> I don't like this change. This makes an existing routine do some\n> extra work for something it is not designed for. You could get this\n> information with a postgres -C command, for example. Now, you cannot\n> use postgres -C because of.. Reasons (1a9d802). But you could use a\n> pg_ctl command for the same purpose. I guess that we'd better\n> refactor this into a new routine of Cluster.pm where we execute a\n> pg_ctl command and return both stdout and stderr back to the caller?\n\nTurns out this change isn't needed anyway (it was copied from the other \npatch). Afterall, the patch is about extracting the restore_command from \nthe config, so...\n\n> The TAP part could be refactored on its own.\n\nAgreed. I've struggled quite a bit even understanding what's happening \nin there ;-)\n\n\n> + In case the <filename>postgresql.conf</filename> of your\n> target cluster is not in the\n> + target pgdata and you want to use the <option>-c /\n> --restore-target-wal</option> option,\n> + provide a (relative or absolute) path to the\n> <filename>postgresql.conf</filename> using this option.\n> \n> This could do a better job to explain in which cases this option can\n> be used for. You could say something like that:\n> \"This option specifies a path to a configuration file to be used when\n> starting the target cluster. This makes easier the use of\n> -c/--restore-target-wal by setting restore_command in the extra\n> configuration file specified via this option.\"\n\nI reviewed the wording and think it is pretty universal now.\n\n\n> Hmm. What about the target cluster started in --single mode as of\n> ensureCleanShutdown()? Should we try to apply this option also in\n> this case?\n\nI'd say so; as far as I can tell, \"postgres\" would deny starting a \n(Debian/non-standard) cluster the way the code is right now.\n\nWhich led me to realize we have\n\n /* find postgres executable */\n rc = find_other_exec(argv0, \"postgres\",\n PG_BACKEND_VERSIONSTR,\n postgres_exec_path);\nin getRestoreCommand _and_\n\n /* locate postgres binary */\n if ((ret = find_other_exec(argv0, \"postgres\", \n\n\t\t\t\t\t\tPG_BACKEND_VERSIONSTR,\n exec_path)) < 0)\nin ensureCleanShutdown.\nBoth with the same error messages, AFAICS. D'oh... can of worms.\n\nSo, how should we call the global \"find the ***** 'postgres' executable \nand boil out if that fails\" function?\n char postgres_exec_path[MAXPGPATH] = findPostgresExec();\n?\n\n\nAnyway, v4 attached so we can settle on the tests and wording...\n\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato","msg_date":"Mon, 28 Feb 2022 20:48:23 +0100","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"On Mon, Feb 28, 2022 at 08:48:23PM +0100, Gunnar \"Nick\" Bluth wrote:\n> So, how should we call the global \"find the ***** 'postgres' executable and\n> boil out if that fails\" function?\n> char postgres_exec_path[MAXPGPATH] = findPostgresExec();\n> ?\n\nThat would mean only a couple of lines gained, and the readability\ngained is minimal, so I'd be fine to keep the code as-is. I am not\nsure about the full patch set yet, but the refactoring of the commands\nto use PQExpBuffer is good by itself, so I have extracted this part of\nthe patch and applied that for now.\n--\nMichael","msg_date":"Tue, 1 Mar 2022 13:18:23 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Am 01.03.22 um 05:18 schrieb Michael Paquier:\n> On Mon, Feb 28, 2022 at 08:48:23PM +0100, Gunnar \"Nick\" Bluth wrote:\n>> So, how should we call the global \"find the ***** 'postgres' executable and\n>> boil out if that fails\" function?\n>> char postgres_exec_path[MAXPGPATH] = findPostgresExec();\n>> ?\n> \n> That would mean only a couple of lines gained, and the readability\n> gained is minimal, so I'd be fine to keep the code as-is. I am not\n> sure about the full patch set yet, but the refactoring of the commands\n> to use PQExpBuffer is good by itself, so I have extracted this part of\n> the patch and applied that for now.\n\nWhich made the remaining patch v5 (which also appends the --config-file \nfor ensureCleanShutdown) much cleaner :)\n\nCheers,\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato","msg_date":"Tue, 1 Mar 2022 12:35:27 +0100","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Heya,\n\nsome minor comments, I didn't look at the added test and I did not test\nthe patch at all, but (as part of the Debian/Ubuntu packaging team) I\nthink this patch is really important:\n\nOn Tue, Mar 01, 2022 at 12:35:27PM +0100, Gunnar \"Nick\" Bluth wrote:\n> diff --git a/doc/src/sgml/ref/pg_rewind.sgml b/doc/src/sgml/ref/pg_rewind.sgml\n> index 33e6bb64ad..d0278e9b46 100644\n> --- a/doc/src/sgml/ref/pg_rewind.sgml\n> +++ b/doc/src/sgml/ref/pg_rewind.sgml\n> @@ -241,6 +241,18 @@ PostgreSQL documentation\n> </listitem>\n> </varlistentry>\n> \n> + <varlistentry>\n> + <term><option>--config-file=<replaceable class=\"parameter\">filepath</replaceable></option></term>\n> + <listitem>\n> + <para>\n> + When using the <option>-c / --restore-target-wal</option> option, the <varname>restore_command</varname>\n> + is extracted from the configuration of the target cluster. If the <filename>postgresql.conf</filename>\n> + of that cluster is not in the target pgdata directory (or if you want to use an alternative config),\n\nI think we usually just say \"data directory\"; pgdata was previously only\nused as part of the option name, not like this in the text.\n\n> diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c\n> index b39b5c1aac..294353a9d5 100644\n> --- a/src/bin/pg_rewind/pg_rewind.c\n> +++ b/src/bin/pg_rewind/pg_rewind.c\n> @@ -61,6 +61,7 @@ char\t *datadir_target = NULL;\n> char\t *datadir_source = NULL;\n> char\t *connstr_source = NULL;\n> char\t *restore_command = NULL;\n> +char\t *config_file = NULL;\n> \n> static bool debug = false;\n> bool\t\tshowprogress = false;\n> @@ -87,6 +88,7 @@ usage(const char *progname)\n> \tprintf(_(\"Options:\\n\"));\n> \tprintf(_(\" -c, --restore-target-wal use restore_command in target configuration to\\n\"\n> \t\t\t \" retrieve WAL files from archives\\n\"));\n> +\tprintf(_(\" --config-file=FILE path to postgresql.conf if outside target-pgdata (for -c)\\n\"));\n\n> \tprintf(_(\" -D, --target-pgdata=DIRECTORY existing data directory to modify\\n\"));\n> \tprintf(_(\" --source-pgdata=DIRECTORY source data directory to synchronize with\\n\"));\n> \tprintf(_(\" --source-server=CONNSTR source server to synchronize with\\n\"));\n\ntarget-pgdata is the name of the option, but maybe it is useful to spell\nout \"target data directory\" here, even if that means we spill over to\ntwo lines (which we might have to do anyway, that new line is pretty\nlong).\n\n> @@ -121,6 +123,7 @@ main(int argc, char **argv)\n> \t\t{\"no-sync\", no_argument, NULL, 'N'},\n> \t\t{\"progress\", no_argument, NULL, 'P'},\n> \t\t{\"debug\", no_argument, NULL, 3},\n> +\t\t{\"config-file\", required_argument, NULL, 5},\n\nNot sure what our policy is, but the way the numbered options are 1, 2,\n4, 3, 5 after this is weird; this was already off before this patch\nthough.\n\n> \t\t{NULL, 0, NULL, 0}\n> \t};\n> \tint\t\t\toption_index;\n> @@ -205,6 +208,10 @@ main(int argc, char **argv)\n> \t\t\tcase 4:\n> \t\t\t\tno_ensure_shutdown = true;\n> \t\t\t\tbreak;\n> +\n> +\t\t\tcase 5:\n> +\t\t\t\tconfig_file = pg_strdup(optarg);\n> +\t\t\t\tbreak;\n> \t\t}\n> \t}\n> \n> @@ -1060,7 +1067,11 @@ getRestoreCommand(const char *argv0)\n> \t/* add -D switch, with properly quoted data directory */\n> \tappendPQExpBufferStr(postgres_cmd, \" -D \");\n> \tappendShellString(postgres_cmd, datadir_target);\n> -\n> +\tif (config_file != NULL)\n> +\t{\n> +\t\tappendPQExpBufferStr(postgres_cmd, \" --config_file=\");\n> +\t\tappendShellString(postgres_cmd, config_file);\n> +\t}\n> \t/* add -C switch, for restore_command */\n\nYou removed an empty line here. Maybe rather prepend this with a comment\non what's going on, and have en empty line before and afterwards.\n\n> \tappendPQExpBufferStr(postgres_cmd, \" -C restore_command\");\n> \n> @@ -1138,6 +1149,11 @@ ensureCleanShutdown(const char *argv0)\n> \t/* add set of options with properly quoted data directory */\n> \tappendPQExpBufferStr(postgres_cmd, \" --single -F -D \");\n> \tappendShellString(postgres_cmd, datadir_target);\n> +\tif (config_file != NULL)\n> +\t{\n> +\t\tappendPQExpBufferStr(postgres_cmd, \" --config_file=\");\n> +\t\tappendShellString(postgres_cmd, config_file);\n> +\t}\n> \n> \t/* finish with the database name, and a properly quoted redirection */\n\nKinde same here.\n\n\nCheers,\n\nMichael\n\n-- \nMichael Banck\nTeamleiter PostgreSQL-Team\nProjektleiter\nTel.: +49 2166 9901-171\nEmail: michael.banck@credativ.de\n\ncredativ GmbH, HRB M�nchengladbach 12080\nUSt-ID-Nummer: DE204566209\nTrompeterallee 108, 41189 M�nchengladbach\nGesch�ftsf�hrung: Dr. Michael Meskes, Geoff Richardson, Peter Lilley\n\nUnser Umgang mit personenbezogenen Daten unterliegt\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\n\n\n","msg_date":"Thu, 10 Mar 2022 14:43:32 +0100","msg_from":"Michael Banck <michael.banck@credativ.de>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Am 10.03.22 um 14:43 schrieb Michael Banck:\n> Heya,\n> \n> some minor comments, I didn't look at the added test and I did not test\n> the patch at all, but (as part of the Debian/Ubuntu packaging team) I\n> think this patch is really important:\n\nMuch appreciated!\n\n[...]\n\n> I think we usually just say \"data directory\"; pgdata was previously only\n> used as part of the option name, not like this in the text.\n\nFixed.\n\n[...]\n\n> target-pgdata is the name of the option, but maybe it is useful to spell\n> out \"target data directory\" here, even if that means we spill over to\n> two lines (which we might have to do anyway, that new line is pretty\n> long).\n\nFixed.\n\n> \n>> @@ -121,6 +123,7 @@ main(int argc, char **argv)\n>> \t\t{\"no-sync\", no_argument, NULL, 'N'},\n>> \t\t{\"progress\", no_argument, NULL, 'P'},\n>> \t\t{\"debug\", no_argument, NULL, 3},\n>> +\t\t{\"config-file\", required_argument, NULL, 5},\n> \n> Not sure what our policy is, but the way the numbered options are 1, 2,\n> 4, 3, 5 after this is weird; this was already off before this patch\n> though.\n\nThat one has been triggering my inner Monk too ;-)\n\nFixed!\n\n\n[...]\n\n> You removed an empty line here. Maybe rather prepend this with a comment\n> on what's going on, and have en empty line before and afterwards.\n\n> Kinde same here.\n\nIt does smell a bit like \"stating the obvious\", but alas! Both fixed.\n\nCheers,\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato","msg_date":"Thu, 10 Mar 2022 15:28:57 +0100","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Hi,\n\nAm Donnerstag, dem 10.03.2022 um 15:28 +0100 schrieb Gunnar \"Nick\"\nBluth:\n> Am 10.03.22 um 14:43 schrieb Michael Banck:\n> > some minor comments, I didn't look at the added test and I did not\n> > test the patch at all, but (as part of the Debian/Ubuntu packaging\n> > team) I think this patch is really important:\n> \n> Much appreciated!\n> \n> [...]\n> \n> Fixed.\n> \n> [...]\n\nThanks for the updated patch.\n\nThe patch applies, make is ok, make check is ok, pg_rewind TAP tests\nare ok.\n\nI did some tests now with Patroni 2.1.3 and the attached patch applied.\n\nThe following test was made:\n\n1. Deploy 3-node (pg1, pg2, pg3) patroni cluster on Debian unstable\n running postgresql-15 (approx. master)\n2. Switch on archive_mode, and set archive_command and restore_command\n3. Switchover so that pg1 is the primary (if not already)\n4. Kill pg1 hard\n5. Wait till a new leader has taken over and the (now 2-node) cluster \n is healthy again\n6. Restart pg1\n\nwithout this patch:\n\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,084 INFO: running pg_rewind from pg3\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,102 INFO: running pg_rewind from dbname=postgres user=postgres_rewind host=10.0.3.227 port=5432 target_session_attrs=read-write\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,135 INFO: pg_rewind exit code=1\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,136 INFO: stdout=\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,136 INFO: stderr=postgres: could not access the server configuration file \"/var/lib/postgresql/15/main/postgresql.conf\": No such file or directory\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: no data was returned by command \"/usr/lib/postgresql/15/bin/postgres -D /var/lib/postgresql/15/main -C restore_command\"\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,149 ERROR: Failed to rewind from healty master: pg3\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,149 WARNING: remove_data_directory_on_diverged_timelines is set. removing...\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,149 INFO: Removing data directory: /var/lib/postgresql/15/main\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,245 INFO: Lock owner: pg3; I am pg1\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,248 INFO: trying to bootstrap from leader 'pg3'\n\nSo pg_rewind fails and Patroni does a re-bootstrap.\n\nwith this patch:\n\n|Apr 03 19:12:35 pg1 patroni@15-main[99]: 2022-04-03 19:12:35,576 INFO: running pg_rewind from pg2\n|Apr 03 19:12:35 pg1 patroni@15-main[99]: 2022-04-03 19:12:35,590 INFO: running pg_rewind from dbname=postgres user=postgres_rewind host=10.0.3.145 port=5432 target_session_attrs=read-write\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,147 INFO: pg_rewind exit code=0\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,148 INFO: stdout=\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,148 INFO: stderr=pg_rewind: servers diverged at WAL location 0/1D000180 on timeline 38\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: pg_rewind: rewinding from last common checkpoint at 0/1D000108 on timeline 38\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: pg_rewind: Done!\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,151 WARNING: Postgresql is not running.\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,151 INFO: Lock owner: pg2; I am pg1\n\nHere, pg_rewind is called and works fine.\n\nSo I think this works as intended, and I'm marking it Ready for\nCommitter.\n\n\nMichael\n\n-- \nMichael Banck\nTeamleiter PostgreSQL-Team\nProjektleiter\nTel.: +49 2166 9901-171\nE-Mail: michael.banck@credativ.de\n\ncredativ GmbH, HRB Mönchengladbach 12080\nUSt-ID-Nummer: DE204566209\nTrompeterallee 108, 41189 Mönchengladbach\nGeschäftsführung: Dr. Michael Meskes, Geoff Richardson, Peter Lilley\n\nUnser Umgang mit personenbezogenen Daten unterliegt\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz","msg_date":"Sun, 03 Apr 2022 21:31:32 +0200","msg_from":"Michael Banck <michael.banck@credativ.de>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Just reposting the previous patches to straighten out the cfbot.","msg_date":"Mon, 4 Apr 2022 11:39:45 -0400","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"On Sun, Apr 03, 2022 at 09:31:32PM +0200, Michael Banck wrote:\n> The patch applies, make is ok, make check is ok, pg_rewind TAP tests\n> are ok.\n\n+ appendPQExpBufferStr(postgres_cmd, \" --config_file=\");\n+ appendShellString(postgres_cmd, config_file);\nNit. This is a valid option for the postgres command, but I'd rather\nuse a -c switch instead, mostly as a matter of self-documentation.\n\nIn the tests, the only difference between the modes \"archive_cli\" and\n\"archive\" is the extra option given to the pg_rewind command, and\nthat's a simple redirect to what pg_rewind would choose by default\nanyway. A more elegant solution would be to have an extra option in\nRewindTest::run_pg_rewind(), where any caller of the routine can feed\nextra options to the pg_rewind command. Now, in this case, we are\nnot going to lose any coverage if the existing \"archive\" mode is\nchanged so as we move away postgresql.conf from the target data folder\nand just use --config-file by default, no? I would make the choice of\nsimplicity, by giving up on \"archive_cli\" and using\nprimary-postgresql.conf.tmp as value for --config-file. There could\nbe an argument for using --config-file for all the modes, as well.\n--\nMichael","msg_date":"Tue, 5 Apr 2022 11:10:30 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"On Tue, Apr 05, 2022 at 11:10:30AM +0900, Michael Paquier wrote:\n> In the tests, the only difference between the modes \"archive_cli\" and\n> \"archive\" is the extra option given to the pg_rewind command, and\n> that's a simple redirect to what pg_rewind would choose by default\n> anyway. A more elegant solution would be to have an extra option in\n> RewindTest::run_pg_rewind(), where any caller of the routine can feed\n> extra options to the pg_rewind command. Now, in this case, we are\n> not going to lose any coverage if the existing \"archive\" mode is\n> changed so as we move away postgresql.conf from the target data folder\n> and just use --config-file by default, no? I would make the choice of\n> simplicity, by giving up on \"archive_cli\" and using\n> primary-postgresql.conf.tmp as value for --config-file. There could\n> be an argument for using --config-file for all the modes, as well.\n\nThe clock is ticking, so I have looked at this patch by myself. I\nhave finished by tweaking the docs, the code and the tests as of the \nattached. One thing that I found annoying is the lack of description\nabout the fact that this option affects pg_rewind when it internally\nstarts the target cluster, as of when we retrieve restore_command and\nwhen we enforce crash recovery to have a target cluster with a clean\nshutdown state. The tests are heavily simplified, having no impact on\nthe run-time while improving the coverage for the code paths changed\nhere.\n--\nMichael","msg_date":"Wed, 6 Apr 2022 21:04:50 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Am 06.04.22 um 14:04 schrieb Michael Paquier:\n> On Tue, Apr 05, 2022 at 11:10:30AM +0900, Michael Paquier wrote:\n>> In the tests, the only difference between the modes \"archive_cli\" and\n>> \"archive\" is the extra option given to the pg_rewind command, and\n>> that's a simple redirect to what pg_rewind would choose by default\n>> anyway. A more elegant solution would be to have an extra option in\n>> RewindTest::run_pg_rewind(), where any caller of the routine can feed\n>> extra options to the pg_rewind command. Now, in this case, we are\n>> not going to lose any coverage if the existing \"archive\" mode is\n>> changed so as we move away postgresql.conf from the target data folder\n>> and just use --config-file by default, no? I would make the choice of\n>> simplicity, by giving up on \"archive_cli\" and using\n>> primary-postgresql.conf.tmp as value for --config-file. There could\n>> be an argument for using --config-file for all the modes, as well.\n> \n> The clock is ticking, so I have looked at this patch by myself. I\n\nTa! Sorry, been ridiculously busy these days, and I do confess that I've\nbeen struggling a bit with those tests before ;-)\n\n> have finished by tweaking the docs, the code and the tests as of the \n> attached. One thing that I found annoying is the lack of description\n> about the fact that this option affects pg_rewind when it internally\n> starts the target cluster, as of when we retrieve restore_command and\n> when we enforce crash recovery to have a target cluster with a clean\n> shutdown state. The tests are heavily simplified, having no impact on\n> the run-time while improving the coverage for the code paths changed\n> here.\n> --\n> Michael\n\nLGTM!\n\nThx again!\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato\n\n\n","msg_date":"Wed, 6 Apr 2022 14:32:44 +0200","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"On Wed, Apr 06, 2022 at 02:32:44PM +0200, Gunnar \"Nick\" Bluth wrote:\n> Ta! Sorry, been ridiculously busy these days, and I do confess that I've\n> been struggling a bit with those tests before ;-)\n\nNo problem. Double-checked this morning and applied, then.\n--\nMichael","msg_date":"Thu, 7 Apr 2022 08:54:29 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"},{"msg_contents":"Am 07.04.22 um 01:54 schrieb Michael Paquier:\n> On Wed, Apr 06, 2022 at 02:32:44PM +0200, Gunnar \"Nick\" Bluth wrote:\n>> Ta! Sorry, been ridiculously busy these days, and I do confess that I've\n>> been struggling a bit with those tests before ;-)\n> \n> No problem. Double-checked this morning and applied, then.\n> --\n> Michael\n\nWonderful!\n\nSo, @cyberdemn, who would have thought we get this fixed so quickly?\nTime to update the Patroni PR once more! ;-)\n\nThanks & best regards,\n-- \nGunnar \"Nick\" Bluth\n\nEimermacherweg 106\nD-48159 Münster\n\nMobil +49 172 8853339\nEmail: gunnar.bluth@pro-open.de\n__________________________________________________________________________\n\"Ceterum censeo SystemD esse delendam\" - Cato\n\n\n","msg_date":"Thu, 7 Apr 2022 09:51:01 +0200","msg_from":"\"Gunnar \\\"Nick\\\" Bluth\" <gunnar.bluth@pro-open.de>","msg_from_op":true,"msg_subject":"Re: PATCH: add \"--config-file=\" option to pg_rewind"}],"string":"[\n {\n \"msg_contents\": \"Hi!\\n\\nDuring a Patroni PR discussion \\n(https://github.com/zalando/patroni/pull/2225), we realised that if one \\nwants to use the \\\"-c\\\" option on a typical Debian/Ubuntu installation \\n(where the config resides below /etc/postgresql/), pg_rewind needs a way \\nto be told where the postgresql.conf actually is.\\n\\nThe attached patch implements this as an option.\\n\\nI'm extremely unhappy that I called it \\\"--config-file\\\" here, while \\n\\\"postgres\\\" itself wants \\\"--config_file\\\". But as the other dashed options \\nof pg_rewind are, well, dashed and not underscored, it seemed to be \\nbetter that the other way...\\n\\nAs the \\\"-c\\\" feature appeared in 12 and existing Debian installations are \\nunable to use it right now, I suggest to even backport the patch.\\n\\nOh, and I'm still not subscribed to -hackers, so the usual \\\"please CC \\nme\\\" applies!\\n\\nBest regards,\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\",\n \"msg_date\": \"Wed, 23 Feb 2022 16:28:09 +0100\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"On Wed, 2022-02-23 at 16:28 +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\r\\n> I'm extremely unhappy that I called it \\\"--config-file\\\" here, while \\r\\n> \\\"postgres\\\" itself wants \\\"--config_file\\\". But as the other dashed options \\r\\n> of pg_rewind are, well, dashed and not underscored, it seemed to be \\r\\n> better that the other way...\\r\\n\\r\\nThe \\\"--\\\"-style options to postgres take either underscores or dashes in\\r\\nthe names (just double-checked with PG14), so this shouldn't be a\\r\\nproblem.\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 17:21:15 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Hi Gunnar,\\n\\n> During a Patroni PR discussion\\n> (https://github.com/zalando/patroni/pull/2225), we realised that if one\\n> wants to use the \\\"-c\\\" option on a typical Debian/Ubuntu installation\\n> (where the config resides below /etc/postgresql/), pg_rewind needs a way\\n> to be told where the postgresql.conf actually is.\\n>\\n> The attached patch implements this as an option.\\n>\\n> [...]\\n\\nGood catch!\\n\\nCould you also implement a TAP test for the new code?\\n\\n-- \\nBest regards,\\nAleksander Alekseev\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 12:02:54 +0300\",\n \"msg_from\": \"Aleksander Alekseev <aleksander@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Hello Gunnar,\\n\\nOn Thu, 24 Feb 2022 at 10:03, Aleksander Alekseev <aleksander@timescale.com>\\nwrote:\\n\\n\\n> > wants to use the \\\"-c\\\" option on a typical Debian/Ubuntu installation\\n> > (where the config resides below /etc/postgresql/), pg_rewind needs a way\\n> > to be told where the postgresql.conf actually is.\\n> >\\n> > The attached patch implements this as an option.\\n> >\\n> > [...]\\n>\\n> Good catch!\\n>\\n\\nYeah, it is a known problem, and there was already another patch trying to\\naddress it [1].\\nHonestly, I like your approach much better, because, the previous make was\\nassuming that the data_directory GUC is specified in the postgresql.conf,\\nwhich is not very flexible.\\n\\n\\n>\\n> Could you also implement a TAP test for the new code?\\n>\\n\\n+1\\n\\n\\n+ <para>\\n+ In case the <filename>postgresql.conf</filename> of your target\\ncluster is not in the\\n+ target pgdata and you want to use the <option>-c</option> option,\\n+ provide a (relative or absolute) path to the\\n<filename>postgresql.conf</filename> using this option.\\n+ </para>\\n\\nIt took me a while to understand the meaning of <option>-c</option>. Maybe\\nchanging it to <option>--restore-target-wal</option> will make it easier to\\nunderstand.\\n\\n[1]\\nhttps://www.postgresql.org/message-id/flat/16982-f12294cccd221480%40postgresql.org\\n\\nRegards,\\n--\\nAlexander Kukushkin\\n\\nHello Gunnar,On Thu, 24 Feb 2022 at 10:03, Aleksander Alekseev <aleksander@timescale.com> wrote: > wants to use the \\\"-c\\\" option on a typical Debian/Ubuntu installation\\n> (where the config resides below /etc/postgresql/), pg_rewind needs a way\\n> to be told where the postgresql.conf actually is.\\n>\\n> The attached patch implements this as an option.\\n>\\n> [...]\\n\\nGood catch!Yeah, it is a known problem, and there was already another patch trying to address it [1].Honestly, I like your approach much better, because, the previous make was assuming that the data_directory GUC is specified in the postgresql.conf, which is not very flexible. \\n\\nCould you also implement a TAP test for the new code?+1 + <para>+ In case the <filename>postgresql.conf</filename> of your target cluster is not in the + target pgdata and you want to use the <option>-c</option> option,+ provide a (relative or absolute) path to the <filename>postgresql.conf</filename> using this option.+ </para>It took me a while to understand the meaning of <option>-c</option>. Maybe changing it to <option>--restore-target-wal</option> will make it easier to understand.[1] https://www.postgresql.org/message-id/flat/16982-f12294cccd221480%40postgresql.orgRegards,--Alexander Kukushkin\",\n \"msg_date\": \"Thu, 24 Feb 2022 10:27:30 +0100\",\n \"msg_from\": \"Alexander Kukushkin <cyberdemn@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Hi hackers,\\n\\n> It took me a while to understand the meaning of <option>-c</option>. Maybe changing it to <option>--restore-target-wal</option> will make it easier to understand.\\n\\n+1, I \\\"stumbled\\\" while reading this too at first.\\n\\n-- \\nBest regards,\\nAleksander Alekseev\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 12:58:33 +0300\",\n \"msg_from\": \"Aleksander Alekseev <aleksander@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"> On 24 Feb 2022, at 10:27, Alexander Kukushkin <cyberdemn@gmail.com> wrote:\\n> \\n> Hello Gunnar,\\n> \\n> On Thu, 24 Feb 2022 at 10:03, Aleksander Alekseev <aleksander@timescale.com> wrote:\\n> \\n> > wants to use the \\\"-c\\\" option on a typical Debian/Ubuntu installation\\n> > (where the config resides below /etc/postgresql/), pg_rewind needs a way\\n> > to be told where the postgresql.conf actually is.\\n> >\\n> > The attached patch implements this as an option.\\n> >\\n> > [...]\\n> \\n> Good catch!\\n> \\n> Yeah, it is a known problem, and there was already another patch trying to address it [1].\\n\\nThere is yet another related patch which provides a parameter to pass in\\nrestore_command in cases where postgresq.conf isn't reachable:\\n\\n https://commitfest.postgresql.org/37/3213/\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 14:08:40 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Am 24.02.2022 um 14:08 schrieb Daniel Gustafsson:\\n>> On 24 Feb 2022, at 10:27, Alexander Kukushkin <cyberdemn@gmail.com> wrote:\\n>>\\n>> Hello Gunnar,\\n>>\\n>> On Thu, 24 Feb 2022 at 10:03, Aleksander Alekseev <aleksander@timescale.com> wrote:\\n>> \\n>>> wants to use the \\\"-c\\\" option on a typical Debian/Ubuntu installation\\n>>> (where the config resides below /etc/postgresql/), pg_rewind needs a way\\n>>> to be told where the postgresql.conf actually is.\\n>>>\\n>>> The attached patch implements this as an option.\\n>>>\\n>>> [...]\\n>>\\n>> Good catch!\\n>>\\n>> Yeah, it is a known problem, and there was already another patch trying to address it [1].\\n> \\n> There is yet another related patch which provides a parameter to pass in\\n> restore_command in cases where postgresq.conf isn't reachable:\\n> \\n> https://commitfest.postgresql.org/37/3213/\\n\\nThat looks just as good to me, and it already has tests, so I'll happily \\nstep down!\\n\\n(Note to myself: check the CF first next time ;-)\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 14:43:17 +0100\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"> On 24 Feb 2022, at 14:43, Gunnar Nick Bluth <gunnar.bluth@pro-open.de> wrote:\\n\\n> That looks just as good to me, and it already has tests, so I'll happily step down!\\n\\nActually, I think this looks like a saner approach. Putting a config setting\\nin two place (postgresql.conf and on the commandline for pg_rewind) is a recipe\\nfor them diverging.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 14:46:49 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Am 24.02.22 um 14:46 schrieb Daniel Gustafsson:\\n>> On 24 Feb 2022, at 14:43, Gunnar Nick Bluth <gunnar.bluth@pro-open.de> wrote:\\n> \\n>> That looks just as good to me, and it already has tests, so I'll happily step down!\\n> \\n> Actually, I think this looks like a saner approach. Putting a config setting\\n> in two place (postgresql.conf and on the commandline for pg_rewind) is a recipe\\n> for them diverging.\\n\\nWell, well,\\n\\nreverted https://commitfest.postgresql.org/37/3562/ back to \\\"needs \\nreview\\\" and moved it to \\\"Replication & Recovery\\\".\\n\\nPatch v2 includes:\\n* doc improvements as suggested\\n* tests shamelessly copied & adapted from \\nhttps://commitfest.postgresql.org/37/3213/\\n\\nSince the need is pretty obvious (I *think* the Debian-ish userbase is \\nquite significant alone), I'd love to see one of the approaches making \\nit into 15!\\n\\nBest regards,\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\",\n \"msg_date\": \"Fri, 25 Feb 2022 10:35:49 +0100\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 10:35:49AM +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\n> Am 24.02.22 um 14:46 schrieb Daniel Gustafsson:\\n>> Actually, I think this looks like a saner approach. Putting a config setting\\n>> in two place (postgresql.conf and on the commandline for pg_rewind) is a recipe\\n>> for them diverging.\\n\\nFWIW, I have a bad feeling about passing down directly a command\\nthrough an option itself part of a command, so what's discussed on\\nthis thread is refreshing.\\n\\n+ } else {\\n+ snprintf(postgres_cmd, sizeof(postgres_cmd),\\n+ \\\"\\\\\\\"%s\\\\\\\" -D \\\\\\\"%s\\\\\\\" --config_file=\\\\\\\"%s\\\\\\\" -C restore_command\\\",\\n+ postgres_exec_path, datadir_target, config_file);\\n+ }\\n\\nShouldn't this one use appendShellString() on config_file?\\n--\\nMichael\",\n \"msg_date\": \"Sat, 26 Feb 2022 14:51:14 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Am 26.02.22 um 06:51 schrieb Michael Paquier:\\n> On Fri, Feb 25, 2022 at 10:35:49AM +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\n>> Am 24.02.22 um 14:46 schrieb Daniel Gustafsson:\\n>>> Actually, I think this looks like a saner approach. Putting a config setting\\n>>> in two place (postgresql.conf and on the commandline for pg_rewind) is a recipe\\n>>> for them diverging.\\n> \\n> FWIW, I have a bad feeling about passing down directly a command\\n> through an option itself part of a command, so what's discussed on\\n> this thread is refreshing.\\n\\nTa.\\n\\n> \\n> + } else {\\n> + snprintf(postgres_cmd, sizeof(postgres_cmd),\\n> + \\\"\\\\\\\"%s\\\\\\\" -D \\\\\\\"%s\\\\\\\" --config_file=\\\\\\\"%s\\\\\\\" -C restore_command\\\",\\n> + postgres_exec_path, datadir_target, config_file);\\n> + }\\n> \\n> Shouldn't this one use appendShellString() on config_file?\\n\\nIt probably should, yes. I don't fancy this repetitive code myself.\\nBut then, pg_rewind as a whole could use an overhaul. I don't see any \\nuse of PQExpBuffer in it, but a lot of char* ...\\n\\nI myself am not a C hacker (as you can probably tell already) and don't \\nreally feel fit (enough) to rewrite pg_rewind.c to use \\nfe_utils/string_utils.h :/\\n\\nI might give it a try anyhow, but that may be a bit... heavy,,, just to \\nadd one option?\\n\\nWhile we're at it (I'm really good at opening cans of worms):\\n09:47 $ grep -r -l --include \\\\*.c fe_utils/string_utils.h src/bin/ | wc -l\\n28\\n09:48 $ grep -r -L --include \\\\*.c fe_utils/string_utils.h src/bin/ | wc -l\\n66\\n\\nGSOC project? ;-)\\n\\nBest,\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 09:55:20 +0100\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 09:55:20AM +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\n> Am 26.02.22 um 06:51 schrieb Michael Paquier:\\n>> Shouldn't this one use appendShellString() on config_file?\\n> \\n> It probably should, yes. I don't fancy this repetitive code myself.\\n> But then, pg_rewind as a whole could use an overhaul. I don't see any use of\\n> PQExpBuffer in it, but a lot of char* ...\\n\\nHaving a lot of char* does not necessarily mean that all of them have\\nto be changed to accomodate with PQExpBuffer. My guess that this is a\\ncase-by-case, where we should apply that in places where it makes the\\ncode cleaner to follow. In the case of your patch though, the two\\nareas changed would make the logic correct, instead, because we have\\nto apply correct quoting rules to any commands executed.\\n\\n> GSOC project? ;-)\\n\\nIt does not seem so, though there are surely more areas that could\\ngain in clarity.\\n--\\nMichael\",\n \"msg_date\": \"Sun, 27 Feb 2022 21:06:59 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Am 27.02.22 um 13:06 schrieb Michael Paquier:\\n> On Sat, Feb 26, 2022 at 09:55:20AM +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\n>> Am 26.02.22 um 06:51 schrieb Michael Paquier:\\n>>> Shouldn't this one use appendShellString() on config_file?\\n>>\\n>> It probably should, yes. I don't fancy this repetitive code myself.\\n>> But then, pg_rewind as a whole could use an overhaul. I don't see any use of\\n>> PQExpBuffer in it, but a lot of char* ...\\n> \\n> Having a lot of char* does not necessarily mean that all of them have\\n> to be changed to accomodate with PQExpBuffer. My guess that this is a\\n> case-by-case, where we should apply that in places where it makes the\\n> code cleaner to follow. In the case of your patch though, the two\\n> areas changed would make the logic correct, instead, because we have\\n> to apply correct quoting rules to any commands executed.\\n\\nAlas! v3 attached.\\n\\n\\n>> GSOC project? ;-)\\n> \\n> It does not seem so, though there are surely more areas that could\\n> gain in clarity.\\n\\nThat's universally true ;-)\\n\\n\\nBest regards,\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\",\n \"msg_date\": \"Sun, 27 Feb 2022 14:30:33 +0100\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 02:30:33PM +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\n> That's universally true ;-)\\n\\n-# Internal routine to enable archive recovery command on a standby node\\n+# Internal routine to enable archive recovery command on a standby node.\\n+# Returns generated restore_command.\\n sub enable_restoring\\n {\\n \\tmy ($self, $root_node, $standby) = @_;\\n@@ -1103,7 +1104,7 @@ restore_command = '$copy_command'\\n \\t{\\n \\t\\t$self->set_recovery_mode();\\n \\t}\\n-\\treturn;\\n+\\treturn $copy_command;\\n\\nI don't like this change. This makes an existing routine do some\\nextra work for something it is not designed for. You could get this\\ninformation with a postgres -C command, for example. Now, you cannot\\nuse postgres -C because of.. Reasons (1a9d802). But you could use a\\npg_ctl command for the same purpose. I guess that we'd better\\nrefactor this into a new routine of Cluster.pm where we execute a\\npg_ctl command and return both stdout and stderr back to the caller?\\n\\nThe TAP part could be refactored on its own.\\n\\n+ In case the <filename>postgresql.conf</filename> of your\\ntarget cluster is not in the\\n+ target pgdata and you want to use the <option>-c /\\n--restore-target-wal</option> option,\\n+ provide a (relative or absolute) path to the\\n<filename>postgresql.conf</filename> using this option.\\n\\nThis could do a better job to explain in which cases this option can\\nbe used for. You could say something like that:\\n\\\"This option specifies a path to a configuration file to be used when\\nstarting the target cluster. This makes easier the use of\\n-c/--restore-target-wal by setting restore_command in the extra\\nconfiguration file specified via this option.\\\"\\n\\nHmm. What about the target cluster started in --single mode as of\\nensureCleanShutdown()? Should we try to apply this option also in\\nthis case?\\n--\\nMichael\",\n \"msg_date\": \"Mon, 28 Feb 2022 20:56:24 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Am 28.02.22 um 12:56 schrieb Michael Paquier:\\n> On Sun, Feb 27, 2022 at 02:30:33PM +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\n>> That's universally true ;-)\\n> \\n> -# Internal routine to enable archive recovery command on a standby node\\n> +# Internal routine to enable archive recovery command on a standby node.\\n> +# Returns generated restore_command.\\n> sub enable_restoring\\n> {\\n> \\tmy ($self, $root_node, $standby) = @_;\\n> @@ -1103,7 +1104,7 @@ restore_command = '$copy_command'\\n> \\t{\\n> \\t\\t$self->set_recovery_mode();\\n> \\t}\\n> -\\treturn;\\n> +\\treturn $copy_command;\\n> \\n> I don't like this change. This makes an existing routine do some\\n> extra work for something it is not designed for. You could get this\\n> information with a postgres -C command, for example. Now, you cannot\\n> use postgres -C because of.. Reasons (1a9d802). But you could use a\\n> pg_ctl command for the same purpose. I guess that we'd better\\n> refactor this into a new routine of Cluster.pm where we execute a\\n> pg_ctl command and return both stdout and stderr back to the caller?\\n\\nTurns out this change isn't needed anyway (it was copied from the other \\npatch). Afterall, the patch is about extracting the restore_command from \\nthe config, so...\\n\\n> The TAP part could be refactored on its own.\\n\\nAgreed. I've struggled quite a bit even understanding what's happening \\nin there ;-)\\n\\n\\n> + In case the <filename>postgresql.conf</filename> of your\\n> target cluster is not in the\\n> + target pgdata and you want to use the <option>-c /\\n> --restore-target-wal</option> option,\\n> + provide a (relative or absolute) path to the\\n> <filename>postgresql.conf</filename> using this option.\\n> \\n> This could do a better job to explain in which cases this option can\\n> be used for. You could say something like that:\\n> \\\"This option specifies a path to a configuration file to be used when\\n> starting the target cluster. This makes easier the use of\\n> -c/--restore-target-wal by setting restore_command in the extra\\n> configuration file specified via this option.\\\"\\n\\nI reviewed the wording and think it is pretty universal now.\\n\\n\\n> Hmm. What about the target cluster started in --single mode as of\\n> ensureCleanShutdown()? Should we try to apply this option also in\\n> this case?\\n\\nI'd say so; as far as I can tell, \\\"postgres\\\" would deny starting a \\n(Debian/non-standard) cluster the way the code is right now.\\n\\nWhich led me to realize we have\\n\\n /* find postgres executable */\\n rc = find_other_exec(argv0, \\\"postgres\\\",\\n PG_BACKEND_VERSIONSTR,\\n postgres_exec_path);\\nin getRestoreCommand _and_\\n\\n /* locate postgres binary */\\n if ((ret = find_other_exec(argv0, \\\"postgres\\\", \\n\\n\\t\\t\\t\\t\\t\\tPG_BACKEND_VERSIONSTR,\\n exec_path)) < 0)\\nin ensureCleanShutdown.\\nBoth with the same error messages, AFAICS. D'oh... can of worms.\\n\\nSo, how should we call the global \\\"find the ***** 'postgres' executable \\nand boil out if that fails\\\" function?\\n char postgres_exec_path[MAXPGPATH] = findPostgresExec();\\n?\\n\\n\\nAnyway, v4 attached so we can settle on the tests and wording...\\n\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\",\n \"msg_date\": \"Mon, 28 Feb 2022 20:48:23 +0100\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 08:48:23PM +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\n> So, how should we call the global \\\"find the ***** 'postgres' executable and\\n> boil out if that fails\\\" function?\\n> char postgres_exec_path[MAXPGPATH] = findPostgresExec();\\n> ?\\n\\nThat would mean only a couple of lines gained, and the readability\\ngained is minimal, so I'd be fine to keep the code as-is. I am not\\nsure about the full patch set yet, but the refactoring of the commands\\nto use PQExpBuffer is good by itself, so I have extracted this part of\\nthe patch and applied that for now.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 1 Mar 2022 13:18:23 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Am 01.03.22 um 05:18 schrieb Michael Paquier:\\n> On Mon, Feb 28, 2022 at 08:48:23PM +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\n>> So, how should we call the global \\\"find the ***** 'postgres' executable and\\n>> boil out if that fails\\\" function?\\n>> char postgres_exec_path[MAXPGPATH] = findPostgresExec();\\n>> ?\\n> \\n> That would mean only a couple of lines gained, and the readability\\n> gained is minimal, so I'd be fine to keep the code as-is. I am not\\n> sure about the full patch set yet, but the refactoring of the commands\\n> to use PQExpBuffer is good by itself, so I have extracted this part of\\n> the patch and applied that for now.\\n\\nWhich made the remaining patch v5 (which also appends the --config-file \\nfor ensureCleanShutdown) much cleaner :)\\n\\nCheers,\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\",\n \"msg_date\": \"Tue, 1 Mar 2022 12:35:27 +0100\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Heya,\\n\\nsome minor comments, I didn't look at the added test and I did not test\\nthe patch at all, but (as part of the Debian/Ubuntu packaging team) I\\nthink this patch is really important:\\n\\nOn Tue, Mar 01, 2022 at 12:35:27PM +0100, Gunnar \\\"Nick\\\" Bluth wrote:\\n> diff --git a/doc/src/sgml/ref/pg_rewind.sgml b/doc/src/sgml/ref/pg_rewind.sgml\\n> index 33e6bb64ad..d0278e9b46 100644\\n> --- a/doc/src/sgml/ref/pg_rewind.sgml\\n> +++ b/doc/src/sgml/ref/pg_rewind.sgml\\n> @@ -241,6 +241,18 @@ PostgreSQL documentation\\n> </listitem>\\n> </varlistentry>\\n> \\n> + <varlistentry>\\n> + <term><option>--config-file=<replaceable class=\\\"parameter\\\">filepath</replaceable></option></term>\\n> + <listitem>\\n> + <para>\\n> + When using the <option>-c / --restore-target-wal</option> option, the <varname>restore_command</varname>\\n> + is extracted from the configuration of the target cluster. If the <filename>postgresql.conf</filename>\\n> + of that cluster is not in the target pgdata directory (or if you want to use an alternative config),\\n\\nI think we usually just say \\\"data directory\\\"; pgdata was previously only\\nused as part of the option name, not like this in the text.\\n\\n> diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c\\n> index b39b5c1aac..294353a9d5 100644\\n> --- a/src/bin/pg_rewind/pg_rewind.c\\n> +++ b/src/bin/pg_rewind/pg_rewind.c\\n> @@ -61,6 +61,7 @@ char\\t *datadir_target = NULL;\\n> char\\t *datadir_source = NULL;\\n> char\\t *connstr_source = NULL;\\n> char\\t *restore_command = NULL;\\n> +char\\t *config_file = NULL;\\n> \\n> static bool debug = false;\\n> bool\\t\\tshowprogress = false;\\n> @@ -87,6 +88,7 @@ usage(const char *progname)\\n> \\tprintf(_(\\\"Options:\\\\n\\\"));\\n> \\tprintf(_(\\\" -c, --restore-target-wal use restore_command in target configuration to\\\\n\\\"\\n> \\t\\t\\t \\\" retrieve WAL files from archives\\\\n\\\"));\\n> +\\tprintf(_(\\\" --config-file=FILE path to postgresql.conf if outside target-pgdata (for -c)\\\\n\\\"));\\n\\n> \\tprintf(_(\\\" -D, --target-pgdata=DIRECTORY existing data directory to modify\\\\n\\\"));\\n> \\tprintf(_(\\\" --source-pgdata=DIRECTORY source data directory to synchronize with\\\\n\\\"));\\n> \\tprintf(_(\\\" --source-server=CONNSTR source server to synchronize with\\\\n\\\"));\\n\\ntarget-pgdata is the name of the option, but maybe it is useful to spell\\nout \\\"target data directory\\\" here, even if that means we spill over to\\ntwo lines (which we might have to do anyway, that new line is pretty\\nlong).\\n\\n> @@ -121,6 +123,7 @@ main(int argc, char **argv)\\n> \\t\\t{\\\"no-sync\\\", no_argument, NULL, 'N'},\\n> \\t\\t{\\\"progress\\\", no_argument, NULL, 'P'},\\n> \\t\\t{\\\"debug\\\", no_argument, NULL, 3},\\n> +\\t\\t{\\\"config-file\\\", required_argument, NULL, 5},\\n\\nNot sure what our policy is, but the way the numbered options are 1, 2,\\n4, 3, 5 after this is weird; this was already off before this patch\\nthough.\\n\\n> \\t\\t{NULL, 0, NULL, 0}\\n> \\t};\\n> \\tint\\t\\t\\toption_index;\\n> @@ -205,6 +208,10 @@ main(int argc, char **argv)\\n> \\t\\t\\tcase 4:\\n> \\t\\t\\t\\tno_ensure_shutdown = true;\\n> \\t\\t\\t\\tbreak;\\n> +\\n> +\\t\\t\\tcase 5:\\n> +\\t\\t\\t\\tconfig_file = pg_strdup(optarg);\\n> +\\t\\t\\t\\tbreak;\\n> \\t\\t}\\n> \\t}\\n> \\n> @@ -1060,7 +1067,11 @@ getRestoreCommand(const char *argv0)\\n> \\t/* add -D switch, with properly quoted data directory */\\n> \\tappendPQExpBufferStr(postgres_cmd, \\\" -D \\\");\\n> \\tappendShellString(postgres_cmd, datadir_target);\\n> -\\n> +\\tif (config_file != NULL)\\n> +\\t{\\n> +\\t\\tappendPQExpBufferStr(postgres_cmd, \\\" --config_file=\\\");\\n> +\\t\\tappendShellString(postgres_cmd, config_file);\\n> +\\t}\\n> \\t/* add -C switch, for restore_command */\\n\\nYou removed an empty line here. Maybe rather prepend this with a comment\\non what's going on, and have en empty line before and afterwards.\\n\\n> \\tappendPQExpBufferStr(postgres_cmd, \\\" -C restore_command\\\");\\n> \\n> @@ -1138,6 +1149,11 @@ ensureCleanShutdown(const char *argv0)\\n> \\t/* add set of options with properly quoted data directory */\\n> \\tappendPQExpBufferStr(postgres_cmd, \\\" --single -F -D \\\");\\n> \\tappendShellString(postgres_cmd, datadir_target);\\n> +\\tif (config_file != NULL)\\n> +\\t{\\n> +\\t\\tappendPQExpBufferStr(postgres_cmd, \\\" --config_file=\\\");\\n> +\\t\\tappendShellString(postgres_cmd, config_file);\\n> +\\t}\\n> \\n> \\t/* finish with the database name, and a properly quoted redirection */\\n\\nKinde same here.\\n\\n\\nCheers,\\n\\nMichael\\n\\n-- \\nMichael Banck\\nTeamleiter PostgreSQL-Team\\nProjektleiter\\nTel.: +49 2166 9901-171\\nEmail: michael.banck@credativ.de\\n\\ncredativ GmbH, HRB M�nchengladbach 12080\\nUSt-ID-Nummer: DE204566209\\nTrompeterallee 108, 41189 M�nchengladbach\\nGesch�ftsf�hrung: Dr. Michael Meskes, Geoff Richardson, Peter Lilley\\n\\nUnser Umgang mit personenbezogenen Daten unterliegt\\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 14:43:32 +0100\",\n \"msg_from\": \"Michael Banck <michael.banck@credativ.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Am 10.03.22 um 14:43 schrieb Michael Banck:\\n> Heya,\\n> \\n> some minor comments, I didn't look at the added test and I did not test\\n> the patch at all, but (as part of the Debian/Ubuntu packaging team) I\\n> think this patch is really important:\\n\\nMuch appreciated!\\n\\n[...]\\n\\n> I think we usually just say \\\"data directory\\\"; pgdata was previously only\\n> used as part of the option name, not like this in the text.\\n\\nFixed.\\n\\n[...]\\n\\n> target-pgdata is the name of the option, but maybe it is useful to spell\\n> out \\\"target data directory\\\" here, even if that means we spill over to\\n> two lines (which we might have to do anyway, that new line is pretty\\n> long).\\n\\nFixed.\\n\\n> \\n>> @@ -121,6 +123,7 @@ main(int argc, char **argv)\\n>> \\t\\t{\\\"no-sync\\\", no_argument, NULL, 'N'},\\n>> \\t\\t{\\\"progress\\\", no_argument, NULL, 'P'},\\n>> \\t\\t{\\\"debug\\\", no_argument, NULL, 3},\\n>> +\\t\\t{\\\"config-file\\\", required_argument, NULL, 5},\\n> \\n> Not sure what our policy is, but the way the numbered options are 1, 2,\\n> 4, 3, 5 after this is weird; this was already off before this patch\\n> though.\\n\\nThat one has been triggering my inner Monk too ;-)\\n\\nFixed!\\n\\n\\n[...]\\n\\n> You removed an empty line here. Maybe rather prepend this with a comment\\n> on what's going on, and have en empty line before and afterwards.\\n\\n> Kinde same here.\\n\\nIt does smell a bit like \\\"stating the obvious\\\", but alas! Both fixed.\\n\\nCheers,\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\",\n \"msg_date\": \"Thu, 10 Mar 2022 15:28:57 +0100\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nAm Donnerstag, dem 10.03.2022 um 15:28 +0100 schrieb Gunnar \\\"Nick\\\"\\nBluth:\\n> Am 10.03.22 um 14:43 schrieb Michael Banck:\\n> > some minor comments, I didn't look at the added test and I did not\\n> > test the patch at all, but (as part of the Debian/Ubuntu packaging\\n> > team) I think this patch is really important:\\n> \\n> Much appreciated!\\n> \\n> [...]\\n> \\n> Fixed.\\n> \\n> [...]\\n\\nThanks for the updated patch.\\n\\nThe patch applies, make is ok, make check is ok, pg_rewind TAP tests\\nare ok.\\n\\nI did some tests now with Patroni 2.1.3 and the attached patch applied.\\n\\nThe following test was made:\\n\\n1. Deploy 3-node (pg1, pg2, pg3) patroni cluster on Debian unstable\\n running postgresql-15 (approx. master)\\n2. Switch on archive_mode, and set archive_command and restore_command\\n3. Switchover so that pg1 is the primary (if not already)\\n4. Kill pg1 hard\\n5. Wait till a new leader has taken over and the (now 2-node) cluster \\n is healthy again\\n6. Restart pg1\\n\\nwithout this patch:\\n\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,084 INFO: running pg_rewind from pg3\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,102 INFO: running pg_rewind from dbname=postgres user=postgres_rewind host=10.0.3.227 port=5432 target_session_attrs=read-write\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,135 INFO: pg_rewind exit code=1\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,136 INFO: stdout=\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,136 INFO: stderr=postgres: could not access the server configuration file \\\"/var/lib/postgresql/15/main/postgresql.conf\\\": No such file or directory\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: no data was returned by command \\\"/usr/lib/postgresql/15/bin/postgres -D /var/lib/postgresql/15/main -C restore_command\\\"\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,149 ERROR: Failed to rewind from healty master: pg3\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,149 WARNING: remove_data_directory_on_diverged_timelines is set. removing...\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,149 INFO: Removing data directory: /var/lib/postgresql/15/main\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,245 INFO: Lock owner: pg3; I am pg1\\n|Apr 03 19:09:01 pg1 patroni@15-main[99]: 2022-04-03 19:09:01,248 INFO: trying to bootstrap from leader 'pg3'\\n\\nSo pg_rewind fails and Patroni does a re-bootstrap.\\n\\nwith this patch:\\n\\n|Apr 03 19:12:35 pg1 patroni@15-main[99]: 2022-04-03 19:12:35,576 INFO: running pg_rewind from pg2\\n|Apr 03 19:12:35 pg1 patroni@15-main[99]: 2022-04-03 19:12:35,590 INFO: running pg_rewind from dbname=postgres user=postgres_rewind host=10.0.3.145 port=5432 target_session_attrs=read-write\\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,147 INFO: pg_rewind exit code=0\\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,148 INFO: stdout=\\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,148 INFO: stderr=pg_rewind: servers diverged at WAL location 0/1D000180 on timeline 38\\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: pg_rewind: rewinding from last common checkpoint at 0/1D000108 on timeline 38\\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: pg_rewind: Done!\\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,151 WARNING: Postgresql is not running.\\n|Apr 03 19:12:37 pg1 patroni@15-main[99]: 2022-04-03 19:12:37,151 INFO: Lock owner: pg2; I am pg1\\n\\nHere, pg_rewind is called and works fine.\\n\\nSo I think this works as intended, and I'm marking it Ready for\\nCommitter.\\n\\n\\nMichael\\n\\n-- \\nMichael Banck\\nTeamleiter PostgreSQL-Team\\nProjektleiter\\nTel.: +49 2166 9901-171\\nE-Mail: michael.banck@credativ.de\\n\\ncredativ GmbH, HRB Mönchengladbach 12080\\nUSt-ID-Nummer: DE204566209\\nTrompeterallee 108, 41189 Mönchengladbach\\nGeschäftsführung: Dr. Michael Meskes, Geoff Richardson, Peter Lilley\\n\\nUnser Umgang mit personenbezogenen Daten unterliegt\\nfolgenden Bestimmungen: https://www.credativ.de/datenschutz\",\n \"msg_date\": \"Sun, 03 Apr 2022 21:31:32 +0200\",\n \"msg_from\": \"Michael Banck <michael.banck@credativ.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Just reposting the previous patches to straighten out the cfbot.\",\n \"msg_date\": \"Mon, 4 Apr 2022 11:39:45 -0400\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"On Sun, Apr 03, 2022 at 09:31:32PM +0200, Michael Banck wrote:\\n> The patch applies, make is ok, make check is ok, pg_rewind TAP tests\\n> are ok.\\n\\n+ appendPQExpBufferStr(postgres_cmd, \\\" --config_file=\\\");\\n+ appendShellString(postgres_cmd, config_file);\\nNit. This is a valid option for the postgres command, but I'd rather\\nuse a -c switch instead, mostly as a matter of self-documentation.\\n\\nIn the tests, the only difference between the modes \\\"archive_cli\\\" and\\n\\\"archive\\\" is the extra option given to the pg_rewind command, and\\nthat's a simple redirect to what pg_rewind would choose by default\\nanyway. A more elegant solution would be to have an extra option in\\nRewindTest::run_pg_rewind(), where any caller of the routine can feed\\nextra options to the pg_rewind command. Now, in this case, we are\\nnot going to lose any coverage if the existing \\\"archive\\\" mode is\\nchanged so as we move away postgresql.conf from the target data folder\\nand just use --config-file by default, no? I would make the choice of\\nsimplicity, by giving up on \\\"archive_cli\\\" and using\\nprimary-postgresql.conf.tmp as value for --config-file. There could\\nbe an argument for using --config-file for all the modes, as well.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 5 Apr 2022 11:10:30 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 05, 2022 at 11:10:30AM +0900, Michael Paquier wrote:\\n> In the tests, the only difference between the modes \\\"archive_cli\\\" and\\n> \\\"archive\\\" is the extra option given to the pg_rewind command, and\\n> that's a simple redirect to what pg_rewind would choose by default\\n> anyway. A more elegant solution would be to have an extra option in\\n> RewindTest::run_pg_rewind(), where any caller of the routine can feed\\n> extra options to the pg_rewind command. Now, in this case, we are\\n> not going to lose any coverage if the existing \\\"archive\\\" mode is\\n> changed so as we move away postgresql.conf from the target data folder\\n> and just use --config-file by default, no? I would make the choice of\\n> simplicity, by giving up on \\\"archive_cli\\\" and using\\n> primary-postgresql.conf.tmp as value for --config-file. There could\\n> be an argument for using --config-file for all the modes, as well.\\n\\nThe clock is ticking, so I have looked at this patch by myself. I\\nhave finished by tweaking the docs, the code and the tests as of the \\nattached. One thing that I found annoying is the lack of description\\nabout the fact that this option affects pg_rewind when it internally\\nstarts the target cluster, as of when we retrieve restore_command and\\nwhen we enforce crash recovery to have a target cluster with a clean\\nshutdown state. The tests are heavily simplified, having no impact on\\nthe run-time while improving the coverage for the code paths changed\\nhere.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 6 Apr 2022 21:04:50 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Am 06.04.22 um 14:04 schrieb Michael Paquier:\\n> On Tue, Apr 05, 2022 at 11:10:30AM +0900, Michael Paquier wrote:\\n>> In the tests, the only difference between the modes \\\"archive_cli\\\" and\\n>> \\\"archive\\\" is the extra option given to the pg_rewind command, and\\n>> that's a simple redirect to what pg_rewind would choose by default\\n>> anyway. A more elegant solution would be to have an extra option in\\n>> RewindTest::run_pg_rewind(), where any caller of the routine can feed\\n>> extra options to the pg_rewind command. Now, in this case, we are\\n>> not going to lose any coverage if the existing \\\"archive\\\" mode is\\n>> changed so as we move away postgresql.conf from the target data folder\\n>> and just use --config-file by default, no? I would make the choice of\\n>> simplicity, by giving up on \\\"archive_cli\\\" and using\\n>> primary-postgresql.conf.tmp as value for --config-file. There could\\n>> be an argument for using --config-file for all the modes, as well.\\n> \\n> The clock is ticking, so I have looked at this patch by myself. I\\n\\nTa! Sorry, been ridiculously busy these days, and I do confess that I've\\nbeen struggling a bit with those tests before ;-)\\n\\n> have finished by tweaking the docs, the code and the tests as of the \\n> attached. One thing that I found annoying is the lack of description\\n> about the fact that this option affects pg_rewind when it internally\\n> starts the target cluster, as of when we retrieve restore_command and\\n> when we enforce crash recovery to have a target cluster with a clean\\n> shutdown state. The tests are heavily simplified, having no impact on\\n> the run-time while improving the coverage for the code paths changed\\n> here.\\n> --\\n> Michael\\n\\nLGTM!\\n\\nThx again!\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Apr 2022 14:32:44 +0200\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 06, 2022 at 02:32:44PM +0200, Gunnar \\\"Nick\\\" Bluth wrote:\\n> Ta! Sorry, been ridiculously busy these days, and I do confess that I've\\n> been struggling a bit with those tests before ;-)\\n\\nNo problem. Double-checked this morning and applied, then.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 7 Apr 2022 08:54:29 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n },\n {\n \"msg_contents\": \"Am 07.04.22 um 01:54 schrieb Michael Paquier:\\n> On Wed, Apr 06, 2022 at 02:32:44PM +0200, Gunnar \\\"Nick\\\" Bluth wrote:\\n>> Ta! Sorry, been ridiculously busy these days, and I do confess that I've\\n>> been struggling a bit with those tests before ;-)\\n> \\n> No problem. Double-checked this morning and applied, then.\\n> --\\n> Michael\\n\\nWonderful!\\n\\nSo, @cyberdemn, who would have thought we get this fixed so quickly?\\nTime to update the Patroni PR once more! ;-)\\n\\nThanks & best regards,\\n-- \\nGunnar \\\"Nick\\\" Bluth\\n\\nEimermacherweg 106\\nD-48159 Münster\\n\\nMobil +49 172 8853339\\nEmail: gunnar.bluth@pro-open.de\\n__________________________________________________________________________\\n\\\"Ceterum censeo SystemD esse delendam\\\" - Cato\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 09:51:01 +0200\",\n \"msg_from\": \"\\\"Gunnar \\\\\\\"Nick\\\\\\\" Bluth\\\" <gunnar.bluth@pro-open.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: PATCH: add \\\"--config-file=\\\" option to pg_rewind\"\n }\n]","isConversational":false}}},{"rowIdx":53566,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello,\r\n\r\nAs part of the NSS work it came up [1] that clients don't have a good\r\nway to ask libpq what SSL library it was compiled with, unless they\r\nalready have a connection pointer so that they can call\r\nPQsslAttribute(conn, \"library\"). This poses a chicken-and-egg problem:\r\nwith the NSS proposal, the client may have to know which library is in\r\nuse before it can construct a proper connection string. For example, I\r\nhave a test suite that needs to set up an NSS database with\r\ncertificates before telling libpq to connect using that database.\r\n\r\nThe simplest proposal was to just allow PQsslAttribute() to take NULL\r\nas the connection parameter when querying the \"library\" attribute, and\r\nthat's what I've done in this patch. In current versions of libpq, the\r\n\"library\" attribute will always be NULL if you pass NULL as the\r\nconnection; a client that needs to know whether this new behavior is\r\npresent can look at the LIBPQ_HAS_SSL_LIBRARY_DETECTION feature macro.\r\n\r\nIf this looks good, I'm not sure how best to test it in the regression\r\nsuite. I see that libpq has an installcheck recipe that compiles a test\r\nexecutable for URI parsing; should I add a simple test alongside that?\r\n\r\nThanks,\r\n--Jacob\r\n\r\n[1] https://www.postgresql.org/message-id/a1798e46d6d801344ebc93672c6947ef5297c8a0.camel%40vmware.com","msg_date":"Wed, 23 Feb 2022 18:38:08 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"[PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"\nOn 2/23/22 13:38, Jacob Champion wrote:\n> Hello,\n>\n> As part of the NSS work it came up [1] that clients don't have a good\n> way to ask libpq what SSL library it was compiled with, unless they\n> already have a connection pointer so that they can call\n> PQsslAttribute(conn, \"library\"). This poses a chicken-and-egg problem:\n> with the NSS proposal, the client may have to know which library is in\n> use before it can construct a proper connection string. For example, I\n> have a test suite that needs to set up an NSS database with\n> certificates before telling libpq to connect using that database.\n>\n> The simplest proposal was to just allow PQsslAttribute() to take NULL\n> as the connection parameter when querying the \"library\" attribute, and\n> that's what I've done in this patch. In current versions of libpq, the\n> \"library\" attribute will always be NULL if you pass NULL as the\n> connection; a client that needs to know whether this new behavior is\n> present can look at the LIBPQ_HAS_SSL_LIBRARY_DETECTION feature macro.\n>\n> If this looks good, I'm not sure how best to test it in the regression\n> suite. I see that libpq has an installcheck recipe that compiles a test\n> executable for URI parsing; should I add a simple test alongside that?\n\n\nCreate a TAP tests that calls a small client?\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 23 Feb 2022 14:11:34 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"On Wed, 2022-02-23 at 14:11 -0500, Andrew Dunstan wrote:\r\n> On 2/23/22 13:38, Jacob Champion wrote:\r\n> > \r\n> > If this looks good, I'm not sure how best to test it in the regression\r\n> > suite. I see that libpq has an installcheck recipe that compiles a test\r\n> > executable for URI parsing; should I add a simple test alongside that?\r\n> \r\n> Create a TAP tests that calls a small client?\r\n\r\nFirst stab in v2-0002. Though I see that Andres is overhauling the\r\ntests in this folder today [1], so I'll need to watch that thread. :)\r\n\r\nThanks!\r\n--Jacob\r\n\r\n[1] https://www.postgresql.org/message-id/20220223203031.ezrd73ohvjgfksow%40alap3.anarazel.de","msg_date":"Wed, 23 Feb 2022 23:20:52 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"On Wed, 2022-02-23 at 23:20 +0000, Jacob Champion wrote:\r\n> First stab in v2-0002. Though I see that Andres is overhauling the\r\n> tests in this folder today [1], so I'll need to watch that thread. :)\r\n\r\nv3 rebases over Andres' changes and actually adds the Perl driver that\r\nI missed the git-add for.\r\n\r\n--Jacob","msg_date":"Mon, 28 Feb 2022 20:21:45 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"On Mon, Feb 28, 2022 at 3:21 PM Jacob Champion <pchampion@vmware.com> wrote:\n> v3 rebases over Andres' changes and actually adds the Perl driver that\n> I missed the git-add for.\n\nThis seems totally reasonable. However, I think it should update the\ndocumentation somehow.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 25 Mar 2022 15:32:23 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"On Fri, 2022-03-25 at 15:32 -0400, Robert Haas wrote:\r\n> On Mon, Feb 28, 2022 at 3:21 PM Jacob Champion <pchampion@vmware.com> wrote:\r\n> > v3 rebases over Andres' changes and actually adds the Perl driver that\r\n> > I missed the git-add for.\r\n> \r\n> This seems totally reasonable. However, I think it should update the\r\n> documentation somehow.\r\n\r\nDone in v4.\r\n\r\nDo I need to merge my tiny test program into the libpq_pipeline tests?\r\nI'm not sure what the roadmap is for those.\r\n\r\nThanks!\r\n--Jacob","msg_date":"Fri, 25 Mar 2022 21:25:37 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"> On 25 Mar 2022, at 22:25, Jacob Champion <pchampion@vmware.com> wrote:\n> On Fri, 2022-03-25 at 15:32 -0400, Robert Haas wrote:\n\n>> This seems totally reasonable. However, I think it should update the\n>> documentation somehow.\n> \n> Done in v4.\n\nI would prefer to not introduce a <note> for this, I think adding it as a\n<para> under PQsslAttribute is better given the rest of the libpq API\ndocumentation. The proposed text reads fine to me.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Fri, 25 Mar 2022 22:48:33 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"Jacob Champion <pchampion@vmware.com> writes:\n> Do I need to merge my tiny test program into the libpq_pipeline tests?\n\nDoesn't seem worth the trouble to me, notably because you'd\nthen have to cope with non-SSL builds too.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 25 Mar 2022 18:00:22 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"On Fri, 2022-03-25 at 18:00 -0400, Tom Lane wrote:\r\n> Jacob Champion <pchampion@vmware.com> writes:\r\n> > Do I need to merge my tiny test program into the libpq_pipeline tests?\r\n> \r\n> Doesn't seem worth the trouble to me, notably because you'd\r\n> then have to cope with non-SSL builds too.\r\n\r\nFine by me.\r\n\r\nv5 moves the docs out of the Note, as requested by Daniel.\r\n\r\nThanks,\r\n--Jacob","msg_date":"Fri, 25 Mar 2022 22:45:33 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"> On 25 Mar 2022, at 23:45, Jacob Champion <pchampion@vmware.com> wrote:\n> \n> On Fri, 2022-03-25 at 18:00 -0400, Tom Lane wrote:\n>> Jacob Champion <pchampion@vmware.com> writes:\n>>> Do I need to merge my tiny test program into the libpq_pipeline tests?\n>> \n>> Doesn't seem worth the trouble to me, notably because you'd\n>> then have to cope with non-SSL builds too.\n> \n> Fine by me.\n> \n> v5 moves the docs out of the Note, as requested by Daniel.\n\nI went over this again and I think this version is ready for committer. Having\ntried to add implement a new TLS library I would add a small comment on\nPQsslAttributeNames() for this, reverse-engineering what to implement is hard\nas it is without special cases easily identifiable. That can easily be done\nwhen pushed, no need for a new version IMHO.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Mon, 28 Mar 2022 22:33:57 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"Pushed with a few small tweaks to make it match project style, thanks!\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Tue, 29 Mar 2022 14:08:26 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"},{"msg_contents":"On Tue, 2022-03-29 at 14:08 +0200, Daniel Gustafsson wrote:\r\n> Pushed with a few small tweaks to make it match project style, thanks!\r\n\r\nThank you!\r\n\r\n--Jacob\r\n","msg_date":"Tue, 29 Mar 2022 15:29:32 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Enable SSL library detection via PQsslAttribute"}],"string":"[\n {\n \"msg_contents\": \"Hello,\\r\\n\\r\\nAs part of the NSS work it came up [1] that clients don't have a good\\r\\nway to ask libpq what SSL library it was compiled with, unless they\\r\\nalready have a connection pointer so that they can call\\r\\nPQsslAttribute(conn, \\\"library\\\"). This poses a chicken-and-egg problem:\\r\\nwith the NSS proposal, the client may have to know which library is in\\r\\nuse before it can construct a proper connection string. For example, I\\r\\nhave a test suite that needs to set up an NSS database with\\r\\ncertificates before telling libpq to connect using that database.\\r\\n\\r\\nThe simplest proposal was to just allow PQsslAttribute() to take NULL\\r\\nas the connection parameter when querying the \\\"library\\\" attribute, and\\r\\nthat's what I've done in this patch. In current versions of libpq, the\\r\\n\\\"library\\\" attribute will always be NULL if you pass NULL as the\\r\\nconnection; a client that needs to know whether this new behavior is\\r\\npresent can look at the LIBPQ_HAS_SSL_LIBRARY_DETECTION feature macro.\\r\\n\\r\\nIf this looks good, I'm not sure how best to test it in the regression\\r\\nsuite. I see that libpq has an installcheck recipe that compiles a test\\r\\nexecutable for URI parsing; should I add a simple test alongside that?\\r\\n\\r\\nThanks,\\r\\n--Jacob\\r\\n\\r\\n[1] https://www.postgresql.org/message-id/a1798e46d6d801344ebc93672c6947ef5297c8a0.camel%40vmware.com\",\n \"msg_date\": \"Wed, 23 Feb 2022 18:38:08 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"\\nOn 2/23/22 13:38, Jacob Champion wrote:\\n> Hello,\\n>\\n> As part of the NSS work it came up [1] that clients don't have a good\\n> way to ask libpq what SSL library it was compiled with, unless they\\n> already have a connection pointer so that they can call\\n> PQsslAttribute(conn, \\\"library\\\"). This poses a chicken-and-egg problem:\\n> with the NSS proposal, the client may have to know which library is in\\n> use before it can construct a proper connection string. For example, I\\n> have a test suite that needs to set up an NSS database with\\n> certificates before telling libpq to connect using that database.\\n>\\n> The simplest proposal was to just allow PQsslAttribute() to take NULL\\n> as the connection parameter when querying the \\\"library\\\" attribute, and\\n> that's what I've done in this patch. In current versions of libpq, the\\n> \\\"library\\\" attribute will always be NULL if you pass NULL as the\\n> connection; a client that needs to know whether this new behavior is\\n> present can look at the LIBPQ_HAS_SSL_LIBRARY_DETECTION feature macro.\\n>\\n> If this looks good, I'm not sure how best to test it in the regression\\n> suite. I see that libpq has an installcheck recipe that compiles a test\\n> executable for URI parsing; should I add a simple test alongside that?\\n\\n\\nCreate a TAP tests that calls a small client?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 14:11:34 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"On Wed, 2022-02-23 at 14:11 -0500, Andrew Dunstan wrote:\\r\\n> On 2/23/22 13:38, Jacob Champion wrote:\\r\\n> > \\r\\n> > If this looks good, I'm not sure how best to test it in the regression\\r\\n> > suite. I see that libpq has an installcheck recipe that compiles a test\\r\\n> > executable for URI parsing; should I add a simple test alongside that?\\r\\n> \\r\\n> Create a TAP tests that calls a small client?\\r\\n\\r\\nFirst stab in v2-0002. Though I see that Andres is overhauling the\\r\\ntests in this folder today [1], so I'll need to watch that thread. :)\\r\\n\\r\\nThanks!\\r\\n--Jacob\\r\\n\\r\\n[1] https://www.postgresql.org/message-id/20220223203031.ezrd73ohvjgfksow%40alap3.anarazel.de\",\n \"msg_date\": \"Wed, 23 Feb 2022 23:20:52 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"On Wed, 2022-02-23 at 23:20 +0000, Jacob Champion wrote:\\r\\n> First stab in v2-0002. Though I see that Andres is overhauling the\\r\\n> tests in this folder today [1], so I'll need to watch that thread. :)\\r\\n\\r\\nv3 rebases over Andres' changes and actually adds the Perl driver that\\r\\nI missed the git-add for.\\r\\n\\r\\n--Jacob\",\n \"msg_date\": \"Mon, 28 Feb 2022 20:21:45 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 3:21 PM Jacob Champion <pchampion@vmware.com> wrote:\\n> v3 rebases over Andres' changes and actually adds the Perl driver that\\n> I missed the git-add for.\\n\\nThis seems totally reasonable. However, I think it should update the\\ndocumentation somehow.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 15:32:23 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"On Fri, 2022-03-25 at 15:32 -0400, Robert Haas wrote:\\r\\n> On Mon, Feb 28, 2022 at 3:21 PM Jacob Champion <pchampion@vmware.com> wrote:\\r\\n> > v3 rebases over Andres' changes and actually adds the Perl driver that\\r\\n> > I missed the git-add for.\\r\\n> \\r\\n> This seems totally reasonable. However, I think it should update the\\r\\n> documentation somehow.\\r\\n\\r\\nDone in v4.\\r\\n\\r\\nDo I need to merge my tiny test program into the libpq_pipeline tests?\\r\\nI'm not sure what the roadmap is for those.\\r\\n\\r\\nThanks!\\r\\n--Jacob\",\n \"msg_date\": \"Fri, 25 Mar 2022 21:25:37 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"> On 25 Mar 2022, at 22:25, Jacob Champion <pchampion@vmware.com> wrote:\\n> On Fri, 2022-03-25 at 15:32 -0400, Robert Haas wrote:\\n\\n>> This seems totally reasonable. However, I think it should update the\\n>> documentation somehow.\\n> \\n> Done in v4.\\n\\nI would prefer to not introduce a <note> for this, I think adding it as a\\n<para> under PQsslAttribute is better given the rest of the libpq API\\ndocumentation. The proposed text reads fine to me.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 22:48:33 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"Jacob Champion <pchampion@vmware.com> writes:\\n> Do I need to merge my tiny test program into the libpq_pipeline tests?\\n\\nDoesn't seem worth the trouble to me, notably because you'd\\nthen have to cope with non-SSL builds too.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 18:00:22 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"On Fri, 2022-03-25 at 18:00 -0400, Tom Lane wrote:\\r\\n> Jacob Champion <pchampion@vmware.com> writes:\\r\\n> > Do I need to merge my tiny test program into the libpq_pipeline tests?\\r\\n> \\r\\n> Doesn't seem worth the trouble to me, notably because you'd\\r\\n> then have to cope with non-SSL builds too.\\r\\n\\r\\nFine by me.\\r\\n\\r\\nv5 moves the docs out of the Note, as requested by Daniel.\\r\\n\\r\\nThanks,\\r\\n--Jacob\",\n \"msg_date\": \"Fri, 25 Mar 2022 22:45:33 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"> On 25 Mar 2022, at 23:45, Jacob Champion <pchampion@vmware.com> wrote:\\n> \\n> On Fri, 2022-03-25 at 18:00 -0400, Tom Lane wrote:\\n>> Jacob Champion <pchampion@vmware.com> writes:\\n>>> Do I need to merge my tiny test program into the libpq_pipeline tests?\\n>> \\n>> Doesn't seem worth the trouble to me, notably because you'd\\n>> then have to cope with non-SSL builds too.\\n> \\n> Fine by me.\\n> \\n> v5 moves the docs out of the Note, as requested by Daniel.\\n\\nI went over this again and I think this version is ready for committer. Having\\ntried to add implement a new TLS library I would add a small comment on\\nPQsslAttributeNames() for this, reverse-engineering what to implement is hard\\nas it is without special cases easily identifiable. That can easily be done\\nwhen pushed, no need for a new version IMHO.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 22:33:57 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"Pushed with a few small tweaks to make it match project style, thanks!\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 14:08:26 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n },\n {\n \"msg_contents\": \"On Tue, 2022-03-29 at 14:08 +0200, Daniel Gustafsson wrote:\\r\\n> Pushed with a few small tweaks to make it match project style, thanks!\\r\\n\\r\\nThank you!\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 15:29:32 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Enable SSL library detection via PQsslAttribute\"\n }\n]","isConversational":false}}},{"rowIdx":53567,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nWhen verifying that the meson port actually runs all perl based tests I came\nacross src/interfaces/libpq/test/regress.pl. Instead of running tests yet\nanother way, it seems better to convert it to a tap test.\n\nI hope others agree?\n\nWhere would we want that test to live? Right now we have the slightly odd\nconvention that some tap tests live in src/test/{misc,modules,...}. But\ne.g. frontend binary ones are below src/bin/.\n\nFor now I've left it in src/interfaces/libpq/test, with the test in\nt/001_uri.pl. But we should at least get rid of the test/...\n\nPerhaps we should just rename src/test/modules/libpq_pipeline to\nsrc/test/modules/libpq and move uri-regress in there as well?\n\n\nThe tap test needs a bit more polish, mostly posted this to get some feedback\nbefore wasting more time :)\n\nGreetings,\n\nAndres Freund","msg_date":"Wed, 23 Feb 2022 12:30:31 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"convert libpq uri-regress tests to tap test"},{"msg_contents":"\nOn 2/23/22 15:30, Andres Freund wrote:\n> Hi,\n>\n> When verifying that the meson port actually runs all perl based tests I came\n> across src/interfaces/libpq/test/regress.pl. Instead of running tests yet\n> another way, it seems better to convert it to a tap test.\n>\n> I hope others agree?\n\n\nyes\n\n\n>\n> Where would we want that test to live? Right now we have the slightly odd\n> convention that some tap tests live in src/test/{misc,modules,...}. But\n> e.g. frontend binary ones are below src/bin/.\n>\n> For now I've left it in src/interfaces/libpq/test, with the test in\n> t/001_uri.pl. But we should at least get rid of the test/...\n>\n> Perhaps we should just rename src/test/modules/libpq_pipeline to\n> src/test/modules/libpq and move uri-regress in there as well?\n\n\n\nWFM\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 23 Feb 2022 15:39:25 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"> On 23 Feb 2022, at 21:30, Andres Freund <andres@anarazel.de> wrote:\n\n> When verifying that the meson port actually runs all perl based tests I came\n> across src/interfaces/libpq/test/regress.pl. Instead of running tests yet\n> another way, it seems better to convert it to a tap test.\n> \n> I hope others agree?\n\nMany +1's on that.\n\n> The tap test needs a bit more polish, mostly posted this to get some feedback\n> before wasting more time :)\n\nSkimming the new test it seems structurally fine to me.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Wed, 23 Feb 2022 21:40:59 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Andrew Dunstan <andrew@dunslane.net> writes:\n> On 2/23/22 15:30, Andres Freund wrote:\n>> Perhaps we should just rename src/test/modules/libpq_pipeline to\n>> src/test/modules/libpq and move uri-regress in there as well?\n\n> WFM\n\n+1\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 23 Feb 2022 15:57:08 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi,\n\nOn 2022-02-23 15:57:08 -0500, Tom Lane wrote:\n> Andrew Dunstan <andrew@dunslane.net> writes:\n> > On 2/23/22 15:30, Andres Freund wrote:\n> >> Perhaps we should just rename src/test/modules/libpq_pipeline to\n> >> src/test/modules/libpq and move uri-regress in there as well?\n> \n> > WFM\n> \n> +1\n\nCool.\n\n\nOne question on making that happen: Right now the makefiles building C stuff\nin src/test/modules all have the contrib-style stanza to build via pgxs. But\nafaics pgxs doesn't support building multiple programs. Which\nsrc/test/modules/libpq would need to.\n\nAaics there's currently no way to have Mkvcbuild.pm build multiple programs in\none dir via its makefile parsing stuff? Andrew, any suggestions for not\nneeding to spell this out in both the makefile and Mkvcbuild.pm?\n\n\nSeparately: I don't really understand why we do the whole if USE_PGXS dance in\nsrc/test/modules?\n\n- Andres\n\n\n","msg_date":"Wed, 23 Feb 2022 13:16:06 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On 2022-Feb-23, Andres Freund wrote:\n\n> When verifying that the meson port actually runs all perl based tests I came\n> across src/interfaces/libpq/test/regress.pl. Instead of running tests yet\n> another way, it seems better to convert it to a tap test.\n> \n> I hope others agree?\n\nWFM.\n\n> Perhaps we should just rename src/test/modules/libpq_pipeline to\n> src/test/modules/libpq and move uri-regress in there as well?\n\nWell, building multiple binaries would require some trickery that might\nbe excessive for this little tool. But +1 to that on principle.\n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\n\"La virtud es el justo medio entre dos defectos\" (Aristóteles)\n\n\n","msg_date":"Wed, 23 Feb 2022 18:44:44 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On 2022-Feb-23, Andres Freund wrote:\n\n> Separately: I don't really understand why we do the whole if USE_PGXS dance in\n> src/test/modules?\n\nYeah, it seems a bit silly. I'm not opposed to making these makefiles\nunconditionally do the PGXS thing -- or the non-PGXS thing, if that\nmakes it easier to build multiple binaries.\n\n-- \nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\n\"La grandeza es una experiencia transitoria. Nunca es consistente.\nDepende en gran parte de la imaginación humana creadora de mitos\"\n(Irulan)\n\n\n","msg_date":"Wed, 23 Feb 2022 18:46:34 -0300","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"\nOn 2/23/22 16:16, Andres Freund wrote:\n> Hi,\n>\n> On 2022-02-23 15:57:08 -0500, Tom Lane wrote:\n>> Andrew Dunstan <andrew@dunslane.net> writes:\n>>> On 2/23/22 15:30, Andres Freund wrote:\n>>>> Perhaps we should just rename src/test/modules/libpq_pipeline to\n>>>> src/test/modules/libpq and move uri-regress in there as well?\n>>> WFM\n>> +1\n> Cool.\n>\n>\n> One question on making that happen: Right now the makefiles building C stuff\n> in src/test/modules all have the contrib-style stanza to build via pgxs. But\n> afaics pgxs doesn't support building multiple programs. Which\n> src/test/modules/libpq would need to.\n\n\nThat's my understanding also.\n\n\n>\n> Aaics there's currently no way to have Mkvcbuild.pm build multiple programs in\n> one dir via its makefile parsing stuff? Andrew, any suggestions for not\n> needing to spell this out in both the makefile and Mkvcbuild.pm?\n\n\nWell, it should be a SMOC. If you can solve the first problem I'm sure I\ncan come up with a solution for Mkvcbuild.pm. But until we see the shape\nof the pgxs changes, planning for Mkcvbuild.pm changes seems remature.\n\n\n>\n>\n> Separately: I don't really understand why we do the whole if USE_PGXS dance in\n> src/test/modules?\n>\n\nENOCLUE\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 23 Feb 2022 16:57:02 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\n> On 2022-Feb-23, Andres Freund wrote:\n>> Separately: I don't really understand why we do the whole if USE_PGXS dance in\n>> src/test/modules?\n\n> Yeah, it seems a bit silly. I'm not opposed to making these makefiles\n> unconditionally do the PGXS thing -- or the non-PGXS thing, if that\n> makes it easier to build multiple binaries.\n\nAgreed, we could easily lose that. I think we only do it in contrib\nto serve as an example of how to use PGXS ... although considering\nwe're not *testing* that build mode, one wonders how much that proves.\nIn any case, src/test/modules doesn't have to do it.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 23 Feb 2022 17:03:56 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On 23.02.22 21:30, Andres Freund wrote:\n> hen verifying that the meson port actually runs all perl based tests I came\n> across src/interfaces/libpq/test/regress.pl. Instead of running tests yet\n> another way, it seems better to convert it to a tap test.\n> \n> I hope others agree?\n> \n> Where would we want that test to live? Right now we have the slightly odd\n> convention that some tap tests live in src/test/{misc,modules,...}. But\n> e.g. frontend binary ones are below src/bin/.\n\nlibpq TAP tests should be in src/interfaces/libpq/t/.\n\nI think there were issues that the build farm wouldn't pick up a test \nlocated there, but that should be fixed rather than worked around.\n\n\n","msg_date":"Wed, 23 Feb 2022 23:53:17 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> On 23.02.22 21:30, Andres Freund wrote:\n>> Where would we want that test to live? Right now we have the slightly odd\n>> convention that some tap tests live in src/test/{misc,modules,...}. But\n>> e.g. frontend binary ones are below src/bin/.\n\n> libpq TAP tests should be in src/interfaces/libpq/t/.\n\n> I think there were issues that the build farm wouldn't pick up a test \n> located there, but that should be fixed rather than worked around.\n\nThat's failing to account for the fact that a libpq test can't\nreally be a pure-perl TAP test; you need some C code to drive the\nlibrary. I don't agree with intermixing such code with libpq\nitself, independently of any buildsystem issues (which there\nmight well be). So I think the design of putting such tests under\nsrc/modules is fine.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 23 Feb 2022 17:58:37 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On 23.02.22 23:58, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>> On 23.02.22 21:30, Andres Freund wrote:\n>>> Where would we want that test to live? Right now we have the slightly odd\n>>> convention that some tap tests live in src/test/{misc,modules,...}. But\n>>> e.g. frontend binary ones are below src/bin/.\n> \n>> libpq TAP tests should be in src/interfaces/libpq/t/.\n> \n>> I think there were issues that the build farm wouldn't pick up a test\n>> located there, but that should be fixed rather than worked around.\n> \n> That's failing to account for the fact that a libpq test can't\n> really be a pure-perl TAP test; you need some C code to drive the\n> library. I don't agree with intermixing such code with libpq\n> itself, independently of any buildsystem issues (which there\n> might well be).\n\nSuch things could be put under src/interfaces/libpq/test, or some other \nsubdirectory. We already have src/interfaces/ecpg/test.\n\n> So I think the design of putting such tests under\n> src/modules is fine.\n\nI don't get what the rationale for that would be. libpq tests are not \n\"modules\" of any kind.\n\nIf I'm working on libpq, I want to do make && make check inside the \nlibpq source directory.\n\n\n","msg_date":"Thu, 24 Feb 2022 00:21:36 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> On 23.02.22 23:58, Tom Lane wrote:\n>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>>> libpq TAP tests should be in src/interfaces/libpq/t/.\n\n>> That's failing to account for the fact that a libpq test can't\n>> really be a pure-perl TAP test; you need some C code to drive the\n>> library.\n\n> Such things could be put under src/interfaces/libpq/test, or some other \n> subdirectory. We already have src/interfaces/ecpg/test.\n\nOK, but then the TAP scripts are under src/interfaces/libpq/test/t,\nwhich isn't what you said. I have no great objection to moving\nsrc/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\nthough, as long as the buildfarm will cope.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 23 Feb 2022 20:52:41 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"\nOn 2/23/22 20:52, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>> On 23.02.22 23:58, Tom Lane wrote:\n>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>>>> libpq TAP tests should be in src/interfaces/libpq/t/.\n>>> That's failing to account for the fact that a libpq test can't\n>>> really be a pure-perl TAP test; you need some C code to drive the\n>>> library.\n>> Such things could be put under src/interfaces/libpq/test, or some other \n>> subdirectory. We already have src/interfaces/ecpg/test.\n> OK, but then the TAP scripts are under src/interfaces/libpq/test/t,\n> which isn't what you said. I have no great objection to moving\n> src/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\n> though, as long as the buildfarm will cope.\n>\n> \t\t\t\n\n\nIt won't without some adjustment.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 24 Feb 2022 07:19:39 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On 24.02.22 02:52, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>> On 23.02.22 23:58, Tom Lane wrote:\n>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>>>> libpq TAP tests should be in src/interfaces/libpq/t/.\n> \n>>> That's failing to account for the fact that a libpq test can't\n>>> really be a pure-perl TAP test; you need some C code to drive the\n>>> library.\n> \n>> Such things could be put under src/interfaces/libpq/test, or some other\n>> subdirectory. We already have src/interfaces/ecpg/test.\n> \n> OK, but then the TAP scripts are under src/interfaces/libpq/test/t,\n> which isn't what you said. I have no great objection to moving\n> src/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\n> though, as long as the buildfarm will cope.\n\nI think the TAP scripts should be in src/interfaces/libpq/t/, as usual. \nThe supporting code snippets could live in some other directory under \nsrc/interfaces/libpq/, which might be called \"test\" or something else, \nnot that important.\n\nI think we should pick a layout that is proper and future-proof and then \nadjust the buildfarm client as necessary. The issue of writing \nlibpq-specific tests has come up a few times recently; I think it would \nbe worth finding a proper solution to this that would facilitate that \nwork in the future.\n\n\n","msg_date":"Thu, 24 Feb 2022 13:31:40 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi,\n\nOn 2022-02-24 13:31:40 +0100, Peter Eisentraut wrote:\n> I think the TAP scripts should be in src/interfaces/libpq/t/, as usual. The\n> supporting code snippets could live in some other directory under\n> src/interfaces/libpq/, which might be called \"test\" or something else, not\n> that important.\n\nWhy not in t/? We can't easily build the test programs in libpq/ itself, but\nlibpq/t should be fairly doable.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 24 Feb 2022 06:43:11 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2022-02-24 13:31:40 +0100, Peter Eisentraut wrote:\n>> I think the TAP scripts should be in src/interfaces/libpq/t/, as usual. The\n>> supporting code snippets could live in some other directory under\n>> src/interfaces/libpq/, which might be called \"test\" or something else, not\n>> that important.\n\n> Why not in t/? We can't easily build the test programs in libpq/ itself, but\n> libpq/t should be fairly doable.\n\nI think that having t/ directories contain only Perl test scripts\nis a good convention that we should stick to. Peter's proposal\nof a separate test/ subdirectory for C test scaffolding is\nprobably fine.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 24 Feb 2022 10:17:28 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi,\n\nOn 2022-02-24 10:17:28 -0500, Tom Lane wrote:\n> Andres Freund <andres@anarazel.de> writes:\n> > On 2022-02-24 13:31:40 +0100, Peter Eisentraut wrote:\n> >> I think the TAP scripts should be in src/interfaces/libpq/t/, as usual. The\n> >> supporting code snippets could live in some other directory under\n> >> src/interfaces/libpq/, which might be called \"test\" or something else, not\n> >> that important.\n>\n> > Why not in t/? We can't easily build the test programs in libpq/ itself, but\n> > libpq/t should be fairly doable.\n>\n> I think that having t/ directories contain only Perl test scripts\n> is a good convention that we should stick to. Peter's proposal\n> of a separate test/ subdirectory for C test scaffolding is\n> probably fine.\n\nThat exists today and continues to exist in the patch upthread, so it's easy\n;). I just need to move the libpq/test/t to libpq/t and adjust the binary\npath.\n\nOne annoying bit is that our current tap invocation infrastructure for msvc\nwon't know how to deal with that. We put the build directory containing t/\nonto PATH, but that won't work for test/. But we also don't want to install\ntest binaries. Not sure what the solution for that is.\n\nEven on !windows, we only know how to find \"test executables\" in tap tests via\nPATH. We're in the source dir, so we can't just do test/executable.\n\nI probably just need another coffee, but right now I don't even see how to add\nanything to PATH given $(prove_check)'s definition - it ends up with multiple\nshells. We can fix that by using && in the definition, which might be a good\nthing anyway?\n\nAttached three patches:\n\n0001: Convert src/interfaces/libpq/test to a tap test\n0002: Add tap test infrastructure to src/interfaces/libpq\n0003: Move libpq_pipeline test into src/interfaces/libpq.\n\nI did 0001 and 0002 in that order because prove errors out with a stacktrace\nif no tap tests exist... It might make more sense to commit them together, but\nfor review it's easier to keep them separate I think.\n\n\nAndrew, do you have an idea about the feasibility of supporting any of this\nwith the msvc build?\n\nI'm mildly inclined to only do 0001 and 0002 for now. We'd not loose msvc\ncoverage, because it already doesn't build the test. Once we've ironed that\nstuff out, we could do 0003?\n\nGreetings,\n\nAndres Freund","msg_date":"Thu, 24 Feb 2022 08:46:23 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On Thu, 2022-02-24 at 08:46 -0800, Andres Freund wrote:\r\n> One annoying bit is that our current tap invocation infrastructure for msvc\r\n> won't know how to deal with that. We put the build directory containing t/\r\n> onto PATH, but that won't work for test/. But we also don't want to install\r\n> test binaries. Not sure what the solution for that is.\r\n\r\nWould it help if the C executable, not Perl, was the thing actually\r\nproducing the TAP output? The binaries built from test/ could be placed\r\ninto t/. Or does that just open up a new set of problems?\r\n\r\n--Jacob\r\n","msg_date":"Thu, 24 Feb 2022 17:03:33 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi,\n\nOn 2022-02-24 17:03:33 +0000, Jacob Champion wrote:\n> On Thu, 2022-02-24 at 08:46 -0800, Andres Freund wrote:\n> > One annoying bit is that our current tap invocation infrastructure for msvc\n> > won't know how to deal with that. We put the build directory containing t/\n> > onto PATH, but that won't work for test/. But we also don't want to install\n> > test binaries. Not sure what the solution for that is.\n> \n> Would it help if the C executable, not Perl, was the thing actually\n> producing the TAP output? The binaries built from test/ could be placed\n> into t/. Or does that just open up a new set of problems?\n\nI don't think it would help that much. And for the libpq pipeline test it\ndoesn't really make sense anyway, because we intentionally use it with\ndifferent traces and such.\n\nI've thought about a few C tap tests that I'd like, but I think that's a\nseparate discussion.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 24 Feb 2022 09:35:43 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On 24.02.22 16:17, Tom Lane wrote:\n> I think that having t/ directories contain only Perl test scripts\n> is a good convention that we should stick to. Peter's proposal\n> of a separate test/ subdirectory for C test scaffolding is\n> probably fine.\n\nI wonder if there are any conventions in the Perl community about where \nto put test support files relative to the \"t\" directory.\n\n\n\n","msg_date":"Fri, 25 Feb 2022 14:32:43 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi,\n\nOn 2022-02-24 08:46:23 -0800, Andres Freund wrote:\n> I'm mildly inclined to only do 0001 and 0002 for now. We'd not loose msvc\n> coverage, because it already doesn't build the test. Once we've ironed that\n> stuff out, we could do 0003?\n\n From what I can see in the buildfarm client, we'd not loose (nor gain) any\nbuildfarm coverage either. It doesn't run the test today.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 25 Feb 2022 09:56:47 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi,\n\nOn 2022-02-25 09:56:47 -0800, Andres Freund wrote:\n> On 2022-02-24 08:46:23 -0800, Andres Freund wrote:\n> > I'm mildly inclined to only do 0001 and 0002 for now. We'd not loose msvc\n> > coverage, because it already doesn't build the test. Once we've ironed that\n> > stuff out, we could do 0003?\n> \n> From what I can see in the buildfarm client, we'd not loose (nor gain) any\n> buildfarm coverage either. It doesn't run the test today.\n\nAttached are rebased patches. I polished 0001, the regress.pl -> 001_uri.pl\nconversion some more (although some of perltidy's changes aren't clearly an\nimprovement).\n\nI'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\n\nGreetings,\n\nAndres Freund","msg_date":"Fri, 25 Feb 2022 17:52:29 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi,\n\nOn 2022-02-25 17:52:29 -0800, Andres Freund wrote:\n> I'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\n\nPushed. Attached is the remainder, 0003, the move of libpq_pipeline to\nsrc/interfaces/libpq that I'm not planning to push for now.\n\nRegards,\n\nAndres","msg_date":"Sat, 26 Feb 2022 17:46:26 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"\nOn 2/24/22 07:19, Andrew Dunstan wrote:\n> On 2/23/22 20:52, Tom Lane wrote:\n>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>>> On 23.02.22 23:58, Tom Lane wrote:\n>>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>>>>> libpq TAP tests should be in src/interfaces/libpq/t/.\n>>>> That's failing to account for the fact that a libpq test can't\n>>>> really be a pure-perl TAP test; you need some C code to drive the\n>>>> library.\n>>> Such things could be put under src/interfaces/libpq/test, or some other \n>>> subdirectory. We already have src/interfaces/ecpg/test.\n>> OK, but then the TAP scripts are under src/interfaces/libpq/test/t,\n>> which isn't what you said. I have no great objection to moving\n>> src/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\n>> though, as long as the buildfarm will cope.\n>>\n>> \t\t\t\n>\n> It won't without some adjustment.\n\n\n\nSee\n<https://github.com/PGBuildFarm/client-code/commit/ffc0fc029877632e9437af51bd99ace308daf0c8>\nand\n<https://buildfarm.postgresql.org/cgi-bin/show_stage_log.pl?nm=crake&dt=2022-03-01%2010%3A47%3A22&stg=module-libpq-check>\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Tue, 1 Mar 2022 10:44:27 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi, \n\nOn March 1, 2022 7:44:27 AM PST, Andrew Dunstan <andrew@dunslane.net> wrote:\n>\n>On 2/24/22 07:19, Andrew Dunstan wrote:\n>> On 2/23/22 20:52, Tom Lane wrote:\n>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>>>> On 23.02.22 23:58, Tom Lane wrote:\n>>>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>>>>>> libpq TAP tests should be in src/interfaces/libpq/t/.\n>>>>> That's failing to account for the fact that a libpq test can't\n>>>>> really be a pure-perl TAP test; you need some C code to drive the\n>>>>> library.\n>>>> Such things could be put under src/interfaces/libpq/test, or some other \n>>>> subdirectory. We already have src/interfaces/ecpg/test.\n>>> OK, but then the TAP scripts are under src/interfaces/libpq/test/t,\n>>> which isn't what you said. I have no great objection to moving\n>>> src/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\n>>> though, as long as the buildfarm will cope.\n>>>\n>>> \t\t\t\n>>\n>> It won't without some adjustment.\n>\n>\n>\n>See\n><https://github.com/PGBuildFarm/client-code/commit/ffc0fc029877632e9437af51bd99ace308daf0c8>\n>and\n><https://buildfarm.postgresql.org/cgi-bin/show_stage_log.pl?nm=crake&dt=2022-03-01%2010%3A47%3A22&stg=module-libpq-check>\n\nThanks! \n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Tue, 01 Mar 2022 07:52:43 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\n> Pushed. Attached is the remainder, 0003, the move of libpq_pipeline to\n> src/interfaces/libpq that I'm not planning to push for now.\n\nI saw that Andrew just pushed something to start building this under MSVC.\n\nIn case it's of any interest, I had done this differently a while back.\nThis probably doesn't apply except on top of some other patches, but you get\nthe idea.\n\ncommit 923f8a1c2cbea35cb01d1599caa2a81e3186181c\nAuthor: Justin Pryzby <pryzbyj@telsasoft.com>\nDate: Mon Feb 28 01:31:10 2022 -0600\n\n f!\n \n ci-os-only: windows\n\ndiff --git a/src/tools/msvc/Mkvcbuild.pm b/src/tools/msvc/Mkvcbuild.pm\nindex 4364ab943fd..71ec747e544 100644\n--- a/src/tools/msvc/Mkvcbuild.pm\n+++ b/src/tools/msvc/Mkvcbuild.pm\n@@ -44,6 +44,7 @@ my $contrib_extraincludes = {};\n my $contrib_extrasource = {\n \t'uri-regress' => ['src/interfaces/libpq/test/uri-regress.c'],\n \t'testclient' => ['src/interfaces/libpq/test/testclient.c'],\n+\t'libpq_pipeline' => ['src/interfaces/libpq/test/libpq_pipeline.c'],\n };\n my @contrib_excludes = (\n \t'bool_plperl', 'commit_ts',\n@@ -475,7 +476,7 @@ sub mkvcbuild\n \t\tpush @contrib_excludes, 'uuid-ossp';\n \t}\n \n-\tforeach my $subdir ('contrib', 'src/test/modules', 'src/interfaces/libpq')\n+\tforeach my $subdir ('contrib', 'src/test/modules') #, 'src/interfaces/libpq')\n \t{\n \t\topendir($D, $subdir) || croak \"Could not opendir on $subdir!\\n\";\n \t\twhile (my $d = readdir($D))\n@@ -804,6 +805,20 @@ sub mkvcbuild\n \t\t$p->AddReference($postgres);\n \t}\n \n+\t$mf = Project::read_file('src/interfaces/libpq/test/Makefile');\n+\t$mf =~ s{\\\\\\r?\\n}{}g;\n+\t$mf =~ m{PROGRAMS\\s*=\\s*(.*)$}m\n+\t || die 'Could not match in src/interfaces/libpq/test/Makefile' . \"\\n\";\n+\tforeach my $prg (split /\\s+/, $1)\n+\t{\n+\t\tmy $proj = $solution->AddProject($prg, 'exe', 'bin');\n+\t\t$proj->AddFile(\"src/interfaces/libpq/test/$prg.c\"); # implicit source file\n+\t\t$proj->AddIncludeDir('src/interfaces/libpq');\n+\t\t# XXX: pipeline needs pgcommon and ws2, but uri-regress doesn't\n+\t\t$proj->AddReference($libpq, $libpgport, $libpgcommon);\n+\t\t$proj->AddLibrary('ws2_32.lib');\n+\t}\n+\n \t$mf = Project::read_file('src/bin/scripts/Makefile');\n \t$mf =~ s{\\\\\\r?\\n}{}g;\n \t$mf =~ m{PROGRAMS\\s*=\\s*(.*)$}m\n\n\n","msg_date":"Sat, 16 Apr 2022 09:44:54 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"\nOn 2022-04-16 Sa 10:44, Justin Pryzby wrote:\n> On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\n>> Pushed. Attached is the remainder, 0003, the move of libpq_pipeline to\n>> src/interfaces/libpq that I'm not planning to push for now.\n> I saw that Andrew just pushed something to start building this under MSVC.\n>\n> In case it's of any interest, I had done this differently a while back.\n> This probably doesn't apply except on top of some other patches, but you get\n> the idea.\n>\n\nI think what I have committed should be quite adequate for now. Once we\nget to building with meson a lot of this ugliness should go away.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Sat, 16 Apr 2022 20:30:28 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\n> On 2022-02-25 17:52:29 -0800, Andres Freund wrote:\n> > I'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\n> \n> Pushed.\n\nIf I'm not wrong, this isn't being run by check-world.\n\ncommit 4dc465207517c4b69a1f2b657a8ad0700c08e34c\nAuthor: Justin Pryzby <pryzbyj@telsasoft.com>\nDate: Sat May 28 22:32:58 2022 -0500\n\n libpq tests were not being run\n \n See also:\n ac25173cdbc40b310a7e72d9557c45a699f1f7b3\n 6b04abdfc5e0653542ac5d586e639185a8c61a39\n\ndiff --git a/GNUmakefile.in b/GNUmakefile.in\nindex 2352fc1171a..cb613086c7c 100644\n--- a/GNUmakefile.in\n+++ b/GNUmakefile.in\n@@ -68,10 +68,10 @@ check check-tests installcheck installcheck-parallel installcheck-tests: CHECKPR\n check check-tests installcheck installcheck-parallel installcheck-tests: submake-generated-headers\n \t$(MAKE) -C src/test/regress $@\n \n-$(call recurse,check-world,src/test src/pl src/interfaces/ecpg contrib src/bin,check)\n-$(call recurse,checkprep, src/test src/pl src/interfaces/ecpg contrib src/bin)\n+$(call recurse,check-world,src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin,check)\n+$(call recurse,checkprep, src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin)\n \n-$(call recurse,installcheck-world,src/test src/pl src/interfaces/ecpg contrib src/bin,installcheck)\n+$(call recurse,installcheck-world,src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin,installcheck)\n $(call recurse,install-tests,src/test/regress,install-tests)\n \n GNUmakefile: GNUmakefile.in $(top_builddir)/config.status\n\n\n","msg_date":"Sun, 29 May 2022 10:18:50 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On Sun, May 29, 2022 at 10:18:50AM -0500, Justin Pryzby wrote:\n> On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\n> > On 2022-02-25 17:52:29 -0800, Andres Freund wrote:\n> > > I'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\n> > \n> > Pushed.\n> \n> If I'm not wrong, this isn't being run by check-world.\n\nYou are right.\n\n> -$(call recurse,check-world,src/test src/pl src/interfaces/ecpg contrib src/bin,check)\n> -$(call recurse,checkprep, src/test src/pl src/interfaces/ecpg contrib src/bin)\n> +$(call recurse,check-world,src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin,check)\n> +$(call recurse,checkprep, src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin)\n> \n> -$(call recurse,installcheck-world,src/test src/pl src/interfaces/ecpg contrib src/bin,installcheck)\n> +$(call recurse,installcheck-world,src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin,installcheck)\n> $(call recurse,install-tests,src/test/regress,install-tests)\n\nWhy don't you just use src/interfaces/ instead of adding a direct\npath to libpq?\n--\nMichael","msg_date":"Tue, 31 May 2022 13:58:25 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On Tue, May 31, 2022 at 01:58:25PM +0900, Michael Paquier wrote:\n> Why don't you just use src/interfaces/ instead of adding a direct\n> path to libpq?\n\nSo, this leads to something like the attached. Does that sound fine\nto you?\n--\nMichael","msg_date":"Wed, 1 Jun 2022 13:59:06 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi,\n\nOn 2022-05-29 10:18:50 -0500, Justin Pryzby wrote:\n> On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\n> > On 2022-02-25 17:52:29 -0800, Andres Freund wrote:\n> > > I'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\n> > \n> > Pushed.\n> \n> If I'm not wrong, this isn't being run by check-world.\n\nOops, yes. Thanks for catching!\n\n\n","msg_date":"Thu, 2 Jun 2022 09:47:22 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"Hi,\n\nOn 2022-06-01 13:59:06 +0900, Michael Paquier wrote:\n> On Tue, May 31, 2022 at 01:58:25PM +0900, Michael Paquier wrote:\n> > Why don't you just use src/interfaces/ instead of adding a direct\n> > path to libpq?\n> \n> So, this leads to something like the attached. Does that sound fine\n> to you?\n\nThat looks reasonable to me. Do you want to apply it or will you?\n\nRegards,\n\nAndres\n\n\n","msg_date":"Thu, 2 Jun 2022 09:48:25 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: convert libpq uri-regress tests to tap test"},{"msg_contents":"On Thu, Jun 02, 2022 at 09:48:25AM -0700, Andres Freund wrote:\n> On 2022-06-01 13:59:06 +0900, Michael Paquier wrote:\n>> So, this leads to something like the attached. Does that sound fine\n>> to you?\n> \n> That looks reasonable to me. Do you want to apply it or will you?\n\nThanks for double-checking! I should be able to take care of that\ntoday.\n--\nMichael","msg_date":"Fri, 3 Jun 2022 09:43:28 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: convert libpq uri-regress tests to tap test"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nWhen verifying that the meson port actually runs all perl based tests I came\\nacross src/interfaces/libpq/test/regress.pl. Instead of running tests yet\\nanother way, it seems better to convert it to a tap test.\\n\\nI hope others agree?\\n\\nWhere would we want that test to live? Right now we have the slightly odd\\nconvention that some tap tests live in src/test/{misc,modules,...}. But\\ne.g. frontend binary ones are below src/bin/.\\n\\nFor now I've left it in src/interfaces/libpq/test, with the test in\\nt/001_uri.pl. But we should at least get rid of the test/...\\n\\nPerhaps we should just rename src/test/modules/libpq_pipeline to\\nsrc/test/modules/libpq and move uri-regress in there as well?\\n\\n\\nThe tap test needs a bit more polish, mostly posted this to get some feedback\\nbefore wasting more time :)\\n\\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Wed, 23 Feb 2022 12:30:31 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"\\nOn 2/23/22 15:30, Andres Freund wrote:\\n> Hi,\\n>\\n> When verifying that the meson port actually runs all perl based tests I came\\n> across src/interfaces/libpq/test/regress.pl. Instead of running tests yet\\n> another way, it seems better to convert it to a tap test.\\n>\\n> I hope others agree?\\n\\n\\nyes\\n\\n\\n>\\n> Where would we want that test to live? Right now we have the slightly odd\\n> convention that some tap tests live in src/test/{misc,modules,...}. But\\n> e.g. frontend binary ones are below src/bin/.\\n>\\n> For now I've left it in src/interfaces/libpq/test, with the test in\\n> t/001_uri.pl. But we should at least get rid of the test/...\\n>\\n> Perhaps we should just rename src/test/modules/libpq_pipeline to\\n> src/test/modules/libpq and move uri-regress in there as well?\\n\\n\\n\\nWFM\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 15:39:25 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"> On 23 Feb 2022, at 21:30, Andres Freund <andres@anarazel.de> wrote:\\n\\n> When verifying that the meson port actually runs all perl based tests I came\\n> across src/interfaces/libpq/test/regress.pl. Instead of running tests yet\\n> another way, it seems better to convert it to a tap test.\\n> \\n> I hope others agree?\\n\\nMany +1's on that.\\n\\n> The tap test needs a bit more polish, mostly posted this to get some feedback\\n> before wasting more time :)\\n\\nSkimming the new test it seems structurally fine to me.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 21:40:59 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew@dunslane.net> writes:\\n> On 2/23/22 15:30, Andres Freund wrote:\\n>> Perhaps we should just rename src/test/modules/libpq_pipeline to\\n>> src/test/modules/libpq and move uri-regress in there as well?\\n\\n> WFM\\n\\n+1\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 15:57:08 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-23 15:57:08 -0500, Tom Lane wrote:\\n> Andrew Dunstan <andrew@dunslane.net> writes:\\n> > On 2/23/22 15:30, Andres Freund wrote:\\n> >> Perhaps we should just rename src/test/modules/libpq_pipeline to\\n> >> src/test/modules/libpq and move uri-regress in there as well?\\n> \\n> > WFM\\n> \\n> +1\\n\\nCool.\\n\\n\\nOne question on making that happen: Right now the makefiles building C stuff\\nin src/test/modules all have the contrib-style stanza to build via pgxs. But\\nafaics pgxs doesn't support building multiple programs. Which\\nsrc/test/modules/libpq would need to.\\n\\nAaics there's currently no way to have Mkvcbuild.pm build multiple programs in\\none dir via its makefile parsing stuff? Andrew, any suggestions for not\\nneeding to spell this out in both the makefile and Mkvcbuild.pm?\\n\\n\\nSeparately: I don't really understand why we do the whole if USE_PGXS dance in\\nsrc/test/modules?\\n\\n- Andres\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 13:16:06 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On 2022-Feb-23, Andres Freund wrote:\\n\\n> When verifying that the meson port actually runs all perl based tests I came\\n> across src/interfaces/libpq/test/regress.pl. Instead of running tests yet\\n> another way, it seems better to convert it to a tap test.\\n> \\n> I hope others agree?\\n\\nWFM.\\n\\n> Perhaps we should just rename src/test/modules/libpq_pipeline to\\n> src/test/modules/libpq and move uri-regress in there as well?\\n\\nWell, building multiple binaries would require some trickery that might\\nbe excessive for this little tool. But +1 to that on principle.\\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\\n\\\"La virtud es el justo medio entre dos defectos\\\" (Aristóteles)\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 18:44:44 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On 2022-Feb-23, Andres Freund wrote:\\n\\n> Separately: I don't really understand why we do the whole if USE_PGXS dance in\\n> src/test/modules?\\n\\nYeah, it seems a bit silly. I'm not opposed to making these makefiles\\nunconditionally do the PGXS thing -- or the non-PGXS thing, if that\\nmakes it easier to build multiple binaries.\\n\\n-- \\nÁlvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/\\n\\\"La grandeza es una experiencia transitoria. Nunca es consistente.\\nDepende en gran parte de la imaginación humana creadora de mitos\\\"\\n(Irulan)\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 18:46:34 -0300\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"\\nOn 2/23/22 16:16, Andres Freund wrote:\\n> Hi,\\n>\\n> On 2022-02-23 15:57:08 -0500, Tom Lane wrote:\\n>> Andrew Dunstan <andrew@dunslane.net> writes:\\n>>> On 2/23/22 15:30, Andres Freund wrote:\\n>>>> Perhaps we should just rename src/test/modules/libpq_pipeline to\\n>>>> src/test/modules/libpq and move uri-regress in there as well?\\n>>> WFM\\n>> +1\\n> Cool.\\n>\\n>\\n> One question on making that happen: Right now the makefiles building C stuff\\n> in src/test/modules all have the contrib-style stanza to build via pgxs. But\\n> afaics pgxs doesn't support building multiple programs. Which\\n> src/test/modules/libpq would need to.\\n\\n\\nThat's my understanding also.\\n\\n\\n>\\n> Aaics there's currently no way to have Mkvcbuild.pm build multiple programs in\\n> one dir via its makefile parsing stuff? Andrew, any suggestions for not\\n> needing to spell this out in both the makefile and Mkvcbuild.pm?\\n\\n\\nWell, it should be a SMOC. If you can solve the first problem I'm sure I\\ncan come up with a solution for Mkvcbuild.pm. But until we see the shape\\nof the pgxs changes, planning for Mkcvbuild.pm changes seems remature.\\n\\n\\n>\\n>\\n> Separately: I don't really understand why we do the whole if USE_PGXS dance in\\n> src/test/modules?\\n>\\n\\nENOCLUE\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 16:57:02 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Alvaro Herrera <alvherre@alvh.no-ip.org> writes:\\n> On 2022-Feb-23, Andres Freund wrote:\\n>> Separately: I don't really understand why we do the whole if USE_PGXS dance in\\n>> src/test/modules?\\n\\n> Yeah, it seems a bit silly. I'm not opposed to making these makefiles\\n> unconditionally do the PGXS thing -- or the non-PGXS thing, if that\\n> makes it easier to build multiple binaries.\\n\\nAgreed, we could easily lose that. I think we only do it in contrib\\nto serve as an example of how to use PGXS ... although considering\\nwe're not *testing* that build mode, one wonders how much that proves.\\nIn any case, src/test/modules doesn't have to do it.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 17:03:56 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On 23.02.22 21:30, Andres Freund wrote:\\n> hen verifying that the meson port actually runs all perl based tests I came\\n> across src/interfaces/libpq/test/regress.pl. Instead of running tests yet\\n> another way, it seems better to convert it to a tap test.\\n> \\n> I hope others agree?\\n> \\n> Where would we want that test to live? Right now we have the slightly odd\\n> convention that some tap tests live in src/test/{misc,modules,...}. But\\n> e.g. frontend binary ones are below src/bin/.\\n\\nlibpq TAP tests should be in src/interfaces/libpq/t/.\\n\\nI think there were issues that the build farm wouldn't pick up a test \\nlocated there, but that should be fixed rather than worked around.\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 23:53:17 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> On 23.02.22 21:30, Andres Freund wrote:\\n>> Where would we want that test to live? Right now we have the slightly odd\\n>> convention that some tap tests live in src/test/{misc,modules,...}. But\\n>> e.g. frontend binary ones are below src/bin/.\\n\\n> libpq TAP tests should be in src/interfaces/libpq/t/.\\n\\n> I think there were issues that the build farm wouldn't pick up a test \\n> located there, but that should be fixed rather than worked around.\\n\\nThat's failing to account for the fact that a libpq test can't\\nreally be a pure-perl TAP test; you need some C code to drive the\\nlibrary. I don't agree with intermixing such code with libpq\\nitself, independently of any buildsystem issues (which there\\nmight well be). So I think the design of putting such tests under\\nsrc/modules is fine.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 17:58:37 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On 23.02.22 23:58, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>> On 23.02.22 21:30, Andres Freund wrote:\\n>>> Where would we want that test to live? Right now we have the slightly odd\\n>>> convention that some tap tests live in src/test/{misc,modules,...}. But\\n>>> e.g. frontend binary ones are below src/bin/.\\n> \\n>> libpq TAP tests should be in src/interfaces/libpq/t/.\\n> \\n>> I think there were issues that the build farm wouldn't pick up a test\\n>> located there, but that should be fixed rather than worked around.\\n> \\n> That's failing to account for the fact that a libpq test can't\\n> really be a pure-perl TAP test; you need some C code to drive the\\n> library. I don't agree with intermixing such code with libpq\\n> itself, independently of any buildsystem issues (which there\\n> might well be).\\n\\nSuch things could be put under src/interfaces/libpq/test, or some other \\nsubdirectory. We already have src/interfaces/ecpg/test.\\n\\n> So I think the design of putting such tests under\\n> src/modules is fine.\\n\\nI don't get what the rationale for that would be. libpq tests are not \\n\\\"modules\\\" of any kind.\\n\\nIf I'm working on libpq, I want to do make && make check inside the \\nlibpq source directory.\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 00:21:36 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> On 23.02.22 23:58, Tom Lane wrote:\\n>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>>> libpq TAP tests should be in src/interfaces/libpq/t/.\\n\\n>> That's failing to account for the fact that a libpq test can't\\n>> really be a pure-perl TAP test; you need some C code to drive the\\n>> library.\\n\\n> Such things could be put under src/interfaces/libpq/test, or some other \\n> subdirectory. We already have src/interfaces/ecpg/test.\\n\\nOK, but then the TAP scripts are under src/interfaces/libpq/test/t,\\nwhich isn't what you said. I have no great objection to moving\\nsrc/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\\nthough, as long as the buildfarm will cope.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 20:52:41 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"\\nOn 2/23/22 20:52, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>> On 23.02.22 23:58, Tom Lane wrote:\\n>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>>>> libpq TAP tests should be in src/interfaces/libpq/t/.\\n>>> That's failing to account for the fact that a libpq test can't\\n>>> really be a pure-perl TAP test; you need some C code to drive the\\n>>> library.\\n>> Such things could be put under src/interfaces/libpq/test, or some other \\n>> subdirectory. We already have src/interfaces/ecpg/test.\\n> OK, but then the TAP scripts are under src/interfaces/libpq/test/t,\\n> which isn't what you said. I have no great objection to moving\\n> src/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\\n> though, as long as the buildfarm will cope.\\n>\\n> \\t\\t\\t\\n\\n\\nIt won't without some adjustment.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 07:19:39 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On 24.02.22 02:52, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>> On 23.02.22 23:58, Tom Lane wrote:\\n>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>>>> libpq TAP tests should be in src/interfaces/libpq/t/.\\n> \\n>>> That's failing to account for the fact that a libpq test can't\\n>>> really be a pure-perl TAP test; you need some C code to drive the\\n>>> library.\\n> \\n>> Such things could be put under src/interfaces/libpq/test, or some other\\n>> subdirectory. We already have src/interfaces/ecpg/test.\\n> \\n> OK, but then the TAP scripts are under src/interfaces/libpq/test/t,\\n> which isn't what you said. I have no great objection to moving\\n> src/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\\n> though, as long as the buildfarm will cope.\\n\\nI think the TAP scripts should be in src/interfaces/libpq/t/, as usual. \\nThe supporting code snippets could live in some other directory under \\nsrc/interfaces/libpq/, which might be called \\\"test\\\" or something else, \\nnot that important.\\n\\nI think we should pick a layout that is proper and future-proof and then \\nadjust the buildfarm client as necessary. The issue of writing \\nlibpq-specific tests has come up a few times recently; I think it would \\nbe worth finding a proper solution to this that would facilitate that \\nwork in the future.\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 13:31:40 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-24 13:31:40 +0100, Peter Eisentraut wrote:\\n> I think the TAP scripts should be in src/interfaces/libpq/t/, as usual. The\\n> supporting code snippets could live in some other directory under\\n> src/interfaces/libpq/, which might be called \\\"test\\\" or something else, not\\n> that important.\\n\\nWhy not in t/? We can't easily build the test programs in libpq/ itself, but\\nlibpq/t should be fairly doable.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 06:43:11 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2022-02-24 13:31:40 +0100, Peter Eisentraut wrote:\\n>> I think the TAP scripts should be in src/interfaces/libpq/t/, as usual. The\\n>> supporting code snippets could live in some other directory under\\n>> src/interfaces/libpq/, which might be called \\\"test\\\" or something else, not\\n>> that important.\\n\\n> Why not in t/? We can't easily build the test programs in libpq/ itself, but\\n> libpq/t should be fairly doable.\\n\\nI think that having t/ directories contain only Perl test scripts\\nis a good convention that we should stick to. Peter's proposal\\nof a separate test/ subdirectory for C test scaffolding is\\nprobably fine.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 10:17:28 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-24 10:17:28 -0500, Tom Lane wrote:\\n> Andres Freund <andres@anarazel.de> writes:\\n> > On 2022-02-24 13:31:40 +0100, Peter Eisentraut wrote:\\n> >> I think the TAP scripts should be in src/interfaces/libpq/t/, as usual. The\\n> >> supporting code snippets could live in some other directory under\\n> >> src/interfaces/libpq/, which might be called \\\"test\\\" or something else, not\\n> >> that important.\\n>\\n> > Why not in t/? We can't easily build the test programs in libpq/ itself, but\\n> > libpq/t should be fairly doable.\\n>\\n> I think that having t/ directories contain only Perl test scripts\\n> is a good convention that we should stick to. Peter's proposal\\n> of a separate test/ subdirectory for C test scaffolding is\\n> probably fine.\\n\\nThat exists today and continues to exist in the patch upthread, so it's easy\\n;). I just need to move the libpq/test/t to libpq/t and adjust the binary\\npath.\\n\\nOne annoying bit is that our current tap invocation infrastructure for msvc\\nwon't know how to deal with that. We put the build directory containing t/\\nonto PATH, but that won't work for test/. But we also don't want to install\\ntest binaries. Not sure what the solution for that is.\\n\\nEven on !windows, we only know how to find \\\"test executables\\\" in tap tests via\\nPATH. We're in the source dir, so we can't just do test/executable.\\n\\nI probably just need another coffee, but right now I don't even see how to add\\nanything to PATH given $(prove_check)'s definition - it ends up with multiple\\nshells. We can fix that by using && in the definition, which might be a good\\nthing anyway?\\n\\nAttached three patches:\\n\\n0001: Convert src/interfaces/libpq/test to a tap test\\n0002: Add tap test infrastructure to src/interfaces/libpq\\n0003: Move libpq_pipeline test into src/interfaces/libpq.\\n\\nI did 0001 and 0002 in that order because prove errors out with a stacktrace\\nif no tap tests exist... It might make more sense to commit them together, but\\nfor review it's easier to keep them separate I think.\\n\\n\\nAndrew, do you have an idea about the feasibility of supporting any of this\\nwith the msvc build?\\n\\nI'm mildly inclined to only do 0001 and 0002 for now. We'd not loose msvc\\ncoverage, because it already doesn't build the test. Once we've ironed that\\nstuff out, we could do 0003?\\n\\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Thu, 24 Feb 2022 08:46:23 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On Thu, 2022-02-24 at 08:46 -0800, Andres Freund wrote:\\r\\n> One annoying bit is that our current tap invocation infrastructure for msvc\\r\\n> won't know how to deal with that. We put the build directory containing t/\\r\\n> onto PATH, but that won't work for test/. But we also don't want to install\\r\\n> test binaries. Not sure what the solution for that is.\\r\\n\\r\\nWould it help if the C executable, not Perl, was the thing actually\\r\\nproducing the TAP output? The binaries built from test/ could be placed\\r\\ninto t/. Or does that just open up a new set of problems?\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 17:03:33 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-24 17:03:33 +0000, Jacob Champion wrote:\\n> On Thu, 2022-02-24 at 08:46 -0800, Andres Freund wrote:\\n> > One annoying bit is that our current tap invocation infrastructure for msvc\\n> > won't know how to deal with that. We put the build directory containing t/\\n> > onto PATH, but that won't work for test/. But we also don't want to install\\n> > test binaries. Not sure what the solution for that is.\\n> \\n> Would it help if the C executable, not Perl, was the thing actually\\n> producing the TAP output? The binaries built from test/ could be placed\\n> into t/. Or does that just open up a new set of problems?\\n\\nI don't think it would help that much. And for the libpq pipeline test it\\ndoesn't really make sense anyway, because we intentionally use it with\\ndifferent traces and such.\\n\\nI've thought about a few C tap tests that I'd like, but I think that's a\\nseparate discussion.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 09:35:43 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On 24.02.22 16:17, Tom Lane wrote:\\n> I think that having t/ directories contain only Perl test scripts\\n> is a good convention that we should stick to. Peter's proposal\\n> of a separate test/ subdirectory for C test scaffolding is\\n> probably fine.\\n\\nI wonder if there are any conventions in the Perl community about where \\nto put test support files relative to the \\\"t\\\" directory.\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 14:32:43 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-24 08:46:23 -0800, Andres Freund wrote:\\n> I'm mildly inclined to only do 0001 and 0002 for now. We'd not loose msvc\\n> coverage, because it already doesn't build the test. Once we've ironed that\\n> stuff out, we could do 0003?\\n\\n From what I can see in the buildfarm client, we'd not loose (nor gain) any\\nbuildfarm coverage either. It doesn't run the test today.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 09:56:47 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-25 09:56:47 -0800, Andres Freund wrote:\\n> On 2022-02-24 08:46:23 -0800, Andres Freund wrote:\\n> > I'm mildly inclined to only do 0001 and 0002 for now. We'd not loose msvc\\n> > coverage, because it already doesn't build the test. Once we've ironed that\\n> > stuff out, we could do 0003?\\n> \\n> From what I can see in the buildfarm client, we'd not loose (nor gain) any\\n> buildfarm coverage either. It doesn't run the test today.\\n\\nAttached are rebased patches. I polished 0001, the regress.pl -> 001_uri.pl\\nconversion some more (although some of perltidy's changes aren't clearly an\\nimprovement).\\n\\nI'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\\n\\nGreetings,\\n\\nAndres Freund\",\n \"msg_date\": \"Fri, 25 Feb 2022 17:52:29 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-25 17:52:29 -0800, Andres Freund wrote:\\n> I'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\\n\\nPushed. Attached is the remainder, 0003, the move of libpq_pipeline to\\nsrc/interfaces/libpq that I'm not planning to push for now.\\n\\nRegards,\\n\\nAndres\",\n \"msg_date\": \"Sat, 26 Feb 2022 17:46:26 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"\\nOn 2/24/22 07:19, Andrew Dunstan wrote:\\n> On 2/23/22 20:52, Tom Lane wrote:\\n>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>>> On 23.02.22 23:58, Tom Lane wrote:\\n>>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>>>>> libpq TAP tests should be in src/interfaces/libpq/t/.\\n>>>> That's failing to account for the fact that a libpq test can't\\n>>>> really be a pure-perl TAP test; you need some C code to drive the\\n>>>> library.\\n>>> Such things could be put under src/interfaces/libpq/test, or some other \\n>>> subdirectory. We already have src/interfaces/ecpg/test.\\n>> OK, but then the TAP scripts are under src/interfaces/libpq/test/t,\\n>> which isn't what you said. I have no great objection to moving\\n>> src/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\\n>> though, as long as the buildfarm will cope.\\n>>\\n>> \\t\\t\\t\\n>\\n> It won't without some adjustment.\\n\\n\\n\\nSee\\n<https://github.com/PGBuildFarm/client-code/commit/ffc0fc029877632e9437af51bd99ace308daf0c8>\\nand\\n<https://buildfarm.postgresql.org/cgi-bin/show_stage_log.pl?nm=crake&dt=2022-03-01%2010%3A47%3A22&stg=module-libpq-check>\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 10:44:27 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi, \\n\\nOn March 1, 2022 7:44:27 AM PST, Andrew Dunstan <andrew@dunslane.net> wrote:\\n>\\n>On 2/24/22 07:19, Andrew Dunstan wrote:\\n>> On 2/23/22 20:52, Tom Lane wrote:\\n>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>>>> On 23.02.22 23:58, Tom Lane wrote:\\n>>>>> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>>>>>> libpq TAP tests should be in src/interfaces/libpq/t/.\\n>>>>> That's failing to account for the fact that a libpq test can't\\n>>>>> really be a pure-perl TAP test; you need some C code to drive the\\n>>>>> library.\\n>>>> Such things could be put under src/interfaces/libpq/test, or some other \\n>>>> subdirectory. We already have src/interfaces/ecpg/test.\\n>>> OK, but then the TAP scripts are under src/interfaces/libpq/test/t,\\n>>> which isn't what you said. I have no great objection to moving\\n>>> src/test/modules/libpq_pipeline/ to src/interfaces/libpq/test/,\\n>>> though, as long as the buildfarm will cope.\\n>>>\\n>>> \\t\\t\\t\\n>>\\n>> It won't without some adjustment.\\n>\\n>\\n>\\n>See\\n><https://github.com/PGBuildFarm/client-code/commit/ffc0fc029877632e9437af51bd99ace308daf0c8>\\n>and\\n><https://buildfarm.postgresql.org/cgi-bin/show_stage_log.pl?nm=crake&dt=2022-03-01%2010%3A47%3A22&stg=module-libpq-check>\\n\\nThanks! \\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Tue, 01 Mar 2022 07:52:43 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\\n> Pushed. Attached is the remainder, 0003, the move of libpq_pipeline to\\n> src/interfaces/libpq that I'm not planning to push for now.\\n\\nI saw that Andrew just pushed something to start building this under MSVC.\\n\\nIn case it's of any interest, I had done this differently a while back.\\nThis probably doesn't apply except on top of some other patches, but you get\\nthe idea.\\n\\ncommit 923f8a1c2cbea35cb01d1599caa2a81e3186181c\\nAuthor: Justin Pryzby <pryzbyj@telsasoft.com>\\nDate: Mon Feb 28 01:31:10 2022 -0600\\n\\n f!\\n \\n ci-os-only: windows\\n\\ndiff --git a/src/tools/msvc/Mkvcbuild.pm b/src/tools/msvc/Mkvcbuild.pm\\nindex 4364ab943fd..71ec747e544 100644\\n--- a/src/tools/msvc/Mkvcbuild.pm\\n+++ b/src/tools/msvc/Mkvcbuild.pm\\n@@ -44,6 +44,7 @@ my $contrib_extraincludes = {};\\n my $contrib_extrasource = {\\n \\t'uri-regress' => ['src/interfaces/libpq/test/uri-regress.c'],\\n \\t'testclient' => ['src/interfaces/libpq/test/testclient.c'],\\n+\\t'libpq_pipeline' => ['src/interfaces/libpq/test/libpq_pipeline.c'],\\n };\\n my @contrib_excludes = (\\n \\t'bool_plperl', 'commit_ts',\\n@@ -475,7 +476,7 @@ sub mkvcbuild\\n \\t\\tpush @contrib_excludes, 'uuid-ossp';\\n \\t}\\n \\n-\\tforeach my $subdir ('contrib', 'src/test/modules', 'src/interfaces/libpq')\\n+\\tforeach my $subdir ('contrib', 'src/test/modules') #, 'src/interfaces/libpq')\\n \\t{\\n \\t\\topendir($D, $subdir) || croak \\\"Could not opendir on $subdir!\\\\n\\\";\\n \\t\\twhile (my $d = readdir($D))\\n@@ -804,6 +805,20 @@ sub mkvcbuild\\n \\t\\t$p->AddReference($postgres);\\n \\t}\\n \\n+\\t$mf = Project::read_file('src/interfaces/libpq/test/Makefile');\\n+\\t$mf =~ s{\\\\\\\\\\\\r?\\\\n}{}g;\\n+\\t$mf =~ m{PROGRAMS\\\\s*=\\\\s*(.*)$}m\\n+\\t || die 'Could not match in src/interfaces/libpq/test/Makefile' . \\\"\\\\n\\\";\\n+\\tforeach my $prg (split /\\\\s+/, $1)\\n+\\t{\\n+\\t\\tmy $proj = $solution->AddProject($prg, 'exe', 'bin');\\n+\\t\\t$proj->AddFile(\\\"src/interfaces/libpq/test/$prg.c\\\"); # implicit source file\\n+\\t\\t$proj->AddIncludeDir('src/interfaces/libpq');\\n+\\t\\t# XXX: pipeline needs pgcommon and ws2, but uri-regress doesn't\\n+\\t\\t$proj->AddReference($libpq, $libpgport, $libpgcommon);\\n+\\t\\t$proj->AddLibrary('ws2_32.lib');\\n+\\t}\\n+\\n \\t$mf = Project::read_file('src/bin/scripts/Makefile');\\n \\t$mf =~ s{\\\\\\\\\\\\r?\\\\n}{}g;\\n \\t$mf =~ m{PROGRAMS\\\\s*=\\\\s*(.*)$}m\\n\\n\\n\",\n \"msg_date\": \"Sat, 16 Apr 2022 09:44:54 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"\\nOn 2022-04-16 Sa 10:44, Justin Pryzby wrote:\\n> On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\\n>> Pushed. Attached is the remainder, 0003, the move of libpq_pipeline to\\n>> src/interfaces/libpq that I'm not planning to push for now.\\n> I saw that Andrew just pushed something to start building this under MSVC.\\n>\\n> In case it's of any interest, I had done this differently a while back.\\n> This probably doesn't apply except on top of some other patches, but you get\\n> the idea.\\n>\\n\\nI think what I have committed should be quite adequate for now. Once we\\nget to building with meson a lot of this ugliness should go away.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 16 Apr 2022 20:30:28 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\\n> On 2022-02-25 17:52:29 -0800, Andres Freund wrote:\\n> > I'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\\n> \\n> Pushed.\\n\\nIf I'm not wrong, this isn't being run by check-world.\\n\\ncommit 4dc465207517c4b69a1f2b657a8ad0700c08e34c\\nAuthor: Justin Pryzby <pryzbyj@telsasoft.com>\\nDate: Sat May 28 22:32:58 2022 -0500\\n\\n libpq tests were not being run\\n \\n See also:\\n ac25173cdbc40b310a7e72d9557c45a699f1f7b3\\n 6b04abdfc5e0653542ac5d586e639185a8c61a39\\n\\ndiff --git a/GNUmakefile.in b/GNUmakefile.in\\nindex 2352fc1171a..cb613086c7c 100644\\n--- a/GNUmakefile.in\\n+++ b/GNUmakefile.in\\n@@ -68,10 +68,10 @@ check check-tests installcheck installcheck-parallel installcheck-tests: CHECKPR\\n check check-tests installcheck installcheck-parallel installcheck-tests: submake-generated-headers\\n \\t$(MAKE) -C src/test/regress $@\\n \\n-$(call recurse,check-world,src/test src/pl src/interfaces/ecpg contrib src/bin,check)\\n-$(call recurse,checkprep, src/test src/pl src/interfaces/ecpg contrib src/bin)\\n+$(call recurse,check-world,src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin,check)\\n+$(call recurse,checkprep, src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin)\\n \\n-$(call recurse,installcheck-world,src/test src/pl src/interfaces/ecpg contrib src/bin,installcheck)\\n+$(call recurse,installcheck-world,src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin,installcheck)\\n $(call recurse,install-tests,src/test/regress,install-tests)\\n \\n GNUmakefile: GNUmakefile.in $(top_builddir)/config.status\\n\\n\\n\",\n \"msg_date\": \"Sun, 29 May 2022 10:18:50 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On Sun, May 29, 2022 at 10:18:50AM -0500, Justin Pryzby wrote:\\n> On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\\n> > On 2022-02-25 17:52:29 -0800, Andres Freund wrote:\\n> > > I'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\\n> > \\n> > Pushed.\\n> \\n> If I'm not wrong, this isn't being run by check-world.\\n\\nYou are right.\\n\\n> -$(call recurse,check-world,src/test src/pl src/interfaces/ecpg contrib src/bin,check)\\n> -$(call recurse,checkprep, src/test src/pl src/interfaces/ecpg contrib src/bin)\\n> +$(call recurse,check-world,src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin,check)\\n> +$(call recurse,checkprep, src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin)\\n> \\n> -$(call recurse,installcheck-world,src/test src/pl src/interfaces/ecpg contrib src/bin,installcheck)\\n> +$(call recurse,installcheck-world,src/test src/pl src/interfaces/ecpg src/interfaces/libpq contrib src/bin,installcheck)\\n> $(call recurse,install-tests,src/test/regress,install-tests)\\n\\nWhy don't you just use src/interfaces/ instead of adding a direct\\npath to libpq?\\n--\\nMichael\",\n \"msg_date\": \"Tue, 31 May 2022 13:58:25 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On Tue, May 31, 2022 at 01:58:25PM +0900, Michael Paquier wrote:\\n> Why don't you just use src/interfaces/ instead of adding a direct\\n> path to libpq?\\n\\nSo, this leads to something like the attached. Does that sound fine\\nto you?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 1 Jun 2022 13:59:06 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-05-29 10:18:50 -0500, Justin Pryzby wrote:\\n> On Sat, Feb 26, 2022 at 05:46:26PM -0800, Andres Freund wrote:\\n> > On 2022-02-25 17:52:29 -0800, Andres Freund wrote:\\n> > > I'd like to commit 0001 and 0002 soon, unless somebody sees a reason not to?\\n> > \\n> > Pushed.\\n> \\n> If I'm not wrong, this isn't being run by check-world.\\n\\nOops, yes. Thanks for catching!\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Jun 2022 09:47:22 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-06-01 13:59:06 +0900, Michael Paquier wrote:\\n> On Tue, May 31, 2022 at 01:58:25PM +0900, Michael Paquier wrote:\\n> > Why don't you just use src/interfaces/ instead of adding a direct\\n> > path to libpq?\\n> \\n> So, this leads to something like the attached. Does that sound fine\\n> to you?\\n\\nThat looks reasonable to me. Do you want to apply it or will you?\\n\\nRegards,\\n\\nAndres\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Jun 2022 09:48:25 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 02, 2022 at 09:48:25AM -0700, Andres Freund wrote:\\n> On 2022-06-01 13:59:06 +0900, Michael Paquier wrote:\\n>> So, this leads to something like the attached. Does that sound fine\\n>> to you?\\n> \\n> That looks reasonable to me. Do you want to apply it or will you?\\n\\nThanks for double-checking! I should be able to take care of that\\ntoday.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 3 Jun 2022 09:43:28 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: convert libpq uri-regress tests to tap test\"\n }\n]","isConversational":false}}},{"rowIdx":53568,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I brought this up in [1] but wanted to start a dedicated thread.\n\nSince 16fa9b2b30a357 GiST indexes are not built in shared buffers.\nHowever, smgrimmedsync() is not done anywhere and skipFsync=true is\nalways passed to smgrwrite() and smgrextend(). So, if a checkpoint\nstarts and finishes after WAL-logging some of the index build pages and\nthere is a crash sometime before the dirty pages make it to permanent\nstorage, won't that data be lost?\n\n(Also, as Heikki has pointed out before, passing skipFsync=false isn't\nsufficient either if checkpoint finishes before\nregister_dirty_segment(), so seems like smgrimmesync() is needed.)\n\nSeems like the following would address this:\n\ndiff --git a/src/backend/access/gist/gistbuild.c\nb/src/backend/access/gist/gistbuild.c\nindex 9854116fca..364a7455ea 100644\n--- a/src/backend/access/gist/gistbuild.c\n+++ b/src/backend/access/gist/gistbuild.c\n@@ -454,9 +454,11 @@ gist_indexsortbuild(GISTBuildState *state)\n smgrwrite(RelationGetSmgr(state->indexrel), MAIN_FORKNUM,\nGIST_ROOT_BLKNO,\n pagestate->page, true);\n if (RelationNeedsWAL(state->indexrel))\n+ {\n log_newpage(&state->indexrel->rd_node, MAIN_FORKNUM,\nGIST_ROOT_BLKNO,\n pagestate->page, true);\n-\n+ smgrimmedsync(RelationGetSmgr(state->indexrel), MAIN_FORKNUM);\n+ }\n pfree(pagestate->page);\n pfree(pagestate);\n }\n\n- Melanie\n\n[1] https://www.postgresql.org/message-id/CAAKRu_ZmkCNz-%3D5U5wZyyi3%3DUsfs1btsTKc_L9r1ceFvbxE3kg%40mail.gmail.com\n\n\n","msg_date":"Wed, 23 Feb 2022 16:30:50 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"GiST index build missing smgrimmedsync()?"},{"msg_contents":"On 23/02/2022 23:30, Melanie Plageman wrote:\n> I brought this up in [1] but wanted to start a dedicated thread.\n> \n> Since 16fa9b2b30a357 GiST indexes are not built in shared buffers.\n> However, smgrimmedsync() is not done anywhere and skipFsync=true is\n> always passed to smgrwrite() and smgrextend(). So, if a checkpoint\n> starts and finishes after WAL-logging some of the index build pages and\n> there is a crash sometime before the dirty pages make it to permanent\n> storage, won't that data be lost?\n\nYes, good catch!\n\n> Seems like the following would address this:\n\nCommitted essentially that, except that I put the smgrimmedsync in a \nseparate if-block, and copied the comment from the similar piece of code \nfrom nbtsort.c to explain the issue.\n\nThanks!\n\n- Heikki\n\n\n","msg_date":"Thu, 24 Feb 2022 16:32:53 +0200","msg_from":"Heikki Linnakangas <hlinnaka@iki.fi>","msg_from_op":false,"msg_subject":"Re: GiST index build missing smgrimmedsync()?"}],"string":"[\n {\n \"msg_contents\": \"I brought this up in [1] but wanted to start a dedicated thread.\\n\\nSince 16fa9b2b30a357 GiST indexes are not built in shared buffers.\\nHowever, smgrimmedsync() is not done anywhere and skipFsync=true is\\nalways passed to smgrwrite() and smgrextend(). So, if a checkpoint\\nstarts and finishes after WAL-logging some of the index build pages and\\nthere is a crash sometime before the dirty pages make it to permanent\\nstorage, won't that data be lost?\\n\\n(Also, as Heikki has pointed out before, passing skipFsync=false isn't\\nsufficient either if checkpoint finishes before\\nregister_dirty_segment(), so seems like smgrimmesync() is needed.)\\n\\nSeems like the following would address this:\\n\\ndiff --git a/src/backend/access/gist/gistbuild.c\\nb/src/backend/access/gist/gistbuild.c\\nindex 9854116fca..364a7455ea 100644\\n--- a/src/backend/access/gist/gistbuild.c\\n+++ b/src/backend/access/gist/gistbuild.c\\n@@ -454,9 +454,11 @@ gist_indexsortbuild(GISTBuildState *state)\\n smgrwrite(RelationGetSmgr(state->indexrel), MAIN_FORKNUM,\\nGIST_ROOT_BLKNO,\\n pagestate->page, true);\\n if (RelationNeedsWAL(state->indexrel))\\n+ {\\n log_newpage(&state->indexrel->rd_node, MAIN_FORKNUM,\\nGIST_ROOT_BLKNO,\\n pagestate->page, true);\\n-\\n+ smgrimmedsync(RelationGetSmgr(state->indexrel), MAIN_FORKNUM);\\n+ }\\n pfree(pagestate->page);\\n pfree(pagestate);\\n }\\n\\n- Melanie\\n\\n[1] https://www.postgresql.org/message-id/CAAKRu_ZmkCNz-%3D5U5wZyyi3%3DUsfs1btsTKc_L9r1ceFvbxE3kg%40mail.gmail.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 16:30:50 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"GiST index build missing smgrimmedsync()?\"\n },\n {\n \"msg_contents\": \"On 23/02/2022 23:30, Melanie Plageman wrote:\\n> I brought this up in [1] but wanted to start a dedicated thread.\\n> \\n> Since 16fa9b2b30a357 GiST indexes are not built in shared buffers.\\n> However, smgrimmedsync() is not done anywhere and skipFsync=true is\\n> always passed to smgrwrite() and smgrextend(). So, if a checkpoint\\n> starts and finishes after WAL-logging some of the index build pages and\\n> there is a crash sometime before the dirty pages make it to permanent\\n> storage, won't that data be lost?\\n\\nYes, good catch!\\n\\n> Seems like the following would address this:\\n\\nCommitted essentially that, except that I put the smgrimmedsync in a \\nseparate if-block, and copied the comment from the similar piece of code \\nfrom nbtsort.c to explain the issue.\\n\\nThanks!\\n\\n- Heikki\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 16:32:53 +0200\",\n \"msg_from\": \"Heikki Linnakangas <hlinnaka@iki.fi>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: GiST index build missing smgrimmedsync()?\"\n }\n]","isConversational":false}}},{"rowIdx":53569,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"In pg_regress we realloc() with the destination and source pointer being equal,\nwithout checking for OOM. While a fairly unlikely source of errors, is there a\nreason not to use pg_realloc() there for hygiene?\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/","msg_date":"Wed, 23 Feb 2022 22:32:30 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"reallocing without oom check in pg_regress"},{"msg_contents":"Daniel Gustafsson <daniel@yesql.se> writes:\n> In pg_regress we realloc() with the destination and source pointer being equal,\n> without checking for OOM. While a fairly unlikely source of errors, is there a\n> reason not to use pg_realloc() there for hygiene?\n\nYeah, looks like oversight to me.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 23 Feb 2022 17:05:15 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: reallocing without oom check in pg_regress"},{"msg_contents":"> On 23 Feb 2022, at 23:05, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> Daniel Gustafsson <daniel@yesql.se> writes:\n>> In pg_regress we realloc() with the destination and source pointer being equal,\n>> without checking for OOM. While a fairly unlikely source of errors, is there a\n>> reason not to use pg_realloc() there for hygiene?\n> \n> Yeah, looks like oversight to me.\n\nThanks for confirming, I've pushed this now after taking it for a spin on the\nCI just in case.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Thu, 24 Feb 2022 21:00:05 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":true,"msg_subject":"Re: reallocing without oom check in pg_regress"}],"string":"[\n {\n \"msg_contents\": \"In pg_regress we realloc() with the destination and source pointer being equal,\\nwithout checking for OOM. While a fairly unlikely source of errors, is there a\\nreason not to use pg_realloc() there for hygiene?\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\",\n \"msg_date\": \"Wed, 23 Feb 2022 22:32:30 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"reallocing without oom check in pg_regress\"\n },\n {\n \"msg_contents\": \"Daniel Gustafsson <daniel@yesql.se> writes:\\n> In pg_regress we realloc() with the destination and source pointer being equal,\\n> without checking for OOM. While a fairly unlikely source of errors, is there a\\n> reason not to use pg_realloc() there for hygiene?\\n\\nYeah, looks like oversight to me.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 17:05:15 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: reallocing without oom check in pg_regress\"\n },\n {\n \"msg_contents\": \"> On 23 Feb 2022, at 23:05, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n>> In pg_regress we realloc() with the destination and source pointer being equal,\\n>> without checking for OOM. While a fairly unlikely source of errors, is there a\\n>> reason not to use pg_realloc() there for hygiene?\\n> \\n> Yeah, looks like oversight to me.\\n\\nThanks for confirming, I've pushed this now after taking it for a spin on the\\nCI just in case.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 21:00:05 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: reallocing without oom check in pg_regress\"\n }\n]","isConversational":false}}},{"rowIdx":53570,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\r\n\r\nStephen pointed out [1] that the authenticated identity that's stored\r\nin MyProcPort can't be retrieved by extensions or triggers. Attached is\r\na patch that provides both a C API and a SQL function for retrieving\r\nit.\r\n\r\nGetAuthenticatedIdentityString() is a mouthful but I wanted to\r\ndifferentiate it from the existing GetAuthenticatedUserId(); better\r\nnames welcome. It only exists as an accessor because I wasn't sure if\r\nextensions outside of contrib were encouraged to rely on the internal\r\nlayout of struct Port. (If they can, then that call can go away\r\nentirely.)\r\n\r\nThanks,\r\n--Jacob\r\n\r\n[1] https://www.postgresql.org/message-id/CAOuzzgpFpuroNRabEvB9kST_TSyS2jFicBNoXvW7G2pZFixyBw%40mail.gmail.com","msg_date":"Thu, 24 Feb 2022 00:15:40 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"[PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, Feb 24, 2022 at 12:15:40AM +0000, Jacob Champion wrote:\n> Stephen pointed out [1] that the authenticated identity that's stored\n> in MyProcPort can't be retrieved by extensions or triggers. Attached is\n> a patch that provides both a C API and a SQL function for retrieving\n> it.\n> \n> GetAuthenticatedIdentityString() is a mouthful but I wanted to\n> differentiate it from the existing GetAuthenticatedUserId(); better\n> names welcome. It only exists as an accessor because I wasn't sure if\n> extensions outside of contrib were encouraged to rely on the internal\n> layout of struct Port. (If they can, then that call can go away\n> entirely.)\n\n+char *\n+GetAuthenticatedIdentityString(void)\n+{\n+ if (!MyProcPort || !MyProcPort->authn_id)\n+ return NULL;\n+\n+ return pstrdup(MyProcPort->authn_id);\n\nI don't quite see the additional value that this API brings as\nMyProcPort is directly accessible, and contrib modules like\npostgres_fdw and sslinfo just use that to find the data of the current\nbackend. Cannot a module like pgaudit, through the elog hook, do its\nwork with what we have already in place?\n\nWhat's the use case for a given session to be able to report back\nonly its own authn through SQL?\n\nI could still see a use case for that at a more global level with\nbeentrys, but it looked like there was not much interest the last time\nI dropped this idea.\n--\nMichael","msg_date":"Thu, 24 Feb 2022 20:39:48 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, 2022-02-24 at 20:39 +0900, Michael Paquier wrote:\r\n> I don't quite see the additional value that this API brings as\r\n> MyProcPort is directly accessible, and contrib modules like\r\n> postgres_fdw and sslinfo just use that to find the data of the current\r\n> backend.\r\n\r\nRight -- I just didn't know if third-party modules were actually able\r\nto rely on the internal layout of struct Port. Is that guaranteed to\r\nremain constant for a major release line? If so, this new API is\r\nsuperfluous.\r\n\r\n> Cannot a module like pgaudit, through the elog hook, do its\r\n> work with what we have already in place?\r\n\r\nGiven the above, I would hope so. Stephen mentioned that pgaudit only\r\nhad access to the logged-in role, and when I proposed a miscadmin.h API\r\nhe said that would help. CC'ing him to see what he meant; I don't know\r\nif pgaudit has additional constraints.\r\n\r\n> What's the use case for a given session to be able to report back\r\n> only its own authn through SQL?\r\n\r\nThat's for triggers to be able to grab the current ID for\r\nlogging/auditing. Is there a better way to expose this for that use\r\ncase?\r\n\r\n> I could still see a use case for that at a more global level with\r\n> beentrys, but it looked like there was not much interest the last time\r\n> I dropped this idea.\r\n\r\nI agree that this would be useful to have in the stats. From my outside\r\nperspective, it seems like it's difficult to get strings of arbitrary\r\nlength in there; is that accurate?\r\n\r\nThanks,\r\n--Jacob\r\n","msg_date":"Thu, 24 Feb 2022 16:50:59 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn Thu, Feb 24, 2022 at 04:50:59PM +0000, Jacob Champion wrote:\n> On Thu, 2022-02-24 at 20:39 +0900, Michael Paquier wrote:\n> > I don't quite see the additional value that this API brings as\n> > MyProcPort is directly accessible, and contrib modules like\n> > postgres_fdw and sslinfo just use that to find the data of the current\n> > backend.\n> \n> Right -- I just didn't know if third-party modules were actually able\n> to rely on the internal layout of struct Port. Is that guaranteed to\n> remain constant for a major release line? If so, this new API is\n> superfluous.\n\nYes, third-party can rely on Port layout. We don't break ABI between minor\nrelease. In some occasions we can add additional fields at the end of a\nstruct, but nothing more.\n\n> > I could still see a use case for that at a more global level with\n> > beentrys, but it looked like there was not much interest the last time\n> > I dropped this idea.\n> \n> I agree that this would be useful to have in the stats. From my outside\n> perspective, it seems like it's difficult to get strings of arbitrary\n> length in there; is that accurate?\n\nYes, as it's all in shared memory. The only workaround is using something like\ntrack_activity_query_size, but it's not great.\n\n\n","msg_date":"Fri, 25 Feb 2022 01:15:38 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, 2022-02-25 at 01:15 +0800, Julien Rouhaud wrote:\r\n> On Thu, Feb 24, 2022 at 04:50:59PM +0000, Jacob Champion wrote:\r\n> > On Thu, 2022-02-24 at 20:39 +0900, Michael Paquier wrote:\r\n> > > I don't quite see the additional value that this API brings as\r\n> > > MyProcPort is directly accessible, and contrib modules like\r\n> > > postgres_fdw and sslinfo just use that to find the data of the current\r\n> > > backend.\r\n> > \r\n> > Right -- I just didn't know if third-party modules were actually able\r\n> > to rely on the internal layout of struct Port. Is that guaranteed to\r\n> > remain constant for a major release line? If so, this new API is\r\n> > superfluous.\r\n> \r\n> Yes, third-party can rely on Port layout. We don't break ABI between minor\r\n> release. In some occasions we can add additional fields at the end of a\r\n> struct, but nothing more.\r\n\r\nThat simplifies things. PFA a smaller v2; if pgaudit can't make use of\r\nlibpq-be.h for some reason, then I guess we can tailor a fix to that\r\nuse case.\r\n\r\n> > > I could still see a use case for that at a more global level with\r\n> > > beentrys, but it looked like there was not much interest the last time\r\n> > > I dropped this idea.\r\n> > \r\n> > I agree that this would be useful to have in the stats. From my outside\r\n> > perspective, it seems like it's difficult to get strings of arbitrary\r\n> > length in there; is that accurate?\r\n> \r\n> Yes, as it's all in shared memory. The only workaround is using something like\r\n> track_activity_query_size, but it's not great.\r\n\r\nYeah... I was following a similar track with the initial work last\r\nyear, but I dropped it when the cost of implementation started to grow\r\nconsiderably. At the time, though, it looked like some overhauls to the\r\nstats framework were being pursued? I should read up on that thread.\r\n\r\n--Jacob","msg_date":"Thu, 24 Feb 2022 20:44:08 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, Feb 24, 2022 at 08:44:08PM +0000, Jacob Champion wrote:\n> \n> Yeah... I was following a similar track with the initial work last\n> year, but I dropped it when the cost of implementation started to grow\n> considerably. At the time, though, it looked like some overhauls to the\n> stats framework were being pursued? I should read up on that thread.\n\nDo you mean the shared memory stats patchset? IIUC this is unrelated, as the\nbeentry stuff Michael was talking about is a different infrastructure\n(backend_status.[ch]), and I don't think there are any plans to move that to\ndynamic shared memory.\n\n\n","msg_date":"Fri, 25 Feb 2022 13:01:26 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 2022-02-25 13:01:26 +0800, Julien Rouhaud wrote:\n> On Thu, Feb 24, 2022 at 08:44:08PM +0000, Jacob Champion wrote:\n> > \n> > Yeah... I was following a similar track with the initial work last\n> > year, but I dropped it when the cost of implementation started to grow\n> > considerably. At the time, though, it looked like some overhauls to the\n> > stats framework were being pursued? I should read up on that thread.\n> \n> Do you mean the shared memory stats patchset? IIUC this is unrelated, as the\n> beentry stuff Michael was talking about is a different infrastructure\n> (backend_status.[ch]), and I don't think there are any plans to move that to\n> dynamic shared memory.\n\nUntil a year ago the backend_status.c stuff was in in pgstat.c - I just split\nthem out because of the shared memory status work - so it'd not be surprising\nfor them to mentally be thrown in one bucket.\n\nBasically the type of stats we're trying to move to dynamic shared memory is\nabout counters that should persist for a while and are accumulated across\nconnections etc. Whereas backend_status.c is more about tracking the current\nstate (what query is a backend running, what user, etc).\n\nThey're not unrelated though: backend_status.c feeds pgstat.c with some\ninformation occasionally.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 24 Feb 2022 21:18:26 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn Thu, Feb 24, 2022 at 09:18:26PM -0800, Andres Freund wrote:\n> \n> On 2022-02-25 13:01:26 +0800, Julien Rouhaud wrote:\n> > On Thu, Feb 24, 2022 at 08:44:08PM +0000, Jacob Champion wrote:\n> > > \n> > > Yeah... I was following a similar track with the initial work last\n> > > year, but I dropped it when the cost of implementation started to grow\n> > > considerably. At the time, though, it looked like some overhauls to the\n> > > stats framework were being pursued? I should read up on that thread.\n> > \n> > Do you mean the shared memory stats patchset? IIUC this is unrelated, as the\n> > beentry stuff Michael was talking about is a different infrastructure\n> > (backend_status.[ch]), and I don't think there are any plans to move that to\n> > dynamic shared memory.\n> \n> Until a year ago the backend_status.c stuff was in in pgstat.c - I just split\n> them out because of the shared memory status work - so it'd not be surprising\n> for them to mentally be thrown in one bucket.\n\nRight.\n\n> Basically the type of stats we're trying to move to dynamic shared memory is\n> about counters that should persist for a while and are accumulated across\n> connections etc. Whereas backend_status.c is more about tracking the current\n> state (what query is a backend running, what user, etc).\n\nBut would it be acceptable to use dynamic shared memory in backend_status and\ne.g. have a dsa_pointer rather than a fixed length array? That seems like a\nbad idea for query text in general, but for authn_id for instance it seems less\nlikely to hold gigantic strings, and also have more or less consistent size\nwhen provided.\n\n\n","msg_date":"Fri, 25 Feb 2022 14:25:52 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, 2022-02-24 at 20:44 +0000, Jacob Champion wrote:\r\n> That simplifies things. PFA a smaller v2; if pgaudit can't make use of\r\n> libpq-be.h for some reason, then I guess we can tailor a fix to that\r\n> use case.\r\n\r\nHa, opr_sanity caught my use of cstring. I'll work on a fix later\r\ntoday.\r\n\r\n--Jacob\r\n","msg_date":"Fri, 25 Feb 2022 16:28:39 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, 2022-02-25 at 16:28 +0000, Jacob Champion wrote:\r\n> Ha, opr_sanity caught my use of cstring. I'll work on a fix later\r\n> today.\r\n\r\nFixed in v3.\r\n\r\n--Jacob","msg_date":"Fri, 25 Feb 2022 20:19:24 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On 2022-02-25 20:19:24 +0000, Jacob Champion wrote:\n> From 2fde60a6bc3739f1894c8c264120e4fa0f04df64 Mon Sep 17 00:00:00 2001\n> From: Jacob Champion <pchampion@vmware.com>\n> Date: Mon, 14 Feb 2022 08:10:53 -0800\n> Subject: [PATCH v3] Add API to retrieve authn_id from SQL\n\n> The authn_id field in MyProcPort is currently only accessible to the\n> backend itself. Add a SQL function, session_authn_id(), to expose the\n> field to triggers that may want to make use of it.\n\nLooks to me like authn_id isn't synchronized to parallel workers right now. So\nthe function will return the wrong thing when executed as part of a parallel\nquery.\n\nI don't think we should add further functions not prefixed with pg_.\n\n\nPerhaps a few tests for less trivial authn_ids could be worthwhile?\nE.g. certificate DNs.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 25 Feb 2022 13:23:49 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 2022-02-25 14:25:52 +0800, Julien Rouhaud wrote:\n> > Basically the type of stats we're trying to move to dynamic shared memory is\n> > about counters that should persist for a while and are accumulated across\n> > connections etc. Whereas backend_status.c is more about tracking the current\n> > state (what query is a backend running, what user, etc).\n> \n> But would it be acceptable to use dynamic shared memory in backend_status and\n> e.g. have a dsa_pointer rather than a fixed length array?\n\nMight be OK, but it does add a fair bit of complexity. Suddenly there's a\nbunch more initialization order dependencies that you don't have right\nnow. I'd not go there for just this.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 25 Feb 2022 13:26:49 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\n> Looks to me like authn_id isn't synchronized to parallel workers right now. So\n> the function will return the wrong thing when executed as part of a parallel\n> query.\n\nFWIW, I am not completely sure what's the use case for being able to\nsee the authn of the current session through a trigger. We expose\nthat when log_connections is enabled, for audit purposes. I can also\nget behind something more central so as one can get a full picture of\nthe authn used by a bunch of session, particularly with complex HBA\npolicies, but this looks rather limited to me in usability. Perhaps\nthat's not enough to stand as an objection, though, and the patch is\ndead simple.\n\n> I don't think we should add further functions not prefixed with pg_.\n\nYep.\n\n> Perhaps a few tests for less trivial authn_ids could be worthwhile?\n> E.g. certificate DNs.\n\nYes, src/test/ssl would handle that just fine. Now, this stuff\nalready looks after authn results with log_connections=on, so that\nfeels like a duplicate.\n--\nMichael","msg_date":"Sat, 26 Feb 2022 14:39:17 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\r\n> Looks to me like authn_id isn't synchronized to parallel workers right now. So\r\n> the function will return the wrong thing when executed as part of a parallel\r\n> query.\r\n\r\nThanks for the catch. It looks like MyProcPort is left empty, and other\r\nfunctions that rely on like inet_server_addr() are marked parallel-\r\nrestricted, so I've done the same in v4.\r\n\r\nOn Sat, 2022-02-26 at 14:39 +0900, Michael Paquier wrote:\r\n> FWIW, I am not completely sure what's the use case for being able to\r\n> see the authn of the current session through a trigger. We expose\r\n> that when log_connections is enabled, for audit purposes. I can also\r\n> get behind something more central so as one can get a full picture of\r\n> the authn used by a bunch of session, particularly with complex HBA\r\n> policies, but this looks rather limited to me in usability. Perhaps\r\n> that's not enough to stand as an objection, though, and the patch is\r\n> dead simple.\r\n\r\nI'm primarily motivated by the linked thread -- if the gap between\r\nbuiltin roles and authn_id are going to be used as ammo against other\r\nsecurity features, then let's close that gap. But I think it's fair to\r\nsay that if someone is already using triggers to exhaustively audit a\r\ntable, it'd be nice to have this info in the same place too.\r\n\r\n> > I don't think we should add further functions not prefixed with pg_.\r\n> \r\n> Yep.\r\n\r\nFixed.\r\n\r\n> > Perhaps a few tests for less trivial authn_ids could be worthwhile?\r\n> > E.g. certificate DNs.\r\n> \r\n> Yes, src/test/ssl would handle that just fine. Now, this stuff\r\n> already looks after authn results with log_connections=on, so that\r\n> feels like a duplicate.\r\n\r\nIt was easy enough to add, so I added it. I suppose it does protect\r\nagainst any reimplementations of pg_session_authn_id() that can't\r\nhandle longer ID strings, though I admit that's a stretch.\r\n\r\nThanks,\r\n--Jacob","msg_date":"Mon, 28 Feb 2022 18:58:00 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Greetings,\n\n* Jacob Champion (pchampion@vmware.com) wrote:\n> On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\n> > Looks to me like authn_id isn't synchronized to parallel workers right now. So\n> > the function will return the wrong thing when executed as part of a parallel\n> > query.\n> \n> Thanks for the catch. It looks like MyProcPort is left empty, and other\n> functions that rely on like inet_server_addr() are marked parallel-\n> restricted, so I've done the same in v4.\n\nThat's probably alright.\n\n> On Sat, 2022-02-26 at 14:39 +0900, Michael Paquier wrote:\n> > FWIW, I am not completely sure what's the use case for being able to\n> > see the authn of the current session through a trigger. We expose\n> > that when log_connections is enabled, for audit purposes. I can also\n> > get behind something more central so as one can get a full picture of\n> > the authn used by a bunch of session, particularly with complex HBA\n> > policies, but this looks rather limited to me in usability. Perhaps\n> > that's not enough to stand as an objection, though, and the patch is\n> > dead simple.\n> \n> I'm primarily motivated by the linked thread -- if the gap between\n> builtin roles and authn_id are going to be used as ammo against other\n> security features, then let's close that gap. But I think it's fair to\n> say that if someone is already using triggers to exhaustively audit a\n> table, it'd be nice to have this info in the same place too.\n\nYeah, we really should make this available to trigger-based auditing\nsystems too and not just through log_connections which involves a great\ndeal more log parsing and work external to the database to figure out\nwho did what.\n\n> > > I don't think we should add further functions not prefixed with pg_.\n> > \n> > Yep.\n> \n> Fixed.\n\nThat's fine.\n\n> > > Perhaps a few tests for less trivial authn_ids could be worthwhile?\n> > > E.g. certificate DNs.\n> > \n> > Yes, src/test/ssl would handle that just fine. Now, this stuff\n> > already looks after authn results with log_connections=on, so that\n> > feels like a duplicate.\n> \n> It was easy enough to add, so I added it. I suppose it does protect\n> against any reimplementations of pg_session_authn_id() that can't\n> handle longer ID strings, though I admit that's a stretch.\n> \n> Thanks,\n> --Jacob\n\n> commit efec9f040843d1de2fc52f5ce0d020478a5bc75d\n> Author: Jacob Champion <pchampion@vmware.com>\n> Date: Mon Feb 28 10:28:51 2022 -0800\n> \n> squash! Add API to retrieve authn_id from SQL\n\nBleh. :) Squash indeed.\n\n> Subject: [PATCH v4] Add API to retrieve authn_id from SQL\n> \n> The authn_id field in MyProcPort is currently only accessible to the\n> backend itself. Add a SQL function, pg_session_authn_id(), to expose\n> the field to triggers that may want to make use of it.\n\nOnly did a quick look but generally looks reasonable to me.\n\nThanks,\n\nStephen","msg_date":"Mon, 28 Feb 2022 16:00:36 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Mon, 2022-02-28 at 16:00 -0500, Stephen Frost wrote:\r\n> > commit efec9f040843d1de2fc52f5ce0d020478a5bc75d\r\n> > Author: Jacob Champion <pchampion@vmware.com>\r\n> > Date: Mon Feb 28 10:28:51 2022 -0800\r\n> > \r\n> > squash! Add API to retrieve authn_id from SQL\r\n> \r\n> Bleh. :) Squash indeed.\r\n\r\nHa, I wasn't sure if anyone read the since-diffs :) I'll start\r\nwordsmithing them more in the future.\r\n\r\n> > Subject: [PATCH v4] Add API to retrieve authn_id from SQL\r\n> > \r\n> > The authn_id field in MyProcPort is currently only accessible to the\r\n> > backend itself. Add a SQL function, pg_session_authn_id(), to expose\r\n> > the field to triggers that may want to make use of it.\r\n> \r\n> Only did a quick look but generally looks reasonable to me.\r\n\r\nThanks!\r\n\r\n--Jacob\r\n","msg_date":"Mon, 28 Feb 2022 21:42:54 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Mon, Feb 28, 2022 at 04:00:36PM -0500, Stephen Frost wrote:\n> * Jacob Champion (pchampion@vmware.com) wrote:\n>> On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\n>>> Looks to me like authn_id isn't synchronized to parallel workers right now. So\n>>> the function will return the wrong thing when executed as part of a parallel\n>>> query.\n>> \n>> Thanks for the catch. It looks like MyProcPort is left empty, and other\n>> functions that rely on like inet_server_addr() are marked parallel-\n>> restricted, so I've done the same in v4.\n> \n> That's probably alright.\n\nI'd say as well that this is right as-is. If it happens that there is\na use-case for making this parallel aware in the future, it could be\ndone. Now, it may be a bit weird to make parallel workers inherit the\nauthn ID of the parent as these did not go through authentication, no?\nLetting this function being run only by the leader feels intuitive.\n\n> Yeah, we really should make this available to trigger-based auditing\n> systems too and not just through log_connections which involves a great\n> deal more log parsing and work external to the database to figure out\n> who did what.\n\nOkay, I won't fight hard on that if all of you think that this is\nuseful for a given session.\n\n>> Subject: [PATCH v4] Add API to retrieve authn_id from SQL\n>> \n>> The authn_id field in MyProcPort is currently only accessible to the\n>> backend itself. Add a SQL function, pg_session_authn_id(), to expose\n>> the field to triggers that may want to make use of it.\n> \n> Only did a quick look but generally looks reasonable to me.\n\nThe function and the test are fine, pgperltidy complains a bit about\nthe format of the tests. \n\nAyway, this function needs to be documented. I think that you should\njust add that in \"Session Information Functions\" in func.sgml, same\narea as current_user(). The last time we talked about the authn ID,\none thing we discussed about was how to describe that in a good way to\nthe user, which is why the section of log_connections was reworked a\nbit. And we don't have yet any references to what an authenticated\nidentity is in the docs.\n\nThere is no need to update catversion.h in the patch, committers\nusually take care of that and that's an area of the code that\nconflicts a lot.\n--\nMichael","msg_date":"Tue, 1 Mar 2022 15:53:02 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Greetings,\n\n* Michael Paquier (michael@paquier.xyz) wrote:\n> On Mon, Feb 28, 2022 at 04:00:36PM -0500, Stephen Frost wrote:\n> > * Jacob Champion (pchampion@vmware.com) wrote:\n> >> On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\n> >>> Looks to me like authn_id isn't synchronized to parallel workers right now. So\n> >>> the function will return the wrong thing when executed as part of a parallel\n> >>> query.\n> >> \n> >> Thanks for the catch. It looks like MyProcPort is left empty, and other\n> >> functions that rely on like inet_server_addr() are marked parallel-\n> >> restricted, so I've done the same in v4.\n> > \n> > That's probably alright.\n> \n> I'd say as well that this is right as-is. If it happens that there is\n> a use-case for making this parallel aware in the future, it could be\n> done. Now, it may be a bit weird to make parallel workers inherit the\n> authn ID of the parent as these did not go through authentication, no?\n> Letting this function being run only by the leader feels intuitive.\n\nI'm not really sure why we're arguing about this, but clearly the authn\nID of the leader process is what should be used because that's the\nauthentication under which the parallel worker is running, just as much\nas the effective role is the authorization. Having this be available in\nworker processes would certainly be good as it would allow more query\nplans to be considered when these functions are used. At this time, I\ndon't think that outweighs the complications around having it and I'm\nnot suggesting that Jacob needs to go do that, but surely it would be\nbetter.\n\n> >> Subject: [PATCH v4] Add API to retrieve authn_id from SQL\n> >> \n> >> The authn_id field in MyProcPort is currently only accessible to the\n> >> backend itself. Add a SQL function, pg_session_authn_id(), to expose\n> >> the field to triggers that may want to make use of it.\n> > \n> > Only did a quick look but generally looks reasonable to me.\n> \n> The function and the test are fine, pgperltidy complains a bit about\n> the format of the tests. \n> \n> Ayway, this function needs to be documented. I think that you should\n> just add that in \"Session Information Functions\" in func.sgml, same\n> area as current_user(). The last time we talked about the authn ID,\n> one thing we discussed about was how to describe that in a good way to\n> the user, which is why the section of log_connections was reworked a\n> bit. And we don't have yet any references to what an authenticated\n> identity is in the docs.\n\nAgreed that it should be documented and that location seems reasonable\nto me.\n\n> There is no need to update catversion.h in the patch, committers\n> usually take care of that and that's an area of the code that\n> conflicts a lot.\n\nYeah, best to let committers handle catversion bumps.\n\nThanks,\n\nStephen","msg_date":"Tue, 1 Mar 2022 08:35:27 -0500","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On 25.02.22 21:19, Jacob Champion wrote:\n> On Fri, 2022-02-25 at 16:28 +0000, Jacob Champion wrote:\n>> Ha, opr_sanity caught my use of cstring. I'll work on a fix later\n>> today.\n> \n> Fixed in v3.\n\nThis patch contains no documentation. I'm having a hard time \nunderstanding what the name \"session_authn_id\" is supposed to convey. \nThe comment for the Port.authn_id field says this is the \"system \nusername\", which sounds like a clearer terminology.\n\n\n","msg_date":"Tue, 1 Mar 2022 19:56:45 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, 2022-03-01 at 08:35 -0500, Stephen Frost wrote:\r\n> * Michael Paquier (michael@paquier.xyz) wrote:\r\n> > \r\n> > Ayway, this function needs to be documented. I think that you should\r\n> > just add that in \"Session Information Functions\" in func.sgml, same\r\n> > area as current_user(). The last time we talked about the authn ID,\r\n> > one thing we discussed about was how to describe that in a good way to\r\n> > the user, which is why the section of log_connections was reworked a\r\n> > bit. And we don't have yet any references to what an authenticated\r\n> > identity is in the docs.\r\n> \r\n> Agreed that it should be documented and that location seems reasonable\r\n> to me.\r\n\r\nAdded a first draft in v5, alongside the perltidy fixups mentioned by\r\nMichael.\r\n\r\n> > There is no need to update catversion.h in the patch, committers\r\n> > usually take care of that and that's an area of the code that\r\n> > conflicts a lot.\r\n> \r\n> Yeah, best to let committers handle catversion bumps.\r\n\r\nHeh, that was added for my benefit -- I was tired of forgetting to\r\ninitdb after switching dev branches -- but I've dropped it from the\r\npatch and will just carry that diff locally.\r\n\r\nThanks,\r\n--Jacob","msg_date":"Tue, 1 Mar 2022 22:03:20 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, 2022-03-01 at 19:56 +0100, Peter Eisentraut wrote:\r\n> This patch contains no documentation. I'm having a hard time \r\n> understanding what the name \"session_authn_id\" is supposed to convey. \r\n> The comment for the Port.authn_id field says this is the \"system \r\n> username\", which sounds like a clearer terminology.\r\n\r\n\"System username\" may help from an internal development perspective,\r\nespecially as it relates to pg_ident.conf, but I don't think that's\r\nlikely to be a useful descriptor to an end user. (I don't think of a\r\nclient certificate's Subject Distinguished Name as a \"system\r\nusername\".) Does my attempt in v5 help?\r\n\r\n--Jacob\r\n","msg_date":"Tue, 1 Mar 2022 22:05:37 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, Mar 01, 2022 at 10:03:20PM +0000, Jacob Champion wrote:\n> Added a first draft in v5, alongside the perltidy fixups mentioned by\n> Michael.\n\n+ The authenticated identity is an immutable identifier for the user\n+ presented during the connection handshake; the exact format depends on\n+ the authentication method in use. (For example, when using the\n+ <literal>scram-sha-256</literal> auth method, the authenticated identity\n+ is simply the username. When using the <literal>cert</literal> auth\n+ method, the authenticated identity is the Distinguished Name of the\n+ client certificate.) Even for auth methods which use the username as\n+ the authenticated identity, this function differs from\n+ <literal>session_user</literal> in that its return value cannot be\n+ changed after login.\n\nThat looks enough seen from here. Thanks!\n\nNit: \"auth method\" would be a first in the documentation, so this had\nbetter be \"authentication method\". (No need to send an updated patch\njust for that).\n\nSo, any comments and/or opinions from others?\n--\nMichael","msg_date":"Wed, 2 Mar 2022 16:09:48 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On 01.03.22 23:05, Jacob Champion wrote:\n> On Tue, 2022-03-01 at 19:56 +0100, Peter Eisentraut wrote:\n>> This patch contains no documentation. I'm having a hard time\n>> understanding what the name \"session_authn_id\" is supposed to convey.\n>> The comment for the Port.authn_id field says this is the \"system\n>> username\", which sounds like a clearer terminology.\n> \n> \"System username\" may help from an internal development perspective,\n> especially as it relates to pg_ident.conf, but I don't think that's\n> likely to be a useful descriptor to an end user. (I don't think of a\n> client certificate's Subject Distinguished Name as a \"system\n> username\".) Does my attempt in v5 help?\n\nYeah, maybe there are better names. But I have no idea what the letter \ncombination \"authn_id\" is supposed to stand for. Is it an \n\"authentication identifier\"? What does it identify? Maybe I'm missing \nsomething here, but I don't find it clear.\n\n\n","msg_date":"Wed, 2 Mar 2022 09:18:17 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 2022-03-01 08:35:27 -0500, Stephen Frost wrote:\n> I'm not really sure why we're arguing about this, but clearly the authn\n> ID of the leader process is what should be used because that's the\n> authentication under which the parallel worker is running, just as much\n> as the effective role is the authorization. Having this be available in\n> worker processes would certainly be good as it would allow more query\n> plans to be considered when these functions are used. At this time, I\n> don't think that outweighs the complications around having it and I'm\n> not suggesting that Jacob needs to go do that, but surely it would be\n> better.\n\nI don't think we should commit this without synchronizing the authn between\nworker / leader (in a separate commit). Too likely that some function that's\nmarked parallel ok queries the authn_id, opening up a security/monitoring hole\nor such because of a bogus return value.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 2 Mar 2022 13:27:40 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, Mar 02, 2022 at 01:27:40PM -0800, Andres Freund wrote:\n> I don't think we should commit this without synchronizing the authn between\n> worker / leader (in a separate commit). Too likely that some function that's\n> marked parallel ok queries the authn_id, opening up a security/monitoring hole\n> or such because of a bogus return value.\n\nHmm, OK. Using the same authn ID for the leader and the workers still\nlooks a bit strange to me as the worker is not the one that does the \nauthentication, only the leader does that. Anyway, FixedParallelState\nincludes some authentication data passed down by the leader when\nspawning a worker. So, if we were to pass down the authn, we are\ngoing to need a new PARALLEL_KEY_* to serialize and restore the data\npassed down via a DSM like any other states as per the business in\nparallel.c. Jacob, what do you think?\n--\nMichael","msg_date":"Thu, 3 Mar 2022 16:45:33 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, 2022-03-02 at 09:18 +0100, Peter Eisentraut wrote:\r\n> On 01.03.22 23:05, Jacob Champion wrote:\r\n> > On Tue, 2022-03-01 at 19:56 +0100, Peter Eisentraut wrote:\r\n> > > This patch contains no documentation. I'm having a hard time\r\n> > > understanding what the name \"session_authn_id\" is supposed to convey.\r\n> > > The comment for the Port.authn_id field says this is the \"system\r\n> > > username\", which sounds like a clearer terminology.\r\n> > \r\n> > \"System username\" may help from an internal development perspective,\r\n> > especially as it relates to pg_ident.conf, but I don't think that's\r\n> > likely to be a useful descriptor to an end user. (I don't think of a\r\n> > client certificate's Subject Distinguished Name as a \"system\r\n> > username\".) Does my attempt in v5 help?\r\n> \r\n> Yeah, maybe there are better names. But I have no idea what the letter \r\n> combination \"authn_id\" is supposed to stand for. Is it an \r\n> \"authentication identifier\"? What does it identify?\r\n\r\nAuthenticated identity, but yeah, that's the gist. (\"AuthN\" being a\r\nstandard-ish way to differentiate authentication from \"AuthZ\"\r\nauthorization.)\r\n\r\nIt's meant to uniquely identify the end user in the case of usermaps,\r\nwhere multiple separate entities might log in using the same role. It\r\nis distinct from the authorized role name, though they might be exactly\r\nthe same in many common setups. And it's not set at all if no\r\nauthentication was done.\r\n\r\n> Maybe I'm missing something here, but I don't find it clear.\r\n\r\nI just used the internal name, but if we want to make it more clear\r\nthen now would be a good time. Do you have any suggestions? Does\r\nexpanding the name (pg_session_authenticated_id, or even\r\npg_session_authenticated_identity) help?\r\n\r\n--Jacob\r\n","msg_date":"Thu, 3 Mar 2022 19:12:11 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, 2022-03-03 at 16:45 +0900, Michael Paquier wrote:\r\n> Anyway, FixedParallelState\r\n> includes some authentication data passed down by the leader when\r\n> spawning a worker. So, if we were to pass down the authn, we are\r\n> going to need a new PARALLEL_KEY_* to serialize and restore the data\r\n> passed down via a DSM like any other states as per the business in\r\n> parallel.c. Jacob, what do you think?\r\n\r\nI guess it depends on what we want MyProcPort to look like in a\r\nparallel worker. Are we comfortable having most of it be blank/useless?\r\nDoes it need to be filled in?\r\n\r\n--Jacob\r\n","msg_date":"Thu, 3 Mar 2022 19:16:17 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, Mar 03, 2022 at 07:16:17PM +0000, Jacob Champion wrote:\n> I guess it depends on what we want MyProcPort to look like in a\n> parallel worker. Are we comfortable having most of it be blank/useless?\n> Does it need to be filled in?\n\nGood question. It depends on the definition of how much\nauthentication information makes sense for the parallel workers to\ninherit from the parent. And as I mentioned upthread, this definition\nis not completely clear to me because the parallel workers do *not* go\nthrough the authentication paths of the parent, they are just spawned\nin their own dedicated paths that the leader invokes. Inheriting all\nthis information from the leader has also an impact on the\nPgBackendStatus entries of the workers as these are reported in\npg_stat_activity as far as I recall, and it could be confusing to see,\nfor example, some SSL or some GSS information for automatically\nspawned processes because these don't use SSL or GSS when they pass\nback data to the leader.\n\nI have been looking at the commit history, and found about 6b7d11f\nthat switched all the functions of sslinfo to be parallel-restricted\nespecially because of this. So if we inherit all this information the\nrestriction on the sslinfo functions could be lifted, though the\ninterest is honestly limited in this case.\n\npostgres_fdw has introduced recently the concept of cached\nconnections, as of v14 with 411ae64 and 708d165, with a set of\nparallel-restricted functions. Some of the code paths related to\nappname look at MyProcPort, so there could be a risk of having some\ninconsistent information if this is accessed in a parallel worker.\nLooking at the code, I don't think that it would happen now but\ncopying some of the data of MyProcPort to the worker could avoid any\nfuture issues if this code gets extended.\n\nAt the end of the day, Port is an interface used for the communication\nbetween the postmaster with the frontends, so I'd like to say that it\nis correct to not apply this concept to parallel workers because they\nare not designed to contact any frontend-side things.\n--\nMichael","msg_date":"Fri, 4 Mar 2022 10:45:24 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, 2022-03-04 at 10:45 +0900, Michael Paquier wrote:\r\n> At the end of the day, Port is an interface used for the communication\r\n> between the postmaster with the frontends, so I'd like to say that it\r\n> is correct to not apply this concept to parallel workers because they\r\n> are not designed to contact any frontend-side things.\r\n\r\nComing back to this late, sorry. I'm not quite sure where to move with\r\nthis. I'm considering copying pieces of Port over just so we can see\r\nwhat it looks like in practice?\r\n\r\nPersonally I think it makes sense for the parallel workers to have the\r\nauthn information for the client -- in fact there's a lot of\r\ninformation that it seems shouldn't be hidden from them -- but there\r\nare other pieces, like the socket handle, that are clearly not useful.\r\n\r\nThanks,\r\n--Jacob\r\n","msg_date":"Thu, 17 Mar 2022 22:02:16 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Jacob Champion <pchampion@vmware.com> writes:\n> On Fri, 2022-03-04 at 10:45 +0900, Michael Paquier wrote:\n>> At the end of the day, Port is an interface used for the communication\n>> between the postmaster with the frontends, so I'd like to say that it\n>> is correct to not apply this concept to parallel workers because they\n>> are not designed to contact any frontend-side things.\n\n> Coming back to this late, sorry. I'm not quite sure where to move with\n> this. I'm considering copying pieces of Port over just so we can see\n> what it looks like in practice?\n\n> Personally I think it makes sense for the parallel workers to have the\n> authn information for the client -- in fact there's a lot of\n> information that it seems shouldn't be hidden from them -- but there\n> are other pieces, like the socket handle, that are clearly not useful.\n\nYeah. It seems to me that putting the auth info into struct Port was\na fairly random thing to do in the first place, and we are now dealing\nwith the fallout of that.\n\nI think what we ought to do here is separate out the data that we think\nparallel workers need access to. It does not seem wise to say \"workers\ncan access fields A,B,C of MyPort but not fields X,Y,Z\". I do not have\na concrete proposal for fixing it though.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 17 Mar 2022 18:33:38 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, 2022-03-17 at 18:33 -0400, Tom Lane wrote:\r\n> Yeah. It seems to me that putting the auth info into struct Port was\r\n> a fairly random thing to do in the first place, and we are now dealing\r\n> with the fallout of that.\r\n> \r\n> I think what we ought to do here is separate out the data that we think\r\n> parallel workers need access to. It does not seem wise to say \"workers\r\n> can access fields A,B,C of MyPort but not fields X,Y,Z\". I do not have\r\n> a concrete proposal for fixing it though.\r\n\r\nv6-0002 has my first attempt at this. I moved authn_id into its own\r\nsubstruct inside Port, which gets serialized with the parallel key\r\nmachinery. (My name selection of \"SharedPort\" is pretty bland.)\r\n\r\nOver time, we could move more fields into the shared struct and fill\r\nout the serialization logic as needed, and then maybe eventually\r\nSharedPort can be broken off into its own thing with its own\r\nallocation. But I don't know if we should do it all at once, yet.\r\n\r\nWDYT?\r\n\r\n--Jacob","msg_date":"Wed, 23 Mar 2022 22:52:22 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Jacob Champion <pchampion@vmware.com> writes:\n> On Thu, 2022-03-17 at 18:33 -0400, Tom Lane wrote:\n>> I think what we ought to do here is separate out the data that we think\n>> parallel workers need access to. It does not seem wise to say \"workers\n>> can access fields A,B,C of MyPort but not fields X,Y,Z\". I do not have\n>> a concrete proposal for fixing it though.\n\n> v6-0002 has my first attempt at this. I moved authn_id into its own\n> substruct inside Port, which gets serialized with the parallel key\n> machinery. (My name selection of \"SharedPort\" is pretty bland.)\n\nHm. I was more envisioning getting the \"sharable\" info out of Port\nentirely, although I'm not quite sure where it should go instead.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 23 Mar 2022 19:00:28 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, 2022-03-23 at 19:00 -0400, Tom Lane wrote:\r\n> Hm. I was more envisioning getting the \"sharable\" info out of Port\r\n> entirely, although I'm not quite sure where it should go instead.\r\n\r\nIf it helps, I can move the substruct out and up to a new global struct\r\n(MyProcShared?). At this point I think it's mostly search-and-replace.\r\n\r\n--Jacob\r\n","msg_date":"Wed, 23 Mar 2022 23:06:14 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 2022-03-23 23:06:14 +0000, Jacob Champion wrote:\n> On Wed, 2022-03-23 at 19:00 -0400, Tom Lane wrote:\n> > Hm. I was more envisioning getting the \"sharable\" info out of Port\n> > entirely, although I'm not quite sure where it should go instead.\n> \n> If it helps, I can move the substruct out and up to a new global struct\n> (MyProcShared?). At this point I think it's mostly search-and-replace.\n\nPerhaps alongside CurrentUserId etc in miscinit.c? It would be nicer if all\nthose were together in a struct, but oh well.\n\nAnother option would be to make it a GUC. With a bit of care it could be\nautomatically synced by the existing parallelism infrastructure...\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 23 Mar 2022 16:54:36 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, 2022-03-23 at 16:54 -0700, Andres Freund wrote:\r\n> On 2022-03-23 23:06:14 +0000, Jacob Champion wrote:\r\n> > On Wed, 2022-03-23 at 19:00 -0400, Tom Lane wrote:\r\n> > > Hm. I was more envisioning getting the \"sharable\" info out of Port\r\n> > > entirely, although I'm not quite sure where it should go instead.\r\n> > \r\n> > If it helps, I can move the substruct out and up to a new global struct\r\n> > (MyProcShared?). At this point I think it's mostly search-and-replace.\r\n> \r\n> Perhaps alongside CurrentUserId etc in miscinit.c? It would be nicer if all\r\n> those were together in a struct, but oh well.\r\n\r\nNext draft in v7. My naming choices probably make even less sense now. Any ideas for names for \"a bag of stuff that we want parallel workers to have too\"?\r\n\r\n> Another option would be to make it a GUC. With a bit of care it could be\r\n> automatically synced by the existing parallelism infrastructure...\r\n\r\nLike a write-once, PGC_INTERNAL setting? I guess I don't have any\r\nintuition on how that would compare to the separate-global-and-accessor \r\napproach. Is the primary advantage that you don't have to maintain the\r\nserialization logic, or is there more to it?\r\n\r\nThanks,\r\n--Jacob","msg_date":"Thu, 24 Mar 2022 17:44:06 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, Mar 2, 2022 at 4:27 PM Andres Freund <andres@anarazel.de> wrote:\n> I don't think we should commit this without synchronizing the authn between\n> worker / leader (in a separate commit). Too likely that some function that's\n> marked parallel ok queries the authn_id, opening up a security/monitoring hole\n> or such because of a bogus return value.\n\nIt is not free to copy data from the leader to the worker. I don't\nthink we should just adopt a policy of copying everything anyone\nthinks of, because then most of the time we'll be copying a bunch of\nstuff that really isn't needed.\n\nMy gut reaction is to think that this is way too marginal to be worth\nmaking parallel-safe, but it is also possible that I just don't know\nenough to understand its true value.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 24 Mar 2022 13:55:29 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 2022-03-24 13:55:29 -0400, Robert Haas wrote:\n> On Wed, Mar 2, 2022 at 4:27 PM Andres Freund <andres@anarazel.de> wrote:\n> > I don't think we should commit this without synchronizing the authn between\n> > worker / leader (in a separate commit). Too likely that some function that's\n> > marked parallel ok queries the authn_id, opening up a security/monitoring hole\n> > or such because of a bogus return value.\n>\n> It is not free to copy data from the leader to the worker. I don't\n> think we should just adopt a policy of copying everything anyone\n> thinks of, because then most of the time we'll be copying a bunch of\n> stuff that really isn't needed.\n\nI agree.\n\n\n> My gut reaction is to think that this is way too marginal to be worth\n> making parallel-safe, but it is also possible that I just don't know\n> enough to understand its true value.\n\nMy problem with that is that as far as I can tell the only real use of the\nfield / function is for stuff like audit logging, RLS etc. Which seems\nproblematic for two reasons:\n\n1) It's likely that the call to the function is nested into other functions,\n \"hiding\" the parallel safety. Then it'd return bogus data silently. At\n the very least we need to make it error out if called in a parallel worker.\n2) If used for the purposes above, there's basically no parallelism possible\n anymore.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 24 Mar 2022 11:19:35 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, Mar 24, 2022 at 05:44:06PM +0000, Jacob Champion wrote:\n> On Wed, 2022-03-23 at 16:54 -0700, Andres Freund wrote:\n>> Another option would be to make it a GUC. With a bit of care it could be\n>> automatically synced by the existing parallelism infrastructure...\n> \n> Like a write-once, PGC_INTERNAL setting? I guess I don't have any\n> intuition on how that would compare to the separate-global-and-accessor \n> approach. Is the primary advantage that you don't have to maintain the\n> serialization logic, or is there more to it?\n\nHmm. That would be a first for a GUC, no? It is not seem natural\ncompared to the other information pieces passed down from the leader\nto the workers.\n\n+extern SharedPort MyProcShared;\n\nThis naming is interesting, and seems to be in line with a couple of\nexecutor structures that share information across workers. Still\nthat's a bit inconsistent as Shared is used once at the beginning and\nonce at the end? I don't have a better idea on top of my mind.\n\nAnyway, wouldn't it be better to reverse the patch order, introducing\nthe shared Proc information first and then build the parallel-safe\nfunction on top of it? \n--\nMichael","msg_date":"Sat, 26 Mar 2022 15:18:59 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 2022-03-26 15:18:59 +0900, Michael Paquier wrote:\n> On Thu, Mar 24, 2022 at 05:44:06PM +0000, Jacob Champion wrote:\n> > On Wed, 2022-03-23 at 16:54 -0700, Andres Freund wrote:\n> >> Another option would be to make it a GUC. With a bit of care it could be\n> >> automatically synced by the existing parallelism infrastructure...\n> >\n> > Like a write-once, PGC_INTERNAL setting?\n\nPerhaps PGC_INTERNAL, perhaps PGC_SU_BACKEND, set with PGC_S_OVERRIDE?\n\n\n> > I guess I don't have any\n> > intuition on how that would compare to the separate-global-and-accessor\n> > approach. Is the primary advantage that you don't have to maintain the\n> > serialization logic, or is there more to it?\n>\n> Hmm. That would be a first for a GUC, no? It is not seem natural\n> compared to the other information pieces passed down from the leader\n> to the workers.\n\nWhat would be the first for a GUC? We have plenty GUCs that are set on a\nper-connection basis to reflect some fact? And there's several authenitcation\nrelated bits of state known to guc.c , think role, session_authorization,\nis_superuser.\n\nSharing per-connection state via GUCs for paralellism? I don't think that is\ntrue either. E.g. application_name, client_encoding.\n\n\n> +extern SharedPort MyProcShared;\n\nI strongly dislike MyProcShared. It's way too easily confused with MyProc\nwhich point to shared memory.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sat, 26 Mar 2022 10:15:50 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2022-03-26 15:18:59 +0900, Michael Paquier wrote:\n>> On Thu, Mar 24, 2022 at 05:44:06PM +0000, Jacob Champion wrote:\n>>> On Wed, 2022-03-23 at 16:54 -0700, Andres Freund wrote:\n>>>> Another option would be to make it a GUC. With a bit of care it could be\n>>>> automatically synced by the existing parallelism infrastructure...\n\n>>> Like a write-once, PGC_INTERNAL setting?\n\n> Perhaps PGC_INTERNAL, perhaps PGC_SU_BACKEND, set with PGC_S_OVERRIDE?\n\nSeems like making it a GUC does nothing to fix the complaint you had about\npassing data to workers whether it's needed or not. Sure, we don't then\nneed to write any new code for it, but it's still new cycles. And it's\nnew cycles all throughout guc.c, too, not just in parallel worker start.\n\nI also note that exposing it as a GUC means we have zero control over\nwho/what can read it. Maybe that's not a problem, but it needs to be\nthought about before we go down that path.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 26 Mar 2022 13:57:39 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 2022-03-26 13:57:39 -0400, Tom Lane wrote:\n> Seems like making it a GUC does nothing to fix the complaint you had about\n> passing data to workers whether it's needed or not.\n\nI don't think that was my complaint. Maybe Robert's?\n\n\n> Sure, we don't then need to write any new code for it, but it's still new\n> cycles.\n\nI think it'd quite possibly less cycles than separately syncing it.\n\nBecause I wanted to know what the overhead be in relation to other things, I\nmade serialize_variable() log whenever it decides to serialize, and it's a bit\ndepressing :/.\n\nserialized DateStyle = ISO, MDY\nserialized default_text_search_config = pg_catalog.english\nserialized force_parallel_mode = on\nserialized lc_messages = en_US.UTF-8\nserialized lc_monetary = en_US.UTF-8\nserialized lc_numeric = en_US.UTF-8\nserialized lc_time = en_US.UTF-8\nserialized log_checkpoints = true\nserialized log_line_prefix = %m [%p][%b][%v:%x][%a]\nserialized log_timezone = America/Los_Angeles\nserialized max_stack_depth = 2048\nserialized max_wal_size = 153600\nserialized min_wal_size = 48\nserialized restart_after_crash = true\nserialized session_authorization = andres\nserialized ssl_cert_file = /home/andres/tmp/pgdev/ssl-cert-snakeoil.pem\nserialized ssl_key_file = /home/andres/tmp/pgdev/ssl-cert-snakeoil.key\nserialized TimeZone = America/Los_Angeles\nserialized timezone_abbreviations = Default\nserialized track_io_timing = true\nserialized transaction_deferrable = false\nserialized transaction_isolation = read committed\nserialized transaction_read_only = false\ntotal serialized guc state is 1324\n\nOf course, compared to the total size of 94784 bytes that's not too\nmuch... FWIW, 65536 of that is for the tuple queues...\n\n\n> I also note that exposing it as a GUC means we have zero control over\n> who/what can read it. Maybe that's not a problem, but it needs to be\n> thought about before we go down that path.\n\nYes, I think that's a fair concern.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sat, 26 Mar 2022 11:36:37 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Sat, 2022-03-26 at 11:36 -0700, Andres Freund wrote:\r\n> > I also note that exposing it as a GUC means we have zero control over\r\n> > who/what can read it. Maybe that's not a problem, but it needs to be\r\n> > thought about before we go down that path.\r\n> \r\n> Yes, I think that's a fair concern.\r\n\r\nI like that there's no builtin way, today, for a superuser to modify\r\nthe internal value; it strengthens the use as an auditing tool. Moving\r\nthis to a PGC_SU_BACKEND GUC seems to weaken that. And it looks like\r\nPGC_INTERNAL is skipped during the serialization, so if we chose that\r\noption, we'd need to write new code anyway?\r\n\r\nWe'd also need to guess whether the GUC system's serialization of NULL\r\nas an empty string is likely to cause problems for any future auth\r\nmethods. My guess is \"no\", to be honest, but I do like maintaining the\r\ndistinction -- it feels safer.\r\n\r\nv8 rebases over the recent SSL changes to get the cfbot green again.\r\n\r\nThanks,\r\n--Jacob","msg_date":"Tue, 29 Mar 2022 23:38:29 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 2022-03-29 23:38:29 +0000, Jacob Champion wrote:\n> On Sat, 2022-03-26 at 11:36 -0700, Andres Freund wrote:\n> > > I also note that exposing it as a GUC means we have zero control over\n> > > who/what can read it. Maybe that's not a problem, but it needs to be\n> > > thought about before we go down that path.\n> > \n> > Yes, I think that's a fair concern.\n> \n> I like that there's no builtin way, today, for a superuser to modify\n> the internal value; it strengthens the use as an auditing tool. Moving\n> this to a PGC_SU_BACKEND GUC seems to weaken that. And it looks like\n> PGC_INTERNAL is skipped during the serialization, so if we chose that\n> option, we'd need to write new code anyway?\n\nIt'd be pretty simple to change can_skip_gucvar()'s selection of what to\nsync. E.g. an additional flag like GUC_PARALLEL_SYNCHRONIZE.\n\nI'm not convinced that a GUC is the answer, to be clear.\n\n\n> We'd also need to guess whether the GUC system's serialization of NULL\n> as an empty string is likely to cause problems for any future auth\n> methods.\n\nYou can't represent a NULL in a postgres 'text' datum, independent of\nparallelism. So the current definition of pg_session_authn_id() already\nprecludes that (and set_authn_id() and ...). Honestly, I can't see a reason\nwhy we should ever allow authn_id to contain a NULL byte.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 29 Mar 2022 16:53:32 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, 2022-03-29 at 16:53 -0700, Andres Freund wrote:\r\n> > We'd also need to guess whether the GUC system's serialization of NULL\r\n> > as an empty string is likely to cause problems for any future auth\r\n> > methods.\r\n> \r\n> You can't represent a NULL in a postgres 'text' datum, independent of\r\n> parallelism. So the current definition of pg_session_authn_id() already\r\n> precludes that (and set_authn_id() and ...). Honestly, I can't see a reason\r\n> why we should ever allow authn_id to contain a NULL byte.\r\n\r\nI don't mean a NULL byte, just a NULL pointer. This part of the\r\nimplementation doesn't distinguish between it and an empty string:\r\n\r\n> /* NULL becomes empty string, see estimate_variable_size() */\r\n> do_serialize(destptr, maxbytes, \"%s\",\r\n> *conf->variable ? *conf->variable : \"\");\r\n\r\nWhether that's a problem in the future entirely depends on whether\r\nthere's some authentication method that considers the empty string a\r\nsane and meaningful identity. We might reasonably decide that the\r\nanswer is \"no\", but I like being able to make that decision as opposed\r\nto delegating it to an existing generic framework.\r\n\r\n(That last point may be my core concern about making it a GUC: I'd like\r\nus to have full control of how and where this particular piece of\r\ninformation gets modified.)\r\n\r\nThanks,\r\n--Jacob\r\n","msg_date":"Wed, 30 Mar 2022 16:02:09 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, 2022-03-29 at 23:38 +0000, Jacob Champion wrote:\r\n> v8 rebases over the recent SSL changes to get the cfbot green again.\r\n\r\nI think the Windows failure [1] is unrelated to this patch, but for\r\nposterity:\r\n\r\n> [03:01:58.925] c:\\cirrus>call \"C:/Program Files/Git/usr/bin/timeout.exe\" -v -k60s 15m perl src/tools/msvc/vcregress.pl recoverycheck \r\n> [03:03:16.106] [03:03:16] t/001_stream_rep.pl .................. ok 76551 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\r\n> [03:03:16.120] [03:03:16] t/002_archiving.pl ................... ok 0 ms ( 0.02 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.02 CPU)\r\n> [03:03:16.128] [03:03:16] t/003_recovery_targets.pl ............ ok 0 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\r\n> [03:03:16.138] [03:03:16] t/004_timeline_switch.pl ............. ok 0 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\r\n> [03:03:16.141] [03:03:16] t/005_replay_delay.pl ................ ok 0 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\r\n> [03:03:24.561] [03:03:24] t/006_logical_decoding.pl ............ ok 8416 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\r\n> [03:03:32.496] [03:03:32] t/007_sync_rep.pl .................... ok 7895 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\r\n> [03:03:32.496] [03:03:32] t/008_fsm_truncation.pl .............. ok 0 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\r\n> [03:16:58.985] /usr/bin/timeout: sending signal TERM to command ‘perl’\r\n\r\nThe server and client logs don't quite match up; it looks like we get\r\npartway through t/018_wal_optimize.pl, maybe to the end of the\r\n`wal_level = minimal, SET TABLESPACE commit subtransaction` test,\r\nbefore hanging.\r\n\r\nI see that there's an active thread about a hang later in the recovery\r\nsuite [2]. That's suspicious since 019 is just the next test, but I\r\ndon't see any evidence in the logs that we actually started test 019 in\r\nthis run.\r\n\r\n--Jacob\r\n\r\n[1] https://cirrus-ci.com/task/5434752374145024\r\n[2] https://www.postgresql.org/message-id/flat/83b46e5f-2a52-86aa-fa6c-8174908174b8%40iki.fi\r\n","msg_date":"Thu, 31 Mar 2022 21:49:00 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, Mar 30, 2022 at 04:02:09PM +0000, Jacob Champion wrote:\n> Whether that's a problem in the future entirely depends on whether\n> there's some authentication method that considers the empty string a\n> sane and meaningful identity. We might reasonably decide that the\n> answer is \"no\", but I like being able to make that decision as opposed\n> to delegating it to an existing generic framework.\n\nMy guess on the matter is that an empty authn holds the same meaning\nas NULL because it has no data, but I can see your point as well to\nmake this distinction. In order to do that, couldn't you just use\nshm_toc_lookup(noError=true)? PARALLEL_KEY_SHAREDPORT could be an\noptional entry in the TOC data.\n\nThe name choice is still an issue, as per Andres' point that\nMyProcShared is confusing as it can refer to shared memory. What we\nwant is a structure name for something that's related to MyProc and\nshared across all parallel workers including the leader. I would\ngive up on the \"Shared\" part, using \"Parallel\" and \"Info\" instead.\nHere are some ideas:\n- ProcParallelInfo\n- ProcInfoParallel\n- ParallelProcInfo\n--\nMichael","msg_date":"Tue, 5 Apr 2022 15:13:41 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, 2022-04-05 at 15:13 +0900, Michael Paquier wrote:\r\n> On Wed, Mar 30, 2022 at 04:02:09PM +0000, Jacob Champion wrote:\r\n> > Whether that's a problem in the future entirely depends on whether\r\n> > there's some authentication method that considers the empty string a\r\n> > sane and meaningful identity. We might reasonably decide that the\r\n> > answer is \"no\", but I like being able to make that decision as opposed\r\n> > to delegating it to an existing generic framework.\r\n> \r\n> My guess on the matter is that an empty authn holds the same meaning\r\n> as NULL because it has no data,\r\n\r\nWhether it holds meaning or not depends entirely on the auth method, I\r\nthink. Hypothetical example -- a system could accept client\r\ncertificates with an empty Subject. What identity that Subject\r\nrepresents would depend on the organization, but it's distinct from\r\nNULL/unauthenticated because the certificate is still signed by a CA.\r\n\r\n(Postgres rejects empty Subjects when using clientname=DN and I'm not\r\nproposing that we change that; I'm haven't actually checked that\r\nthey're RFC-legal. But it's possible that a future auth method could\r\nhave a reasonable standard definition for an empty identifier.)\r\n\r\n> but I can see your point as well to\r\n> make this distinction. In order to do that, couldn't you just use\r\n> shm_toc_lookup(noError=true)? PARALLEL_KEY_SHAREDPORT could be an\r\n> optional entry in the TOC data.\r\n\r\nThe current patch already handles NULL with a byte of overhead; is\r\nthere any advantage to using noError? (It might make things messier\r\nonce a second member gets added to the struct.) My concern was directed\r\nat the GUC proposal.\r\n\r\n> The name choice is still an issue, as per Andres' point that\r\n> MyProcShared is confusing as it can refer to shared memory. What we\r\n> want is a structure name for something that's related to MyProc and\r\n> shared across all parallel workers including the leader. I would\r\n> give up on the \"Shared\" part, using \"Parallel\" and \"Info\" instead.\r\n> Here are some ideas:\r\n> - ProcParallelInfo\r\n> - ProcInfoParallel\r\n> - ParallelProcInfo\r\n\r\nI like ParallelProcInfo; it reads nicely. I've used that in v9.\r\n\r\nThanks!\r\n--Jacob","msg_date":"Tue, 5 Apr 2022 18:23:06 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, Apr 05, 2022 at 06:23:06PM +0000, Jacob Champion wrote:\n> Whether it holds meaning or not depends entirely on the auth method, I\n> think. Hypothetical example -- a system could accept client\n> certificates with an empty Subject. What identity that Subject\n> represents would depend on the organization, but it's distinct from\n> NULL/unauthenticated because the certificate is still signed by a CA.\n\nInteresting point.\n\n> The current patch already handles NULL with a byte of overhead; is\n> there any advantage to using noError? (It might make things messier\n> once a second member gets added to the struct.) My concern was directed\n> at the GUC proposal.\n\nFWIW, I am a bit concerned by this approach because it feels\ninconsistent with any other conditional fields passed down from the\nparallel leader to its workers. And if we need to add more fields to \nParallelProcInfo in the future, it will be cleaner to use different\nTOC keys to pass down different fields anyway, no?\n--\nMichael","msg_date":"Wed, 6 Apr 2022 20:09:36 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, 2022-04-06 at 20:09 +0900, Michael Paquier wrote:\r\n> > The current patch already handles NULL with a byte of overhead; is\r\n> > there any advantage to using noError? (It might make things messier\r\n> > once a second member gets added to the struct.) My concern was directed\r\n> > at the GUC proposal.\r\n> \r\n> FWIW, I am a bit concerned by this approach because it feels\r\n> inconsistent with any other conditional fields passed down from the\r\n> parallel leader to its workers. And if we need to add more fields to \r\n> ParallelProcInfo in the future, it will be cleaner to use different\r\n> TOC keys to pass down different fields anyway, no?\r\n\r\nI assumed that we would follow the existing model of \"(de)serialize a\r\nwhole struct\", rather than pulling it apart into many separate keys. If\r\nit got too complicated then we could consider introducing a\r\nSerializedParallelProcInfo struct like some of the other functions do.\r\nMaybe that wouldn't work so well if many of the fields are strings?\r\n\r\nIs there a significant cost to changing this later, if one approach\r\nturns out to be wrong?\r\n\r\n--Jacob\r\n","msg_date":"Wed, 6 Apr 2022 19:16:43 +0000","msg_from":"Jacob Champion <pchampion@vmware.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, Apr 06, 2022 at 07:16:43PM +0000, Jacob Champion wrote:\n> I assumed that we would follow the existing model of \"(de)serialize a\n> whole struct\", rather than pulling it apart into many separate keys. If\n> it got too complicated then we could consider introducing a\n> SerializedParallelProcInfo struct like some of the other functions do.\n> Maybe that wouldn't work so well if many of the fields are strings?\n> \n> Is there a significant cost to changing this later, if one approach\n> turns out to be wrong?\n\nI don't think this is going to be an issue as long as we don't change\nthe definitions of MyParallelProcInfo, Port or PARALLEL_KEY_* in the\nstable branch. My guess is that we are fine to switch to one approach\nor the other as this is just some internal communication logic between\nthe parallel leader and its workers.\n\nWhat is the feeling of others about this patch and the introduction of\nParallelProcInfo (or ParallelPortInfo?) to store the authn coming from\nPort? The feature freeze is very close.\n--\nMichael","msg_date":"Thu, 7 Apr 2022 09:13:32 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"v10 is rebased over latest; I've also added a PGDLLIMPORT to the new global.\n\n--Jacob","msg_date":"Tue, 31 May 2022 15:21:03 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, May 31, 2022 at 6:21 PM Jacob Champion <jchampion@timescale.com> wrote:\n> v10 is rebased over latest; I've also added a PGDLLIMPORT to the new global.\n\nI took a quick look at this and it doesn't seem crazy to me, except\nthat I think ParallelProcInfo is a bad name for it. It's kind of\ngeneric, because neither \"proc\" nor \"info\" means a whole lot. It's\nalso kind of wrong, because I think \"parallel\" should be things that\nhave to do with parallelism, not just things that happen to be\nsynchronized across processes when parallelism is in use. It doesn't\nmake sense to me to have something called a ParallelProcInfo that is\nused for every single connection in the universe even if parallelism\nis completely disabled on the system.\n\nI'm not sure what it SHOULD be called, exactly: that's one of the hard\nproblems in computer science.[1]\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n[1] https://martinfowler.com/bliki/TwoHardThings.html\n\n\n","msg_date":"Thu, 2 Jun 2022 09:51:51 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, Jun 2, 2022 at 6:52 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> I'm not sure what it SHOULD be called, exactly: that's one of the hard\n> problems in computer science.[1]\n\nYeah...\n\nAll right, here's the full list of previous suggestions, I think:\n\n- SharedPort\n- MyProcShared\n- ProcParallelInfo\n- ProcInfoParallel\n- ParallelProcInfo\n- ParallelPortInfo\n\nI have a few new proposals:\n\n- GlobalPortInfo\n- GlobalConnInfo\n- Synced[Port/Conn]Info\n- Worker[Port/Conn]Info (but I think this suffers from the exact same\nproblem as Parallel)\n- ThingsThatHappenToBeSynchronizedAcrossProcessesWhenParallelismIsInUse\n- OrderImporterConsumerTemplateBeanFactory\n\nI am struggling to come up with a single adjective that captures this\nconcept of sometimes-synchronized, that isn't conflicting with\nexisting uses (like Shared). Other suggestions are very welcome.\n\nThanks,\n--Jacob\n\n\n","msg_date":"Thu, 2 Jun 2022 15:56:28 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, Jun 02, 2022 at 03:56:28PM -0700, Jacob Champion wrote:\n> All right, here's the full list of previous suggestions, I think:\n> \n> - SharedPort\n> - MyProcShared\n> - ProcParallelInfo\n> - ProcInfoParallel\n> - ParallelProcInfo\n> - ParallelPortInfo\n> \n> I have a few new proposals:\n> \n> - GlobalPortInfo\n> - GlobalConnInfo\n> - Synced[Port/Conn]Info\n> - Worker[Port/Conn]Info (but I think this suffers from the exact same\n> problem as Parallel)\n> - ThingsThatHappenToBeSynchronizedAcrossProcessesWhenParallelismIsInUse\n> - OrderImporterConsumerTemplateBeanFactory\n\nParallelPortInfo sounds kind of right for the job to me in this set of\nproposals, as the data is from the Port, and that's some information\nshared between all the parallel workers and the leader.\n--\nMichael","msg_date":"Fri, 3 Jun 2022 09:38:35 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> ParallelPortInfo sounds kind of right for the job to me in this set of\n> proposals, as the data is from the Port, and that's some information\n> shared between all the parallel workers and the leader.\n\nI agree with Robert's complaint that Parallel is far too generic\na term here. Also, the fact that this data is currently in struct\nPort seems like an artifact.\n\nDon't we have a term for the set of processes comprising a leader\nplus parallel workers? If we called that set FooGroup, then\nsomething like FooGroupSharedInfo would be on-point.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 03 Jun 2022 10:04:12 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, Jun 03, 2022 at 10:04:12AM -0400, Tom Lane wrote:\n> I agree with Robert's complaint that Parallel is far too generic\n> a term here. Also, the fact that this data is currently in struct\n> Port seems like an artifact.\n> \n> Don't we have a term for the set of processes comprising a leader\n> plus parallel workers? If we called that set FooGroup, then\n> something like FooGroupSharedInfo would be on-point.\n\nAs far as I know, proc.h includes the term \"group members\", which\nincludes the leader and its workers (see CLOG and lock part)?\n--\nMichael","msg_date":"Sat, 4 Jun 2022 11:36:10 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, Jun 3, 2022 at 7:36 PM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Fri, Jun 03, 2022 at 10:04:12AM -0400, Tom Lane wrote:\n> > I agree with Robert's complaint that Parallel is far too generic\n> > a term here. Also, the fact that this data is currently in struct\n> > Port seems like an artifact.\n> >\n> > Don't we have a term for the set of processes comprising a leader\n> > plus parallel workers? If we called that set FooGroup, then\n> > something like FooGroupSharedInfo would be on-point.\n>\n> As far as I know, proc.h includes the term \"group members\", which\n> includes the leader and its workers (see CLOG and lock part)?\n\nlmgr/README also refers to \"gangs of related processes\" and \"parallel\ngroups\". So\n\n- GroupSharedInfo\n- ParallelGroupSharedInfo\n- GangSharedInfo\n- SharedLeaderInfo\n\n?\n\n--Jacob\n\n\n","msg_date":"Mon, 6 Jun 2022 08:44:32 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, Jun 3, 2022 at 10:04 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> I agree with Robert's complaint that Parallel is far too generic\n> a term here. Also, the fact that this data is currently in struct\n> Port seems like an artifact.\n\nWhy do we call this thing a Port, anyway?\n\nI think I'd feel more comfortable here if we were defining what went\ninto which struct on some semantic basis rather than being like, OK,\nso all the stuff we want to serialize goes into struct #1, and the\nstuff we don't want to serialize goes into struct #2. I suppose if\nit's just based on whether or not we want to serialize it, then the\nplacement of future additions will just be based on how people happen\nto feel about the thing they're adding right at that moment, and there\nwon't be any consistency.\n\nOne could imagine dividing the Port struct into a couple of different\nstructs, e.g.\n\nAuthenticationState: stuff that is needed only during authentication\nand can be discarded thereafter (e.g. the HBA line, at least if the\ncomment is to be believed)\nClientCommunicationState: stuff that is used to communicate with the\nclient but doesn't need to be or can't be shared (e.g. the SSL object\nitself)\nClientConnectionInfo: stuff that someone might want to look at for\ninformation purposes at any time (e.g. authn_id, apparently)\n\nThen we could serialize the third of these, keep the second around but\nnot serialize it, and free the first once connection setup is\ncomplete.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 6 Jun 2022 14:43:51 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Mon, Jun 6, 2022 at 11:44 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> I think I'd feel more comfortable here if we were defining what went\n> into which struct on some semantic basis rather than being like, OK,\n> so all the stuff we want to serialize goes into struct #1, and the\n> stuff we don't want to serialize goes into struct #2. I suppose if\n> it's just based on whether or not we want to serialize it, then the\n> placement of future additions will just be based on how people happen\n> to feel about the thing they're adding right at that moment, and there\n> won't be any consistency.\n\n\"This struct contains connection fields that are explicitly safe for\nworkers to access\" _is_ a useful semantic, in my opinion. And it seems\nlike it'd make it easier to determine what needs to be included in the\nstruct; I'm not sure I follow why it would result in less consistency.\n\nBut to your suggestion, if we just called the new struct\n\"ClientConnectionInfo\", would it be a useful step towards your\nproposed three-bucket state? I guess I'm having trouble understanding\nwhy a struct that is defined as \"this stuff *doesn't* get serialized\"\nis materially different from having one that's the opposite.\n\nThanks,\n--Jacob\n\n\n","msg_date":"Tue, 7 Jun 2022 15:54:38 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, Jun 7, 2022 at 6:54 PM Jacob Champion <jchampion@timescale.com> wrote:\n> \"This struct contains connection fields that are explicitly safe for\n> workers to access\" _is_ a useful semantic, in my opinion. And it seems\n> like it'd make it easier to determine what needs to be included in the\n> struct; I'm not sure I follow why it would result in less consistency.\n>\n> But to your suggestion, if we just called the new struct\n> \"ClientConnectionInfo\", would it be a useful step towards your\n> proposed three-bucket state? I guess I'm having trouble understanding\n> why a struct that is defined as \"this stuff *doesn't* get serialized\"\n> is materially different from having one that's the opposite.\n\nWell, it isn't, and if my proposal boils down to that, which perhaps\nit does, then my proposal isn't that great, honestly. Let me try again\nto explain, though, and maybe it will seem less arbitrary with a\nsecond explanation -- or maybe it won't.\n\nIf we say \"this struct contains authentication-related information\nthat we got from the client and which functions may want to look at\nlater,\" then I feel like the chances are good that when someone adds a\nnew thing to the system in the future, they will know whether or not\nthat new thing falls into that category or not. If the definition of a\nstruct is \"everything that should be serialized,\" then I feel like the\nchances are less good that everyone will know whether a new thing they\nare adding falls into that category or not. Perhaps that is\nill-founded, but I don't think \"should be serialized\" is necessarily\nsomething that everybody is going to have the same view on, or even\nknow what it means.\n\nAlso, I don't think we want to end up with a situation where we have a\nstruct that contains wildly unrelated things that all need to be\nserialized. If the definition of the struct is \"stuff we should\nserialize and send to the worker,\" well then maybe the transaction\nsnapshot ought to go in there! Well, no. I mean, we already have a\nseparate place for that, but suppose somehow we didn't. It doesn't\nbelong here, because yes the things in this struct get serialized, but\nit's not only any old thing that needs serializing, it's more specific\nthan that.\n\nI guess what this boils down to is that I really want this thing to\nhave a meaningful name by means of which a future developer can make a\nguess as to whether some field they're adding ought to go in there. I\ntheorize that SharedPort is not too great because (a) Port is already\na bad name and (b) how am I supposed to know whether my stuff ought to\nbe shared or not? I like something like ClientConnectionInfo better\nbecause it seems to describe what the stuff in the struct *is* rather\nthan what we *do* with it.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 7 Jun 2022 20:45:15 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, Jun 7, 2022 at 5:45 PM Robert Haas <robertmhaas@gmail.com> wrote:\n> Perhaps that is\n> ill-founded, but I don't think \"should be serialized\" is necessarily\n> something that everybody is going to have the same view on, or even\n> know what it means.\n\nUsing this thread as an example, once it was decided that the parallel\nworkers needed the additional info, the need for serialization\nfollowed directly from that. I don't get the feeling that developers\nare going to jump through the hoops of writing serialization logic for\na new field in the struct just by accident; they should know why\nthey're writing that code, and hopefully it would be easy for\nreviewers to catch a patch that showed up with pointless\nserialization.\n\n> Also, I don't think we want to end up with a situation where we have a\n> struct that contains wildly unrelated things that all need to be\n> serialized. If the definition of the struct is \"stuff we should\n> serialize and send to the worker,\" well then maybe the transaction\n> snapshot ought to go in there! Well, no. I mean, we already have a\n> separate place for that, but suppose somehow we didn't. It doesn't\n> belong here, because yes the things in this struct get serialized, but\n> it's not only any old thing that needs serializing, it's more specific\n> than that.\n\nI completely agree with you here -- the name should not be so generic\nthat it's just a catch-all for any serialized fields that exist.\n\n> I guess what this boils down to is that I really want this thing to\n> have a meaningful name by means of which a future developer can make a\n> guess as to whether some field they're adding ought to go in there. I\n> theorize that SharedPort is not too great because (a) Port is already\n> a bad name and (b) how am I supposed to know whether my stuff ought to\n> be shared or not? I like something like ClientConnectionInfo better\n> because it seems to describe what the stuff in the struct *is* rather\n> than what we *do* with it.\n\nI think having both would be useful in this case -- what the stuff is,\nso that it's clear what doesn't belong in it, and what we do with it,\nso it's clear that you have to write serialization code if you add new\nthings. The nature of the struct is such that I think you _have_ to\nfigure out whether or not your stuff ought to be shared before you\nhave any business adding it.\n\nBut I don't have any better ideas for how to achieve both. I'm fine\nwith your suggestion of ClientConnectionInfo, if that sounds good to\nothers; the doc comment can clarify why it differs from Port? Or add\none of the Shared-/Gang-/Group- prefixes to it, maybe?\n\nThanks,\n--Jacob\n\n\n","msg_date":"Wed, 8 Jun 2022 16:53:08 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, Jun 8, 2022 at 7:53 PM Jacob Champion <jchampion@timescale.com> wrote:\n> But I don't have any better ideas for how to achieve both. I'm fine\n> with your suggestion of ClientConnectionInfo, if that sounds good to\n> others; the doc comment can clarify why it differs from Port? Or add\n> one of the Shared-/Gang-/Group- prefixes to it, maybe?\n\nI don't like the prefixes, so I'd prefer explaining it in the struct comment.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 9 Jun 2022 09:23:12 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Thu, Jun 9, 2022 at 6:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\n> On Wed, Jun 8, 2022 at 7:53 PM Jacob Champion <jchampion@timescale.com> wrote:\n> > But I don't have any better ideas for how to achieve both. I'm fine\n> > with your suggestion of ClientConnectionInfo, if that sounds good to\n> > others; the doc comment can clarify why it differs from Port? Or add\n> > one of the Shared-/Gang-/Group- prefixes to it, maybe?\n>\n> I don't like the prefixes, so I'd prefer explaining it in the struct comment.\n\nDone that way in v11.\n\nThanks!\n--Jacob","msg_date":"Fri, 10 Jun 2022 10:58:09 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 6/10/22 7:58 PM, Jacob Champion wrote:\n> CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.\n>\n>\n>\n> On Thu, Jun 9, 2022 at 6:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\n>> On Wed, Jun 8, 2022 at 7:53 PM Jacob Champion <jchampion@timescale.com> wrote:\n>>> But I don't have any better ideas for how to achieve both. I'm fine\n>>> with your suggestion of ClientConnectionInfo, if that sounds good to\n>>> others; the doc comment can clarify why it differs from Port? Or add\n>>> one of the Shared-/Gang-/Group- prefixes to it, maybe?\n>> I don't like the prefixes, so I'd prefer explaining it in the struct comment.\n> Done that way in v11.\n>\n> Thanks!\n> --Jacob\n\nFWIW, I just created a new thread to expose the port->authn_id through \nthe SYSTEM_USER sql reserved word.\n\nRegards,\n\nBertrand\n\n\n\n\n","msg_date":"Wed, 22 Jun 2022 15:31:16 +0200","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On 6/22/22 06:31, Drouvot, Bertrand wrote:\n> FWIW, I just created a new thread to expose the port->authn_id through \n> the SYSTEM_USER sql reserved word.\n\nReview for both seems to have dried up a bit. I'm not particularly\ninvested in my code, but I do want to see *a* solution go in. So if it\nhelps the review momentum for me to withdraw this patch and put my\neffort into SYSTEM_USER, I can do that no problem.\n\nThoughts from prior reviewers? Is SYSTEM_USER the way to go?\n\n--Jacob\n\n\n","msg_date":"Tue, 2 Aug 2022 14:57:38 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 8/2/22 11:57 PM, Jacob Champion wrote:\n> On 6/22/22 06:31, Drouvot, Bertrand wrote:\n>> FWIW, I just created a new thread to expose the port->authn_id through\n>> the SYSTEM_USER sql reserved word.\n> Review for both seems to have dried up a bit. I'm not particularly\n> invested in my code, but I do want to see *a* solution go in. So if it\n> helps the review momentum for me to withdraw this patch and put my\n> effort into SYSTEM_USER, I can do that no problem.\n>\n> Thoughts from prior reviewers? Is SYSTEM_USER the way to go?\n\nI did not look in detail to this thread, but if the goal is \"only\" to \nexpose authn_id (as the subject describes) then it seems to me that \nSYSTEM_USER [1] is the way to go.\n\n[1]: https://commitfest.postgresql.org/39/3703/\n\n-- \nBertrand Drouvot\nAmazon Web Services: https://aws.amazon.com\n\n\n\n","msg_date":"Fri, 5 Aug 2022 12:48:33 +0200","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, Aug 05, 2022 at 12:48:33PM +0200, Drouvot, Bertrand wrote:\n> On 8/2/22 11:57 PM, Jacob Champion wrote:\n>> Thoughts from prior reviewers? Is SYSTEM_USER the way to go?\n\nReading through the other thread, there is a clear parallel between\nboth in concept to provide this information at SQL level, indeed.\n\n> I did not look in detail to this thread, but if the goal is \"only\" to expose\n> authn_id (as the subject describes) then it seems to me that SYSTEM_USER [1]\n> is the way to go.\n> \n> [1]: https://commitfest.postgresql.org/39/3703/\n\nHowever, I am not sure if the suggestion of auth_method:authn as\noutput generated by SYSTEM_USER would be correct according to the SQL\nspecification, either. The spec being not really talkative about the\ndetails of what an external module should be opens up for a lot of\ninterpretation, something that both thread are dealing with.\n\nAnyway, we are talking about two different problems on this thread:\n1) Provide the authn/SYSTEM_USER through some SQL interface, which is\nwhat 0001 is an attempt of, SYSTEM_USER is a different attempt.\n2) Move authn out of Port into its own structure, named\nClientConnectionInfo, and pass it down to the parallel workers.\n\nSYSTEM_USER overlaps with 0001, but I see no reason to not do 0002 in\nall cases. Even if we are not sure yet of how to expose that at SQL\nlevel, we still want to pass down this information to the parallel\nworkers and we still want to not have that in Port. An extension\ncould also do easily the job once 0002 is done, so I see a good\nargument about doing it anyway. The name ClientConnectionInfo from\nRobert looks like the compromise we have for the new structure holding\nthe information about the client information passed down to the\nworkers.\n--\nMichael","msg_date":"Sat, 6 Aug 2022 15:26:12 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On 8/6/22 02:26, Michael Paquier wrote:\n> On Fri, Aug 05, 2022 at 12:48:33PM +0200, Drouvot, Bertrand wrote:\n>> On 8/2/22 11:57 PM, Jacob Champion wrote:\n>>> Thoughts from prior reviewers? Is SYSTEM_USER the way to go?\n> \n> Reading through the other thread, there is a clear parallel between\n> both in concept to provide this information at SQL level, indeed.\n> \n>> I did not look in detail to this thread, but if the goal is \"only\" to expose\n>> authn_id (as the subject describes) then it seems to me that SYSTEM_USER [1]\n>> is the way to go.\n>> \n>> [1]: https://commitfest.postgresql.org/39/3703/\n> \n> However, I am not sure if the suggestion of auth_method:authn as\n> output generated by SYSTEM_USER would be correct according to the SQL\n> specification, either. The spec being not really talkative about the\n> details of what an external module should be opens up for a lot of\n> interpretation, something that both thread are dealing with.\n\nAs I pointed out here [ \nhttps://www.postgresql.org/message-id/28b4a9ef-5103-f117-99e1-99ae5a86a6e8%40joeconway.com \n] both the SQL Server and Oracle interpretations are similar to the one \nprovided by Bertrand's patch:\n\nSQL Server:\n \"If the current user is logged in to SQL Server by\n using Windows Authentication, SYSTEM_USER returns the\n Windows login identification name in the form:\n DOMAIN\\user_login_name. However, if the current user\n is logged in to SQL Server by using SQL Server\n Authentication, SYSTEM_USER returns the SQL Server\n login identification name\"\n\nOracle:\n \"SYSTEM_USER\n Returns the name of the current data store user as\n identified by the operating system.\"\n\nI am not sure how else we should interpret SYSTEM_USER -- if it isn't \nport->authn_id what else would you propose it should be?\n\n-- \nJoe Conway\nRDS Open Source Databases\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Sat, 6 Aug 2022 10:59:26 -0400","msg_from":"Joe Conway <mail@joeconway.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Sat, Aug 06, 2022 at 10:59:26AM -0400, Joe Conway wrote:\n> I am not sure how else we should interpret SYSTEM_USER -- if it isn't\n> port->authn_id what else would you propose it should be?\n\nWhat you say sounds rather right, but I was wondering mainly what\nOracle and SQL server report when it comes to other authentication\nmethods like SSPI or a cert, where we don't use a user name but some\ndata dependent on the auth method. And I have no experience with\nthese.\n\nAnyway, I was looking at Bertrand's patch, and I can see that it is\ndoing nothing to move away the connection information that we have in\nPort away to a different structure passed down to the parallel\nworkers, which is what I understand is a cleanup worth on its own\nbased on the discussion of this thread. Hence, I still see a good\nargument for the introduction of ClientConnectionInfo that gets passed\ndown to the workers. Based on that, I think that we'd better finish\nv11-0002 (only ClientConnectionInfo, no SQL interface) as a first step\nto build for the next ones, with authn being the first piece of\ninformation given to the workers. With a separate structure, the\nauth_method can also be a second member in ClientConnectionInfo,\ncompleting what would be needed to build SYSTEM_USER as the workers\nwould have access to it.\n\nAm I getting that right?\n--\nMichael","msg_date":"Sun, 7 Aug 2022 16:41:08 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 8/7/22 9:41 AM, Michael Paquier wrote:\n> Anyway, I was looking at Bertrand's patch, and I can see that it is\n> doing nothing to move away the connection information that we have in\n> Port away to a different structure passed down to the parallel\n> workers,\n\nThanks for looking at it!\n\nThat's right. The main reason is that in the v2-0003 SYSTEM_USER patch \nwhat is passed down to the parallel workers is not Port->authn_id but a \nnew \"SystemUser\" (defined in miscinit.c with CurrentUserId and friends).\n\n> which is what I understand is a cleanup worth on its own\n> based on the discussion of this thread. Hence, I still see a good\n> argument for the introduction of ClientConnectionInfo that gets passed\n> down to the workers.\n\nI agree that it could it be useful too.\n\n> Based on that, I think that we'd better finish\n> v11-0002 (only ClientConnectionInfo, no SQL interface)\n\nI agree.\n\n> as a first step\n> to build for the next ones, with authn being the first piece of\n> information given to the workers. With a separate structure, the\n> auth_method can also be a second member in ClientConnectionInfo,\n> completing what would be needed to build SYSTEM_USER as the workers\n> would have access to it.\n\nbut I'm not sure we should do it as a first step (given the fact that \nthis is not Port->authn_id that is passed down to the parallel workers \nin the SYSTEM_USER patch).\n\nWhat do you think about working on both (aka a) v11-002 only \nClientConnectionInfo and b) SYSTEM_USER) in parallel?\n\nThanks\n\n-- \nBertrand Drouvot\nAmazon Web Services: https://aws.amazon.com\n\n\n\n","msg_date":"Mon, 8 Aug 2022 12:43:14 +0200","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Mon, Aug 08, 2022 at 12:43:14PM +0200, Drouvot, Bertrand wrote:\n> but I'm not sure we should do it as a first step (given the fact that this\n> is not Port->authn_id that is passed down to the parallel workers in the\n> SYSTEM_USER patch).\n> \n> What do you think about working on both (aka a) v11-002 only\n> ClientConnectionInfo and b) SYSTEM_USER) in parallel?\n\nIt seems to me that completing ClientConnectionInfo first has the\nadvantage of not having to tweak twice the interface we are going to\nuse when passing down the full structure to the workers, so I would\nchoose for doing it first (with one field for the authn, and a second\nfield for the auth method so as the the workers can build SYSTEM_USER\nby themselves when required).\n--\nMichael","msg_date":"Tue, 9 Aug 2022 18:17:55 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 8/9/22 11:17 AM, Michael Paquier wrote:\n> On Mon, Aug 08, 2022 at 12:43:14PM +0200, Drouvot, Bertrand wrote:\n>> but I'm not sure we should do it as a first step (given the fact that this\n>> is not Port->authn_id that is passed down to the parallel workers in the\n>> SYSTEM_USER patch).\n>>\n>> What do you think about working on both (aka a) v11-002 only\n>> ClientConnectionInfo and b) SYSTEM_USER) in parallel?\n> It seems to me that completing ClientConnectionInfo first has the\n> advantage of not having to tweak twice the interface we are going to\n> use when passing down the full structure to the workers, so I would\n> choose for doing it first (with one field for the authn, and a second\n> field for the auth method so as the the workers can build SYSTEM_USER\n> by themselves when required).\n\nYeah fair point.\n\nAgree that it makes sense to work on those patches in this particular \norder then.\n\nThanks,\n\n-- \n\nBertrand Drouvot\nAmazon Web Services: https://aws.amazon.com\n\n\n\n","msg_date":"Tue, 9 Aug 2022 12:39:02 +0200","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, Aug 9, 2022 at 3:39 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\n> Agree that it makes sense to work on those patches in this particular\n> order then.\n\nSounds good. The ClientConnectionInfo patch (previously 0002) is\nattached, with the SQL function removed.\n\nThanks,\n--Jacob","msg_date":"Wed, 10 Aug 2022 08:09:58 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 8/10/22 5:09 PM, Jacob Champion wrote:\n> On Tue, Aug 9, 2022 at 3:39 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\n>> Agree that it makes sense to work on those patches in this particular\n>> order then.\n> Sounds good. The ClientConnectionInfo patch (previously 0002) is\n> attached, with the SQL function removed.\n\nThanks for the patch!\n\nLooking at:\n\n+typedef struct\n+{\n+ /*\n+ * Authenticated identity. The meaning of this identifier is \ndependent on\n+ * hba->auth_method; it is the identity (if any) that the user \npresented\n+ * during the authentication cycle, before they were assigned a \ndatabase\n+ * role. (It is effectively the \"SYSTEM-USERNAME\" of a pg_ident \nusermap\n+ * -- though the exact string in use may be different, depending on \npg_hba\n+ * options.)\n+ *\n+ * authn_id is NULL if the user has not actually been \nauthenticated, for\n+ * example if the \"trust\" auth method is in use.\n+ */\n+ const char *authn_id;\n+} ClientConnectionInfo;\n\nWhat do you think about adding a second field in ClientConnectionInfo \nfor the auth method (as suggested by Michael upthread)?\n\nThat will be needed by the SYSTEM_USER patch (that its current version \nimplements as \"auth_method:identity\").\n\nThanks,\n\n-- \nBertrand Drouvot\nAmazon Web Services: https://aws.amazon.com\n\n\n\n","msg_date":"Thu, 11 Aug 2022 07:47:39 +0200","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, Aug 10, 2022 at 10:48 PM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\n> What do you think about adding a second field in ClientConnectionInfo\n> for the auth method (as suggested by Michael upthread)?\n\nSure -- without a followup patch, it's not really tested, though.\n\nv2 adjusts set_authn_id() to copy the auth_method over as well. It\n\"passes tests\" but is otherwise unexercised.\n\nThanks,\n--Jacob","msg_date":"Thu, 11 Aug 2022 15:28:31 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 8/12/22 12:28 AM, Jacob Champion wrote:\n> On Wed, Aug 10, 2022 at 10:48 PM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\n>> What do you think about adding a second field in ClientConnectionInfo\n>> for the auth method (as suggested by Michael upthread)?\n> Sure -- without a followup patch, it's not really tested, though.\n>\n> v2 adjusts set_authn_id() to copy the auth_method over as well. It\n> \"passes tests\" but is otherwise unexercised.\n\nThank you!\n\nTo help with the testing I've just provided a new version (aka \nv2-0004-system_user-implementation.patch) of the SYSTEM_USER patch in \n[1] that can be applied on top of \n\"v2-0001-Allow-parallel-workers-to-read-authn_id.patch\".\n\nBut for this to work, the first comment below on your patch needs to be \naddressed.\n\nOnce the first comment is addressed and the new SYSTEM_USER patch \napplied (that adds new tap tests) then we can test the propagation to \nthe parallel workers with:\n\nmake -C src/test/kerberos check PROVE_TESTS=t/001_auth.pl PROVE_FLAGS=-v\n\nand\n\nmake -C src/test/authentication check PROVE_TESTS=t/001_password.pl \nPROVE_FLAGS=-v\n\nBoth are currently successful.\n\nRegarding the comments on \nv2-0001-Allow-parallel-workers-to-read-authn_id.patch:\n\n1)\n\nThis is the one to be applied before adding the new SYSTEM_USER one on top:\n\n+typedef struct\n+{\n+ /*\n+ * Authenticated identity. The meaning of this identifier is \ndependent on\n\nhas to be replaced by:\n\n+typedef struct ClientConnectionInfo\n+{\n+ /*\n+ * Authenticated identity. The meaning of this identifier is \ndependent on\n\n2)\n\n+ * Authenticated identity. The meaning of this identifier is \ndependent on\n\nThere is one extra space before \"The\"\n\n3)\n\n+SerializeClientConnectionInfo(Size maxsize, char *start_address)\n+{\n+ /*\n+ * First byte is an indication of whether or not authn_id has \nbeen set to\n+ * non-NULL, to differentiate that case from the empty string.\n+ */\n\nis authn_id being an empty string possible?\n\n4)\n\n+ */\n+\n+ClientConnectionInfo MyClientConnectionInfo;\n+\n+/*\n+ * Calculate the space needed to serialize MyClientConnectionInfo.\n+ */\n+Size\n+EstimateClientConnectionInfoSpace(void)\n\n From a coding style point of view, shouldn't \"ClientConnectionInfo \nMyClientConnectionInfo;\" be moved to the top of the file?\n\n\n[1]: https://commitfest.postgresql.org/39/3703/\n\nRegards,\n\n-- \nBertrand Drouvot\nAmazon Web Services: https://aws.amazon.com\n\n\n\n","msg_date":"Fri, 12 Aug 2022 15:34:04 +0200","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Fri, Aug 12, 2022 at 03:34:04PM +0200, Drouvot, Bertrand wrote:\n> 3)\n> \n> +SerializeClientConnectionInfo(Size maxsize, char *start_address)\n> +{\n> + /*\n> + * First byte is an indication of whether or not authn_id has been\n> set to\n> + * non-NULL, to differentiate that case from the empty string.\n> + */\n> \n> is authn_id being an empty string possible?\n\nI don't recall that this can be the case yet, but we cannot discard\nit. One thing was itching me about the serialization and\ndeserialization logic though: could it be more readable if we used an\nintermediate structure to store the length of the serialized strings?\nWe use this approach in other areas, like for the snapshot data in\nsnapmgr.c. This would handle the case of an empty and NULL string, by\nstoring -1 as length for NULL and >= 0 for the string length if there\nis something set, while making the addition of more fields a\nno-brainer.\n--\nMichael","msg_date":"Sun, 14 Aug 2022 18:57:40 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 8/14/22 11:57 AM, Michael Paquier wrote:\n> On Fri, Aug 12, 2022 at 03:34:04PM +0200, Drouvot, Bertrand wrote:\n>> 3)\n>>\n>> +SerializeClientConnectionInfo(Size maxsize, char *start_address)\n>> +{\n>> + /*\n>> + * First byte is an indication of whether or not authn_id has been\n>> set to\n>> + * non-NULL, to differentiate that case from the empty string.\n>> + */\n>>\n>> is authn_id being an empty string possible?\n> I don't recall that this can be the case yet, but we cannot discard\n> it.\nFair point.\n> One thing was itching me about the serialization and\n> deserialization logic though: could it be more readable if we used an\n> intermediate structure to store the length of the serialized strings?\n> We use this approach in other areas, like for the snapshot data in\n> snapmgr.c. This would handle the case of an empty and NULL string, by\n> storing -1 as length for NULL and >= 0 for the string length if there\n> is something set, while making the addition of more fields a\n> no-brainer.\n\nI think that's a good idea and I think that would be more readable (as \ncompare to storing a \"hint\" in the first byte).\n\nRegards,\n\n-- \nBertrand Drouvot\nAmazon Web Services: https://aws.amazon.com\n\n\n\n","msg_date":"Tue, 16 Aug 2022 11:02:30 +0200","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hello,\n\nOn Fri, Aug 12, 2022 at 6:34 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\n> +typedef struct\n> +{\n> + /*\n> + * Authenticated identity. The meaning of this identifier is\n> dependent on\n>\n> has to be replaced by:\n>\n> +typedef struct ClientConnectionInfo\n> +{\n> + /*\n> + * Authenticated identity. The meaning of this identifier is\n> dependent on\n\nOkay, will do in the next patch (coming soon).\n\n> + * Authenticated identity. The meaning of this identifier is\n> dependent on\n>\n> There is one extra space before \"The\"\n\nThis comment block was just moved verbatim; the double-spaced\nsentences were there before.\n\n> From a coding style point of view, shouldn't \"ClientConnectionInfo\n> MyClientConnectionInfo;\" be moved to the top of the file?\n\nThe style in this file seems to be to declare the variables at the top\nof the section in which they're used. See the sections for \"User ID\nstate\" and \"Library preload support\".\n\nThanks,\n--Jacob\n\n\n","msg_date":"Tue, 16 Aug 2022 09:09:03 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, Aug 16, 2022 at 2:02 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\n> On 8/14/22 11:57 AM, Michael Paquier wrote:\n> > One thing was itching me about the serialization and\n> > deserialization logic though: could it be more readable if we used an\n> > intermediate structure to store the length of the serialized strings?\n> > We use this approach in other areas, like for the snapshot data in\n> > snapmgr.c. This would handle the case of an empty and NULL string, by\n> > storing -1 as length for NULL and >= 0 for the string length if there\n> > is something set, while making the addition of more fields a\n> > no-brainer.\n>\n> I think that's a good idea and I think that would be more readable (as\n> compare to storing a \"hint\" in the first byte).\n\nSounds good. v3, attached, should make the requested changes:\n- declare `struct ClientConnectionInfo`\n- use an intermediate serialization struct\n- switch to length-\"prefixing\" for the string\n\nI do like the way this reads compared to before.\n\nThanks,\n--Jacob","msg_date":"Tue, 16 Aug 2022 09:58:28 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 8/16/22 6:58 PM, Jacob Champion wrote:\n> Sounds good. v3, attached, should make the requested changes:\n> - declare `struct ClientConnectionInfo`\n> - use an intermediate serialization struct\n> - switch to length-\"prefixing\" for the string\n>\n> I do like the way this reads compared to before.\n\nThanks for the new version!\n\n+ /* Copy authn_id into the space after the struct. */\n+ if (serialized.authn_id_len >= 0)\n\nMaybe remove the \".\" at the end of the comment? (to be consistent with \nthe other comment just above)\n\n+/*\n+ * Restore MyClientConnectionInfo from its serialized representation.\n+ */\n+void\n+RestoreClientConnectionInfo(char *conninfo)\n+{\n+ SerializedClientConnectionInfo serialized;\n+ char *authn_id;\n\nMove \"char *authn_id;\" in the \"if (serialized.authn_id_len >= 0)\" \nbelow?\n\n+\n+ memcpy(&serialized, conninfo, sizeof(serialized));\n+ authn_id = conninfo + sizeof(serialized);\n\nMove \"authn_id = conninfo + sizeof(serialized)\" in the \"if \n(serialized.authn_id_len >= 0)\" below?\n\n+\n+ /* Copy the fields back into place. */\n\nRemove the \".\" at the end of the comment?\n\n+ MyClientConnectionInfo.authn_id = NULL;\n+ MyClientConnectionInfo.auth_method = serialized.auth_method;\n+\n+ if (serialized.authn_id_len >= 0)\n+ MyClientConnectionInfo.authn_id = \nMemoryContextStrdup(TopMemoryContext,\n+ authn_id);\n\nThis instead?\n\n if (serialized.authn_id_len >= 0)\n {\n char *authn_id;\n authn_id = conninfo + sizeof(serialized);\n MyClientConnectionInfo.authn_id = \nMemoryContextStrdup(TopMemoryContext,\nauthn_id);\n }\n\n+ src/backend/utils/init/miscinit.c:RestoreClientConnectionInfo(char \n*conninfo)\n+ src/include/miscadmin.h:extern void RestoreClientConnectionInfo(char \n*procinfo);\n\nconninfo in both to be consistent?\n\nApart from the comments above, that looks good to me.\n\nRegards,\n\n-- \nBertrand Drouvot\nAmazon Web Services: https://aws.amazon.com\n\n\n\n","msg_date":"Wed, 17 Aug 2022 09:53:45 +0200","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Wed, Aug 17, 2022 at 09:53:45AM +0200, Drouvot, Bertrand wrote:\n> Thanks for the new version!\n> \n> + /* Copy authn_id into the space after the struct. */\n> + if (serialized.authn_id_len >= 0)\n> \n> Maybe remove the \".\" at the end of the comment? (to be consistent with the\n> other comment just above)\n\nWhen it comes to such things, I usually apply the rule of consistency\nwith the surroundings, which sounds right here.\n\n> + memcpy(&serialized, conninfo, sizeof(serialized));\n> + authn_id = conninfo + sizeof(serialized);\n> \n> Move \"authn_id = conninfo + sizeof(serialized)\" in the \"if\n> (serialized.authn_id_len >= 0)\" below?\n\nMakes sense, so as never have something pointing to an area should\nshould not look at. This should just be used when we know that there\nis going to be a string.\n\n> + src/backend/utils/init/miscinit.c:RestoreClientConnectionInfo(char\n> *conninfo)\n> + src/include/miscadmin.h:extern void RestoreClientConnectionInfo(char\n> *procinfo);\n> \n> conninfo in both to be consistent?\n\nYep. Looks like a copy-pasto, seen from here.\n\nBy the way, I have looked at the patch, tweaked a couple of things\nwith comments and the style, but overval that's fine. First, I have\nintended to apply this stuff today but I have lacked the time to do\nso. I should be able to get this wrapped tomorrow, though.\n--\nMichael","msg_date":"Mon, 22 Aug 2022 20:32:40 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Mon, Aug 22, 2022 at 4:32 AM Michael Paquier <michael@paquier.xyz> wrote:\n> By the way, I have looked at the patch, tweaked a couple of things\n> with comments and the style, but overval that's fine. First, I have\n> intended to apply this stuff today but I have lacked the time to do\n> so. I should be able to get this wrapped tomorrow, though.\n\nThank you both for the reviews! Let me know if it would help for me to\nissue a new patchset, otherwise I'll sit tight.\n\n--Jacob\n\n\n","msg_date":"Mon, 22 Aug 2022 08:10:10 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Mon, Aug 22, 2022 at 08:10:10AM -0700, Jacob Champion wrote:\n> otherwise I'll sit tight.\n\nSo am I. I have done an extra round of checks around the\nserialization/deserialization logic where I put some elog()'s to look \nat the output passed down with some workers and a couple of auth\nmethods, and after an indentation and some comment polishing I finish\nwith the attached.\n\nThere was one thing that annoyed me with the patch, though, as of the\nlack of initialization of MyClientConnectionInfo at backend startup,\nas we may finish by not calling set_authn() to fill in some of its\ndata, so I have placed an extra memset(0) in InitProcessGlobals()\n(note that Port does a calloc() much earlier than that, but I think\nthat we don't really want to do more in such code paths, especially\nfor the parallelized client information).\n\nI have written a commit message, while on it. Does that look fine to\nyou?\n--\nMichael","msg_date":"Tue, 23 Aug 2022 11:25:51 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nOn 8/23/22 4:25 AM, Michael Paquier wrote:\n> On Mon, Aug 22, 2022 at 08:10:10AM -0700, Jacob Champion wrote:\n>> otherwise I'll sit tight.\n> So am I. I have done an extra round of checks around the\n> serialization/deserialization logic where I put some elog()'s to look\n> at the output passed down with some workers and a couple of auth\n> methods, and after an indentation and some comment polishing I finish\n> with the attached.\n>\n> There was one thing that annoyed me with the patch, though, as of the\n> lack of initialization of MyClientConnectionInfo at backend startup,\n> as we may finish by not calling set_authn() to fill in some of its\n> data, so I have placed an extra memset(0) in InitProcessGlobals()\nFair point.\n> (note that Port does a calloc() much earlier than that, but I think\n> that we don't really want to do more in such code paths, especially\n> for the parallelized client information).\n>\n> I have written a commit message, while on it. Does that look fine to\n> you?\nThanks!\n\nThat sounds all good to me, except a typo for the author in the commit \nmessage: s/Jocob/Jacob/\n\nRegards,\n\n-- \nBertrand Drouvot\nAmazon Web Services: https://aws.amazon.com\n\n\n\n","msg_date":"Tue, 23 Aug 2022 10:53:14 +0200","msg_from":"\"Drouvot, Bertrand\" <bdrouvot@amazon.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On 8/23/22 01:53, Drouvot, Bertrand wrote:\n> That sounds all good to me, except a typo for the author in the commit \n> message: s/Jocob/Jacob/\n\nThanks, I missed that on my readthrough! :D\n\nPatch looks good to me, too, with one question:\n\n> @@ -2688,6 +2689,7 @@ InitProcessGlobals(void) \n> MyProcPid = getpid(); \n> MyStartTimestamp = GetCurrentTimestamp(); \n> MyStartTime = timestamptz_to_time_t(MyStartTimestamp); \n> + memset(&MyClientConnectionInfo, 0, sizeof(MyClientConnectionInfo)); \n> \n> /* \n> * Set a different global seed in every process. We want something \n\nWhen can we rely on static initialization, and when can't we? Is there a\nconcern that the memory could have been polluted from before the\npostmaster's fork?\n\nThanks,\n--Jacob\n\n\n","msg_date":"Tue, 23 Aug 2022 10:04:30 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"On Tue, Aug 23, 2022 at 10:04:30AM -0700, Jacob Champion wrote:\n> On 8/23/22 01:53, Drouvot, Bertrand wrote:\n>> @@ -2688,6 +2689,7 @@ InitProcessGlobals(void) \n>> MyProcPid = getpid(); \n>> MyStartTimestamp = GetCurrentTimestamp(); \n>> MyStartTime = timestamptz_to_time_t(MyStartTimestamp); \n>> + memset(&MyClientConnectionInfo, 0, sizeof(MyClientConnectionInfo)); \n>> \n>> /* \n>> * Set a different global seed in every process. We want something \n> \n> When can we rely on static initialization, and when can't we? Is there a\n> concern that the memory could have been polluted from before the\n> postmaster's fork?\n\nMy main worry here is EXEC_BACKEND, where we would just use our own\nimplementation of fork(), and it is a bad idea at the end to leave\nthat untouched while we could have code paths that attempt to access\nit. At the end, I have moved the initialization at the same place as\nwhere we set MyProcPort for a backend in BackendInitialize(), mainly\nas a matter of consistency because ClientConnectionInfo is aimed at\nbeing a subset of that. And applied.\n--\nMichael","msg_date":"Wed, 24 Aug 2022 13:06:01 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> On Tue, Aug 23, 2022 at 10:04:30AM -0700, Jacob Champion wrote:\n>> When can we rely on static initialization, and when can't we? Is there a\n>> concern that the memory could have been polluted from before the\n>> postmaster's fork?\n\n> My main worry here is EXEC_BACKEND, where we would just use our own\n> implementation of fork(), and it is a bad idea at the end to leave\n> that untouched while we could have code paths that attempt to access\n> it.\n\nUh ... what? EXEC_BACKEND is even more certain to correctly initialize\nstatic/global variables in a child process. I agree with Jacob that\nthis memset is probably useless, and therefore confusing.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 24 Aug 2022 00:40:54 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"},{"msg_contents":"Hi,\n\nMichael Paquier <michael@paquier.xyz> writes:\n\n> [[PGP Signed Part:Undecided]]\n> On Tue, Aug 23, 2022 at 10:04:30AM -0700, Jacob Champion wrote:\n>\n> My main worry here is EXEC_BACKEND, where we would just use our own\n> implementation of fork(), and it is a bad idea at the end to leave\n> that untouched while we could have code paths that attempt to access\n> it. At the end, I have moved the initialization at the same place as\n> where we set MyProcPort for a backend in BackendInitialize(), mainly\n> as a matter of consistency because ClientConnectionInfo is aimed at\n> being a subset of that. And applied.\n\nI found a compiler complaint of this patch. The attached fix that.\n\n-- \nBest Regards\nAndy Fan","msg_date":"Thu, 15 Feb 2024 16:53:30 +0800","msg_from":"Andy Fan <zhihuifan1213@163.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] Expose port->authn_id to extensions and triggers"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\r\\n\\r\\nStephen pointed out [1] that the authenticated identity that's stored\\r\\nin MyProcPort can't be retrieved by extensions or triggers. Attached is\\r\\na patch that provides both a C API and a SQL function for retrieving\\r\\nit.\\r\\n\\r\\nGetAuthenticatedIdentityString() is a mouthful but I wanted to\\r\\ndifferentiate it from the existing GetAuthenticatedUserId(); better\\r\\nnames welcome. It only exists as an accessor because I wasn't sure if\\r\\nextensions outside of contrib were encouraged to rely on the internal\\r\\nlayout of struct Port. (If they can, then that call can go away\\r\\nentirely.)\\r\\n\\r\\nThanks,\\r\\n--Jacob\\r\\n\\r\\n[1] https://www.postgresql.org/message-id/CAOuzzgpFpuroNRabEvB9kST_TSyS2jFicBNoXvW7G2pZFixyBw%40mail.gmail.com\",\n \"msg_date\": \"Thu, 24 Feb 2022 00:15:40 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 24, 2022 at 12:15:40AM +0000, Jacob Champion wrote:\\n> Stephen pointed out [1] that the authenticated identity that's stored\\n> in MyProcPort can't be retrieved by extensions or triggers. Attached is\\n> a patch that provides both a C API and a SQL function for retrieving\\n> it.\\n> \\n> GetAuthenticatedIdentityString() is a mouthful but I wanted to\\n> differentiate it from the existing GetAuthenticatedUserId(); better\\n> names welcome. It only exists as an accessor because I wasn't sure if\\n> extensions outside of contrib were encouraged to rely on the internal\\n> layout of struct Port. (If they can, then that call can go away\\n> entirely.)\\n\\n+char *\\n+GetAuthenticatedIdentityString(void)\\n+{\\n+ if (!MyProcPort || !MyProcPort->authn_id)\\n+ return NULL;\\n+\\n+ return pstrdup(MyProcPort->authn_id);\\n\\nI don't quite see the additional value that this API brings as\\nMyProcPort is directly accessible, and contrib modules like\\npostgres_fdw and sslinfo just use that to find the data of the current\\nbackend. Cannot a module like pgaudit, through the elog hook, do its\\nwork with what we have already in place?\\n\\nWhat's the use case for a given session to be able to report back\\nonly its own authn through SQL?\\n\\nI could still see a use case for that at a more global level with\\nbeentrys, but it looked like there was not much interest the last time\\nI dropped this idea.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 24 Feb 2022 20:39:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, 2022-02-24 at 20:39 +0900, Michael Paquier wrote:\\r\\n> I don't quite see the additional value that this API brings as\\r\\n> MyProcPort is directly accessible, and contrib modules like\\r\\n> postgres_fdw and sslinfo just use that to find the data of the current\\r\\n> backend.\\r\\n\\r\\nRight -- I just didn't know if third-party modules were actually able\\r\\nto rely on the internal layout of struct Port. Is that guaranteed to\\r\\nremain constant for a major release line? If so, this new API is\\r\\nsuperfluous.\\r\\n\\r\\n> Cannot a module like pgaudit, through the elog hook, do its\\r\\n> work with what we have already in place?\\r\\n\\r\\nGiven the above, I would hope so. Stephen mentioned that pgaudit only\\r\\nhad access to the logged-in role, and when I proposed a miscadmin.h API\\r\\nhe said that would help. CC'ing him to see what he meant; I don't know\\r\\nif pgaudit has additional constraints.\\r\\n\\r\\n> What's the use case for a given session to be able to report back\\r\\n> only its own authn through SQL?\\r\\n\\r\\nThat's for triggers to be able to grab the current ID for\\r\\nlogging/auditing. Is there a better way to expose this for that use\\r\\ncase?\\r\\n\\r\\n> I could still see a use case for that at a more global level with\\r\\n> beentrys, but it looked like there was not much interest the last time\\r\\n> I dropped this idea.\\r\\n\\r\\nI agree that this would be useful to have in the stats. From my outside\\r\\nperspective, it seems like it's difficult to get strings of arbitrary\\r\\nlength in there; is that accurate?\\r\\n\\r\\nThanks,\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 16:50:59 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Thu, Feb 24, 2022 at 04:50:59PM +0000, Jacob Champion wrote:\\n> On Thu, 2022-02-24 at 20:39 +0900, Michael Paquier wrote:\\n> > I don't quite see the additional value that this API brings as\\n> > MyProcPort is directly accessible, and contrib modules like\\n> > postgres_fdw and sslinfo just use that to find the data of the current\\n> > backend.\\n> \\n> Right -- I just didn't know if third-party modules were actually able\\n> to rely on the internal layout of struct Port. Is that guaranteed to\\n> remain constant for a major release line? If so, this new API is\\n> superfluous.\\n\\nYes, third-party can rely on Port layout. We don't break ABI between minor\\nrelease. In some occasions we can add additional fields at the end of a\\nstruct, but nothing more.\\n\\n> > I could still see a use case for that at a more global level with\\n> > beentrys, but it looked like there was not much interest the last time\\n> > I dropped this idea.\\n> \\n> I agree that this would be useful to have in the stats. From my outside\\n> perspective, it seems like it's difficult to get strings of arbitrary\\n> length in there; is that accurate?\\n\\nYes, as it's all in shared memory. The only workaround is using something like\\ntrack_activity_query_size, but it's not great.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 01:15:38 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, 2022-02-25 at 01:15 +0800, Julien Rouhaud wrote:\\r\\n> On Thu, Feb 24, 2022 at 04:50:59PM +0000, Jacob Champion wrote:\\r\\n> > On Thu, 2022-02-24 at 20:39 +0900, Michael Paquier wrote:\\r\\n> > > I don't quite see the additional value that this API brings as\\r\\n> > > MyProcPort is directly accessible, and contrib modules like\\r\\n> > > postgres_fdw and sslinfo just use that to find the data of the current\\r\\n> > > backend.\\r\\n> > \\r\\n> > Right -- I just didn't know if third-party modules were actually able\\r\\n> > to rely on the internal layout of struct Port. Is that guaranteed to\\r\\n> > remain constant for a major release line? If so, this new API is\\r\\n> > superfluous.\\r\\n> \\r\\n> Yes, third-party can rely on Port layout. We don't break ABI between minor\\r\\n> release. In some occasions we can add additional fields at the end of a\\r\\n> struct, but nothing more.\\r\\n\\r\\nThat simplifies things. PFA a smaller v2; if pgaudit can't make use of\\r\\nlibpq-be.h for some reason, then I guess we can tailor a fix to that\\r\\nuse case.\\r\\n\\r\\n> > > I could still see a use case for that at a more global level with\\r\\n> > > beentrys, but it looked like there was not much interest the last time\\r\\n> > > I dropped this idea.\\r\\n> > \\r\\n> > I agree that this would be useful to have in the stats. From my outside\\r\\n> > perspective, it seems like it's difficult to get strings of arbitrary\\r\\n> > length in there; is that accurate?\\r\\n> \\r\\n> Yes, as it's all in shared memory. The only workaround is using something like\\r\\n> track_activity_query_size, but it's not great.\\r\\n\\r\\nYeah... I was following a similar track with the initial work last\\r\\nyear, but I dropped it when the cost of implementation started to grow\\r\\nconsiderably. At the time, though, it looked like some overhauls to the\\r\\nstats framework were being pursued? I should read up on that thread.\\r\\n\\r\\n--Jacob\",\n \"msg_date\": \"Thu, 24 Feb 2022 20:44:08 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 24, 2022 at 08:44:08PM +0000, Jacob Champion wrote:\\n> \\n> Yeah... I was following a similar track with the initial work last\\n> year, but I dropped it when the cost of implementation started to grow\\n> considerably. At the time, though, it looked like some overhauls to the\\n> stats framework were being pursued? I should read up on that thread.\\n\\nDo you mean the shared memory stats patchset? IIUC this is unrelated, as the\\nbeentry stuff Michael was talking about is a different infrastructure\\n(backend_status.[ch]), and I don't think there are any plans to move that to\\ndynamic shared memory.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 13:01:26 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-25 13:01:26 +0800, Julien Rouhaud wrote:\\n> On Thu, Feb 24, 2022 at 08:44:08PM +0000, Jacob Champion wrote:\\n> > \\n> > Yeah... I was following a similar track with the initial work last\\n> > year, but I dropped it when the cost of implementation started to grow\\n> > considerably. At the time, though, it looked like some overhauls to the\\n> > stats framework were being pursued? I should read up on that thread.\\n> \\n> Do you mean the shared memory stats patchset? IIUC this is unrelated, as the\\n> beentry stuff Michael was talking about is a different infrastructure\\n> (backend_status.[ch]), and I don't think there are any plans to move that to\\n> dynamic shared memory.\\n\\nUntil a year ago the backend_status.c stuff was in in pgstat.c - I just split\\nthem out because of the shared memory status work - so it'd not be surprising\\nfor them to mentally be thrown in one bucket.\\n\\nBasically the type of stats we're trying to move to dynamic shared memory is\\nabout counters that should persist for a while and are accumulated across\\nconnections etc. Whereas backend_status.c is more about tracking the current\\nstate (what query is a backend running, what user, etc).\\n\\nThey're not unrelated though: backend_status.c feeds pgstat.c with some\\ninformation occasionally.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 21:18:26 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Thu, Feb 24, 2022 at 09:18:26PM -0800, Andres Freund wrote:\\n> \\n> On 2022-02-25 13:01:26 +0800, Julien Rouhaud wrote:\\n> > On Thu, Feb 24, 2022 at 08:44:08PM +0000, Jacob Champion wrote:\\n> > > \\n> > > Yeah... I was following a similar track with the initial work last\\n> > > year, but I dropped it when the cost of implementation started to grow\\n> > > considerably. At the time, though, it looked like some overhauls to the\\n> > > stats framework were being pursued? I should read up on that thread.\\n> > \\n> > Do you mean the shared memory stats patchset? IIUC this is unrelated, as the\\n> > beentry stuff Michael was talking about is a different infrastructure\\n> > (backend_status.[ch]), and I don't think there are any plans to move that to\\n> > dynamic shared memory.\\n> \\n> Until a year ago the backend_status.c stuff was in in pgstat.c - I just split\\n> them out because of the shared memory status work - so it'd not be surprising\\n> for them to mentally be thrown in one bucket.\\n\\nRight.\\n\\n> Basically the type of stats we're trying to move to dynamic shared memory is\\n> about counters that should persist for a while and are accumulated across\\n> connections etc. Whereas backend_status.c is more about tracking the current\\n> state (what query is a backend running, what user, etc).\\n\\nBut would it be acceptable to use dynamic shared memory in backend_status and\\ne.g. have a dsa_pointer rather than a fixed length array? That seems like a\\nbad idea for query text in general, but for authn_id for instance it seems less\\nlikely to hold gigantic strings, and also have more or less consistent size\\nwhen provided.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 14:25:52 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, 2022-02-24 at 20:44 +0000, Jacob Champion wrote:\\r\\n> That simplifies things. PFA a smaller v2; if pgaudit can't make use of\\r\\n> libpq-be.h for some reason, then I guess we can tailor a fix to that\\r\\n> use case.\\r\\n\\r\\nHa, opr_sanity caught my use of cstring. I'll work on a fix later\\r\\ntoday.\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 16:28:39 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, 2022-02-25 at 16:28 +0000, Jacob Champion wrote:\\r\\n> Ha, opr_sanity caught my use of cstring. I'll work on a fix later\\r\\n> today.\\r\\n\\r\\nFixed in v3.\\r\\n\\r\\n--Jacob\",\n \"msg_date\": \"Fri, 25 Feb 2022 20:19:24 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On 2022-02-25 20:19:24 +0000, Jacob Champion wrote:\\n> From 2fde60a6bc3739f1894c8c264120e4fa0f04df64 Mon Sep 17 00:00:00 2001\\n> From: Jacob Champion <pchampion@vmware.com>\\n> Date: Mon, 14 Feb 2022 08:10:53 -0800\\n> Subject: [PATCH v3] Add API to retrieve authn_id from SQL\\n\\n> The authn_id field in MyProcPort is currently only accessible to the\\n> backend itself. Add a SQL function, session_authn_id(), to expose the\\n> field to triggers that may want to make use of it.\\n\\nLooks to me like authn_id isn't synchronized to parallel workers right now. So\\nthe function will return the wrong thing when executed as part of a parallel\\nquery.\\n\\nI don't think we should add further functions not prefixed with pg_.\\n\\n\\nPerhaps a few tests for less trivial authn_ids could be worthwhile?\\nE.g. certificate DNs.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 13:23:49 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-25 14:25:52 +0800, Julien Rouhaud wrote:\\n> > Basically the type of stats we're trying to move to dynamic shared memory is\\n> > about counters that should persist for a while and are accumulated across\\n> > connections etc. Whereas backend_status.c is more about tracking the current\\n> > state (what query is a backend running, what user, etc).\\n> \\n> But would it be acceptable to use dynamic shared memory in backend_status and\\n> e.g. have a dsa_pointer rather than a fixed length array?\\n\\nMight be OK, but it does add a fair bit of complexity. Suddenly there's a\\nbunch more initialization order dependencies that you don't have right\\nnow. I'd not go there for just this.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 13:26:49 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\\n> Looks to me like authn_id isn't synchronized to parallel workers right now. So\\n> the function will return the wrong thing when executed as part of a parallel\\n> query.\\n\\nFWIW, I am not completely sure what's the use case for being able to\\nsee the authn of the current session through a trigger. We expose\\nthat when log_connections is enabled, for audit purposes. I can also\\nget behind something more central so as one can get a full picture of\\nthe authn used by a bunch of session, particularly with complex HBA\\npolicies, but this looks rather limited to me in usability. Perhaps\\nthat's not enough to stand as an objection, though, and the patch is\\ndead simple.\\n\\n> I don't think we should add further functions not prefixed with pg_.\\n\\nYep.\\n\\n> Perhaps a few tests for less trivial authn_ids could be worthwhile?\\n> E.g. certificate DNs.\\n\\nYes, src/test/ssl would handle that just fine. Now, this stuff\\nalready looks after authn results with log_connections=on, so that\\nfeels like a duplicate.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 26 Feb 2022 14:39:17 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\\r\\n> Looks to me like authn_id isn't synchronized to parallel workers right now. So\\r\\n> the function will return the wrong thing when executed as part of a parallel\\r\\n> query.\\r\\n\\r\\nThanks for the catch. It looks like MyProcPort is left empty, and other\\r\\nfunctions that rely on like inet_server_addr() are marked parallel-\\r\\nrestricted, so I've done the same in v4.\\r\\n\\r\\nOn Sat, 2022-02-26 at 14:39 +0900, Michael Paquier wrote:\\r\\n> FWIW, I am not completely sure what's the use case for being able to\\r\\n> see the authn of the current session through a trigger. We expose\\r\\n> that when log_connections is enabled, for audit purposes. I can also\\r\\n> get behind something more central so as one can get a full picture of\\r\\n> the authn used by a bunch of session, particularly with complex HBA\\r\\n> policies, but this looks rather limited to me in usability. Perhaps\\r\\n> that's not enough to stand as an objection, though, and the patch is\\r\\n> dead simple.\\r\\n\\r\\nI'm primarily motivated by the linked thread -- if the gap between\\r\\nbuiltin roles and authn_id are going to be used as ammo against other\\r\\nsecurity features, then let's close that gap. But I think it's fair to\\r\\nsay that if someone is already using triggers to exhaustively audit a\\r\\ntable, it'd be nice to have this info in the same place too.\\r\\n\\r\\n> > I don't think we should add further functions not prefixed with pg_.\\r\\n> \\r\\n> Yep.\\r\\n\\r\\nFixed.\\r\\n\\r\\n> > Perhaps a few tests for less trivial authn_ids could be worthwhile?\\r\\n> > E.g. certificate DNs.\\r\\n> \\r\\n> Yes, src/test/ssl would handle that just fine. Now, this stuff\\r\\n> already looks after authn results with log_connections=on, so that\\r\\n> feels like a duplicate.\\r\\n\\r\\nIt was easy enough to add, so I added it. I suppose it does protect\\r\\nagainst any reimplementations of pg_session_authn_id() that can't\\r\\nhandle longer ID strings, though I admit that's a stretch.\\r\\n\\r\\nThanks,\\r\\n--Jacob\",\n \"msg_date\": \"Mon, 28 Feb 2022 18:58:00 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Jacob Champion (pchampion@vmware.com) wrote:\\n> On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\\n> > Looks to me like authn_id isn't synchronized to parallel workers right now. So\\n> > the function will return the wrong thing when executed as part of a parallel\\n> > query.\\n> \\n> Thanks for the catch. It looks like MyProcPort is left empty, and other\\n> functions that rely on like inet_server_addr() are marked parallel-\\n> restricted, so I've done the same in v4.\\n\\nThat's probably alright.\\n\\n> On Sat, 2022-02-26 at 14:39 +0900, Michael Paquier wrote:\\n> > FWIW, I am not completely sure what's the use case for being able to\\n> > see the authn of the current session through a trigger. We expose\\n> > that when log_connections is enabled, for audit purposes. I can also\\n> > get behind something more central so as one can get a full picture of\\n> > the authn used by a bunch of session, particularly with complex HBA\\n> > policies, but this looks rather limited to me in usability. Perhaps\\n> > that's not enough to stand as an objection, though, and the patch is\\n> > dead simple.\\n> \\n> I'm primarily motivated by the linked thread -- if the gap between\\n> builtin roles and authn_id are going to be used as ammo against other\\n> security features, then let's close that gap. But I think it's fair to\\n> say that if someone is already using triggers to exhaustively audit a\\n> table, it'd be nice to have this info in the same place too.\\n\\nYeah, we really should make this available to trigger-based auditing\\nsystems too and not just through log_connections which involves a great\\ndeal more log parsing and work external to the database to figure out\\nwho did what.\\n\\n> > > I don't think we should add further functions not prefixed with pg_.\\n> > \\n> > Yep.\\n> \\n> Fixed.\\n\\nThat's fine.\\n\\n> > > Perhaps a few tests for less trivial authn_ids could be worthwhile?\\n> > > E.g. certificate DNs.\\n> > \\n> > Yes, src/test/ssl would handle that just fine. Now, this stuff\\n> > already looks after authn results with log_connections=on, so that\\n> > feels like a duplicate.\\n> \\n> It was easy enough to add, so I added it. I suppose it does protect\\n> against any reimplementations of pg_session_authn_id() that can't\\n> handle longer ID strings, though I admit that's a stretch.\\n> \\n> Thanks,\\n> --Jacob\\n\\n> commit efec9f040843d1de2fc52f5ce0d020478a5bc75d\\n> Author: Jacob Champion <pchampion@vmware.com>\\n> Date: Mon Feb 28 10:28:51 2022 -0800\\n> \\n> squash! Add API to retrieve authn_id from SQL\\n\\nBleh. :) Squash indeed.\\n\\n> Subject: [PATCH v4] Add API to retrieve authn_id from SQL\\n> \\n> The authn_id field in MyProcPort is currently only accessible to the\\n> backend itself. Add a SQL function, pg_session_authn_id(), to expose\\n> the field to triggers that may want to make use of it.\\n\\nOnly did a quick look but generally looks reasonable to me.\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Mon, 28 Feb 2022 16:00:36 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Mon, 2022-02-28 at 16:00 -0500, Stephen Frost wrote:\\r\\n> > commit efec9f040843d1de2fc52f5ce0d020478a5bc75d\\r\\n> > Author: Jacob Champion <pchampion@vmware.com>\\r\\n> > Date: Mon Feb 28 10:28:51 2022 -0800\\r\\n> > \\r\\n> > squash! Add API to retrieve authn_id from SQL\\r\\n> \\r\\n> Bleh. :) Squash indeed.\\r\\n\\r\\nHa, I wasn't sure if anyone read the since-diffs :) I'll start\\r\\nwordsmithing them more in the future.\\r\\n\\r\\n> > Subject: [PATCH v4] Add API to retrieve authn_id from SQL\\r\\n> > \\r\\n> > The authn_id field in MyProcPort is currently only accessible to the\\r\\n> > backend itself. Add a SQL function, pg_session_authn_id(), to expose\\r\\n> > the field to triggers that may want to make use of it.\\r\\n> \\r\\n> Only did a quick look but generally looks reasonable to me.\\r\\n\\r\\nThanks!\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 21:42:54 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 04:00:36PM -0500, Stephen Frost wrote:\\n> * Jacob Champion (pchampion@vmware.com) wrote:\\n>> On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\\n>>> Looks to me like authn_id isn't synchronized to parallel workers right now. So\\n>>> the function will return the wrong thing when executed as part of a parallel\\n>>> query.\\n>> \\n>> Thanks for the catch. It looks like MyProcPort is left empty, and other\\n>> functions that rely on like inet_server_addr() are marked parallel-\\n>> restricted, so I've done the same in v4.\\n> \\n> That's probably alright.\\n\\nI'd say as well that this is right as-is. If it happens that there is\\na use-case for making this parallel aware in the future, it could be\\ndone. Now, it may be a bit weird to make parallel workers inherit the\\nauthn ID of the parent as these did not go through authentication, no?\\nLetting this function being run only by the leader feels intuitive.\\n\\n> Yeah, we really should make this available to trigger-based auditing\\n> systems too and not just through log_connections which involves a great\\n> deal more log parsing and work external to the database to figure out\\n> who did what.\\n\\nOkay, I won't fight hard on that if all of you think that this is\\nuseful for a given session.\\n\\n>> Subject: [PATCH v4] Add API to retrieve authn_id from SQL\\n>> \\n>> The authn_id field in MyProcPort is currently only accessible to the\\n>> backend itself. Add a SQL function, pg_session_authn_id(), to expose\\n>> the field to triggers that may want to make use of it.\\n> \\n> Only did a quick look but generally looks reasonable to me.\\n\\nThe function and the test are fine, pgperltidy complains a bit about\\nthe format of the tests. \\n\\nAyway, this function needs to be documented. I think that you should\\njust add that in \\\"Session Information Functions\\\" in func.sgml, same\\narea as current_user(). The last time we talked about the authn ID,\\none thing we discussed about was how to describe that in a good way to\\nthe user, which is why the section of log_connections was reworked a\\nbit. And we don't have yet any references to what an authenticated\\nidentity is in the docs.\\n\\nThere is no need to update catversion.h in the patch, committers\\nusually take care of that and that's an area of the code that\\nconflicts a lot.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 1 Mar 2022 15:53:02 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Michael Paquier (michael@paquier.xyz) wrote:\\n> On Mon, Feb 28, 2022 at 04:00:36PM -0500, Stephen Frost wrote:\\n> > * Jacob Champion (pchampion@vmware.com) wrote:\\n> >> On Fri, Feb 25, 2022 at 01:23:49PM -0800, Andres Freund wrote:\\n> >>> Looks to me like authn_id isn't synchronized to parallel workers right now. So\\n> >>> the function will return the wrong thing when executed as part of a parallel\\n> >>> query.\\n> >> \\n> >> Thanks for the catch. It looks like MyProcPort is left empty, and other\\n> >> functions that rely on like inet_server_addr() are marked parallel-\\n> >> restricted, so I've done the same in v4.\\n> > \\n> > That's probably alright.\\n> \\n> I'd say as well that this is right as-is. If it happens that there is\\n> a use-case for making this parallel aware in the future, it could be\\n> done. Now, it may be a bit weird to make parallel workers inherit the\\n> authn ID of the parent as these did not go through authentication, no?\\n> Letting this function being run only by the leader feels intuitive.\\n\\nI'm not really sure why we're arguing about this, but clearly the authn\\nID of the leader process is what should be used because that's the\\nauthentication under which the parallel worker is running, just as much\\nas the effective role is the authorization. Having this be available in\\nworker processes would certainly be good as it would allow more query\\nplans to be considered when these functions are used. At this time, I\\ndon't think that outweighs the complications around having it and I'm\\nnot suggesting that Jacob needs to go do that, but surely it would be\\nbetter.\\n\\n> >> Subject: [PATCH v4] Add API to retrieve authn_id from SQL\\n> >> \\n> >> The authn_id field in MyProcPort is currently only accessible to the\\n> >> backend itself. Add a SQL function, pg_session_authn_id(), to expose\\n> >> the field to triggers that may want to make use of it.\\n> > \\n> > Only did a quick look but generally looks reasonable to me.\\n> \\n> The function and the test are fine, pgperltidy complains a bit about\\n> the format of the tests. \\n> \\n> Ayway, this function needs to be documented. I think that you should\\n> just add that in \\\"Session Information Functions\\\" in func.sgml, same\\n> area as current_user(). The last time we talked about the authn ID,\\n> one thing we discussed about was how to describe that in a good way to\\n> the user, which is why the section of log_connections was reworked a\\n> bit. And we don't have yet any references to what an authenticated\\n> identity is in the docs.\\n\\nAgreed that it should be documented and that location seems reasonable\\nto me.\\n\\n> There is no need to update catversion.h in the patch, committers\\n> usually take care of that and that's an area of the code that\\n> conflicts a lot.\\n\\nYeah, best to let committers handle catversion bumps.\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Tue, 1 Mar 2022 08:35:27 -0500\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On 25.02.22 21:19, Jacob Champion wrote:\\n> On Fri, 2022-02-25 at 16:28 +0000, Jacob Champion wrote:\\n>> Ha, opr_sanity caught my use of cstring. I'll work on a fix later\\n>> today.\\n> \\n> Fixed in v3.\\n\\nThis patch contains no documentation. I'm having a hard time \\nunderstanding what the name \\\"session_authn_id\\\" is supposed to convey. \\nThe comment for the Port.authn_id field says this is the \\\"system \\nusername\\\", which sounds like a clearer terminology.\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 19:56:45 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, 2022-03-01 at 08:35 -0500, Stephen Frost wrote:\\r\\n> * Michael Paquier (michael@paquier.xyz) wrote:\\r\\n> > \\r\\n> > Ayway, this function needs to be documented. I think that you should\\r\\n> > just add that in \\\"Session Information Functions\\\" in func.sgml, same\\r\\n> > area as current_user(). The last time we talked about the authn ID,\\r\\n> > one thing we discussed about was how to describe that in a good way to\\r\\n> > the user, which is why the section of log_connections was reworked a\\r\\n> > bit. And we don't have yet any references to what an authenticated\\r\\n> > identity is in the docs.\\r\\n> \\r\\n> Agreed that it should be documented and that location seems reasonable\\r\\n> to me.\\r\\n\\r\\nAdded a first draft in v5, alongside the perltidy fixups mentioned by\\r\\nMichael.\\r\\n\\r\\n> > There is no need to update catversion.h in the patch, committers\\r\\n> > usually take care of that and that's an area of the code that\\r\\n> > conflicts a lot.\\r\\n> \\r\\n> Yeah, best to let committers handle catversion bumps.\\r\\n\\r\\nHeh, that was added for my benefit -- I was tired of forgetting to\\r\\ninitdb after switching dev branches -- but I've dropped it from the\\r\\npatch and will just carry that diff locally.\\r\\n\\r\\nThanks,\\r\\n--Jacob\",\n \"msg_date\": \"Tue, 1 Mar 2022 22:03:20 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, 2022-03-01 at 19:56 +0100, Peter Eisentraut wrote:\\r\\n> This patch contains no documentation. I'm having a hard time \\r\\n> understanding what the name \\\"session_authn_id\\\" is supposed to convey. \\r\\n> The comment for the Port.authn_id field says this is the \\\"system \\r\\n> username\\\", which sounds like a clearer terminology.\\r\\n\\r\\n\\\"System username\\\" may help from an internal development perspective,\\r\\nespecially as it relates to pg_ident.conf, but I don't think that's\\r\\nlikely to be a useful descriptor to an end user. (I don't think of a\\r\\nclient certificate's Subject Distinguished Name as a \\\"system\\r\\nusername\\\".) Does my attempt in v5 help?\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 22:05:37 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 01, 2022 at 10:03:20PM +0000, Jacob Champion wrote:\\n> Added a first draft in v5, alongside the perltidy fixups mentioned by\\n> Michael.\\n\\n+ The authenticated identity is an immutable identifier for the user\\n+ presented during the connection handshake; the exact format depends on\\n+ the authentication method in use. (For example, when using the\\n+ <literal>scram-sha-256</literal> auth method, the authenticated identity\\n+ is simply the username. When using the <literal>cert</literal> auth\\n+ method, the authenticated identity is the Distinguished Name of the\\n+ client certificate.) Even for auth methods which use the username as\\n+ the authenticated identity, this function differs from\\n+ <literal>session_user</literal> in that its return value cannot be\\n+ changed after login.\\n\\nThat looks enough seen from here. Thanks!\\n\\nNit: \\\"auth method\\\" would be a first in the documentation, so this had\\nbetter be \\\"authentication method\\\". (No need to send an updated patch\\njust for that).\\n\\nSo, any comments and/or opinions from others?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 2 Mar 2022 16:09:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On 01.03.22 23:05, Jacob Champion wrote:\\n> On Tue, 2022-03-01 at 19:56 +0100, Peter Eisentraut wrote:\\n>> This patch contains no documentation. I'm having a hard time\\n>> understanding what the name \\\"session_authn_id\\\" is supposed to convey.\\n>> The comment for the Port.authn_id field says this is the \\\"system\\n>> username\\\", which sounds like a clearer terminology.\\n> \\n> \\\"System username\\\" may help from an internal development perspective,\\n> especially as it relates to pg_ident.conf, but I don't think that's\\n> likely to be a useful descriptor to an end user. (I don't think of a\\n> client certificate's Subject Distinguished Name as a \\\"system\\n> username\\\".) Does my attempt in v5 help?\\n\\nYeah, maybe there are better names. But I have no idea what the letter \\ncombination \\\"authn_id\\\" is supposed to stand for. Is it an \\n\\\"authentication identifier\\\"? What does it identify? Maybe I'm missing \\nsomething here, but I don't find it clear.\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 09:18:17 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-01 08:35:27 -0500, Stephen Frost wrote:\\n> I'm not really sure why we're arguing about this, but clearly the authn\\n> ID of the leader process is what should be used because that's the\\n> authentication under which the parallel worker is running, just as much\\n> as the effective role is the authorization. Having this be available in\\n> worker processes would certainly be good as it would allow more query\\n> plans to be considered when these functions are used. At this time, I\\n> don't think that outweighs the complications around having it and I'm\\n> not suggesting that Jacob needs to go do that, but surely it would be\\n> better.\\n\\nI don't think we should commit this without synchronizing the authn between\\nworker / leader (in a separate commit). Too likely that some function that's\\nmarked parallel ok queries the authn_id, opening up a security/monitoring hole\\nor such because of a bogus return value.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 13:27:40 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 02, 2022 at 01:27:40PM -0800, Andres Freund wrote:\\n> I don't think we should commit this without synchronizing the authn between\\n> worker / leader (in a separate commit). Too likely that some function that's\\n> marked parallel ok queries the authn_id, opening up a security/monitoring hole\\n> or such because of a bogus return value.\\n\\nHmm, OK. Using the same authn ID for the leader and the workers still\\nlooks a bit strange to me as the worker is not the one that does the \\nauthentication, only the leader does that. Anyway, FixedParallelState\\nincludes some authentication data passed down by the leader when\\nspawning a worker. So, if we were to pass down the authn, we are\\ngoing to need a new PARALLEL_KEY_* to serialize and restore the data\\npassed down via a DSM like any other states as per the business in\\nparallel.c. Jacob, what do you think?\\n--\\nMichael\",\n \"msg_date\": \"Thu, 3 Mar 2022 16:45:33 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, 2022-03-02 at 09:18 +0100, Peter Eisentraut wrote:\\r\\n> On 01.03.22 23:05, Jacob Champion wrote:\\r\\n> > On Tue, 2022-03-01 at 19:56 +0100, Peter Eisentraut wrote:\\r\\n> > > This patch contains no documentation. I'm having a hard time\\r\\n> > > understanding what the name \\\"session_authn_id\\\" is supposed to convey.\\r\\n> > > The comment for the Port.authn_id field says this is the \\\"system\\r\\n> > > username\\\", which sounds like a clearer terminology.\\r\\n> > \\r\\n> > \\\"System username\\\" may help from an internal development perspective,\\r\\n> > especially as it relates to pg_ident.conf, but I don't think that's\\r\\n> > likely to be a useful descriptor to an end user. (I don't think of a\\r\\n> > client certificate's Subject Distinguished Name as a \\\"system\\r\\n> > username\\\".) Does my attempt in v5 help?\\r\\n> \\r\\n> Yeah, maybe there are better names. But I have no idea what the letter \\r\\n> combination \\\"authn_id\\\" is supposed to stand for. Is it an \\r\\n> \\\"authentication identifier\\\"? What does it identify?\\r\\n\\r\\nAuthenticated identity, but yeah, that's the gist. (\\\"AuthN\\\" being a\\r\\nstandard-ish way to differentiate authentication from \\\"AuthZ\\\"\\r\\nauthorization.)\\r\\n\\r\\nIt's meant to uniquely identify the end user in the case of usermaps,\\r\\nwhere multiple separate entities might log in using the same role. It\\r\\nis distinct from the authorized role name, though they might be exactly\\r\\nthe same in many common setups. And it's not set at all if no\\r\\nauthentication was done.\\r\\n\\r\\n> Maybe I'm missing something here, but I don't find it clear.\\r\\n\\r\\nI just used the internal name, but if we want to make it more clear\\r\\nthen now would be a good time. Do you have any suggestions? Does\\r\\nexpanding the name (pg_session_authenticated_id, or even\\r\\npg_session_authenticated_identity) help?\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Thu, 3 Mar 2022 19:12:11 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, 2022-03-03 at 16:45 +0900, Michael Paquier wrote:\\r\\n> Anyway, FixedParallelState\\r\\n> includes some authentication data passed down by the leader when\\r\\n> spawning a worker. So, if we were to pass down the authn, we are\\r\\n> going to need a new PARALLEL_KEY_* to serialize and restore the data\\r\\n> passed down via a DSM like any other states as per the business in\\r\\n> parallel.c. Jacob, what do you think?\\r\\n\\r\\nI guess it depends on what we want MyProcPort to look like in a\\r\\nparallel worker. Are we comfortable having most of it be blank/useless?\\r\\nDoes it need to be filled in?\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Thu, 3 Mar 2022 19:16:17 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 03, 2022 at 07:16:17PM +0000, Jacob Champion wrote:\\n> I guess it depends on what we want MyProcPort to look like in a\\n> parallel worker. Are we comfortable having most of it be blank/useless?\\n> Does it need to be filled in?\\n\\nGood question. It depends on the definition of how much\\nauthentication information makes sense for the parallel workers to\\ninherit from the parent. And as I mentioned upthread, this definition\\nis not completely clear to me because the parallel workers do *not* go\\nthrough the authentication paths of the parent, they are just spawned\\nin their own dedicated paths that the leader invokes. Inheriting all\\nthis information from the leader has also an impact on the\\nPgBackendStatus entries of the workers as these are reported in\\npg_stat_activity as far as I recall, and it could be confusing to see,\\nfor example, some SSL or some GSS information for automatically\\nspawned processes because these don't use SSL or GSS when they pass\\nback data to the leader.\\n\\nI have been looking at the commit history, and found about 6b7d11f\\nthat switched all the functions of sslinfo to be parallel-restricted\\nespecially because of this. So if we inherit all this information the\\nrestriction on the sslinfo functions could be lifted, though the\\ninterest is honestly limited in this case.\\n\\npostgres_fdw has introduced recently the concept of cached\\nconnections, as of v14 with 411ae64 and 708d165, with a set of\\nparallel-restricted functions. Some of the code paths related to\\nappname look at MyProcPort, so there could be a risk of having some\\ninconsistent information if this is accessed in a parallel worker.\\nLooking at the code, I don't think that it would happen now but\\ncopying some of the data of MyProcPort to the worker could avoid any\\nfuture issues if this code gets extended.\\n\\nAt the end of the day, Port is an interface used for the communication\\nbetween the postmaster with the frontends, so I'd like to say that it\\nis correct to not apply this concept to parallel workers because they\\nare not designed to contact any frontend-side things.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 4 Mar 2022 10:45:24 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, 2022-03-04 at 10:45 +0900, Michael Paquier wrote:\\r\\n> At the end of the day, Port is an interface used for the communication\\r\\n> between the postmaster with the frontends, so I'd like to say that it\\r\\n> is correct to not apply this concept to parallel workers because they\\r\\n> are not designed to contact any frontend-side things.\\r\\n\\r\\nComing back to this late, sorry. I'm not quite sure where to move with\\r\\nthis. I'm considering copying pieces of Port over just so we can see\\r\\nwhat it looks like in practice?\\r\\n\\r\\nPersonally I think it makes sense for the parallel workers to have the\\r\\nauthn information for the client -- in fact there's a lot of\\r\\ninformation that it seems shouldn't be hidden from them -- but there\\r\\nare other pieces, like the socket handle, that are clearly not useful.\\r\\n\\r\\nThanks,\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 22:02:16 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Jacob Champion <pchampion@vmware.com> writes:\\n> On Fri, 2022-03-04 at 10:45 +0900, Michael Paquier wrote:\\n>> At the end of the day, Port is an interface used for the communication\\n>> between the postmaster with the frontends, so I'd like to say that it\\n>> is correct to not apply this concept to parallel workers because they\\n>> are not designed to contact any frontend-side things.\\n\\n> Coming back to this late, sorry. I'm not quite sure where to move with\\n> this. I'm considering copying pieces of Port over just so we can see\\n> what it looks like in practice?\\n\\n> Personally I think it makes sense for the parallel workers to have the\\n> authn information for the client -- in fact there's a lot of\\n> information that it seems shouldn't be hidden from them -- but there\\n> are other pieces, like the socket handle, that are clearly not useful.\\n\\nYeah. It seems to me that putting the auth info into struct Port was\\na fairly random thing to do in the first place, and we are now dealing\\nwith the fallout of that.\\n\\nI think what we ought to do here is separate out the data that we think\\nparallel workers need access to. It does not seem wise to say \\\"workers\\ncan access fields A,B,C of MyPort but not fields X,Y,Z\\\". I do not have\\na concrete proposal for fixing it though.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 18:33:38 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, 2022-03-17 at 18:33 -0400, Tom Lane wrote:\\r\\n> Yeah. It seems to me that putting the auth info into struct Port was\\r\\n> a fairly random thing to do in the first place, and we are now dealing\\r\\n> with the fallout of that.\\r\\n> \\r\\n> I think what we ought to do here is separate out the data that we think\\r\\n> parallel workers need access to. It does not seem wise to say \\\"workers\\r\\n> can access fields A,B,C of MyPort but not fields X,Y,Z\\\". I do not have\\r\\n> a concrete proposal for fixing it though.\\r\\n\\r\\nv6-0002 has my first attempt at this. I moved authn_id into its own\\r\\nsubstruct inside Port, which gets serialized with the parallel key\\r\\nmachinery. (My name selection of \\\"SharedPort\\\" is pretty bland.)\\r\\n\\r\\nOver time, we could move more fields into the shared struct and fill\\r\\nout the serialization logic as needed, and then maybe eventually\\r\\nSharedPort can be broken off into its own thing with its own\\r\\nallocation. But I don't know if we should do it all at once, yet.\\r\\n\\r\\nWDYT?\\r\\n\\r\\n--Jacob\",\n \"msg_date\": \"Wed, 23 Mar 2022 22:52:22 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Jacob Champion <pchampion@vmware.com> writes:\\n> On Thu, 2022-03-17 at 18:33 -0400, Tom Lane wrote:\\n>> I think what we ought to do here is separate out the data that we think\\n>> parallel workers need access to. It does not seem wise to say \\\"workers\\n>> can access fields A,B,C of MyPort but not fields X,Y,Z\\\". I do not have\\n>> a concrete proposal for fixing it though.\\n\\n> v6-0002 has my first attempt at this. I moved authn_id into its own\\n> substruct inside Port, which gets serialized with the parallel key\\n> machinery. (My name selection of \\\"SharedPort\\\" is pretty bland.)\\n\\nHm. I was more envisioning getting the \\\"sharable\\\" info out of Port\\nentirely, although I'm not quite sure where it should go instead.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Mar 2022 19:00:28 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, 2022-03-23 at 19:00 -0400, Tom Lane wrote:\\r\\n> Hm. I was more envisioning getting the \\\"sharable\\\" info out of Port\\r\\n> entirely, although I'm not quite sure where it should go instead.\\r\\n\\r\\nIf it helps, I can move the substruct out and up to a new global struct\\r\\n(MyProcShared?). At this point I think it's mostly search-and-replace.\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Wed, 23 Mar 2022 23:06:14 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-23 23:06:14 +0000, Jacob Champion wrote:\\n> On Wed, 2022-03-23 at 19:00 -0400, Tom Lane wrote:\\n> > Hm. I was more envisioning getting the \\\"sharable\\\" info out of Port\\n> > entirely, although I'm not quite sure where it should go instead.\\n> \\n> If it helps, I can move the substruct out and up to a new global struct\\n> (MyProcShared?). At this point I think it's mostly search-and-replace.\\n\\nPerhaps alongside CurrentUserId etc in miscinit.c? It would be nicer if all\\nthose were together in a struct, but oh well.\\n\\nAnother option would be to make it a GUC. With a bit of care it could be\\nautomatically synced by the existing parallelism infrastructure...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Mar 2022 16:54:36 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, 2022-03-23 at 16:54 -0700, Andres Freund wrote:\\r\\n> On 2022-03-23 23:06:14 +0000, Jacob Champion wrote:\\r\\n> > On Wed, 2022-03-23 at 19:00 -0400, Tom Lane wrote:\\r\\n> > > Hm. I was more envisioning getting the \\\"sharable\\\" info out of Port\\r\\n> > > entirely, although I'm not quite sure where it should go instead.\\r\\n> > \\r\\n> > If it helps, I can move the substruct out and up to a new global struct\\r\\n> > (MyProcShared?). At this point I think it's mostly search-and-replace.\\r\\n> \\r\\n> Perhaps alongside CurrentUserId etc in miscinit.c? It would be nicer if all\\r\\n> those were together in a struct, but oh well.\\r\\n\\r\\nNext draft in v7. My naming choices probably make even less sense now. Any ideas for names for \\\"a bag of stuff that we want parallel workers to have too\\\"?\\r\\n\\r\\n> Another option would be to make it a GUC. With a bit of care it could be\\r\\n> automatically synced by the existing parallelism infrastructure...\\r\\n\\r\\nLike a write-once, PGC_INTERNAL setting? I guess I don't have any\\r\\nintuition on how that would compare to the separate-global-and-accessor \\r\\napproach. Is the primary advantage that you don't have to maintain the\\r\\nserialization logic, or is there more to it?\\r\\n\\r\\nThanks,\\r\\n--Jacob\",\n \"msg_date\": \"Thu, 24 Mar 2022 17:44:06 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 4:27 PM Andres Freund <andres@anarazel.de> wrote:\\n> I don't think we should commit this without synchronizing the authn between\\n> worker / leader (in a separate commit). Too likely that some function that's\\n> marked parallel ok queries the authn_id, opening up a security/monitoring hole\\n> or such because of a bogus return value.\\n\\nIt is not free to copy data from the leader to the worker. I don't\\nthink we should just adopt a policy of copying everything anyone\\nthinks of, because then most of the time we'll be copying a bunch of\\nstuff that really isn't needed.\\n\\nMy gut reaction is to think that this is way too marginal to be worth\\nmaking parallel-safe, but it is also possible that I just don't know\\nenough to understand its true value.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Mar 2022 13:55:29 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-24 13:55:29 -0400, Robert Haas wrote:\\n> On Wed, Mar 2, 2022 at 4:27 PM Andres Freund <andres@anarazel.de> wrote:\\n> > I don't think we should commit this without synchronizing the authn between\\n> > worker / leader (in a separate commit). Too likely that some function that's\\n> > marked parallel ok queries the authn_id, opening up a security/monitoring hole\\n> > or such because of a bogus return value.\\n>\\n> It is not free to copy data from the leader to the worker. I don't\\n> think we should just adopt a policy of copying everything anyone\\n> thinks of, because then most of the time we'll be copying a bunch of\\n> stuff that really isn't needed.\\n\\nI agree.\\n\\n\\n> My gut reaction is to think that this is way too marginal to be worth\\n> making parallel-safe, but it is also possible that I just don't know\\n> enough to understand its true value.\\n\\nMy problem with that is that as far as I can tell the only real use of the\\nfield / function is for stuff like audit logging, RLS etc. Which seems\\nproblematic for two reasons:\\n\\n1) It's likely that the call to the function is nested into other functions,\\n \\\"hiding\\\" the parallel safety. Then it'd return bogus data silently. At\\n the very least we need to make it error out if called in a parallel worker.\\n2) If used for the purposes above, there's basically no parallelism possible\\n anymore.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Mar 2022 11:19:35 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 24, 2022 at 05:44:06PM +0000, Jacob Champion wrote:\\n> On Wed, 2022-03-23 at 16:54 -0700, Andres Freund wrote:\\n>> Another option would be to make it a GUC. With a bit of care it could be\\n>> automatically synced by the existing parallelism infrastructure...\\n> \\n> Like a write-once, PGC_INTERNAL setting? I guess I don't have any\\n> intuition on how that would compare to the separate-global-and-accessor \\n> approach. Is the primary advantage that you don't have to maintain the\\n> serialization logic, or is there more to it?\\n\\nHmm. That would be a first for a GUC, no? It is not seem natural\\ncompared to the other information pieces passed down from the leader\\nto the workers.\\n\\n+extern SharedPort MyProcShared;\\n\\nThis naming is interesting, and seems to be in line with a couple of\\nexecutor structures that share information across workers. Still\\nthat's a bit inconsistent as Shared is used once at the beginning and\\nonce at the end? I don't have a better idea on top of my mind.\\n\\nAnyway, wouldn't it be better to reverse the patch order, introducing\\nthe shared Proc information first and then build the parallel-safe\\nfunction on top of it? \\n--\\nMichael\",\n \"msg_date\": \"Sat, 26 Mar 2022 15:18:59 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-26 15:18:59 +0900, Michael Paquier wrote:\\n> On Thu, Mar 24, 2022 at 05:44:06PM +0000, Jacob Champion wrote:\\n> > On Wed, 2022-03-23 at 16:54 -0700, Andres Freund wrote:\\n> >> Another option would be to make it a GUC. With a bit of care it could be\\n> >> automatically synced by the existing parallelism infrastructure...\\n> >\\n> > Like a write-once, PGC_INTERNAL setting?\\n\\nPerhaps PGC_INTERNAL, perhaps PGC_SU_BACKEND, set with PGC_S_OVERRIDE?\\n\\n\\n> > I guess I don't have any\\n> > intuition on how that would compare to the separate-global-and-accessor\\n> > approach. Is the primary advantage that you don't have to maintain the\\n> > serialization logic, or is there more to it?\\n>\\n> Hmm. That would be a first for a GUC, no? It is not seem natural\\n> compared to the other information pieces passed down from the leader\\n> to the workers.\\n\\nWhat would be the first for a GUC? We have plenty GUCs that are set on a\\nper-connection basis to reflect some fact? And there's several authenitcation\\nrelated bits of state known to guc.c , think role, session_authorization,\\nis_superuser.\\n\\nSharing per-connection state via GUCs for paralellism? I don't think that is\\ntrue either. E.g. application_name, client_encoding.\\n\\n\\n> +extern SharedPort MyProcShared;\\n\\nI strongly dislike MyProcShared. It's way too easily confused with MyProc\\nwhich point to shared memory.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Mar 2022 10:15:50 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2022-03-26 15:18:59 +0900, Michael Paquier wrote:\\n>> On Thu, Mar 24, 2022 at 05:44:06PM +0000, Jacob Champion wrote:\\n>>> On Wed, 2022-03-23 at 16:54 -0700, Andres Freund wrote:\\n>>>> Another option would be to make it a GUC. With a bit of care it could be\\n>>>> automatically synced by the existing parallelism infrastructure...\\n\\n>>> Like a write-once, PGC_INTERNAL setting?\\n\\n> Perhaps PGC_INTERNAL, perhaps PGC_SU_BACKEND, set with PGC_S_OVERRIDE?\\n\\nSeems like making it a GUC does nothing to fix the complaint you had about\\npassing data to workers whether it's needed or not. Sure, we don't then\\nneed to write any new code for it, but it's still new cycles. And it's\\nnew cycles all throughout guc.c, too, not just in parallel worker start.\\n\\nI also note that exposing it as a GUC means we have zero control over\\nwho/what can read it. Maybe that's not a problem, but it needs to be\\nthought about before we go down that path.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Mar 2022 13:57:39 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-26 13:57:39 -0400, Tom Lane wrote:\\n> Seems like making it a GUC does nothing to fix the complaint you had about\\n> passing data to workers whether it's needed or not.\\n\\nI don't think that was my complaint. Maybe Robert's?\\n\\n\\n> Sure, we don't then need to write any new code for it, but it's still new\\n> cycles.\\n\\nI think it'd quite possibly less cycles than separately syncing it.\\n\\nBecause I wanted to know what the overhead be in relation to other things, I\\nmade serialize_variable() log whenever it decides to serialize, and it's a bit\\ndepressing :/.\\n\\nserialized DateStyle = ISO, MDY\\nserialized default_text_search_config = pg_catalog.english\\nserialized force_parallel_mode = on\\nserialized lc_messages = en_US.UTF-8\\nserialized lc_monetary = en_US.UTF-8\\nserialized lc_numeric = en_US.UTF-8\\nserialized lc_time = en_US.UTF-8\\nserialized log_checkpoints = true\\nserialized log_line_prefix = %m [%p][%b][%v:%x][%a]\\nserialized log_timezone = America/Los_Angeles\\nserialized max_stack_depth = 2048\\nserialized max_wal_size = 153600\\nserialized min_wal_size = 48\\nserialized restart_after_crash = true\\nserialized session_authorization = andres\\nserialized ssl_cert_file = /home/andres/tmp/pgdev/ssl-cert-snakeoil.pem\\nserialized ssl_key_file = /home/andres/tmp/pgdev/ssl-cert-snakeoil.key\\nserialized TimeZone = America/Los_Angeles\\nserialized timezone_abbreviations = Default\\nserialized track_io_timing = true\\nserialized transaction_deferrable = false\\nserialized transaction_isolation = read committed\\nserialized transaction_read_only = false\\ntotal serialized guc state is 1324\\n\\nOf course, compared to the total size of 94784 bytes that's not too\\nmuch... FWIW, 65536 of that is for the tuple queues...\\n\\n\\n> I also note that exposing it as a GUC means we have zero control over\\n> who/what can read it. Maybe that's not a problem, but it needs to be\\n> thought about before we go down that path.\\n\\nYes, I think that's a fair concern.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Mar 2022 11:36:37 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Sat, 2022-03-26 at 11:36 -0700, Andres Freund wrote:\\r\\n> > I also note that exposing it as a GUC means we have zero control over\\r\\n> > who/what can read it. Maybe that's not a problem, but it needs to be\\r\\n> > thought about before we go down that path.\\r\\n> \\r\\n> Yes, I think that's a fair concern.\\r\\n\\r\\nI like that there's no builtin way, today, for a superuser to modify\\r\\nthe internal value; it strengthens the use as an auditing tool. Moving\\r\\nthis to a PGC_SU_BACKEND GUC seems to weaken that. And it looks like\\r\\nPGC_INTERNAL is skipped during the serialization, so if we chose that\\r\\noption, we'd need to write new code anyway?\\r\\n\\r\\nWe'd also need to guess whether the GUC system's serialization of NULL\\r\\nas an empty string is likely to cause problems for any future auth\\r\\nmethods. My guess is \\\"no\\\", to be honest, but I do like maintaining the\\r\\ndistinction -- it feels safer.\\r\\n\\r\\nv8 rebases over the recent SSL changes to get the cfbot green again.\\r\\n\\r\\nThanks,\\r\\n--Jacob\",\n \"msg_date\": \"Tue, 29 Mar 2022 23:38:29 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-29 23:38:29 +0000, Jacob Champion wrote:\\n> On Sat, 2022-03-26 at 11:36 -0700, Andres Freund wrote:\\n> > > I also note that exposing it as a GUC means we have zero control over\\n> > > who/what can read it. Maybe that's not a problem, but it needs to be\\n> > > thought about before we go down that path.\\n> > \\n> > Yes, I think that's a fair concern.\\n> \\n> I like that there's no builtin way, today, for a superuser to modify\\n> the internal value; it strengthens the use as an auditing tool. Moving\\n> this to a PGC_SU_BACKEND GUC seems to weaken that. And it looks like\\n> PGC_INTERNAL is skipped during the serialization, so if we chose that\\n> option, we'd need to write new code anyway?\\n\\nIt'd be pretty simple to change can_skip_gucvar()'s selection of what to\\nsync. E.g. an additional flag like GUC_PARALLEL_SYNCHRONIZE.\\n\\nI'm not convinced that a GUC is the answer, to be clear.\\n\\n\\n> We'd also need to guess whether the GUC system's serialization of NULL\\n> as an empty string is likely to cause problems for any future auth\\n> methods.\\n\\nYou can't represent a NULL in a postgres 'text' datum, independent of\\nparallelism. So the current definition of pg_session_authn_id() already\\nprecludes that (and set_authn_id() and ...). Honestly, I can't see a reason\\nwhy we should ever allow authn_id to contain a NULL byte.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 16:53:32 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, 2022-03-29 at 16:53 -0700, Andres Freund wrote:\\r\\n> > We'd also need to guess whether the GUC system's serialization of NULL\\r\\n> > as an empty string is likely to cause problems for any future auth\\r\\n> > methods.\\r\\n> \\r\\n> You can't represent a NULL in a postgres 'text' datum, independent of\\r\\n> parallelism. So the current definition of pg_session_authn_id() already\\r\\n> precludes that (and set_authn_id() and ...). Honestly, I can't see a reason\\r\\n> why we should ever allow authn_id to contain a NULL byte.\\r\\n\\r\\nI don't mean a NULL byte, just a NULL pointer. This part of the\\r\\nimplementation doesn't distinguish between it and an empty string:\\r\\n\\r\\n> /* NULL becomes empty string, see estimate_variable_size() */\\r\\n> do_serialize(destptr, maxbytes, \\\"%s\\\",\\r\\n> *conf->variable ? *conf->variable : \\\"\\\");\\r\\n\\r\\nWhether that's a problem in the future entirely depends on whether\\r\\nthere's some authentication method that considers the empty string a\\r\\nsane and meaningful identity. We might reasonably decide that the\\r\\nanswer is \\\"no\\\", but I like being able to make that decision as opposed\\r\\nto delegating it to an existing generic framework.\\r\\n\\r\\n(That last point may be my core concern about making it a GUC: I'd like\\r\\nus to have full control of how and where this particular piece of\\r\\ninformation gets modified.)\\r\\n\\r\\nThanks,\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 16:02:09 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, 2022-03-29 at 23:38 +0000, Jacob Champion wrote:\\r\\n> v8 rebases over the recent SSL changes to get the cfbot green again.\\r\\n\\r\\nI think the Windows failure [1] is unrelated to this patch, but for\\r\\nposterity:\\r\\n\\r\\n> [03:01:58.925] c:\\\\cirrus>call \\\"C:/Program Files/Git/usr/bin/timeout.exe\\\" -v -k60s 15m perl src/tools/msvc/vcregress.pl recoverycheck \\r\\n> [03:03:16.106] [03:03:16] t/001_stream_rep.pl .................. ok 76551 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\\r\\n> [03:03:16.120] [03:03:16] t/002_archiving.pl ................... ok 0 ms ( 0.02 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.02 CPU)\\r\\n> [03:03:16.128] [03:03:16] t/003_recovery_targets.pl ............ ok 0 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\\r\\n> [03:03:16.138] [03:03:16] t/004_timeline_switch.pl ............. ok 0 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\\r\\n> [03:03:16.141] [03:03:16] t/005_replay_delay.pl ................ ok 0 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\\r\\n> [03:03:24.561] [03:03:24] t/006_logical_decoding.pl ............ ok 8416 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\\r\\n> [03:03:32.496] [03:03:32] t/007_sync_rep.pl .................... ok 7895 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\\r\\n> [03:03:32.496] [03:03:32] t/008_fsm_truncation.pl .............. ok 0 ms ( 0.00 usr 0.00 sys + 0.00 cusr 0.00 csys = 0.00 CPU)\\r\\n> [03:16:58.985] /usr/bin/timeout: sending signal TERM to command ‘perl’\\r\\n\\r\\nThe server and client logs don't quite match up; it looks like we get\\r\\npartway through t/018_wal_optimize.pl, maybe to the end of the\\r\\n`wal_level = minimal, SET TABLESPACE commit subtransaction` test,\\r\\nbefore hanging.\\r\\n\\r\\nI see that there's an active thread about a hang later in the recovery\\r\\nsuite [2]. That's suspicious since 019 is just the next test, but I\\r\\ndon't see any evidence in the logs that we actually started test 019 in\\r\\nthis run.\\r\\n\\r\\n--Jacob\\r\\n\\r\\n[1] https://cirrus-ci.com/task/5434752374145024\\r\\n[2] https://www.postgresql.org/message-id/flat/83b46e5f-2a52-86aa-fa6c-8174908174b8%40iki.fi\\r\\n\",\n \"msg_date\": \"Thu, 31 Mar 2022 21:49:00 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2022 at 04:02:09PM +0000, Jacob Champion wrote:\\n> Whether that's a problem in the future entirely depends on whether\\n> there's some authentication method that considers the empty string a\\n> sane and meaningful identity. We might reasonably decide that the\\n> answer is \\\"no\\\", but I like being able to make that decision as opposed\\n> to delegating it to an existing generic framework.\\n\\nMy guess on the matter is that an empty authn holds the same meaning\\nas NULL because it has no data, but I can see your point as well to\\nmake this distinction. In order to do that, couldn't you just use\\nshm_toc_lookup(noError=true)? PARALLEL_KEY_SHAREDPORT could be an\\noptional entry in the TOC data.\\n\\nThe name choice is still an issue, as per Andres' point that\\nMyProcShared is confusing as it can refer to shared memory. What we\\nwant is a structure name for something that's related to MyProc and\\nshared across all parallel workers including the leader. I would\\ngive up on the \\\"Shared\\\" part, using \\\"Parallel\\\" and \\\"Info\\\" instead.\\nHere are some ideas:\\n- ProcParallelInfo\\n- ProcInfoParallel\\n- ParallelProcInfo\\n--\\nMichael\",\n \"msg_date\": \"Tue, 5 Apr 2022 15:13:41 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, 2022-04-05 at 15:13 +0900, Michael Paquier wrote:\\r\\n> On Wed, Mar 30, 2022 at 04:02:09PM +0000, Jacob Champion wrote:\\r\\n> > Whether that's a problem in the future entirely depends on whether\\r\\n> > there's some authentication method that considers the empty string a\\r\\n> > sane and meaningful identity. We might reasonably decide that the\\r\\n> > answer is \\\"no\\\", but I like being able to make that decision as opposed\\r\\n> > to delegating it to an existing generic framework.\\r\\n> \\r\\n> My guess on the matter is that an empty authn holds the same meaning\\r\\n> as NULL because it has no data,\\r\\n\\r\\nWhether it holds meaning or not depends entirely on the auth method, I\\r\\nthink. Hypothetical example -- a system could accept client\\r\\ncertificates with an empty Subject. What identity that Subject\\r\\nrepresents would depend on the organization, but it's distinct from\\r\\nNULL/unauthenticated because the certificate is still signed by a CA.\\r\\n\\r\\n(Postgres rejects empty Subjects when using clientname=DN and I'm not\\r\\nproposing that we change that; I'm haven't actually checked that\\r\\nthey're RFC-legal. But it's possible that a future auth method could\\r\\nhave a reasonable standard definition for an empty identifier.)\\r\\n\\r\\n> but I can see your point as well to\\r\\n> make this distinction. In order to do that, couldn't you just use\\r\\n> shm_toc_lookup(noError=true)? PARALLEL_KEY_SHAREDPORT could be an\\r\\n> optional entry in the TOC data.\\r\\n\\r\\nThe current patch already handles NULL with a byte of overhead; is\\r\\nthere any advantage to using noError? (It might make things messier\\r\\nonce a second member gets added to the struct.) My concern was directed\\r\\nat the GUC proposal.\\r\\n\\r\\n> The name choice is still an issue, as per Andres' point that\\r\\n> MyProcShared is confusing as it can refer to shared memory. What we\\r\\n> want is a structure name for something that's related to MyProc and\\r\\n> shared across all parallel workers including the leader. I would\\r\\n> give up on the \\\"Shared\\\" part, using \\\"Parallel\\\" and \\\"Info\\\" instead.\\r\\n> Here are some ideas:\\r\\n> - ProcParallelInfo\\r\\n> - ProcInfoParallel\\r\\n> - ParallelProcInfo\\r\\n\\r\\nI like ParallelProcInfo; it reads nicely. I've used that in v9.\\r\\n\\r\\nThanks!\\r\\n--Jacob\",\n \"msg_date\": \"Tue, 5 Apr 2022 18:23:06 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 05, 2022 at 06:23:06PM +0000, Jacob Champion wrote:\\n> Whether it holds meaning or not depends entirely on the auth method, I\\n> think. Hypothetical example -- a system could accept client\\n> certificates with an empty Subject. What identity that Subject\\n> represents would depend on the organization, but it's distinct from\\n> NULL/unauthenticated because the certificate is still signed by a CA.\\n\\nInteresting point.\\n\\n> The current patch already handles NULL with a byte of overhead; is\\n> there any advantage to using noError? (It might make things messier\\n> once a second member gets added to the struct.) My concern was directed\\n> at the GUC proposal.\\n\\nFWIW, I am a bit concerned by this approach because it feels\\ninconsistent with any other conditional fields passed down from the\\nparallel leader to its workers. And if we need to add more fields to \\nParallelProcInfo in the future, it will be cleaner to use different\\nTOC keys to pass down different fields anyway, no?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 6 Apr 2022 20:09:36 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, 2022-04-06 at 20:09 +0900, Michael Paquier wrote:\\r\\n> > The current patch already handles NULL with a byte of overhead; is\\r\\n> > there any advantage to using noError? (It might make things messier\\r\\n> > once a second member gets added to the struct.) My concern was directed\\r\\n> > at the GUC proposal.\\r\\n> \\r\\n> FWIW, I am a bit concerned by this approach because it feels\\r\\n> inconsistent with any other conditional fields passed down from the\\r\\n> parallel leader to its workers. And if we need to add more fields to \\r\\n> ParallelProcInfo in the future, it will be cleaner to use different\\r\\n> TOC keys to pass down different fields anyway, no?\\r\\n\\r\\nI assumed that we would follow the existing model of \\\"(de)serialize a\\r\\nwhole struct\\\", rather than pulling it apart into many separate keys. If\\r\\nit got too complicated then we could consider introducing a\\r\\nSerializedParallelProcInfo struct like some of the other functions do.\\r\\nMaybe that wouldn't work so well if many of the fields are strings?\\r\\n\\r\\nIs there a significant cost to changing this later, if one approach\\r\\nturns out to be wrong?\\r\\n\\r\\n--Jacob\\r\\n\",\n \"msg_date\": \"Wed, 6 Apr 2022 19:16:43 +0000\",\n \"msg_from\": \"Jacob Champion <pchampion@vmware.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 06, 2022 at 07:16:43PM +0000, Jacob Champion wrote:\\n> I assumed that we would follow the existing model of \\\"(de)serialize a\\n> whole struct\\\", rather than pulling it apart into many separate keys. If\\n> it got too complicated then we could consider introducing a\\n> SerializedParallelProcInfo struct like some of the other functions do.\\n> Maybe that wouldn't work so well if many of the fields are strings?\\n> \\n> Is there a significant cost to changing this later, if one approach\\n> turns out to be wrong?\\n\\nI don't think this is going to be an issue as long as we don't change\\nthe definitions of MyParallelProcInfo, Port or PARALLEL_KEY_* in the\\nstable branch. My guess is that we are fine to switch to one approach\\nor the other as this is just some internal communication logic between\\nthe parallel leader and its workers.\\n\\nWhat is the feeling of others about this patch and the introduction of\\nParallelProcInfo (or ParallelPortInfo?) to store the authn coming from\\nPort? The feature freeze is very close.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 7 Apr 2022 09:13:32 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"v10 is rebased over latest; I've also added a PGDLLIMPORT to the new global.\\n\\n--Jacob\",\n \"msg_date\": \"Tue, 31 May 2022 15:21:03 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, May 31, 2022 at 6:21 PM Jacob Champion <jchampion@timescale.com> wrote:\\n> v10 is rebased over latest; I've also added a PGDLLIMPORT to the new global.\\n\\nI took a quick look at this and it doesn't seem crazy to me, except\\nthat I think ParallelProcInfo is a bad name for it. It's kind of\\ngeneric, because neither \\\"proc\\\" nor \\\"info\\\" means a whole lot. It's\\nalso kind of wrong, because I think \\\"parallel\\\" should be things that\\nhave to do with parallelism, not just things that happen to be\\nsynchronized across processes when parallelism is in use. It doesn't\\nmake sense to me to have something called a ParallelProcInfo that is\\nused for every single connection in the universe even if parallelism\\nis completely disabled on the system.\\n\\nI'm not sure what it SHOULD be called, exactly: that's one of the hard\\nproblems in computer science.[1]\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n[1] https://martinfowler.com/bliki/TwoHardThings.html\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Jun 2022 09:51:51 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 2, 2022 at 6:52 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> I'm not sure what it SHOULD be called, exactly: that's one of the hard\\n> problems in computer science.[1]\\n\\nYeah...\\n\\nAll right, here's the full list of previous suggestions, I think:\\n\\n- SharedPort\\n- MyProcShared\\n- ProcParallelInfo\\n- ProcInfoParallel\\n- ParallelProcInfo\\n- ParallelPortInfo\\n\\nI have a few new proposals:\\n\\n- GlobalPortInfo\\n- GlobalConnInfo\\n- Synced[Port/Conn]Info\\n- Worker[Port/Conn]Info (but I think this suffers from the exact same\\nproblem as Parallel)\\n- ThingsThatHappenToBeSynchronizedAcrossProcessesWhenParallelismIsInUse\\n- OrderImporterConsumerTemplateBeanFactory\\n\\nI am struggling to come up with a single adjective that captures this\\nconcept of sometimes-synchronized, that isn't conflicting with\\nexisting uses (like Shared). Other suggestions are very welcome.\\n\\nThanks,\\n--Jacob\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Jun 2022 15:56:28 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 02, 2022 at 03:56:28PM -0700, Jacob Champion wrote:\\n> All right, here's the full list of previous suggestions, I think:\\n> \\n> - SharedPort\\n> - MyProcShared\\n> - ProcParallelInfo\\n> - ProcInfoParallel\\n> - ParallelProcInfo\\n> - ParallelPortInfo\\n> \\n> I have a few new proposals:\\n> \\n> - GlobalPortInfo\\n> - GlobalConnInfo\\n> - Synced[Port/Conn]Info\\n> - Worker[Port/Conn]Info (but I think this suffers from the exact same\\n> problem as Parallel)\\n> - ThingsThatHappenToBeSynchronizedAcrossProcessesWhenParallelismIsInUse\\n> - OrderImporterConsumerTemplateBeanFactory\\n\\nParallelPortInfo sounds kind of right for the job to me in this set of\\nproposals, as the data is from the Port, and that's some information\\nshared between all the parallel workers and the leader.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 3 Jun 2022 09:38:35 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> ParallelPortInfo sounds kind of right for the job to me in this set of\\n> proposals, as the data is from the Port, and that's some information\\n> shared between all the parallel workers and the leader.\\n\\nI agree with Robert's complaint that Parallel is far too generic\\na term here. Also, the fact that this data is currently in struct\\nPort seems like an artifact.\\n\\nDon't we have a term for the set of processes comprising a leader\\nplus parallel workers? If we called that set FooGroup, then\\nsomething like FooGroupSharedInfo would be on-point.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 03 Jun 2022 10:04:12 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, Jun 03, 2022 at 10:04:12AM -0400, Tom Lane wrote:\\n> I agree with Robert's complaint that Parallel is far too generic\\n> a term here. Also, the fact that this data is currently in struct\\n> Port seems like an artifact.\\n> \\n> Don't we have a term for the set of processes comprising a leader\\n> plus parallel workers? If we called that set FooGroup, then\\n> something like FooGroupSharedInfo would be on-point.\\n\\nAs far as I know, proc.h includes the term \\\"group members\\\", which\\nincludes the leader and its workers (see CLOG and lock part)?\\n--\\nMichael\",\n \"msg_date\": \"Sat, 4 Jun 2022 11:36:10 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, Jun 3, 2022 at 7:36 PM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Fri, Jun 03, 2022 at 10:04:12AM -0400, Tom Lane wrote:\\n> > I agree with Robert's complaint that Parallel is far too generic\\n> > a term here. Also, the fact that this data is currently in struct\\n> > Port seems like an artifact.\\n> >\\n> > Don't we have a term for the set of processes comprising a leader\\n> > plus parallel workers? If we called that set FooGroup, then\\n> > something like FooGroupSharedInfo would be on-point.\\n>\\n> As far as I know, proc.h includes the term \\\"group members\\\", which\\n> includes the leader and its workers (see CLOG and lock part)?\\n\\nlmgr/README also refers to \\\"gangs of related processes\\\" and \\\"parallel\\ngroups\\\". So\\n\\n- GroupSharedInfo\\n- ParallelGroupSharedInfo\\n- GangSharedInfo\\n- SharedLeaderInfo\\n\\n?\\n\\n--Jacob\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 Jun 2022 08:44:32 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, Jun 3, 2022 at 10:04 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> I agree with Robert's complaint that Parallel is far too generic\\n> a term here. Also, the fact that this data is currently in struct\\n> Port seems like an artifact.\\n\\nWhy do we call this thing a Port, anyway?\\n\\nI think I'd feel more comfortable here if we were defining what went\\ninto which struct on some semantic basis rather than being like, OK,\\nso all the stuff we want to serialize goes into struct #1, and the\\nstuff we don't want to serialize goes into struct #2. I suppose if\\nit's just based on whether or not we want to serialize it, then the\\nplacement of future additions will just be based on how people happen\\nto feel about the thing they're adding right at that moment, and there\\nwon't be any consistency.\\n\\nOne could imagine dividing the Port struct into a couple of different\\nstructs, e.g.\\n\\nAuthenticationState: stuff that is needed only during authentication\\nand can be discarded thereafter (e.g. the HBA line, at least if the\\ncomment is to be believed)\\nClientCommunicationState: stuff that is used to communicate with the\\nclient but doesn't need to be or can't be shared (e.g. the SSL object\\nitself)\\nClientConnectionInfo: stuff that someone might want to look at for\\ninformation purposes at any time (e.g. authn_id, apparently)\\n\\nThen we could serialize the third of these, keep the second around but\\nnot serialize it, and free the first once connection setup is\\ncomplete.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 6 Jun 2022 14:43:51 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Mon, Jun 6, 2022 at 11:44 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> I think I'd feel more comfortable here if we were defining what went\\n> into which struct on some semantic basis rather than being like, OK,\\n> so all the stuff we want to serialize goes into struct #1, and the\\n> stuff we don't want to serialize goes into struct #2. I suppose if\\n> it's just based on whether or not we want to serialize it, then the\\n> placement of future additions will just be based on how people happen\\n> to feel about the thing they're adding right at that moment, and there\\n> won't be any consistency.\\n\\n\\\"This struct contains connection fields that are explicitly safe for\\nworkers to access\\\" _is_ a useful semantic, in my opinion. And it seems\\nlike it'd make it easier to determine what needs to be included in the\\nstruct; I'm not sure I follow why it would result in less consistency.\\n\\nBut to your suggestion, if we just called the new struct\\n\\\"ClientConnectionInfo\\\", would it be a useful step towards your\\nproposed three-bucket state? I guess I'm having trouble understanding\\nwhy a struct that is defined as \\\"this stuff *doesn't* get serialized\\\"\\nis materially different from having one that's the opposite.\\n\\nThanks,\\n--Jacob\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Jun 2022 15:54:38 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, Jun 7, 2022 at 6:54 PM Jacob Champion <jchampion@timescale.com> wrote:\\n> \\\"This struct contains connection fields that are explicitly safe for\\n> workers to access\\\" _is_ a useful semantic, in my opinion. And it seems\\n> like it'd make it easier to determine what needs to be included in the\\n> struct; I'm not sure I follow why it would result in less consistency.\\n>\\n> But to your suggestion, if we just called the new struct\\n> \\\"ClientConnectionInfo\\\", would it be a useful step towards your\\n> proposed three-bucket state? I guess I'm having trouble understanding\\n> why a struct that is defined as \\\"this stuff *doesn't* get serialized\\\"\\n> is materially different from having one that's the opposite.\\n\\nWell, it isn't, and if my proposal boils down to that, which perhaps\\nit does, then my proposal isn't that great, honestly. Let me try again\\nto explain, though, and maybe it will seem less arbitrary with a\\nsecond explanation -- or maybe it won't.\\n\\nIf we say \\\"this struct contains authentication-related information\\nthat we got from the client and which functions may want to look at\\nlater,\\\" then I feel like the chances are good that when someone adds a\\nnew thing to the system in the future, they will know whether or not\\nthat new thing falls into that category or not. If the definition of a\\nstruct is \\\"everything that should be serialized,\\\" then I feel like the\\nchances are less good that everyone will know whether a new thing they\\nare adding falls into that category or not. Perhaps that is\\nill-founded, but I don't think \\\"should be serialized\\\" is necessarily\\nsomething that everybody is going to have the same view on, or even\\nknow what it means.\\n\\nAlso, I don't think we want to end up with a situation where we have a\\nstruct that contains wildly unrelated things that all need to be\\nserialized. If the definition of the struct is \\\"stuff we should\\nserialize and send to the worker,\\\" well then maybe the transaction\\nsnapshot ought to go in there! Well, no. I mean, we already have a\\nseparate place for that, but suppose somehow we didn't. It doesn't\\nbelong here, because yes the things in this struct get serialized, but\\nit's not only any old thing that needs serializing, it's more specific\\nthan that.\\n\\nI guess what this boils down to is that I really want this thing to\\nhave a meaningful name by means of which a future developer can make a\\nguess as to whether some field they're adding ought to go in there. I\\ntheorize that SharedPort is not too great because (a) Port is already\\na bad name and (b) how am I supposed to know whether my stuff ought to\\nbe shared or not? I like something like ClientConnectionInfo better\\nbecause it seems to describe what the stuff in the struct *is* rather\\nthan what we *do* with it.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Jun 2022 20:45:15 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, Jun 7, 2022 at 5:45 PM Robert Haas <robertmhaas@gmail.com> wrote:\\n> Perhaps that is\\n> ill-founded, but I don't think \\\"should be serialized\\\" is necessarily\\n> something that everybody is going to have the same view on, or even\\n> know what it means.\\n\\nUsing this thread as an example, once it was decided that the parallel\\nworkers needed the additional info, the need for serialization\\nfollowed directly from that. I don't get the feeling that developers\\nare going to jump through the hoops of writing serialization logic for\\na new field in the struct just by accident; they should know why\\nthey're writing that code, and hopefully it would be easy for\\nreviewers to catch a patch that showed up with pointless\\nserialization.\\n\\n> Also, I don't think we want to end up with a situation where we have a\\n> struct that contains wildly unrelated things that all need to be\\n> serialized. If the definition of the struct is \\\"stuff we should\\n> serialize and send to the worker,\\\" well then maybe the transaction\\n> snapshot ought to go in there! Well, no. I mean, we already have a\\n> separate place for that, but suppose somehow we didn't. It doesn't\\n> belong here, because yes the things in this struct get serialized, but\\n> it's not only any old thing that needs serializing, it's more specific\\n> than that.\\n\\nI completely agree with you here -- the name should not be so generic\\nthat it's just a catch-all for any serialized fields that exist.\\n\\n> I guess what this boils down to is that I really want this thing to\\n> have a meaningful name by means of which a future developer can make a\\n> guess as to whether some field they're adding ought to go in there. I\\n> theorize that SharedPort is not too great because (a) Port is already\\n> a bad name and (b) how am I supposed to know whether my stuff ought to\\n> be shared or not? I like something like ClientConnectionInfo better\\n> because it seems to describe what the stuff in the struct *is* rather\\n> than what we *do* with it.\\n\\nI think having both would be useful in this case -- what the stuff is,\\nso that it's clear what doesn't belong in it, and what we do with it,\\nso it's clear that you have to write serialization code if you add new\\nthings. The nature of the struct is such that I think you _have_ to\\nfigure out whether or not your stuff ought to be shared before you\\nhave any business adding it.\\n\\nBut I don't have any better ideas for how to achieve both. I'm fine\\nwith your suggestion of ClientConnectionInfo, if that sounds good to\\nothers; the doc comment can clarify why it differs from Port? Or add\\none of the Shared-/Gang-/Group- prefixes to it, maybe?\\n\\nThanks,\\n--Jacob\\n\\n\\n\",\n \"msg_date\": \"Wed, 8 Jun 2022 16:53:08 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, Jun 8, 2022 at 7:53 PM Jacob Champion <jchampion@timescale.com> wrote:\\n> But I don't have any better ideas for how to achieve both. I'm fine\\n> with your suggestion of ClientConnectionInfo, if that sounds good to\\n> others; the doc comment can clarify why it differs from Port? Or add\\n> one of the Shared-/Gang-/Group- prefixes to it, maybe?\\n\\nI don't like the prefixes, so I'd prefer explaining it in the struct comment.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Jun 2022 09:23:12 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 9, 2022 at 6:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n> On Wed, Jun 8, 2022 at 7:53 PM Jacob Champion <jchampion@timescale.com> wrote:\\n> > But I don't have any better ideas for how to achieve both. I'm fine\\n> > with your suggestion of ClientConnectionInfo, if that sounds good to\\n> > others; the doc comment can clarify why it differs from Port? Or add\\n> > one of the Shared-/Gang-/Group- prefixes to it, maybe?\\n>\\n> I don't like the prefixes, so I'd prefer explaining it in the struct comment.\\n\\nDone that way in v11.\\n\\nThanks!\\n--Jacob\",\n \"msg_date\": \"Fri, 10 Jun 2022 10:58:09 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 6/10/22 7:58 PM, Jacob Champion wrote:\\n> CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.\\n>\\n>\\n>\\n> On Thu, Jun 9, 2022 at 6:23 AM Robert Haas <robertmhaas@gmail.com> wrote:\\n>> On Wed, Jun 8, 2022 at 7:53 PM Jacob Champion <jchampion@timescale.com> wrote:\\n>>> But I don't have any better ideas for how to achieve both. I'm fine\\n>>> with your suggestion of ClientConnectionInfo, if that sounds good to\\n>>> others; the doc comment can clarify why it differs from Port? Or add\\n>>> one of the Shared-/Gang-/Group- prefixes to it, maybe?\\n>> I don't like the prefixes, so I'd prefer explaining it in the struct comment.\\n> Done that way in v11.\\n>\\n> Thanks!\\n> --Jacob\\n\\nFWIW, I just created a new thread to expose the port->authn_id through \\nthe SYSTEM_USER sql reserved word.\\n\\nRegards,\\n\\nBertrand\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 22 Jun 2022 15:31:16 +0200\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On 6/22/22 06:31, Drouvot, Bertrand wrote:\\n> FWIW, I just created a new thread to expose the port->authn_id through \\n> the SYSTEM_USER sql reserved word.\\n\\nReview for both seems to have dried up a bit. I'm not particularly\\ninvested in my code, but I do want to see *a* solution go in. So if it\\nhelps the review momentum for me to withdraw this patch and put my\\neffort into SYSTEM_USER, I can do that no problem.\\n\\nThoughts from prior reviewers? Is SYSTEM_USER the way to go?\\n\\n--Jacob\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Aug 2022 14:57:38 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 8/2/22 11:57 PM, Jacob Champion wrote:\\n> On 6/22/22 06:31, Drouvot, Bertrand wrote:\\n>> FWIW, I just created a new thread to expose the port->authn_id through\\n>> the SYSTEM_USER sql reserved word.\\n> Review for both seems to have dried up a bit. I'm not particularly\\n> invested in my code, but I do want to see *a* solution go in. So if it\\n> helps the review momentum for me to withdraw this patch and put my\\n> effort into SYSTEM_USER, I can do that no problem.\\n>\\n> Thoughts from prior reviewers? Is SYSTEM_USER the way to go?\\n\\nI did not look in detail to this thread, but if the goal is \\\"only\\\" to \\nexpose authn_id (as the subject describes) then it seems to me that \\nSYSTEM_USER [1] is the way to go.\\n\\n[1]: https://commitfest.postgresql.org/39/3703/\\n\\n-- \\nBertrand Drouvot\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 Aug 2022 12:48:33 +0200\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 05, 2022 at 12:48:33PM +0200, Drouvot, Bertrand wrote:\\n> On 8/2/22 11:57 PM, Jacob Champion wrote:\\n>> Thoughts from prior reviewers? Is SYSTEM_USER the way to go?\\n\\nReading through the other thread, there is a clear parallel between\\nboth in concept to provide this information at SQL level, indeed.\\n\\n> I did not look in detail to this thread, but if the goal is \\\"only\\\" to expose\\n> authn_id (as the subject describes) then it seems to me that SYSTEM_USER [1]\\n> is the way to go.\\n> \\n> [1]: https://commitfest.postgresql.org/39/3703/\\n\\nHowever, I am not sure if the suggestion of auth_method:authn as\\noutput generated by SYSTEM_USER would be correct according to the SQL\\nspecification, either. The spec being not really talkative about the\\ndetails of what an external module should be opens up for a lot of\\ninterpretation, something that both thread are dealing with.\\n\\nAnyway, we are talking about two different problems on this thread:\\n1) Provide the authn/SYSTEM_USER through some SQL interface, which is\\nwhat 0001 is an attempt of, SYSTEM_USER is a different attempt.\\n2) Move authn out of Port into its own structure, named\\nClientConnectionInfo, and pass it down to the parallel workers.\\n\\nSYSTEM_USER overlaps with 0001, but I see no reason to not do 0002 in\\nall cases. Even if we are not sure yet of how to expose that at SQL\\nlevel, we still want to pass down this information to the parallel\\nworkers and we still want to not have that in Port. An extension\\ncould also do easily the job once 0002 is done, so I see a good\\nargument about doing it anyway. The name ClientConnectionInfo from\\nRobert looks like the compromise we have for the new structure holding\\nthe information about the client information passed down to the\\nworkers.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 6 Aug 2022 15:26:12 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On 8/6/22 02:26, Michael Paquier wrote:\\n> On Fri, Aug 05, 2022 at 12:48:33PM +0200, Drouvot, Bertrand wrote:\\n>> On 8/2/22 11:57 PM, Jacob Champion wrote:\\n>>> Thoughts from prior reviewers? Is SYSTEM_USER the way to go?\\n> \\n> Reading through the other thread, there is a clear parallel between\\n> both in concept to provide this information at SQL level, indeed.\\n> \\n>> I did not look in detail to this thread, but if the goal is \\\"only\\\" to expose\\n>> authn_id (as the subject describes) then it seems to me that SYSTEM_USER [1]\\n>> is the way to go.\\n>> \\n>> [1]: https://commitfest.postgresql.org/39/3703/\\n> \\n> However, I am not sure if the suggestion of auth_method:authn as\\n> output generated by SYSTEM_USER would be correct according to the SQL\\n> specification, either. The spec being not really talkative about the\\n> details of what an external module should be opens up for a lot of\\n> interpretation, something that both thread are dealing with.\\n\\nAs I pointed out here [ \\nhttps://www.postgresql.org/message-id/28b4a9ef-5103-f117-99e1-99ae5a86a6e8%40joeconway.com \\n] both the SQL Server and Oracle interpretations are similar to the one \\nprovided by Bertrand's patch:\\n\\nSQL Server:\\n \\\"If the current user is logged in to SQL Server by\\n using Windows Authentication, SYSTEM_USER returns the\\n Windows login identification name in the form:\\n DOMAIN\\\\user_login_name. However, if the current user\\n is logged in to SQL Server by using SQL Server\\n Authentication, SYSTEM_USER returns the SQL Server\\n login identification name\\\"\\n\\nOracle:\\n \\\"SYSTEM_USER\\n Returns the name of the current data store user as\\n identified by the operating system.\\\"\\n\\nI am not sure how else we should interpret SYSTEM_USER -- if it isn't \\nport->authn_id what else would you propose it should be?\\n\\n-- \\nJoe Conway\\nRDS Open Source Databases\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Sat, 6 Aug 2022 10:59:26 -0400\",\n \"msg_from\": \"Joe Conway <mail@joeconway.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Sat, Aug 06, 2022 at 10:59:26AM -0400, Joe Conway wrote:\\n> I am not sure how else we should interpret SYSTEM_USER -- if it isn't\\n> port->authn_id what else would you propose it should be?\\n\\nWhat you say sounds rather right, but I was wondering mainly what\\nOracle and SQL server report when it comes to other authentication\\nmethods like SSPI or a cert, where we don't use a user name but some\\ndata dependent on the auth method. And I have no experience with\\nthese.\\n\\nAnyway, I was looking at Bertrand's patch, and I can see that it is\\ndoing nothing to move away the connection information that we have in\\nPort away to a different structure passed down to the parallel\\nworkers, which is what I understand is a cleanup worth on its own\\nbased on the discussion of this thread. Hence, I still see a good\\nargument for the introduction of ClientConnectionInfo that gets passed\\ndown to the workers. Based on that, I think that we'd better finish\\nv11-0002 (only ClientConnectionInfo, no SQL interface) as a first step\\nto build for the next ones, with authn being the first piece of\\ninformation given to the workers. With a separate structure, the\\nauth_method can also be a second member in ClientConnectionInfo,\\ncompleting what would be needed to build SYSTEM_USER as the workers\\nwould have access to it.\\n\\nAm I getting that right?\\n--\\nMichael\",\n \"msg_date\": \"Sun, 7 Aug 2022 16:41:08 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 8/7/22 9:41 AM, Michael Paquier wrote:\\n> Anyway, I was looking at Bertrand's patch, and I can see that it is\\n> doing nothing to move away the connection information that we have in\\n> Port away to a different structure passed down to the parallel\\n> workers,\\n\\nThanks for looking at it!\\n\\nThat's right. The main reason is that in the v2-0003 SYSTEM_USER patch \\nwhat is passed down to the parallel workers is not Port->authn_id but a \\nnew \\\"SystemUser\\\" (defined in miscinit.c with CurrentUserId and friends).\\n\\n> which is what I understand is a cleanup worth on its own\\n> based on the discussion of this thread. Hence, I still see a good\\n> argument for the introduction of ClientConnectionInfo that gets passed\\n> down to the workers.\\n\\nI agree that it could it be useful too.\\n\\n> Based on that, I think that we'd better finish\\n> v11-0002 (only ClientConnectionInfo, no SQL interface)\\n\\nI agree.\\n\\n> as a first step\\n> to build for the next ones, with authn being the first piece of\\n> information given to the workers. With a separate structure, the\\n> auth_method can also be a second member in ClientConnectionInfo,\\n> completing what would be needed to build SYSTEM_USER as the workers\\n> would have access to it.\\n\\nbut I'm not sure we should do it as a first step (given the fact that \\nthis is not Port->authn_id that is passed down to the parallel workers \\nin the SYSTEM_USER patch).\\n\\nWhat do you think about working on both (aka a) v11-002 only \\nClientConnectionInfo and b) SYSTEM_USER) in parallel?\\n\\nThanks\\n\\n-- \\nBertrand Drouvot\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 Aug 2022 12:43:14 +0200\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 08, 2022 at 12:43:14PM +0200, Drouvot, Bertrand wrote:\\n> but I'm not sure we should do it as a first step (given the fact that this\\n> is not Port->authn_id that is passed down to the parallel workers in the\\n> SYSTEM_USER patch).\\n> \\n> What do you think about working on both (aka a) v11-002 only\\n> ClientConnectionInfo and b) SYSTEM_USER) in parallel?\\n\\nIt seems to me that completing ClientConnectionInfo first has the\\nadvantage of not having to tweak twice the interface we are going to\\nuse when passing down the full structure to the workers, so I would\\nchoose for doing it first (with one field for the authn, and a second\\nfield for the auth method so as the the workers can build SYSTEM_USER\\nby themselves when required).\\n--\\nMichael\",\n \"msg_date\": \"Tue, 9 Aug 2022 18:17:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 8/9/22 11:17 AM, Michael Paquier wrote:\\n> On Mon, Aug 08, 2022 at 12:43:14PM +0200, Drouvot, Bertrand wrote:\\n>> but I'm not sure we should do it as a first step (given the fact that this\\n>> is not Port->authn_id that is passed down to the parallel workers in the\\n>> SYSTEM_USER patch).\\n>>\\n>> What do you think about working on both (aka a) v11-002 only\\n>> ClientConnectionInfo and b) SYSTEM_USER) in parallel?\\n> It seems to me that completing ClientConnectionInfo first has the\\n> advantage of not having to tweak twice the interface we are going to\\n> use when passing down the full structure to the workers, so I would\\n> choose for doing it first (with one field for the authn, and a second\\n> field for the auth method so as the the workers can build SYSTEM_USER\\n> by themselves when required).\\n\\nYeah fair point.\\n\\nAgree that it makes sense to work on those patches in this particular \\norder then.\\n\\nThanks,\\n\\n-- \\n\\nBertrand Drouvot\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 Aug 2022 12:39:02 +0200\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 9, 2022 at 3:39 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\\n> Agree that it makes sense to work on those patches in this particular\\n> order then.\\n\\nSounds good. The ClientConnectionInfo patch (previously 0002) is\\nattached, with the SQL function removed.\\n\\nThanks,\\n--Jacob\",\n \"msg_date\": \"Wed, 10 Aug 2022 08:09:58 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 8/10/22 5:09 PM, Jacob Champion wrote:\\n> On Tue, Aug 9, 2022 at 3:39 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\\n>> Agree that it makes sense to work on those patches in this particular\\n>> order then.\\n> Sounds good. The ClientConnectionInfo patch (previously 0002) is\\n> attached, with the SQL function removed.\\n\\nThanks for the patch!\\n\\nLooking at:\\n\\n+typedef struct\\n+{\\n+ /*\\n+ * Authenticated identity. The meaning of this identifier is \\ndependent on\\n+ * hba->auth_method; it is the identity (if any) that the user \\npresented\\n+ * during the authentication cycle, before they were assigned a \\ndatabase\\n+ * role. (It is effectively the \\\"SYSTEM-USERNAME\\\" of a pg_ident \\nusermap\\n+ * -- though the exact string in use may be different, depending on \\npg_hba\\n+ * options.)\\n+ *\\n+ * authn_id is NULL if the user has not actually been \\nauthenticated, for\\n+ * example if the \\\"trust\\\" auth method is in use.\\n+ */\\n+ const char *authn_id;\\n+} ClientConnectionInfo;\\n\\nWhat do you think about adding a second field in ClientConnectionInfo \\nfor the auth method (as suggested by Michael upthread)?\\n\\nThat will be needed by the SYSTEM_USER patch (that its current version \\nimplements as \\\"auth_method:identity\\\").\\n\\nThanks,\\n\\n-- \\nBertrand Drouvot\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 11 Aug 2022 07:47:39 +0200\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 10, 2022 at 10:48 PM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\\n> What do you think about adding a second field in ClientConnectionInfo\\n> for the auth method (as suggested by Michael upthread)?\\n\\nSure -- without a followup patch, it's not really tested, though.\\n\\nv2 adjusts set_authn_id() to copy the auth_method over as well. It\\n\\\"passes tests\\\" but is otherwise unexercised.\\n\\nThanks,\\n--Jacob\",\n \"msg_date\": \"Thu, 11 Aug 2022 15:28:31 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 8/12/22 12:28 AM, Jacob Champion wrote:\\n> On Wed, Aug 10, 2022 at 10:48 PM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\\n>> What do you think about adding a second field in ClientConnectionInfo\\n>> for the auth method (as suggested by Michael upthread)?\\n> Sure -- without a followup patch, it's not really tested, though.\\n>\\n> v2 adjusts set_authn_id() to copy the auth_method over as well. It\\n> \\\"passes tests\\\" but is otherwise unexercised.\\n\\nThank you!\\n\\nTo help with the testing I've just provided a new version (aka \\nv2-0004-system_user-implementation.patch) of the SYSTEM_USER patch in \\n[1] that can be applied on top of \\n\\\"v2-0001-Allow-parallel-workers-to-read-authn_id.patch\\\".\\n\\nBut for this to work, the first comment below on your patch needs to be \\naddressed.\\n\\nOnce the first comment is addressed and the new SYSTEM_USER patch \\napplied (that adds new tap tests) then we can test the propagation to \\nthe parallel workers with:\\n\\nmake -C src/test/kerberos check PROVE_TESTS=t/001_auth.pl PROVE_FLAGS=-v\\n\\nand\\n\\nmake -C src/test/authentication check PROVE_TESTS=t/001_password.pl \\nPROVE_FLAGS=-v\\n\\nBoth are currently successful.\\n\\nRegarding the comments on \\nv2-0001-Allow-parallel-workers-to-read-authn_id.patch:\\n\\n1)\\n\\nThis is the one to be applied before adding the new SYSTEM_USER one on top:\\n\\n+typedef struct\\n+{\\n+ /*\\n+ * Authenticated identity. The meaning of this identifier is \\ndependent on\\n\\nhas to be replaced by:\\n\\n+typedef struct ClientConnectionInfo\\n+{\\n+ /*\\n+ * Authenticated identity. The meaning of this identifier is \\ndependent on\\n\\n2)\\n\\n+ * Authenticated identity. The meaning of this identifier is \\ndependent on\\n\\nThere is one extra space before \\\"The\\\"\\n\\n3)\\n\\n+SerializeClientConnectionInfo(Size maxsize, char *start_address)\\n+{\\n+ /*\\n+ * First byte is an indication of whether or not authn_id has \\nbeen set to\\n+ * non-NULL, to differentiate that case from the empty string.\\n+ */\\n\\nis authn_id being an empty string possible?\\n\\n4)\\n\\n+ */\\n+\\n+ClientConnectionInfo MyClientConnectionInfo;\\n+\\n+/*\\n+ * Calculate the space needed to serialize MyClientConnectionInfo.\\n+ */\\n+Size\\n+EstimateClientConnectionInfoSpace(void)\\n\\n From a coding style point of view, shouldn't \\\"ClientConnectionInfo \\nMyClientConnectionInfo;\\\" be moved to the top of the file?\\n\\n\\n[1]: https://commitfest.postgresql.org/39/3703/\\n\\nRegards,\\n\\n-- \\nBertrand Drouvot\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 12 Aug 2022 15:34:04 +0200\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 12, 2022 at 03:34:04PM +0200, Drouvot, Bertrand wrote:\\n> 3)\\n> \\n> +SerializeClientConnectionInfo(Size maxsize, char *start_address)\\n> +{\\n> + /*\\n> + * First byte is an indication of whether or not authn_id has been\\n> set to\\n> + * non-NULL, to differentiate that case from the empty string.\\n> + */\\n> \\n> is authn_id being an empty string possible?\\n\\nI don't recall that this can be the case yet, but we cannot discard\\nit. One thing was itching me about the serialization and\\ndeserialization logic though: could it be more readable if we used an\\nintermediate structure to store the length of the serialized strings?\\nWe use this approach in other areas, like for the snapshot data in\\nsnapmgr.c. This would handle the case of an empty and NULL string, by\\nstoring -1 as length for NULL and >= 0 for the string length if there\\nis something set, while making the addition of more fields a\\nno-brainer.\\n--\\nMichael\",\n \"msg_date\": \"Sun, 14 Aug 2022 18:57:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 8/14/22 11:57 AM, Michael Paquier wrote:\\n> On Fri, Aug 12, 2022 at 03:34:04PM +0200, Drouvot, Bertrand wrote:\\n>> 3)\\n>>\\n>> +SerializeClientConnectionInfo(Size maxsize, char *start_address)\\n>> +{\\n>> + /*\\n>> + * First byte is an indication of whether or not authn_id has been\\n>> set to\\n>> + * non-NULL, to differentiate that case from the empty string.\\n>> + */\\n>>\\n>> is authn_id being an empty string possible?\\n> I don't recall that this can be the case yet, but we cannot discard\\n> it.\\nFair point.\\n> One thing was itching me about the serialization and\\n> deserialization logic though: could it be more readable if we used an\\n> intermediate structure to store the length of the serialized strings?\\n> We use this approach in other areas, like for the snapshot data in\\n> snapmgr.c. This would handle the case of an empty and NULL string, by\\n> storing -1 as length for NULL and >= 0 for the string length if there\\n> is something set, while making the addition of more fields a\\n> no-brainer.\\n\\nI think that's a good idea and I think that would be more readable (as \\ncompare to storing a \\\"hint\\\" in the first byte).\\n\\nRegards,\\n\\n-- \\nBertrand Drouvot\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Aug 2022 11:02:30 +0200\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hello,\\n\\nOn Fri, Aug 12, 2022 at 6:34 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\\n> +typedef struct\\n> +{\\n> + /*\\n> + * Authenticated identity. The meaning of this identifier is\\n> dependent on\\n>\\n> has to be replaced by:\\n>\\n> +typedef struct ClientConnectionInfo\\n> +{\\n> + /*\\n> + * Authenticated identity. The meaning of this identifier is\\n> dependent on\\n\\nOkay, will do in the next patch (coming soon).\\n\\n> + * Authenticated identity. The meaning of this identifier is\\n> dependent on\\n>\\n> There is one extra space before \\\"The\\\"\\n\\nThis comment block was just moved verbatim; the double-spaced\\nsentences were there before.\\n\\n> From a coding style point of view, shouldn't \\\"ClientConnectionInfo\\n> MyClientConnectionInfo;\\\" be moved to the top of the file?\\n\\nThe style in this file seems to be to declare the variables at the top\\nof the section in which they're used. See the sections for \\\"User ID\\nstate\\\" and \\\"Library preload support\\\".\\n\\nThanks,\\n--Jacob\\n\\n\\n\",\n \"msg_date\": \"Tue, 16 Aug 2022 09:09:03 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 16, 2022 at 2:02 AM Drouvot, Bertrand <bdrouvot@amazon.com> wrote:\\n> On 8/14/22 11:57 AM, Michael Paquier wrote:\\n> > One thing was itching me about the serialization and\\n> > deserialization logic though: could it be more readable if we used an\\n> > intermediate structure to store the length of the serialized strings?\\n> > We use this approach in other areas, like for the snapshot data in\\n> > snapmgr.c. This would handle the case of an empty and NULL string, by\\n> > storing -1 as length for NULL and >= 0 for the string length if there\\n> > is something set, while making the addition of more fields a\\n> > no-brainer.\\n>\\n> I think that's a good idea and I think that would be more readable (as\\n> compare to storing a \\\"hint\\\" in the first byte).\\n\\nSounds good. v3, attached, should make the requested changes:\\n- declare `struct ClientConnectionInfo`\\n- use an intermediate serialization struct\\n- switch to length-\\\"prefixing\\\" for the string\\n\\nI do like the way this reads compared to before.\\n\\nThanks,\\n--Jacob\",\n \"msg_date\": \"Tue, 16 Aug 2022 09:58:28 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 8/16/22 6:58 PM, Jacob Champion wrote:\\n> Sounds good. v3, attached, should make the requested changes:\\n> - declare `struct ClientConnectionInfo`\\n> - use an intermediate serialization struct\\n> - switch to length-\\\"prefixing\\\" for the string\\n>\\n> I do like the way this reads compared to before.\\n\\nThanks for the new version!\\n\\n+ /* Copy authn_id into the space after the struct. */\\n+ if (serialized.authn_id_len >= 0)\\n\\nMaybe remove the \\\".\\\" at the end of the comment? (to be consistent with \\nthe other comment just above)\\n\\n+/*\\n+ * Restore MyClientConnectionInfo from its serialized representation.\\n+ */\\n+void\\n+RestoreClientConnectionInfo(char *conninfo)\\n+{\\n+ SerializedClientConnectionInfo serialized;\\n+ char *authn_id;\\n\\nMove \\\"char *authn_id;\\\" in the \\\"if (serialized.authn_id_len >= 0)\\\" \\nbelow?\\n\\n+\\n+ memcpy(&serialized, conninfo, sizeof(serialized));\\n+ authn_id = conninfo + sizeof(serialized);\\n\\nMove \\\"authn_id = conninfo + sizeof(serialized)\\\" in the \\\"if \\n(serialized.authn_id_len >= 0)\\\" below?\\n\\n+\\n+ /* Copy the fields back into place. */\\n\\nRemove the \\\".\\\" at the end of the comment?\\n\\n+ MyClientConnectionInfo.authn_id = NULL;\\n+ MyClientConnectionInfo.auth_method = serialized.auth_method;\\n+\\n+ if (serialized.authn_id_len >= 0)\\n+ MyClientConnectionInfo.authn_id = \\nMemoryContextStrdup(TopMemoryContext,\\n+ authn_id);\\n\\nThis instead?\\n\\n if (serialized.authn_id_len >= 0)\\n {\\n char *authn_id;\\n authn_id = conninfo + sizeof(serialized);\\n MyClientConnectionInfo.authn_id = \\nMemoryContextStrdup(TopMemoryContext,\\nauthn_id);\\n }\\n\\n+ src/backend/utils/init/miscinit.c:RestoreClientConnectionInfo(char \\n*conninfo)\\n+ src/include/miscadmin.h:extern void RestoreClientConnectionInfo(char \\n*procinfo);\\n\\nconninfo in both to be consistent?\\n\\nApart from the comments above, that looks good to me.\\n\\nRegards,\\n\\n-- \\nBertrand Drouvot\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 Aug 2022 09:53:45 +0200\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Wed, Aug 17, 2022 at 09:53:45AM +0200, Drouvot, Bertrand wrote:\\n> Thanks for the new version!\\n> \\n> + /* Copy authn_id into the space after the struct. */\\n> + if (serialized.authn_id_len >= 0)\\n> \\n> Maybe remove the \\\".\\\" at the end of the comment? (to be consistent with the\\n> other comment just above)\\n\\nWhen it comes to such things, I usually apply the rule of consistency\\nwith the surroundings, which sounds right here.\\n\\n> + memcpy(&serialized, conninfo, sizeof(serialized));\\n> + authn_id = conninfo + sizeof(serialized);\\n> \\n> Move \\\"authn_id = conninfo + sizeof(serialized)\\\" in the \\\"if\\n> (serialized.authn_id_len >= 0)\\\" below?\\n\\nMakes sense, so as never have something pointing to an area should\\nshould not look at. This should just be used when we know that there\\nis going to be a string.\\n\\n> + src/backend/utils/init/miscinit.c:RestoreClientConnectionInfo(char\\n> *conninfo)\\n> + src/include/miscadmin.h:extern void RestoreClientConnectionInfo(char\\n> *procinfo);\\n> \\n> conninfo in both to be consistent?\\n\\nYep. Looks like a copy-pasto, seen from here.\\n\\nBy the way, I have looked at the patch, tweaked a couple of things\\nwith comments and the style, but overval that's fine. First, I have\\nintended to apply this stuff today but I have lacked the time to do\\nso. I should be able to get this wrapped tomorrow, though.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 22 Aug 2022 20:32:40 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 22, 2022 at 4:32 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> By the way, I have looked at the patch, tweaked a couple of things\\n> with comments and the style, but overval that's fine. First, I have\\n> intended to apply this stuff today but I have lacked the time to do\\n> so. I should be able to get this wrapped tomorrow, though.\\n\\nThank you both for the reviews! Let me know if it would help for me to\\nissue a new patchset, otherwise I'll sit tight.\\n\\n--Jacob\\n\\n\\n\",\n \"msg_date\": \"Mon, 22 Aug 2022 08:10:10 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Mon, Aug 22, 2022 at 08:10:10AM -0700, Jacob Champion wrote:\\n> otherwise I'll sit tight.\\n\\nSo am I. I have done an extra round of checks around the\\nserialization/deserialization logic where I put some elog()'s to look \\nat the output passed down with some workers and a couple of auth\\nmethods, and after an indentation and some comment polishing I finish\\nwith the attached.\\n\\nThere was one thing that annoyed me with the patch, though, as of the\\nlack of initialization of MyClientConnectionInfo at backend startup,\\nas we may finish by not calling set_authn() to fill in some of its\\ndata, so I have placed an extra memset(0) in InitProcessGlobals()\\n(note that Port does a calloc() much earlier than that, but I think\\nthat we don't really want to do more in such code paths, especially\\nfor the parallelized client information).\\n\\nI have written a commit message, while on it. Does that look fine to\\nyou?\\n--\\nMichael\",\n \"msg_date\": \"Tue, 23 Aug 2022 11:25:51 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 8/23/22 4:25 AM, Michael Paquier wrote:\\n> On Mon, Aug 22, 2022 at 08:10:10AM -0700, Jacob Champion wrote:\\n>> otherwise I'll sit tight.\\n> So am I. I have done an extra round of checks around the\\n> serialization/deserialization logic where I put some elog()'s to look\\n> at the output passed down with some workers and a couple of auth\\n> methods, and after an indentation and some comment polishing I finish\\n> with the attached.\\n>\\n> There was one thing that annoyed me with the patch, though, as of the\\n> lack of initialization of MyClientConnectionInfo at backend startup,\\n> as we may finish by not calling set_authn() to fill in some of its\\n> data, so I have placed an extra memset(0) in InitProcessGlobals()\\nFair point.\\n> (note that Port does a calloc() much earlier than that, but I think\\n> that we don't really want to do more in such code paths, especially\\n> for the parallelized client information).\\n>\\n> I have written a commit message, while on it. Does that look fine to\\n> you?\\nThanks!\\n\\nThat sounds all good to me, except a typo for the author in the commit \\nmessage: s/Jocob/Jacob/\\n\\nRegards,\\n\\n-- \\nBertrand Drouvot\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Aug 2022 10:53:14 +0200\",\n \"msg_from\": \"\\\"Drouvot, Bertrand\\\" <bdrouvot@amazon.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On 8/23/22 01:53, Drouvot, Bertrand wrote:\\n> That sounds all good to me, except a typo for the author in the commit \\n> message: s/Jocob/Jacob/\\n\\nThanks, I missed that on my readthrough! :D\\n\\nPatch looks good to me, too, with one question:\\n\\n> @@ -2688,6 +2689,7 @@ InitProcessGlobals(void) \\n> MyProcPid = getpid(); \\n> MyStartTimestamp = GetCurrentTimestamp(); \\n> MyStartTime = timestamptz_to_time_t(MyStartTimestamp); \\n> + memset(&MyClientConnectionInfo, 0, sizeof(MyClientConnectionInfo)); \\n> \\n> /* \\n> * Set a different global seed in every process. We want something \\n\\nWhen can we rely on static initialization, and when can't we? Is there a\\nconcern that the memory could have been polluted from before the\\npostmaster's fork?\\n\\nThanks,\\n--Jacob\\n\\n\\n\",\n \"msg_date\": \"Tue, 23 Aug 2022 10:04:30 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"On Tue, Aug 23, 2022 at 10:04:30AM -0700, Jacob Champion wrote:\\n> On 8/23/22 01:53, Drouvot, Bertrand wrote:\\n>> @@ -2688,6 +2689,7 @@ InitProcessGlobals(void) \\n>> MyProcPid = getpid(); \\n>> MyStartTimestamp = GetCurrentTimestamp(); \\n>> MyStartTime = timestamptz_to_time_t(MyStartTimestamp); \\n>> + memset(&MyClientConnectionInfo, 0, sizeof(MyClientConnectionInfo)); \\n>> \\n>> /* \\n>> * Set a different global seed in every process. We want something \\n> \\n> When can we rely on static initialization, and when can't we? Is there a\\n> concern that the memory could have been polluted from before the\\n> postmaster's fork?\\n\\nMy main worry here is EXEC_BACKEND, where we would just use our own\\nimplementation of fork(), and it is a bad idea at the end to leave\\nthat untouched while we could have code paths that attempt to access\\nit. At the end, I have moved the initialization at the same place as\\nwhere we set MyProcPort for a backend in BackendInitialize(), mainly\\nas a matter of consistency because ClientConnectionInfo is aimed at\\nbeing a subset of that. And applied.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 24 Aug 2022 13:06:01 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> On Tue, Aug 23, 2022 at 10:04:30AM -0700, Jacob Champion wrote:\\n>> When can we rely on static initialization, and when can't we? Is there a\\n>> concern that the memory could have been polluted from before the\\n>> postmaster's fork?\\n\\n> My main worry here is EXEC_BACKEND, where we would just use our own\\n> implementation of fork(), and it is a bad idea at the end to leave\\n> that untouched while we could have code paths that attempt to access\\n> it.\\n\\nUh ... what? EXEC_BACKEND is even more certain to correctly initialize\\nstatic/global variables in a child process. I agree with Jacob that\\nthis memset is probably useless, and therefore confusing.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 24 Aug 2022 00:40:54 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nMichael Paquier <michael@paquier.xyz> writes:\\n\\n> [[PGP Signed Part:Undecided]]\\n> On Tue, Aug 23, 2022 at 10:04:30AM -0700, Jacob Champion wrote:\\n>\\n> My main worry here is EXEC_BACKEND, where we would just use our own\\n> implementation of fork(), and it is a bad idea at the end to leave\\n> that untouched while we could have code paths that attempt to access\\n> it. At the end, I have moved the initialization at the same place as\\n> where we set MyProcPort for a backend in BackendInitialize(), mainly\\n> as a matter of consistency because ClientConnectionInfo is aimed at\\n> being a subset of that. And applied.\\n\\nI found a compiler complaint of this patch. The attached fix that.\\n\\n-- \\nBest Regards\\nAndy Fan\",\n \"msg_date\": \"Thu, 15 Feb 2024 16:53:30 +0800\",\n \"msg_from\": \"Andy Fan <zhihuifan1213@163.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] Expose port->authn_id to extensions and triggers\"\n }\n]","isConversational":false}}},{"rowIdx":53571,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nI noticed something odd when going through some\nof the Interval code. The DAYS_PER_YEAR constant\nis defined in src/include/datatype/timestamp.h.\n> #define DAYS_PER_YEAR 365.25 /* assumes leap year every four years */\n\nWe execute the EXTRACT and date_part functions in\nsrc/backend/utils/adt/timestamp.c in\n> static Datum\n> interval_part_common(PG_FUNCTION_ARGS, bool retnumeric)\n\nWhen executing date_part we multiply the total\nyears in the Interval by DAYS_PER_YEAR\n> result += ((double) DAYS_PER_YEAR * SECS_PER_DAY) * (interval->month / MONTHS_PER_YEAR);\n> result += ((double) DAYS_PER_MONTH * SECS_PER_DAY) * (interval->month % MONTHS_PER_YEAR);\n> result += ((double) SECS_PER_DAY) * interval->day;\n\n\nHowever when executing EXTRACT we first truncate\nDAYS_PER_YEAR to an integer, and then multiply it\nby the total years in the Interval\n/* this always fits into int64 */\n> secs_from_day_month = ((int64) DAYS_PER_YEAR * (interval->month / MONTHS_PER_YEAR) +\n> (int64) DAYS_PER_MONTH * (interval->month % MONTHS_PER_YEAR) +\n> interval->day) * SECS_PER_DAY;\n\nIs this truncation on purpose? It seems like\nEXTRACT is not accounting for leap years in\nit's calculation.\n\n- Joe Koshakow\n\n\n","msg_date":"Wed, 23 Feb 2022 19:42:12 -0500","msg_from":"Joseph Koshakow <koshy44@gmail.com>","msg_from_op":true,"msg_subject":"Extract epoch from Interval weird behavior"},{"msg_contents":"On Wed, Feb 23, 2022 at 7:42 PM Joseph Koshakow <koshy44@gmail.com> wrote:\n>\n> Hi all,\n>\n> I noticed something odd when going through some\n> of the Interval code. The DAYS_PER_YEAR constant\n> is defined in src/include/datatype/timestamp.h.\n> > #define DAYS_PER_YEAR 365.25 /* assumes leap year every four years */\n>\n> We execute the EXTRACT and date_part functions in\n> src/backend/utils/adt/timestamp.c in\n> > static Datum\n> > interval_part_common(PG_FUNCTION_ARGS, bool retnumeric)\n>\n> When executing date_part we multiply the total\n> years in the Interval by DAYS_PER_YEAR\n> > result += ((double) DAYS_PER_YEAR * SECS_PER_DAY) * (interval->month / MONTHS_PER_YEAR);\n> > result += ((double) DAYS_PER_MONTH * SECS_PER_DAY) * (interval->month % MONTHS_PER_YEAR);\n> > result += ((double) SECS_PER_DAY) * interval->day;\n>\n>\n> However when executing EXTRACT we first truncate\n> DAYS_PER_YEAR to an integer, and then multiply it\n> by the total years in the Interval\n> /* this always fits into int64 */\n> > secs_from_day_month = ((int64) DAYS_PER_YEAR * (interval->month / MONTHS_PER_YEAR) +\n> > (int64) DAYS_PER_MONTH * (interval->month % MONTHS_PER_YEAR) +\n> > interval->day) * SECS_PER_DAY;\n>\n> Is this truncation on purpose? It seems like\n> EXTRACT is not accounting for leap years in\n> it's calculation.\n>\n> - Joe Koshakow\n\nOops I sent that to the wrong email. If this isn't intented I've created a patch\nthat fixes it, with the following two open questions\n * DAYS_PER_YEAR_NUM is recalculated every time. Is there anyway\nto convert a float directly to a numeric to avoid this?\n * For some reason the change adds a lot of trailing zeros to the result. I'm\nnot sure why that is.\n\n- Joe Koshakow","msg_date":"Wed, 23 Feb 2022 21:35:11 -0500","msg_from":"Joseph Koshakow <koshy44@gmail.com>","msg_from_op":true,"msg_subject":"Re: Extract epoch from Interval weird behavior"},{"msg_contents":"Hi Joseph,\n\n> > Is this truncation on purpose? It seems like\n> > EXTRACT is not accounting for leap years in\n> > it's calculation.\n\nExtracting an epoch from an interval is quite a strange case since\nintervals are not connected to any specific dates.\n\nFor instance:\n\nselect extract('epoch' from interval '1 month')\n\n.. returns 2592000 = 30*24*60*60. But what if the month is February? Should\nwe account for the different number of days for intervals like 6 months or\n24 months?\n\nAlso, leap years don't just happen every 4 years. Here is the actual logic:\n\nbool is_leap_year(int Y) {\n if(Y % 400 == 0) return true;\n else if(Y % 100 == 0) return false;\n else if(Y % 4 == 0) return true;\n else return false;\n}\n\nAnd what about leap seconds?\n\nAll in all, I don't think that the benefit of the proposed change outweighs\nthe fact that it will break the previous behavior for the users who may\nrely on it. I suggest keeping it simple, i.e. the way it is now. What I\nthink we could do instead is explicitly document this behavior in [1].\n\n[1]: https://www.postgresql.org/docs/current/functions-datetime.html\n\n-- \nBest regards,\nAleksander Alekseev\n\nHi Joseph,> > Is this truncation on purpose? It seems like> > EXTRACT is not accounting for leap years in> > it's calculation.Extracting an epoch from an interval is quite a strange case since intervals are not connected to any specific dates.For instance:select extract('epoch' from interval '1 month').. returns 2592000 = 30*24*60*60. But what if the month is February? Should we account for the different number of days for intervals like 6 months or 24 months?Also, leap years don't just happen every 4 years. Here is the actual logic:bool is_leap_year(int Y) { if(Y % 400 == 0) return true; else if(Y % 100 == 0) return false; else if(Y % 4 == 0) return true; else return false;}And what about leap seconds?All in all, I don't think that the benefit of the proposed change outweighs the fact that it will break the previous behavior for the users who may rely on it. I suggest keeping it simple, i.e. the way it is now. What I think we could do instead is explicitly document this behavior in [1].[1]: https://www.postgresql.org/docs/current/functions-datetime.html-- Best regards,Aleksander Alekseev","msg_date":"Thu, 24 Feb 2022 12:47:36 +0300","msg_from":"Aleksander Alekseev <aleksander@timescale.com>","msg_from_op":false,"msg_subject":"Re: Extract epoch from Interval weird behavior"},{"msg_contents":"On Thu, Feb 24, 2022 at 4:47 AM Aleksander Alekseev\n<aleksander@timescale.com> wrote:\n> Extracting an epoch from an interval is quite a strange case since intervals are not connected to any specific dates.\n\nI agree, I think it's a weird use case and that it's probably not\nworth fixing. Though it was fun for me to try.\n\n>\n> All in all, I don't think that the benefit of the proposed change outweighs the fact that it will break the previous behavior for the users who may rely on it. I suggest keeping it simple, i.e. the way it is now. What I think we could do instead is explicitly document this behavior in [1].\n>\n> [1]: https://www.postgresql.org/docs/current/functions-datetime.html\n\nI do want to briefly mention, if I'm understanding the history of\nEXTRACT correctly, that the previous behavior\nactually was to multiply by 365.25, not 365. However The commit that\nchanged the return type from numeric [1]\nchanged that behavior. Looking through the discussions [2], I don't\nsee any mention of it, which makes me think\nit was a mistake. However there is a lot of discussion around numeric\nperformance and being able to optimize\nnumeric division because every divisor was a power of 10. Fixing this\nissue would break that assumption and\ncause some performance degradations which probably isn't worth it.\n\n[1]: https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=a2da77cdb4661826482ebf2ddba1f953bc74afe4\n[2]: https://www.postgresql.org/message-id/flat/42b73d2d-da12-ba9f-570a-420e0cce19d9@phystech.edu\n\n- Joe Koshakow\n\n\n","msg_date":"Thu, 24 Feb 2022 10:17:27 -0500","msg_from":"Joseph Koshakow <koshy44@gmail.com>","msg_from_op":true,"msg_subject":"Re: Extract epoch from Interval weird behavior"},{"msg_contents":"Joseph Koshakow <koshy44@gmail.com> writes:\n> I do want to briefly mention, if I'm understanding the history of\n> EXTRACT correctly, that the previous behavior\n> actually was to multiply by 365.25, not 365. However The commit that\n> changed the return type from numeric [1]\n> changed that behavior. Looking through the discussions [2], I don't\n> see any mention of it, which makes me think\n> it was a mistake.\n\nIndeed ... Peter?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 24 Feb 2022 11:27:37 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Extract epoch from Interval weird behavior"},{"msg_contents":"On 24.02.22 03:35, Joseph Koshakow wrote:\n>> However when executing EXTRACT we first truncate\n>> DAYS_PER_YEAR to an integer, and then multiply it\n>> by the total years in the Interval\n>> /* this always fits into int64 */\n>>> secs_from_day_month = ((int64) DAYS_PER_YEAR * (interval->month / MONTHS_PER_YEAR) +\n>>> (int64) DAYS_PER_MONTH * (interval->month % MONTHS_PER_YEAR) +\n>>> interval->day) * SECS_PER_DAY;\n>> Is this truncation on purpose? It seems like\n>> EXTRACT is not accounting for leap years in\n>> it's calculation.\n\nThis was not intentional. The cast is only to make the multiplication \nhappen in int64; it didn't mean to drop any fractional parts.\n\n> Oops I sent that to the wrong email. If this isn't intented I've created a patch\n> that fixes it, with the following two open questions\n> * DAYS_PER_YEAR_NUM is recalculated every time. Is there anyway\n> to convert a float directly to a numeric to avoid this?\n\nWe really wanted to avoid doing calculations in numeric as much as \npossible. So we should figure out a different way to write this. The \nattached patch works for me. It's a bit ugly since it hardcodes some \nfactors. Maybe we can rephrase it a bit more elegantly.","msg_date":"Fri, 8 Apr 2022 13:38:48 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Extract epoch from Interval weird behavior"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> We really wanted to avoid doing calculations in numeric as much as \n> possible. So we should figure out a different way to write this. The \n> attached patch works for me. It's a bit ugly since it hardcodes some \n> factors. Maybe we can rephrase it a bit more elegantly.\n\nI think it's fine but needs some commentary. Maybe about like\n\"To do this calculation in integer arithmetic even though\nDAYS_PER_YEAR is fractional, multiply everything by 4\nand then divide by 4 again at the end. This relies on\nDAYS_PER_YEAR being a multiple of 0.25 and on SECS_PER_DAY\nbeing a multiple of 4.\"\n\nBTW, it might be good to parenthesize as\n\n(... big calculation ...) * (SECS_PER_DAY/4)\n\nto eliminate any question of whether the value could overflow\nbefore the final division by 4.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 08 Apr 2022 09:10:46 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Extract epoch from Interval weird behavior"},{"msg_contents":"On 08.04.22 15:10, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>> We really wanted to avoid doing calculations in numeric as much as\n>> possible. So we should figure out a different way to write this. The\n>> attached patch works for me. It's a bit ugly since it hardcodes some\n>> factors. Maybe we can rephrase it a bit more elegantly.\n> \n> I think it's fine but needs some commentary. Maybe about like\n> \"To do this calculation in integer arithmetic even though\n> DAYS_PER_YEAR is fractional, multiply everything by 4\n> and then divide by 4 again at the end. This relies on\n> DAYS_PER_YEAR being a multiple of 0.25 and on SECS_PER_DAY\n> being a multiple of 4.\"\n> \n> BTW, it might be good to parenthesize as\n> \n> (... big calculation ...) * (SECS_PER_DAY/4)\n> \n> to eliminate any question of whether the value could overflow\n> before the final division by 4.\n\ndone that way\n\n\n","msg_date":"Tue, 19 Apr 2022 21:12:52 +0200","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Extract epoch from Interval weird behavior"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nI noticed something odd when going through some\\nof the Interval code. The DAYS_PER_YEAR constant\\nis defined in src/include/datatype/timestamp.h.\\n> #define DAYS_PER_YEAR 365.25 /* assumes leap year every four years */\\n\\nWe execute the EXTRACT and date_part functions in\\nsrc/backend/utils/adt/timestamp.c in\\n> static Datum\\n> interval_part_common(PG_FUNCTION_ARGS, bool retnumeric)\\n\\nWhen executing date_part we multiply the total\\nyears in the Interval by DAYS_PER_YEAR\\n> result += ((double) DAYS_PER_YEAR * SECS_PER_DAY) * (interval->month / MONTHS_PER_YEAR);\\n> result += ((double) DAYS_PER_MONTH * SECS_PER_DAY) * (interval->month % MONTHS_PER_YEAR);\\n> result += ((double) SECS_PER_DAY) * interval->day;\\n\\n\\nHowever when executing EXTRACT we first truncate\\nDAYS_PER_YEAR to an integer, and then multiply it\\nby the total years in the Interval\\n/* this always fits into int64 */\\n> secs_from_day_month = ((int64) DAYS_PER_YEAR * (interval->month / MONTHS_PER_YEAR) +\\n> (int64) DAYS_PER_MONTH * (interval->month % MONTHS_PER_YEAR) +\\n> interval->day) * SECS_PER_DAY;\\n\\nIs this truncation on purpose? It seems like\\nEXTRACT is not accounting for leap years in\\nit's calculation.\\n\\n- Joe Koshakow\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Feb 2022 19:42:12 -0500\",\n \"msg_from\": \"Joseph Koshakow <koshy44@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Extract epoch from Interval weird behavior\"\n },\n {\n \"msg_contents\": \"On Wed, Feb 23, 2022 at 7:42 PM Joseph Koshakow <koshy44@gmail.com> wrote:\\n>\\n> Hi all,\\n>\\n> I noticed something odd when going through some\\n> of the Interval code. The DAYS_PER_YEAR constant\\n> is defined in src/include/datatype/timestamp.h.\\n> > #define DAYS_PER_YEAR 365.25 /* assumes leap year every four years */\\n>\\n> We execute the EXTRACT and date_part functions in\\n> src/backend/utils/adt/timestamp.c in\\n> > static Datum\\n> > interval_part_common(PG_FUNCTION_ARGS, bool retnumeric)\\n>\\n> When executing date_part we multiply the total\\n> years in the Interval by DAYS_PER_YEAR\\n> > result += ((double) DAYS_PER_YEAR * SECS_PER_DAY) * (interval->month / MONTHS_PER_YEAR);\\n> > result += ((double) DAYS_PER_MONTH * SECS_PER_DAY) * (interval->month % MONTHS_PER_YEAR);\\n> > result += ((double) SECS_PER_DAY) * interval->day;\\n>\\n>\\n> However when executing EXTRACT we first truncate\\n> DAYS_PER_YEAR to an integer, and then multiply it\\n> by the total years in the Interval\\n> /* this always fits into int64 */\\n> > secs_from_day_month = ((int64) DAYS_PER_YEAR * (interval->month / MONTHS_PER_YEAR) +\\n> > (int64) DAYS_PER_MONTH * (interval->month % MONTHS_PER_YEAR) +\\n> > interval->day) * SECS_PER_DAY;\\n>\\n> Is this truncation on purpose? It seems like\\n> EXTRACT is not accounting for leap years in\\n> it's calculation.\\n>\\n> - Joe Koshakow\\n\\nOops I sent that to the wrong email. If this isn't intented I've created a patch\\nthat fixes it, with the following two open questions\\n * DAYS_PER_YEAR_NUM is recalculated every time. Is there anyway\\nto convert a float directly to a numeric to avoid this?\\n * For some reason the change adds a lot of trailing zeros to the result. I'm\\nnot sure why that is.\\n\\n- Joe Koshakow\",\n \"msg_date\": \"Wed, 23 Feb 2022 21:35:11 -0500\",\n \"msg_from\": \"Joseph Koshakow <koshy44@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extract epoch from Interval weird behavior\"\n },\n {\n \"msg_contents\": \"Hi Joseph,\\n\\n> > Is this truncation on purpose? It seems like\\n> > EXTRACT is not accounting for leap years in\\n> > it's calculation.\\n\\nExtracting an epoch from an interval is quite a strange case since\\nintervals are not connected to any specific dates.\\n\\nFor instance:\\n\\nselect extract('epoch' from interval '1 month')\\n\\n.. returns 2592000 = 30*24*60*60. But what if the month is February? Should\\nwe account for the different number of days for intervals like 6 months or\\n24 months?\\n\\nAlso, leap years don't just happen every 4 years. Here is the actual logic:\\n\\nbool is_leap_year(int Y) {\\n if(Y % 400 == 0) return true;\\n else if(Y % 100 == 0) return false;\\n else if(Y % 4 == 0) return true;\\n else return false;\\n}\\n\\nAnd what about leap seconds?\\n\\nAll in all, I don't think that the benefit of the proposed change outweighs\\nthe fact that it will break the previous behavior for the users who may\\nrely on it. I suggest keeping it simple, i.e. the way it is now. What I\\nthink we could do instead is explicitly document this behavior in [1].\\n\\n[1]: https://www.postgresql.org/docs/current/functions-datetime.html\\n\\n-- \\nBest regards,\\nAleksander Alekseev\\n\\nHi Joseph,> > Is this truncation on purpose? It seems like> > EXTRACT is not accounting for leap years in> > it's calculation.Extracting an epoch from an interval is quite a strange case since intervals are not connected to any specific dates.For instance:select extract('epoch' from interval '1 month').. returns 2592000 = 30*24*60*60. But what if the month is February? Should we account for the different number of days for intervals like 6 months or 24 months?Also, leap years don't just happen every 4 years. Here is the actual logic:bool is_leap_year(int Y) { if(Y % 400 == 0) return true; else if(Y % 100 == 0) return false; else if(Y % 4 == 0) return true; else return false;}And what about leap seconds?All in all, I don't think that the benefit of the proposed change outweighs the fact that it will break the previous behavior for the users who may rely on it. I suggest keeping it simple, i.e. the way it is now. What I think we could do instead is explicitly document this behavior in [1].[1]: https://www.postgresql.org/docs/current/functions-datetime.html-- Best regards,Aleksander Alekseev\",\n \"msg_date\": \"Thu, 24 Feb 2022 12:47:36 +0300\",\n \"msg_from\": \"Aleksander Alekseev <aleksander@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extract epoch from Interval weird behavior\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 24, 2022 at 4:47 AM Aleksander Alekseev\\n<aleksander@timescale.com> wrote:\\n> Extracting an epoch from an interval is quite a strange case since intervals are not connected to any specific dates.\\n\\nI agree, I think it's a weird use case and that it's probably not\\nworth fixing. Though it was fun for me to try.\\n\\n>\\n> All in all, I don't think that the benefit of the proposed change outweighs the fact that it will break the previous behavior for the users who may rely on it. I suggest keeping it simple, i.e. the way it is now. What I think we could do instead is explicitly document this behavior in [1].\\n>\\n> [1]: https://www.postgresql.org/docs/current/functions-datetime.html\\n\\nI do want to briefly mention, if I'm understanding the history of\\nEXTRACT correctly, that the previous behavior\\nactually was to multiply by 365.25, not 365. However The commit that\\nchanged the return type from numeric [1]\\nchanged that behavior. Looking through the discussions [2], I don't\\nsee any mention of it, which makes me think\\nit was a mistake. However there is a lot of discussion around numeric\\nperformance and being able to optimize\\nnumeric division because every divisor was a power of 10. Fixing this\\nissue would break that assumption and\\ncause some performance degradations which probably isn't worth it.\\n\\n[1]: https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=a2da77cdb4661826482ebf2ddba1f953bc74afe4\\n[2]: https://www.postgresql.org/message-id/flat/42b73d2d-da12-ba9f-570a-420e0cce19d9@phystech.edu\\n\\n- Joe Koshakow\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 10:17:27 -0500\",\n \"msg_from\": \"Joseph Koshakow <koshy44@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Extract epoch from Interval weird behavior\"\n },\n {\n \"msg_contents\": \"Joseph Koshakow <koshy44@gmail.com> writes:\\n> I do want to briefly mention, if I'm understanding the history of\\n> EXTRACT correctly, that the previous behavior\\n> actually was to multiply by 365.25, not 365. However The commit that\\n> changed the return type from numeric [1]\\n> changed that behavior. Looking through the discussions [2], I don't\\n> see any mention of it, which makes me think\\n> it was a mistake.\\n\\nIndeed ... Peter?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 11:27:37 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extract epoch from Interval weird behavior\"\n },\n {\n \"msg_contents\": \"On 24.02.22 03:35, Joseph Koshakow wrote:\\n>> However when executing EXTRACT we first truncate\\n>> DAYS_PER_YEAR to an integer, and then multiply it\\n>> by the total years in the Interval\\n>> /* this always fits into int64 */\\n>>> secs_from_day_month = ((int64) DAYS_PER_YEAR * (interval->month / MONTHS_PER_YEAR) +\\n>>> (int64) DAYS_PER_MONTH * (interval->month % MONTHS_PER_YEAR) +\\n>>> interval->day) * SECS_PER_DAY;\\n>> Is this truncation on purpose? It seems like\\n>> EXTRACT is not accounting for leap years in\\n>> it's calculation.\\n\\nThis was not intentional. The cast is only to make the multiplication \\nhappen in int64; it didn't mean to drop any fractional parts.\\n\\n> Oops I sent that to the wrong email. If this isn't intented I've created a patch\\n> that fixes it, with the following two open questions\\n> * DAYS_PER_YEAR_NUM is recalculated every time. Is there anyway\\n> to convert a float directly to a numeric to avoid this?\\n\\nWe really wanted to avoid doing calculations in numeric as much as \\npossible. So we should figure out a different way to write this. The \\nattached patch works for me. It's a bit ugly since it hardcodes some \\nfactors. Maybe we can rephrase it a bit more elegantly.\",\n \"msg_date\": \"Fri, 8 Apr 2022 13:38:48 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extract epoch from Interval weird behavior\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> We really wanted to avoid doing calculations in numeric as much as \\n> possible. So we should figure out a different way to write this. The \\n> attached patch works for me. It's a bit ugly since it hardcodes some \\n> factors. Maybe we can rephrase it a bit more elegantly.\\n\\nI think it's fine but needs some commentary. Maybe about like\\n\\\"To do this calculation in integer arithmetic even though\\nDAYS_PER_YEAR is fractional, multiply everything by 4\\nand then divide by 4 again at the end. This relies on\\nDAYS_PER_YEAR being a multiple of 0.25 and on SECS_PER_DAY\\nbeing a multiple of 4.\\\"\\n\\nBTW, it might be good to parenthesize as\\n\\n(... big calculation ...) * (SECS_PER_DAY/4)\\n\\nto eliminate any question of whether the value could overflow\\nbefore the final division by 4.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 08 Apr 2022 09:10:46 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extract epoch from Interval weird behavior\"\n },\n {\n \"msg_contents\": \"On 08.04.22 15:10, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>> We really wanted to avoid doing calculations in numeric as much as\\n>> possible. So we should figure out a different way to write this. The\\n>> attached patch works for me. It's a bit ugly since it hardcodes some\\n>> factors. Maybe we can rephrase it a bit more elegantly.\\n> \\n> I think it's fine but needs some commentary. Maybe about like\\n> \\\"To do this calculation in integer arithmetic even though\\n> DAYS_PER_YEAR is fractional, multiply everything by 4\\n> and then divide by 4 again at the end. This relies on\\n> DAYS_PER_YEAR being a multiple of 0.25 and on SECS_PER_DAY\\n> being a multiple of 4.\\\"\\n> \\n> BTW, it might be good to parenthesize as\\n> \\n> (... big calculation ...) * (SECS_PER_DAY/4)\\n> \\n> to eliminate any question of whether the value could overflow\\n> before the final division by 4.\\n\\ndone that way\\n\\n\\n\",\n \"msg_date\": \"Tue, 19 Apr 2022 21:12:52 +0200\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Extract epoch from Interval weird behavior\"\n }\n]","isConversational":false}}},{"rowIdx":53572,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI found messages inserted a space before the \"%\" in pgbench.\nI think this is typo because there are no space before the \"%\" in other messages.\nWhat do you think? \n\ndiff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c\nindex f166a77e3a..4ebe5e6ea4 100644\n--- a/src/bin/pgbench/pgbench.c\n+++ b/src/bin/pgbench/pgbench.c\n@@ -5598,11 +5598,11 @@ printResults(StatsData *total,\n \t\treturn;\n \n \tif (throttle_delay && latency_limit)\n-\t\tprintf(\"number of transactions skipped: \" INT64_FORMAT \" (%.3f %%)\\n\",\n+\t\tprintf(\"number of transactions skipped: \" INT64_FORMAT \" (%.3f%%)\\n\",\n \t\t\t total->skipped, 100.0 * total->skipped / total->cnt);\n \n \tif (latency_limit)\n-\t\tprintf(\"number of transactions above the %.1f ms latency limit: \" INT64_FORMAT \"/\" INT64_FORMAT \" (%.3f %%)\\n\",\n+\t\tprintf(\"number of transactions above the %.1f ms latency limit: \" INT64_FORMAT \"/\" INT64_FORMAT \" (%.3f%%)\\n\",\n \t\t\t latency_limit / 1000.0, latency_late, ntx,\n \t\t\t (ntx > 0) ? 100.0 * latency_late / ntx : 0.0);\n\n-- \nKAWAMOTO Masaya <kawamoto@sraoss.co.jp>\nSRA OSS, Inc. Japan\n\n\n","msg_date":"Thu, 24 Feb 2022 11:56:22 +0900","msg_from":"KAWAMOTO Masaya <kawamoto@sraoss.co.jp>","msg_from_op":true,"msg_subject":"Typo in pgbench messages."},{"msg_contents":"> Hi,\n> \n> I found messages inserted a space before the \"%\" in pgbench.\n> I think this is typo because there are no space before the \"%\" in other messages.\n> What do you think? \n\nI think you are right. In English there's should be no space between number and \"%\".\nAFAIK other parts of PostgreSQL follow the rule.\n\n> diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c\n> index f166a77e3a..4ebe5e6ea4 100644\n> --- a/src/bin/pgbench/pgbench.c\n> +++ b/src/bin/pgbench/pgbench.c\n> @@ -5598,11 +5598,11 @@ printResults(StatsData *total,\n> \t\treturn;\n> \n> \tif (throttle_delay && latency_limit)\n> -\t\tprintf(\"number of transactions skipped: \" INT64_FORMAT \" (%.3f %%)\\n\",\n> +\t\tprintf(\"number of transactions skipped: \" INT64_FORMAT \" (%.3f%%)\\n\",\n> \t\t\t total->skipped, 100.0 * total->skipped / total->cnt);\n> \n> \tif (latency_limit)\n> -\t\tprintf(\"number of transactions above the %.1f ms latency limit: \" INT64_FORMAT \"/\" INT64_FORMAT \" (%.3f %%)\\n\",\n> +\t\tprintf(\"number of transactions above the %.1f ms latency limit: \" INT64_FORMAT \"/\" INT64_FORMAT \" (%.3f%%)\\n\",\n> \t\t\t latency_limit / 1000.0, latency_late, ntx,\n> \t\t\t (ntx > 0) ? 100.0 * latency_late / ntx : 0.0);\n\nLooks good to me.\n\nBest reagards,\n--\nTatsuo Ishii\nSRA OSS, Inc. Japan\nEnglish: http://www.sraoss.co.jp/index_en.php\nJapanese:http://www.sraoss.co.jp\n\n\n","msg_date":"Thu, 24 Feb 2022 12:15:55 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@sraoss.co.jp>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":"Thanks for your comment!\n\nSorry, I did not attach the patch file.\nThis patch focas on master branch.\n\nBest regards,\n\nOn Thu, 24 Feb 2022 12:15:55 +0900 (JST)\nTatsuo Ishii <ishii@sraoss.co.jp> wrote:\n\n> > Hi,\n> > \n> > I found messages inserted a space before the \"%\" in pgbench.\n> > I think this is typo because there are no space before the \"%\" in other messages.\n> > What do you think? \n> \n> I think you are right. In English there's should be no space between number and \"%\".\n> AFAIK other parts of PostgreSQL follow the rule.\n> \n> > diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c\n> > index f166a77e3a..4ebe5e6ea4 100644\n> > --- a/src/bin/pgbench/pgbench.c\n> > +++ b/src/bin/pgbench/pgbench.c\n> > @@ -5598,11 +5598,11 @@ printResults(StatsData *total,\n> > \t\treturn;\n> > \n> > \tif (throttle_delay && latency_limit)\n> > -\t\tprintf(\"number of transactions skipped: \" INT64_FORMAT \" (%.3f %%)\\n\",\n> > +\t\tprintf(\"number of transactions skipped: \" INT64_FORMAT \" (%.3f%%)\\n\",\n> > \t\t\t total->skipped, 100.0 * total->skipped / total->cnt);\n> > \n> > \tif (latency_limit)\n> > -\t\tprintf(\"number of transactions above the %.1f ms latency limit: \" INT64_FORMAT \"/\" INT64_FORMAT \" (%.3f %%)\\n\",\n> > +\t\tprintf(\"number of transactions above the %.1f ms latency limit: \" INT64_FORMAT \"/\" INT64_FORMAT \" (%.3f%%)\\n\",\n> > \t\t\t latency_limit / 1000.0, latency_late, ntx,\n> > \t\t\t (ntx > 0) ? 100.0 * latency_late / ntx : 0.0);\n> \n> Looks good to me.\n> \n> Best reagards,\n> --\n> Tatsuo Ishii\n> SRA OSS, Inc. Japan\n> English: http://www.sraoss.co.jp/index_en.php\n> Japanese:http://www.sraoss.co.jp\n> \n> \n\n-- \nKAWAMOTO Masaya <kawamoto@sraoss.co.jp>\nSRA OSS, Inc. Japan","msg_date":"Thu, 24 Feb 2022 14:00:26 +0900","msg_from":"KAWAMOTO Masaya <kawamoto@sraoss.co.jp>","msg_from_op":true,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":">> I think you are right. In English there's should be no space between number and \"%\".\n>> AFAIK other parts of PostgreSQL follow the rule.\n\nI think it's better to back-patch this to stable branches if there's\nno objection. Thought?\n\nBest reagards,\n--\nTatsuo Ishii\nSRA OSS, Inc. Japan\nEnglish: http://www.sraoss.co.jp/index_en.php\nJapanese:http://www.sraoss.co.jp\n\n\n","msg_date":"Thu, 24 Feb 2022 18:38:57 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@sraoss.co.jp>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":"On Thu, Feb 24, 2022 at 06:38:57PM +0900, Tatsuo Ishii wrote:\n> >> I think you are right. In English there's should be no space between number and \"%\".\n> >> AFAIK other parts of PostgreSQL follow the rule.\n> \n> I think it's better to back-patch this to stable branches if there's\n> no objection. Thought?\n\nThat's only cosmetic, so I would just bother about HEAD if I were to\nchange something like that (I would not bother at all, personally).\n\nOne argument against a backpatch is that this would be disruptive with\ntools that parse and analyze the output generated by pgbench. Fabien,\ndon't you have some tools and/or wrappers doing exactly that?\n--\nMichael","msg_date":"Thu, 24 Feb 2022 20:44:57 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":"Bonjour Michaï¿œl,\n\n>> I think it's better to back-patch this to stable branches if there's\n>> no objection. Thought?\n>\n> That's only cosmetic, so I would just bother about HEAD if I were to\n> change something like that (I would not bother at all, personally).\n>\n> One argument against a backpatch is that this would be disruptive with\n> tools that parse and analyze the output generated by pgbench. Fabien,\n> don't you have some tools and/or wrappers doing exactly that?\n\nYep, I like to \"| cut -d' ' -fx\" and other \"line.split(' ')\" or whatever.\n\nI think that the break of typographical rules is intentional to allow such \nsimplistic low-level stream handling through pipes or scripts. I'd prefer \nthat the format is not changed. Maybe a comment could be added to explain \nthe reason behind it.\n\n-- \nFabien.","msg_date":"Thu, 24 Feb 2022 13:58:44 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":"> On 24 Feb 2022, at 13:58, Fabien COELHO <coelho@cri.ensmp.fr> wrote:\n\n>> One argument against a backpatch is that this would be disruptive with\n>> tools that parse and analyze the output generated by pgbench. Fabien,\n>> don't you have some tools and/or wrappers doing exactly that?\n> \n> Yep, I like to \"| cut -d' ' -fx\" and other \"line.split(' ')\" or whatever.\n> \n> I think that the break of typographical rules is intentional to allow such simplistic low-level stream handling through pipes or scripts. I'd prefer that the format is not changed. Maybe a comment could be added to explain the reason behind it.\n\nThat doesn't sound like an overwhelmingly convincing argument to print some\nmessages with 'X %' and others with 'X%'.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Thu, 24 Feb 2022 14:44:05 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":"On Thu, 24 Feb 2022 14:44:05 +0100\nDaniel Gustafsson <daniel@yesql.se> wrote:\n\n> > On 24 Feb 2022, at 13:58, Fabien COELHO <coelho@cri.ensmp.fr> wrote:\n> \n> >> One argument against a backpatch is that this would be disruptive with\n> >> tools that parse and analyze the output generated by pgbench. Fabien,\n> >> don't you have some tools and/or wrappers doing exactly that?\n> > \n> > Yep, I like to \"| cut -d' ' -fx\" and other \"line.split(' ')\" or whatever.\n> > \n> > I think that the break of typographical rules is intentional to allow such simplistic low-level stream handling through pipes or scripts. I'd prefer that the format is not changed. Maybe a comment could be added to explain the reason behind it.\n> \n> That doesn't sound like an overwhelmingly convincing argument to print some\n> messages with 'X %' and others with 'X%'.\n\nI agree with Daniel. If inserting a space in front of % was intentional for\nhandling the result in such tools, we should fix other places where 'X%' is\nused in the pgbench output.\n\nRegards,\nYugo Nagata\n\n-- \nYugo NAGATA <nagata@sraoss.co.jp>\n\n\n","msg_date":"Fri, 25 Feb 2022 15:39:24 +0900","msg_from":"Yugo NAGATA <nagata@sraoss.co.jp>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":"\nHello Daniel,\n\n>> I think that the break of typographical rules is intentional to allow \n>> such simplistic low-level stream handling through pipes or scripts. I'd \n>> prefer that the format is not changed. Maybe a comment could be added \n>> to explain the reason behind it.\n>\n> That doesn't sound like an overwhelmingly convincing argument to print \n> some messages with 'X %' and others with 'X%'.\n\nIndeed. The no-space % are for database loading progress and the final \nreport which I happen not to process through pipes and are more \nuser-facing interfaces/reports. The stream piping need applies more to \nrepeated lines such as those output by progress, which happen to avoid \npercentages anyway so the questions does not arise.\n\nSo fine with me wrt having a more homogeneous report.\n\n-- \nFabien.\n\n\n","msg_date":"Fri, 25 Feb 2022 13:07:46 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":"Hi Fabien,\n\n>> That doesn't sound like an overwhelmingly convincing argument to print\n>> some messages with 'X %' and others with 'X%'.\n> \n> Indeed. The no-space % are for database loading progress and the final\n> report which I happen not to process through pipes and are more\n> user-facing interfaces/reports. The stream piping need applies more to\n> repeated lines such as those output by progress, which happen to avoid\n> percentages anyway so the questions does not arise.\n> \n> So fine with me wrt having a more homogeneous report.\n\nSo are you fine with Kawamoto-san's patch?\n\nBest reagards,\n--\nTatsuo Ishii\nSRA OSS, Inc. Japan\nEnglish: http://www.sraoss.co.jp/index_en.php\nJapanese:http://www.sraoss.co.jp\n\n\n","msg_date":"Mon, 28 Feb 2022 17:32:09 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@sraoss.co.jp>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":"\n>> So fine with me wrt having a more homogeneous report.\n>\n> So are you fine with Kawamoto-san's patch?\n\nYes.\n\nPatch applies cleanly (hmmm, it would have been better to have it as an \nattachement). Make & make check ok. Fine with me.\n\n-- \nFabien.\n\n\n","msg_date":"Tue, 1 Mar 2022 11:46:24 +0100 (CET)","msg_from":"Fabien COELHO <coelho@cri.ensmp.fr>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."},{"msg_contents":">> So are you fine with Kawamoto-san's patch?\n> \n> Yes.\n> \n> Patch applies cleanly (hmmm, it would have been better to have it as\n> an attachement). Make & make check ok. Fine with me.\n\nThank you for the review. Fix pushed to master branch.\n\nBest reagards,\n--\nTatsuo Ishii\nSRA OSS, Inc. Japan\nEnglish: http://www.sraoss.co.jp/index_en.php\nJapanese:http://www.sraoss.co.jp\n\n\n","msg_date":"Wed, 02 Mar 2022 08:35:35 +0900 (JST)","msg_from":"Tatsuo Ishii <ishii@sraoss.co.jp>","msg_from_op":false,"msg_subject":"Re: Typo in pgbench messages."}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI found messages inserted a space before the \\\"%\\\" in pgbench.\\nI think this is typo because there are no space before the \\\"%\\\" in other messages.\\nWhat do you think? \\n\\ndiff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c\\nindex f166a77e3a..4ebe5e6ea4 100644\\n--- a/src/bin/pgbench/pgbench.c\\n+++ b/src/bin/pgbench/pgbench.c\\n@@ -5598,11 +5598,11 @@ printResults(StatsData *total,\\n \\t\\treturn;\\n \\n \\tif (throttle_delay && latency_limit)\\n-\\t\\tprintf(\\\"number of transactions skipped: \\\" INT64_FORMAT \\\" (%.3f %%)\\\\n\\\",\\n+\\t\\tprintf(\\\"number of transactions skipped: \\\" INT64_FORMAT \\\" (%.3f%%)\\\\n\\\",\\n \\t\\t\\t total->skipped, 100.0 * total->skipped / total->cnt);\\n \\n \\tif (latency_limit)\\n-\\t\\tprintf(\\\"number of transactions above the %.1f ms latency limit: \\\" INT64_FORMAT \\\"/\\\" INT64_FORMAT \\\" (%.3f %%)\\\\n\\\",\\n+\\t\\tprintf(\\\"number of transactions above the %.1f ms latency limit: \\\" INT64_FORMAT \\\"/\\\" INT64_FORMAT \\\" (%.3f%%)\\\\n\\\",\\n \\t\\t\\t latency_limit / 1000.0, latency_late, ntx,\\n \\t\\t\\t (ntx > 0) ? 100.0 * latency_late / ntx : 0.0);\\n\\n-- \\nKAWAMOTO Masaya <kawamoto@sraoss.co.jp>\\nSRA OSS, Inc. Japan\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 11:56:22 +0900\",\n \"msg_from\": \"KAWAMOTO Masaya <kawamoto@sraoss.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \"> Hi,\\n> \\n> I found messages inserted a space before the \\\"%\\\" in pgbench.\\n> I think this is typo because there are no space before the \\\"%\\\" in other messages.\\n> What do you think? \\n\\nI think you are right. In English there's should be no space between number and \\\"%\\\".\\nAFAIK other parts of PostgreSQL follow the rule.\\n\\n> diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c\\n> index f166a77e3a..4ebe5e6ea4 100644\\n> --- a/src/bin/pgbench/pgbench.c\\n> +++ b/src/bin/pgbench/pgbench.c\\n> @@ -5598,11 +5598,11 @@ printResults(StatsData *total,\\n> \\t\\treturn;\\n> \\n> \\tif (throttle_delay && latency_limit)\\n> -\\t\\tprintf(\\\"number of transactions skipped: \\\" INT64_FORMAT \\\" (%.3f %%)\\\\n\\\",\\n> +\\t\\tprintf(\\\"number of transactions skipped: \\\" INT64_FORMAT \\\" (%.3f%%)\\\\n\\\",\\n> \\t\\t\\t total->skipped, 100.0 * total->skipped / total->cnt);\\n> \\n> \\tif (latency_limit)\\n> -\\t\\tprintf(\\\"number of transactions above the %.1f ms latency limit: \\\" INT64_FORMAT \\\"/\\\" INT64_FORMAT \\\" (%.3f %%)\\\\n\\\",\\n> +\\t\\tprintf(\\\"number of transactions above the %.1f ms latency limit: \\\" INT64_FORMAT \\\"/\\\" INT64_FORMAT \\\" (%.3f%%)\\\\n\\\",\\n> \\t\\t\\t latency_limit / 1000.0, latency_late, ntx,\\n> \\t\\t\\t (ntx > 0) ? 100.0 * latency_late / ntx : 0.0);\\n\\nLooks good to me.\\n\\nBest reagards,\\n--\\nTatsuo Ishii\\nSRA OSS, Inc. Japan\\nEnglish: http://www.sraoss.co.jp/index_en.php\\nJapanese:http://www.sraoss.co.jp\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 12:15:55 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@sraoss.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \"Thanks for your comment!\\n\\nSorry, I did not attach the patch file.\\nThis patch focas on master branch.\\n\\nBest regards,\\n\\nOn Thu, 24 Feb 2022 12:15:55 +0900 (JST)\\nTatsuo Ishii <ishii@sraoss.co.jp> wrote:\\n\\n> > Hi,\\n> > \\n> > I found messages inserted a space before the \\\"%\\\" in pgbench.\\n> > I think this is typo because there are no space before the \\\"%\\\" in other messages.\\n> > What do you think? \\n> \\n> I think you are right. In English there's should be no space between number and \\\"%\\\".\\n> AFAIK other parts of PostgreSQL follow the rule.\\n> \\n> > diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c\\n> > index f166a77e3a..4ebe5e6ea4 100644\\n> > --- a/src/bin/pgbench/pgbench.c\\n> > +++ b/src/bin/pgbench/pgbench.c\\n> > @@ -5598,11 +5598,11 @@ printResults(StatsData *total,\\n> > \\t\\treturn;\\n> > \\n> > \\tif (throttle_delay && latency_limit)\\n> > -\\t\\tprintf(\\\"number of transactions skipped: \\\" INT64_FORMAT \\\" (%.3f %%)\\\\n\\\",\\n> > +\\t\\tprintf(\\\"number of transactions skipped: \\\" INT64_FORMAT \\\" (%.3f%%)\\\\n\\\",\\n> > \\t\\t\\t total->skipped, 100.0 * total->skipped / total->cnt);\\n> > \\n> > \\tif (latency_limit)\\n> > -\\t\\tprintf(\\\"number of transactions above the %.1f ms latency limit: \\\" INT64_FORMAT \\\"/\\\" INT64_FORMAT \\\" (%.3f %%)\\\\n\\\",\\n> > +\\t\\tprintf(\\\"number of transactions above the %.1f ms latency limit: \\\" INT64_FORMAT \\\"/\\\" INT64_FORMAT \\\" (%.3f%%)\\\\n\\\",\\n> > \\t\\t\\t latency_limit / 1000.0, latency_late, ntx,\\n> > \\t\\t\\t (ntx > 0) ? 100.0 * latency_late / ntx : 0.0);\\n> \\n> Looks good to me.\\n> \\n> Best reagards,\\n> --\\n> Tatsuo Ishii\\n> SRA OSS, Inc. Japan\\n> English: http://www.sraoss.co.jp/index_en.php\\n> Japanese:http://www.sraoss.co.jp\\n> \\n> \\n\\n-- \\nKAWAMOTO Masaya <kawamoto@sraoss.co.jp>\\nSRA OSS, Inc. Japan\",\n \"msg_date\": \"Thu, 24 Feb 2022 14:00:26 +0900\",\n \"msg_from\": \"KAWAMOTO Masaya <kawamoto@sraoss.co.jp>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \">> I think you are right. In English there's should be no space between number and \\\"%\\\".\\n>> AFAIK other parts of PostgreSQL follow the rule.\\n\\nI think it's better to back-patch this to stable branches if there's\\nno objection. Thought?\\n\\nBest reagards,\\n--\\nTatsuo Ishii\\nSRA OSS, Inc. Japan\\nEnglish: http://www.sraoss.co.jp/index_en.php\\nJapanese:http://www.sraoss.co.jp\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 18:38:57 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@sraoss.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 24, 2022 at 06:38:57PM +0900, Tatsuo Ishii wrote:\\n> >> I think you are right. In English there's should be no space between number and \\\"%\\\".\\n> >> AFAIK other parts of PostgreSQL follow the rule.\\n> \\n> I think it's better to back-patch this to stable branches if there's\\n> no objection. Thought?\\n\\nThat's only cosmetic, so I would just bother about HEAD if I were to\\nchange something like that (I would not bother at all, personally).\\n\\nOne argument against a backpatch is that this would be disruptive with\\ntools that parse and analyze the output generated by pgbench. Fabien,\\ndon't you have some tools and/or wrappers doing exactly that?\\n--\\nMichael\",\n \"msg_date\": \"Thu, 24 Feb 2022 20:44:57 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \"Bonjour Michaï¿œl,\\n\\n>> I think it's better to back-patch this to stable branches if there's\\n>> no objection. Thought?\\n>\\n> That's only cosmetic, so I would just bother about HEAD if I were to\\n> change something like that (I would not bother at all, personally).\\n>\\n> One argument against a backpatch is that this would be disruptive with\\n> tools that parse and analyze the output generated by pgbench. Fabien,\\n> don't you have some tools and/or wrappers doing exactly that?\\n\\nYep, I like to \\\"| cut -d' ' -fx\\\" and other \\\"line.split(' ')\\\" or whatever.\\n\\nI think that the break of typographical rules is intentional to allow such \\nsimplistic low-level stream handling through pipes or scripts. I'd prefer \\nthat the format is not changed. Maybe a comment could be added to explain \\nthe reason behind it.\\n\\n-- \\nFabien.\",\n \"msg_date\": \"Thu, 24 Feb 2022 13:58:44 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \"> On 24 Feb 2022, at 13:58, Fabien COELHO <coelho@cri.ensmp.fr> wrote:\\n\\n>> One argument against a backpatch is that this would be disruptive with\\n>> tools that parse and analyze the output generated by pgbench. Fabien,\\n>> don't you have some tools and/or wrappers doing exactly that?\\n> \\n> Yep, I like to \\\"| cut -d' ' -fx\\\" and other \\\"line.split(' ')\\\" or whatever.\\n> \\n> I think that the break of typographical rules is intentional to allow such simplistic low-level stream handling through pipes or scripts. I'd prefer that the format is not changed. Maybe a comment could be added to explain the reason behind it.\\n\\nThat doesn't sound like an overwhelmingly convincing argument to print some\\nmessages with 'X %' and others with 'X%'.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 14:44:05 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \"On Thu, 24 Feb 2022 14:44:05 +0100\\nDaniel Gustafsson <daniel@yesql.se> wrote:\\n\\n> > On 24 Feb 2022, at 13:58, Fabien COELHO <coelho@cri.ensmp.fr> wrote:\\n> \\n> >> One argument against a backpatch is that this would be disruptive with\\n> >> tools that parse and analyze the output generated by pgbench. Fabien,\\n> >> don't you have some tools and/or wrappers doing exactly that?\\n> > \\n> > Yep, I like to \\\"| cut -d' ' -fx\\\" and other \\\"line.split(' ')\\\" or whatever.\\n> > \\n> > I think that the break of typographical rules is intentional to allow such simplistic low-level stream handling through pipes or scripts. I'd prefer that the format is not changed. Maybe a comment could be added to explain the reason behind it.\\n> \\n> That doesn't sound like an overwhelmingly convincing argument to print some\\n> messages with 'X %' and others with 'X%'.\\n\\nI agree with Daniel. If inserting a space in front of % was intentional for\\nhandling the result in such tools, we should fix other places where 'X%' is\\nused in the pgbench output.\\n\\nRegards,\\nYugo Nagata\\n\\n-- \\nYugo NAGATA <nagata@sraoss.co.jp>\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 15:39:24 +0900\",\n \"msg_from\": \"Yugo NAGATA <nagata@sraoss.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \"\\nHello Daniel,\\n\\n>> I think that the break of typographical rules is intentional to allow \\n>> such simplistic low-level stream handling through pipes or scripts. I'd \\n>> prefer that the format is not changed. Maybe a comment could be added \\n>> to explain the reason behind it.\\n>\\n> That doesn't sound like an overwhelmingly convincing argument to print \\n> some messages with 'X %' and others with 'X%'.\\n\\nIndeed. The no-space % are for database loading progress and the final \\nreport which I happen not to process through pipes and are more \\nuser-facing interfaces/reports. The stream piping need applies more to \\nrepeated lines such as those output by progress, which happen to avoid \\npercentages anyway so the questions does not arise.\\n\\nSo fine with me wrt having a more homogeneous report.\\n\\n-- \\nFabien.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 13:07:46 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \"Hi Fabien,\\n\\n>> That doesn't sound like an overwhelmingly convincing argument to print\\n>> some messages with 'X %' and others with 'X%'.\\n> \\n> Indeed. The no-space % are for database loading progress and the final\\n> report which I happen not to process through pipes and are more\\n> user-facing interfaces/reports. The stream piping need applies more to\\n> repeated lines such as those output by progress, which happen to avoid\\n> percentages anyway so the questions does not arise.\\n> \\n> So fine with me wrt having a more homogeneous report.\\n\\nSo are you fine with Kawamoto-san's patch?\\n\\nBest reagards,\\n--\\nTatsuo Ishii\\nSRA OSS, Inc. Japan\\nEnglish: http://www.sraoss.co.jp/index_en.php\\nJapanese:http://www.sraoss.co.jp\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 17:32:09 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@sraoss.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \"\\n>> So fine with me wrt having a more homogeneous report.\\n>\\n> So are you fine with Kawamoto-san's patch?\\n\\nYes.\\n\\nPatch applies cleanly (hmmm, it would have been better to have it as an \\nattachement). Make & make check ok. Fine with me.\\n\\n-- \\nFabien.\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 11:46:24 +0100 (CET)\",\n \"msg_from\": \"Fabien COELHO <coelho@cri.ensmp.fr>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n },\n {\n \"msg_contents\": \">> So are you fine with Kawamoto-san's patch?\\n> \\n> Yes.\\n> \\n> Patch applies cleanly (hmmm, it would have been better to have it as\\n> an attachement). Make & make check ok. Fine with me.\\n\\nThank you for the review. Fix pushed to master branch.\\n\\nBest reagards,\\n--\\nTatsuo Ishii\\nSRA OSS, Inc. Japan\\nEnglish: http://www.sraoss.co.jp/index_en.php\\nJapanese:http://www.sraoss.co.jp\\n\\n\\n\",\n \"msg_date\": \"Wed, 02 Mar 2022 08:35:35 +0900 (JST)\",\n \"msg_from\": \"Tatsuo Ishii <ishii@sraoss.co.jp>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Typo in pgbench messages.\"\n }\n]","isConversational":false}}},{"rowIdx":53573,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Thinking about poor performance in the case where the data fits in\nRAM, but the working set is too big for shared_buffers, I notice a\ncouple of things that seem bad in BufMgr, but don't understand why\nthey are like that.\n\n1. If we need to allocate a buffer to a new block we do this in one\nstep, while holding both partition locks for the old and the new tag.\nSurely it would cause less contention to make the old block/tag\ninvalid (after flushing), drop the old partition lock and then switch\nto the new one? i.e. just hold one mapping partition lock at a time.\nIs there a specific reason we do it this way?\n\n2. Possibly connected to the above, we issue BufTableInsert() BEFORE\nwe issue BufTableDelete(). That means we need extra entries in the\nbuffer mapping hash table to allow us to hold both the old and the new\nat the same time, for a short period. The way dynahash.c works, we try\nto allocate an entry from the freelist and if that doesn't work, we\nbegin searching ALL the freelists for free entries to steal. So if we\nget enough people trying to do virtual I/O at the same time, then we\nwill hit a \"freelist storm\" where everybody is chasing the last few\nentries. It would make more sense if we could do BufTableDelete()\nfirst, then hold onto the buffer mapping entry rather than add it to\nthe freelist, so we can use it again when we do BufTableInsert() very\nshortly afterwards. That way we wouldn't need to search the freelist\nat all. What is the benefit or reason of doing the Delete after the\nInsert?\n\nPut that another way, it looks like BufTable functions are used in a\nway that mismatches against the way dynahash is designed.\n\nThoughts?\n\n-- \nSimon Riggs http://www.EnterpriseDB.com/\n\n\n","msg_date":"Thu, 24 Feb 2022 12:58:23 +0000","msg_from":"Simon Riggs <simon.riggs@enterprisedb.com>","msg_from_op":true,"msg_subject":"Buffer Manager and Contention"},{"msg_contents":"(I added Yura, as the author of a related patch)\n\nAt Thu, 24 Feb 2022 12:58:23 +0000, Simon Riggs <simon.riggs@enterprisedb.com> wrote in \n> Thinking about poor performance in the case where the data fits in\n> RAM, but the working set is too big for shared_buffers, I notice a\n> couple of things that seem bad in BufMgr, but don't understand why\n> they are like that.\n> \n> 1. If we need to allocate a buffer to a new block we do this in one\n> step, while holding both partition locks for the old and the new tag.\n> Surely it would cause less contention to make the old block/tag\n> invalid (after flushing), drop the old partition lock and then switch\n> to the new one? i.e. just hold one mapping partition lock at a time.\n> Is there a specific reason we do it this way?\n\nI'm not sure but I guess the developer wanted to make the operation\natomic.\n\nYura Sokolov is proposing a patch to separte the two partition locks.\n\nhttps://www.postgresql.org/message-id/1edbb61981fe1d99c3f20e3d56d6c88999f4227c.camel%40postgrespro.ru\n\nAnd it seems to me viable for me and a benchmarking in the thread\nshowed a good result. I'd appreciate your input on that thread.\n\n> 2. Possibly connected to the above, we issue BufTableInsert() BEFORE\n> we issue BufTableDelete(). That means we need extra entries in the\n> buffer mapping hash table to allow us to hold both the old and the new\n> at the same time, for a short period. The way dynahash.c works, we try\n\nYes.\n\n> to allocate an entry from the freelist and if that doesn't work, we\n> begin searching ALL the freelists for free entries to steal. So if we\n> get enough people trying to do virtual I/O at the same time, then we\n> will hit a \"freelist storm\" where everybody is chasing the last few\n> entries. It would make more sense if we could do BufTableDelete()\n\nTo reduce that overhead, Yura proposed a surgically modification on\ndynahash, but I didn't like that and the latest patch doesn't contain\nthat part.\n\n> first, then hold onto the buffer mapping entry rather than add it to\n> the freelist, so we can use it again when we do BufTableInsert() very\n> shortly afterwards. That way we wouldn't need to search the freelist\n> at all. What is the benefit or reason of doing the Delete after the\n> Insert?\n\nHmm. something like hash_swap_key() or hash_reinsert_entry()? That\nsounds reasonable. (Yura's proposal was taking out an entry from hash\nthen attach it with a new key again.)\n\n> Put that another way, it looks like BufTable functions are used in a\n> way that mismatches against the way dynahash is designed.\n> \n> Thoughts?\n\nOn the first part, I think Yura's patch works. On the second point,\nYura already showed it gives a certain amount of gain if we do that.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Fri, 25 Feb 2022 10:20:25 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Buffer Manager and Contention"},{"msg_contents":"At Fri, 25 Feb 2022 10:20:25 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in \n> (I added Yura, as the author of a related patch)\n> \n> At Thu, 24 Feb 2022 12:58:23 +0000, Simon Riggs <simon.riggs@enterprisedb.com> wrote in \n> > Thinking about poor performance in the case where the data fits in\n> > RAM, but the working set is too big for shared_buffers, I notice a\n> > couple of things that seem bad in BufMgr, but don't understand why\n> > they are like that.\n> > \n> > 1. If we need to allocate a buffer to a new block we do this in one\n> > step, while holding both partition locks for the old and the new tag.\n> > Surely it would cause less contention to make the old block/tag\n> > invalid (after flushing), drop the old partition lock and then switch\n> > to the new one? i.e. just hold one mapping partition lock at a time.\n> > Is there a specific reason we do it this way?\n> \n> I'm not sure but I guess the developer wanted to make the operation\n> atomic.\n> \n> Yura Sokolov is proposing a patch to separte the two partition locks.\n> \n> https://www.postgresql.org/message-id/1edbb61981fe1d99c3f20e3d56d6c88999f4227c.camel%40postgrespro.ru\n> \n> And it seems to me viable for me and a benchmarking in the thread\n> showed a good result. I'd appreciate your input on that thread.\n> \n> > 2. Possibly connected to the above, we issue BufTableInsert() BEFORE\n> > we issue BufTableDelete(). That means we need extra entries in the\n> > buffer mapping hash table to allow us to hold both the old and the new\n> > at the same time, for a short period. The way dynahash.c works, we try\n> \n> Yes.\n> \n> > to allocate an entry from the freelist and if that doesn't work, we\n> > begin searching ALL the freelists for free entries to steal. So if we\n> > get enough people trying to do virtual I/O at the same time, then we\n> > will hit a \"freelist storm\" where everybody is chasing the last few\n> > entries. It would make more sense if we could do BufTableDelete()\n> \n> To reduce that overhead, Yura proposed a surgically modification on\n> dynahash, but I didn't like that and the latest patch doesn't contain\n> that part.\n> \n> > first, then hold onto the buffer mapping entry rather than add it to\n> > the freelist, so we can use it again when we do BufTableInsert() very\n> > shortly afterwards. That way we wouldn't need to search the freelist\n> > at all. What is the benefit or reason of doing the Delete after the\n> > Insert?\n> \n> Hmm. something like hash_swap_key() or hash_reinsert_entry()? That\n> sounds reasonable. (Yura's proposal was taking out an entry from hash\n> then attach it with a new key again.)\n> \n> > Put that another way, it looks like BufTable functions are used in a\n> > way that mismatches against the way dynahash is designed.\n> > \n> > Thoughts?\n> \n> On the first part, I think Yura's patch works. On the second point,\n> Yura already showed it gives a certain amount of gain if we do that.\n\nOn second thought, even if we have a new dynahash API to atomically\nreplace hash key, we need to hold two partition locks to do that. That\nis contradicting our objective.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Fri, 25 Feb 2022 12:29:16 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Buffer Manager and Contention"},{"msg_contents":"On Fri, 25 Feb 2022 at 01:20, Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote:\n\n> Yura Sokolov is proposing a patch to separte the two partition locks.\n>\n> https://www.postgresql.org/message-id/1edbb61981fe1d99c3f20e3d56d6c88999f4227c.camel%40postgrespro.ru\n>\n> And it seems to me viable for me and a benchmarking in the thread\n> showed a good result. I'd appreciate your input on that thread.\n\nCertainly, I will stop this thread and contribute to Yura's thread.\n\n-- \nSimon Riggs http://www.EnterpriseDB.com/\n\n\n","msg_date":"Fri, 25 Feb 2022 03:34:27 +0000","msg_from":"Simon Riggs <simon.riggs@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: Buffer Manager and Contention"}],"string":"[\n {\n \"msg_contents\": \"Thinking about poor performance in the case where the data fits in\\nRAM, but the working set is too big for shared_buffers, I notice a\\ncouple of things that seem bad in BufMgr, but don't understand why\\nthey are like that.\\n\\n1. If we need to allocate a buffer to a new block we do this in one\\nstep, while holding both partition locks for the old and the new tag.\\nSurely it would cause less contention to make the old block/tag\\ninvalid (after flushing), drop the old partition lock and then switch\\nto the new one? i.e. just hold one mapping partition lock at a time.\\nIs there a specific reason we do it this way?\\n\\n2. Possibly connected to the above, we issue BufTableInsert() BEFORE\\nwe issue BufTableDelete(). That means we need extra entries in the\\nbuffer mapping hash table to allow us to hold both the old and the new\\nat the same time, for a short period. The way dynahash.c works, we try\\nto allocate an entry from the freelist and if that doesn't work, we\\nbegin searching ALL the freelists for free entries to steal. So if we\\nget enough people trying to do virtual I/O at the same time, then we\\nwill hit a \\\"freelist storm\\\" where everybody is chasing the last few\\nentries. It would make more sense if we could do BufTableDelete()\\nfirst, then hold onto the buffer mapping entry rather than add it to\\nthe freelist, so we can use it again when we do BufTableInsert() very\\nshortly afterwards. That way we wouldn't need to search the freelist\\nat all. What is the benefit or reason of doing the Delete after the\\nInsert?\\n\\nPut that another way, it looks like BufTable functions are used in a\\nway that mismatches against the way dynahash is designed.\\n\\nThoughts?\\n\\n-- \\nSimon Riggs http://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 12:58:23 +0000\",\n \"msg_from\": \"Simon Riggs <simon.riggs@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Buffer Manager and Contention\"\n },\n {\n \"msg_contents\": \"(I added Yura, as the author of a related patch)\\n\\nAt Thu, 24 Feb 2022 12:58:23 +0000, Simon Riggs <simon.riggs@enterprisedb.com> wrote in \\n> Thinking about poor performance in the case where the data fits in\\n> RAM, but the working set is too big for shared_buffers, I notice a\\n> couple of things that seem bad in BufMgr, but don't understand why\\n> they are like that.\\n> \\n> 1. If we need to allocate a buffer to a new block we do this in one\\n> step, while holding both partition locks for the old and the new tag.\\n> Surely it would cause less contention to make the old block/tag\\n> invalid (after flushing), drop the old partition lock and then switch\\n> to the new one? i.e. just hold one mapping partition lock at a time.\\n> Is there a specific reason we do it this way?\\n\\nI'm not sure but I guess the developer wanted to make the operation\\natomic.\\n\\nYura Sokolov is proposing a patch to separte the two partition locks.\\n\\nhttps://www.postgresql.org/message-id/1edbb61981fe1d99c3f20e3d56d6c88999f4227c.camel%40postgrespro.ru\\n\\nAnd it seems to me viable for me and a benchmarking in the thread\\nshowed a good result. I'd appreciate your input on that thread.\\n\\n> 2. Possibly connected to the above, we issue BufTableInsert() BEFORE\\n> we issue BufTableDelete(). That means we need extra entries in the\\n> buffer mapping hash table to allow us to hold both the old and the new\\n> at the same time, for a short period. The way dynahash.c works, we try\\n\\nYes.\\n\\n> to allocate an entry from the freelist and if that doesn't work, we\\n> begin searching ALL the freelists for free entries to steal. So if we\\n> get enough people trying to do virtual I/O at the same time, then we\\n> will hit a \\\"freelist storm\\\" where everybody is chasing the last few\\n> entries. It would make more sense if we could do BufTableDelete()\\n\\nTo reduce that overhead, Yura proposed a surgically modification on\\ndynahash, but I didn't like that and the latest patch doesn't contain\\nthat part.\\n\\n> first, then hold onto the buffer mapping entry rather than add it to\\n> the freelist, so we can use it again when we do BufTableInsert() very\\n> shortly afterwards. That way we wouldn't need to search the freelist\\n> at all. What is the benefit or reason of doing the Delete after the\\n> Insert?\\n\\nHmm. something like hash_swap_key() or hash_reinsert_entry()? That\\nsounds reasonable. (Yura's proposal was taking out an entry from hash\\nthen attach it with a new key again.)\\n\\n> Put that another way, it looks like BufTable functions are used in a\\n> way that mismatches against the way dynahash is designed.\\n> \\n> Thoughts?\\n\\nOn the first part, I think Yura's patch works. On the second point,\\nYura already showed it gives a certain amount of gain if we do that.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 10:20:25 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Buffer Manager and Contention\"\n },\n {\n \"msg_contents\": \"At Fri, 25 Feb 2022 10:20:25 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in \\n> (I added Yura, as the author of a related patch)\\n> \\n> At Thu, 24 Feb 2022 12:58:23 +0000, Simon Riggs <simon.riggs@enterprisedb.com> wrote in \\n> > Thinking about poor performance in the case where the data fits in\\n> > RAM, but the working set is too big for shared_buffers, I notice a\\n> > couple of things that seem bad in BufMgr, but don't understand why\\n> > they are like that.\\n> > \\n> > 1. If we need to allocate a buffer to a new block we do this in one\\n> > step, while holding both partition locks for the old and the new tag.\\n> > Surely it would cause less contention to make the old block/tag\\n> > invalid (after flushing), drop the old partition lock and then switch\\n> > to the new one? i.e. just hold one mapping partition lock at a time.\\n> > Is there a specific reason we do it this way?\\n> \\n> I'm not sure but I guess the developer wanted to make the operation\\n> atomic.\\n> \\n> Yura Sokolov is proposing a patch to separte the two partition locks.\\n> \\n> https://www.postgresql.org/message-id/1edbb61981fe1d99c3f20e3d56d6c88999f4227c.camel%40postgrespro.ru\\n> \\n> And it seems to me viable for me and a benchmarking in the thread\\n> showed a good result. I'd appreciate your input on that thread.\\n> \\n> > 2. Possibly connected to the above, we issue BufTableInsert() BEFORE\\n> > we issue BufTableDelete(). That means we need extra entries in the\\n> > buffer mapping hash table to allow us to hold both the old and the new\\n> > at the same time, for a short period. The way dynahash.c works, we try\\n> \\n> Yes.\\n> \\n> > to allocate an entry from the freelist and if that doesn't work, we\\n> > begin searching ALL the freelists for free entries to steal. So if we\\n> > get enough people trying to do virtual I/O at the same time, then we\\n> > will hit a \\\"freelist storm\\\" where everybody is chasing the last few\\n> > entries. It would make more sense if we could do BufTableDelete()\\n> \\n> To reduce that overhead, Yura proposed a surgically modification on\\n> dynahash, but I didn't like that and the latest patch doesn't contain\\n> that part.\\n> \\n> > first, then hold onto the buffer mapping entry rather than add it to\\n> > the freelist, so we can use it again when we do BufTableInsert() very\\n> > shortly afterwards. That way we wouldn't need to search the freelist\\n> > at all. What is the benefit or reason of doing the Delete after the\\n> > Insert?\\n> \\n> Hmm. something like hash_swap_key() or hash_reinsert_entry()? That\\n> sounds reasonable. (Yura's proposal was taking out an entry from hash\\n> then attach it with a new key again.)\\n> \\n> > Put that another way, it looks like BufTable functions are used in a\\n> > way that mismatches against the way dynahash is designed.\\n> > \\n> > Thoughts?\\n> \\n> On the first part, I think Yura's patch works. On the second point,\\n> Yura already showed it gives a certain amount of gain if we do that.\\n\\nOn second thought, even if we have a new dynahash API to atomically\\nreplace hash key, we need to hold two partition locks to do that. That\\nis contradicting our objective.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 12:29:16 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Buffer Manager and Contention\"\n },\n {\n \"msg_contents\": \"On Fri, 25 Feb 2022 at 01:20, Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote:\\n\\n> Yura Sokolov is proposing a patch to separte the two partition locks.\\n>\\n> https://www.postgresql.org/message-id/1edbb61981fe1d99c3f20e3d56d6c88999f4227c.camel%40postgrespro.ru\\n>\\n> And it seems to me viable for me and a benchmarking in the thread\\n> showed a good result. I'd appreciate your input on that thread.\\n\\nCertainly, I will stop this thread and contribute to Yura's thread.\\n\\n-- \\nSimon Riggs http://www.EnterpriseDB.com/\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 03:34:27 +0000\",\n \"msg_from\": \"Simon Riggs <simon.riggs@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Buffer Manager and Contention\"\n }\n]","isConversational":false}}},{"rowIdx":53574,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi hackers,\n\nI noticed that create_index test (make installcheck) fails on my laptop\nbecause different query plans are chosen:\n\n- QUERY PLAN\n--------------------------------------------------------\n- Index Only Scan using tenk1_unique1 on tenk1\n- Index Cond: (unique1 = ANY ('{1,42,7}'::integer[]))\n-(2 rows)\n+ QUERY PLAN\n+-------------------------------------------------------------------\n+ Sort\n+ Sort Key: unique1\n+ -> Bitmap Heap Scan on tenk1\n+ Recheck Cond: (unique1 = ANY ('{1,42,7}'::integer[]))\n+ -> Bitmap Index Scan on tenk1_unique1\n+ Index Cond: (unique1 = ANY ('{1,42,7}'::integer[]))\n+(6 rows)\n\n...\n\n- QUERY PLAN\n--------------------------------------------------------\n- Index Only Scan using tenk1_thous_tenthous on tenk1\n- Index Cond: (thousand < 2)\n- Filter: (tenthous = ANY ('{1001,3000}'::integer[]))\n-(3 rows)\n+ QUERY PLAN\n+--------------------------------------------------------------------------------------\n+ Sort\n+ Sort Key: thousand\n+ -> Index Only Scan using tenk1_thous_tenthous on tenk1\n+ Index Cond: ((thousand < 2) AND (tenthous = ANY\n('{1001,3000}'::integer[])))\n+(4 rows)\n\nInvestigation showed that it happens only with `synchronous_commit = off`.\nOnly the master branch is affected, starting from cc50080a82. This is a\ndebug build. I'm running MacOS Monterey 12.2.1 @ x64.\n\nI didn't investigate further. Do we assume that `make installcheck` suppose\nto pass with a different postgresql.conf options?\n\n-- \nBest regards,\nAleksander Alekseev\n\nHi hackers,I noticed that create_index test (make installcheck) fails on my laptop because different query plans are chosen:- QUERY PLAN--------------------------------------------------------- Index Only Scan using tenk1_unique1 on tenk1- Index Cond: (unique1 = ANY ('{1,42,7}'::integer[]))-(2 rows)+ QUERY PLAN+-------------------------------------------------------------------+ Sort+ Sort Key: unique1+ -> Bitmap Heap Scan on tenk1+ Recheck Cond: (unique1 = ANY ('{1,42,7}'::integer[]))+ -> Bitmap Index Scan on tenk1_unique1+ Index Cond: (unique1 = ANY ('{1,42,7}'::integer[]))+(6 rows)...- QUERY PLAN--------------------------------------------------------- Index Only Scan using tenk1_thous_tenthous on tenk1- Index Cond: (thousand < 2)- Filter: (tenthous = ANY ('{1001,3000}'::integer[]))-(3 rows)+ QUERY PLAN+--------------------------------------------------------------------------------------+ Sort+ Sort Key: thousand+ -> Index Only Scan using tenk1_thous_tenthous on tenk1+ Index Cond: ((thousand < 2) AND (tenthous = ANY ('{1001,3000}'::integer[])))+(4 rows)Investigation showed that it happens only with `synchronous_commit = off`. Only the master branch is affected, starting from cc50080a82. This is a debug build. I'm running MacOS Monterey 12.2.1 @ x64.I didn't investigate further. Do we assume that `make installcheck` suppose to pass with a different postgresql.conf options?-- Best regards,Aleksander Alekseev","msg_date":"Thu, 24 Feb 2022 16:47:25 +0300","msg_from":"Aleksander Alekseev <aleksander@timescale.com>","msg_from_op":true,"msg_subject":"create_index test fails when synchronous_commit = off @ master"},{"msg_contents":"Hi,\n\nOn 2022-02-24 16:47:25 +0300, Aleksander Alekseev wrote:\n> - QUERY PLAN\n> --------------------------------------------------------\n> - Index Only Scan using tenk1_thous_tenthous on tenk1\n> - Index Cond: (thousand < 2)\n> - Filter: (tenthous = ANY ('{1001,3000}'::integer[]))\n> -(3 rows)\n> + QUERY PLAN\n> +--------------------------------------------------------------------------------------\n> + Sort\n> + Sort Key: thousand\n> + -> Index Only Scan using tenk1_thous_tenthous on tenk1\n> + Index Cond: ((thousand < 2) AND (tenthous = ANY\n> ('{1001,3000}'::integer[])))\n> +(4 rows)\n\nHeh. We've been having a lot of fights with exactly this plan change in the\nAIO branch, before cc50080a82, and without synchronous_commit =\noff. Interestingly near-exclusively with the regression run within\npg_upgrade's tests.\n\nFor aio we (David did a lot of that IIRC) finally hunted it down to be due\nvacuum skipping pages due to inability to get a cleanup lock. If that happens\nenough, pg_class.relallvisible changes enough to lead to the different plan.\n\n\nI first was going to suggest that we should just use VACUUM FREEZE to prevent\nthe issue.\n\nBut in this instance the cause isn't cleanup locks, probably that we can't yet\nset hint bits due to synchronous_commit=off? But I don't *fully* understand\nhow it leads to this.\n\nI added the SELECT relpages, reltuples, relallvisible FROM pg_class WHERE oid = 'tenk1'::regclass;\njust after the\nVACUUM ANALYZE tenk1;\n\nsynchronous_commit=on\n+ relpages | reltuples | relallvisible\n+----------+-----------+---------------\n+ 345 | 10000 | 345\n+(1 row)\n\nsynchronous_commit=off\n+ relpages | reltuples | relallvisible\n+----------+-----------+---------------\n+ 345 | 10000 | 0\n+(1 row)\n\nSo it clearly is the explanation for the issue.\n\n\nObviously we can locally work around it by adding a\nSET LOCAL synchronous_commit = local;\nto the COPY. But I'd like to fully understand what's going on.\n\n\n> I didn't investigate further. Do we assume that `make installcheck` suppose\n> to pass with a different postgresql.conf options?\n\nDepends on the option, I think... There's some where it's interesting to run\ntests with different options and where the effort required is reasonable. And\nsome cases where it's not... synchronous_commit=off worked until recently, and\nI think we should keep it working.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 24 Feb 2022 07:33:39 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: create_index test fails when synchronous_commit = off @ master"},{"msg_contents":"Hi,\n\nOn 2022-02-24 07:33:39 -0800, Andres Freund wrote:\n> I added the SELECT relpages, reltuples, relallvisible FROM pg_class WHERE oid = 'tenk1'::regclass;\n> just after the\n> VACUUM ANALYZE tenk1;\n>\n> synchronous_commit=on\n> + relpages | reltuples | relallvisible\n> +----------+-----------+---------------\n> + 345 | 10000 | 345\n> +(1 row)\n>\n> synchronous_commit=off\n> + relpages | reltuples | relallvisible\n> +----------+-----------+---------------\n> + 345 | 10000 | 0\n> +(1 row)\n>\n> So it clearly is the explanation for the issue.\n>\n>\n> Obviously we can locally work around it by adding a\n> SET LOCAL synchronous_commit = local;\n> to the COPY. But I'd like to fully understand what's going on.\n\nIt is the hint bit sets delayed by asynchronous commit. I traced execution and\nwe do end up not setting all visible due to reaching the\n!HeapTupleHeaderXminCommitted() path in lazy_scan_prune()\n\n\t\t\tcase HEAPTUPLE_LIVE:\n...\n\t\t\t\t/*\n\t\t\t\t * Is the tuple definitely visible to all transactions?\n\t\t\t\t *\n\t\t\t\t * NB: Like with per-tuple hint bits, we can't set the\n\t\t\t\t * PD_ALL_VISIBLE flag if the inserter committed\n\t\t\t\t * asynchronously. See SetHintBits for more info. Check that\n\t\t\t\t * the tuple is hinted xmin-committed because of that.\n\t\t\t\t */\n\t\t\t\tif (prunestate->all_visible)\n\t\t\t\t{\n\t\t\t\t\tTransactionId xmin;\n\n\t\t\t\t\tif (!HeapTupleHeaderXminCommitted(tuple.t_data))\n\n\nSo it might be reasonable to use synchronous_commit=on in test_setup.sql?\n\nIt's not super satisfying, but i don't immediately see what else could prevent\nall-visible to be set in this case? There's no dead rows, so concurrent\nsnapshots shouldn't prevent cleanup.\n\nI guess we could alternatively try doing something like flushing pending async\ncommits at the start of vacuum. But that probably isn't warranted.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 24 Feb 2022 07:46:48 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: create_index test fails when synchronous_commit = off @ master"},{"msg_contents":"Hi Andres,\n\n> So it might be reasonable to use synchronous_commit=on in test_setup.sql?\n\n> It's not super satisfying, but I don't immediately see what else could prevent\n> all-visible to be set in this case?\n\nI don't see what else can be done either. Here is the corresponding patch.\n\n-- \nBest regards,\nAleksander Alekseev","msg_date":"Mon, 28 Feb 2022 16:31:59 +0300","msg_from":"Aleksander Alekseev <aleksander@timescale.com>","msg_from_op":true,"msg_subject":"Re: create_index test fails when synchronous_commit = off @ master"},{"msg_contents":"Hi hackers,\n\n> I don't see what else can be done either. Here is the corresponding patch.\n\nHere is an updated patch that includes the commit message.\n\n-- \nBest regards,\nAleksander Alekseev","msg_date":"Wed, 2 Mar 2022 13:03:31 +0300","msg_from":"Aleksander Alekseev <aleksander@timescale.com>","msg_from_op":true,"msg_subject":"Re: create_index test fails when synchronous_commit = off @ master"},{"msg_contents":"Hi hackers,\n\n> Here is an updated patch that includes the commit message.\n\nSince this is a trivial change and no one seems to object to it so far, I'm\ngoing to re-check the patch against the recent `master` and change its\nstatus to \"Ready for Committer\" somewhere next week.\n\n-- \nBest regards,\nAleksander Alekseev\n\nHi hackers,> Here is an updated patch that includes the commit message.Since this is a trivial change and no one seems to object to it so far, I'm going to re-check the patch against the recent `master` and change its status to \"Ready for Committer\" somewhere next week.-- Best regards,Aleksander Alekseev","msg_date":"Fri, 11 Mar 2022 10:38:22 +0300","msg_from":"Aleksander Alekseev <aleksander@timescale.com>","msg_from_op":true,"msg_subject":"Re: create_index test fails when synchronous_commit = off @ master"},{"msg_contents":"On 2022-03-11 10:38:22 +0300, Aleksander Alekseev wrote:\n> > Here is an updated patch that includes the commit message.\n> \n> Since this is a trivial change and no one seems to object to it so far, I'm\n> going to re-check the patch against the recent `master` and change its\n> status to \"Ready for Committer\" somewhere next week.\n\nPushed it now.\n\n- Andres\n\n\n","msg_date":"Sat, 12 Mar 2022 14:17:38 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: create_index test fails when synchronous_commit = off @ master"}],"string":"[\n {\n \"msg_contents\": \"Hi hackers,\\n\\nI noticed that create_index test (make installcheck) fails on my laptop\\nbecause different query plans are chosen:\\n\\n- QUERY PLAN\\n--------------------------------------------------------\\n- Index Only Scan using tenk1_unique1 on tenk1\\n- Index Cond: (unique1 = ANY ('{1,42,7}'::integer[]))\\n-(2 rows)\\n+ QUERY PLAN\\n+-------------------------------------------------------------------\\n+ Sort\\n+ Sort Key: unique1\\n+ -> Bitmap Heap Scan on tenk1\\n+ Recheck Cond: (unique1 = ANY ('{1,42,7}'::integer[]))\\n+ -> Bitmap Index Scan on tenk1_unique1\\n+ Index Cond: (unique1 = ANY ('{1,42,7}'::integer[]))\\n+(6 rows)\\n\\n...\\n\\n- QUERY PLAN\\n--------------------------------------------------------\\n- Index Only Scan using tenk1_thous_tenthous on tenk1\\n- Index Cond: (thousand < 2)\\n- Filter: (tenthous = ANY ('{1001,3000}'::integer[]))\\n-(3 rows)\\n+ QUERY PLAN\\n+--------------------------------------------------------------------------------------\\n+ Sort\\n+ Sort Key: thousand\\n+ -> Index Only Scan using tenk1_thous_tenthous on tenk1\\n+ Index Cond: ((thousand < 2) AND (tenthous = ANY\\n('{1001,3000}'::integer[])))\\n+(4 rows)\\n\\nInvestigation showed that it happens only with `synchronous_commit = off`.\\nOnly the master branch is affected, starting from cc50080a82. This is a\\ndebug build. I'm running MacOS Monterey 12.2.1 @ x64.\\n\\nI didn't investigate further. Do we assume that `make installcheck` suppose\\nto pass with a different postgresql.conf options?\\n\\n-- \\nBest regards,\\nAleksander Alekseev\\n\\nHi hackers,I noticed that create_index test (make installcheck) fails on my laptop because different query plans are chosen:- QUERY PLAN--------------------------------------------------------- Index Only Scan using tenk1_unique1 on tenk1- Index Cond: (unique1 = ANY ('{1,42,7}'::integer[]))-(2 rows)+ QUERY PLAN+-------------------------------------------------------------------+ Sort+ Sort Key: unique1+ -> Bitmap Heap Scan on tenk1+ Recheck Cond: (unique1 = ANY ('{1,42,7}'::integer[]))+ -> Bitmap Index Scan on tenk1_unique1+ Index Cond: (unique1 = ANY ('{1,42,7}'::integer[]))+(6 rows)...- QUERY PLAN--------------------------------------------------------- Index Only Scan using tenk1_thous_tenthous on tenk1- Index Cond: (thousand < 2)- Filter: (tenthous = ANY ('{1001,3000}'::integer[]))-(3 rows)+ QUERY PLAN+--------------------------------------------------------------------------------------+ Sort+ Sort Key: thousand+ -> Index Only Scan using tenk1_thous_tenthous on tenk1+ Index Cond: ((thousand < 2) AND (tenthous = ANY ('{1001,3000}'::integer[])))+(4 rows)Investigation showed that it happens only with `synchronous_commit = off`. Only the master branch is affected, starting from cc50080a82. This is a debug build. I'm running MacOS Monterey 12.2.1 @ x64.I didn't investigate further. Do we assume that `make installcheck` suppose to pass with a different postgresql.conf options?-- Best regards,Aleksander Alekseev\",\n \"msg_date\": \"Thu, 24 Feb 2022 16:47:25 +0300\",\n \"msg_from\": \"Aleksander Alekseev <aleksander@timescale.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"create_index test fails when synchronous_commit = off @ master\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-24 16:47:25 +0300, Aleksander Alekseev wrote:\\n> - QUERY PLAN\\n> --------------------------------------------------------\\n> - Index Only Scan using tenk1_thous_tenthous on tenk1\\n> - Index Cond: (thousand < 2)\\n> - Filter: (tenthous = ANY ('{1001,3000}'::integer[]))\\n> -(3 rows)\\n> + QUERY PLAN\\n> +--------------------------------------------------------------------------------------\\n> + Sort\\n> + Sort Key: thousand\\n> + -> Index Only Scan using tenk1_thous_tenthous on tenk1\\n> + Index Cond: ((thousand < 2) AND (tenthous = ANY\\n> ('{1001,3000}'::integer[])))\\n> +(4 rows)\\n\\nHeh. We've been having a lot of fights with exactly this plan change in the\\nAIO branch, before cc50080a82, and without synchronous_commit =\\noff. Interestingly near-exclusively with the regression run within\\npg_upgrade's tests.\\n\\nFor aio we (David did a lot of that IIRC) finally hunted it down to be due\\nvacuum skipping pages due to inability to get a cleanup lock. If that happens\\nenough, pg_class.relallvisible changes enough to lead to the different plan.\\n\\n\\nI first was going to suggest that we should just use VACUUM FREEZE to prevent\\nthe issue.\\n\\nBut in this instance the cause isn't cleanup locks, probably that we can't yet\\nset hint bits due to synchronous_commit=off? But I don't *fully* understand\\nhow it leads to this.\\n\\nI added the SELECT relpages, reltuples, relallvisible FROM pg_class WHERE oid = 'tenk1'::regclass;\\njust after the\\nVACUUM ANALYZE tenk1;\\n\\nsynchronous_commit=on\\n+ relpages | reltuples | relallvisible\\n+----------+-----------+---------------\\n+ 345 | 10000 | 345\\n+(1 row)\\n\\nsynchronous_commit=off\\n+ relpages | reltuples | relallvisible\\n+----------+-----------+---------------\\n+ 345 | 10000 | 0\\n+(1 row)\\n\\nSo it clearly is the explanation for the issue.\\n\\n\\nObviously we can locally work around it by adding a\\nSET LOCAL synchronous_commit = local;\\nto the COPY. But I'd like to fully understand what's going on.\\n\\n\\n> I didn't investigate further. Do we assume that `make installcheck` suppose\\n> to pass with a different postgresql.conf options?\\n\\nDepends on the option, I think... There's some where it's interesting to run\\ntests with different options and where the effort required is reasonable. And\\nsome cases where it's not... synchronous_commit=off worked until recently, and\\nI think we should keep it working.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 07:33:39 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: create_index test fails when synchronous_commit = off @ master\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-24 07:33:39 -0800, Andres Freund wrote:\\n> I added the SELECT relpages, reltuples, relallvisible FROM pg_class WHERE oid = 'tenk1'::regclass;\\n> just after the\\n> VACUUM ANALYZE tenk1;\\n>\\n> synchronous_commit=on\\n> + relpages | reltuples | relallvisible\\n> +----------+-----------+---------------\\n> + 345 | 10000 | 345\\n> +(1 row)\\n>\\n> synchronous_commit=off\\n> + relpages | reltuples | relallvisible\\n> +----------+-----------+---------------\\n> + 345 | 10000 | 0\\n> +(1 row)\\n>\\n> So it clearly is the explanation for the issue.\\n>\\n>\\n> Obviously we can locally work around it by adding a\\n> SET LOCAL synchronous_commit = local;\\n> to the COPY. But I'd like to fully understand what's going on.\\n\\nIt is the hint bit sets delayed by asynchronous commit. I traced execution and\\nwe do end up not setting all visible due to reaching the\\n!HeapTupleHeaderXminCommitted() path in lazy_scan_prune()\\n\\n\\t\\t\\tcase HEAPTUPLE_LIVE:\\n...\\n\\t\\t\\t\\t/*\\n\\t\\t\\t\\t * Is the tuple definitely visible to all transactions?\\n\\t\\t\\t\\t *\\n\\t\\t\\t\\t * NB: Like with per-tuple hint bits, we can't set the\\n\\t\\t\\t\\t * PD_ALL_VISIBLE flag if the inserter committed\\n\\t\\t\\t\\t * asynchronously. See SetHintBits for more info. Check that\\n\\t\\t\\t\\t * the tuple is hinted xmin-committed because of that.\\n\\t\\t\\t\\t */\\n\\t\\t\\t\\tif (prunestate->all_visible)\\n\\t\\t\\t\\t{\\n\\t\\t\\t\\t\\tTransactionId xmin;\\n\\n\\t\\t\\t\\t\\tif (!HeapTupleHeaderXminCommitted(tuple.t_data))\\n\\n\\nSo it might be reasonable to use synchronous_commit=on in test_setup.sql?\\n\\nIt's not super satisfying, but i don't immediately see what else could prevent\\nall-visible to be set in this case? There's no dead rows, so concurrent\\nsnapshots shouldn't prevent cleanup.\\n\\nI guess we could alternatively try doing something like flushing pending async\\ncommits at the start of vacuum. But that probably isn't warranted.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 07:46:48 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: create_index test fails when synchronous_commit = off @ master\"\n },\n {\n \"msg_contents\": \"Hi Andres,\\n\\n> So it might be reasonable to use synchronous_commit=on in test_setup.sql?\\n\\n> It's not super satisfying, but I don't immediately see what else could prevent\\n> all-visible to be set in this case?\\n\\nI don't see what else can be done either. Here is the corresponding patch.\\n\\n-- \\nBest regards,\\nAleksander Alekseev\",\n \"msg_date\": \"Mon, 28 Feb 2022 16:31:59 +0300\",\n \"msg_from\": \"Aleksander Alekseev <aleksander@timescale.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: create_index test fails when synchronous_commit = off @ master\"\n },\n {\n \"msg_contents\": \"Hi hackers,\\n\\n> I don't see what else can be done either. Here is the corresponding patch.\\n\\nHere is an updated patch that includes the commit message.\\n\\n-- \\nBest regards,\\nAleksander Alekseev\",\n \"msg_date\": \"Wed, 2 Mar 2022 13:03:31 +0300\",\n \"msg_from\": \"Aleksander Alekseev <aleksander@timescale.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: create_index test fails when synchronous_commit = off @ master\"\n },\n {\n \"msg_contents\": \"Hi hackers,\\n\\n> Here is an updated patch that includes the commit message.\\n\\nSince this is a trivial change and no one seems to object to it so far, I'm\\ngoing to re-check the patch against the recent `master` and change its\\nstatus to \\\"Ready for Committer\\\" somewhere next week.\\n\\n-- \\nBest regards,\\nAleksander Alekseev\\n\\nHi hackers,> Here is an updated patch that includes the commit message.Since this is a trivial change and no one seems to object to it so far, I'm going to re-check the patch against the recent `master` and change its status to \\\"Ready for Committer\\\" somewhere next week.-- Best regards,Aleksander Alekseev\",\n \"msg_date\": \"Fri, 11 Mar 2022 10:38:22 +0300\",\n \"msg_from\": \"Aleksander Alekseev <aleksander@timescale.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: create_index test fails when synchronous_commit = off @ master\"\n },\n {\n \"msg_contents\": \"On 2022-03-11 10:38:22 +0300, Aleksander Alekseev wrote:\\n> > Here is an updated patch that includes the commit message.\\n> \\n> Since this is a trivial change and no one seems to object to it so far, I'm\\n> going to re-check the patch against the recent `master` and change its\\n> status to \\\"Ready for Committer\\\" somewhere next week.\\n\\nPushed it now.\\n\\n- Andres\\n\\n\\n\",\n \"msg_date\": \"Sat, 12 Mar 2022 14:17:38 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: create_index test fails when synchronous_commit = off @ master\"\n }\n]","isConversational":false}}},{"rowIdx":53575,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Greetings,\n\nThis patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\npg_waldump records to only those which match the given criteria. This should be more performant\nthan `pg_waldump | grep` as well as more reliable given specific variations in output style\ndepending on how the blocks are specified.\n\nThis currently affects only the main fork, but we could presumably add the option to filter by fork\nas well, if that is considered useful.\n\nBest,\n\nDavid\n\n\n\n\n--","msg_date":"Thu, 24 Feb 2022 13:02:33 -0600","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"[PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Thu, Feb 24, 2022 at 11:06 AM David Christensen\n<david.christensen@crunchydata.com> wrote:\n> This patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\n> pg_waldump records to only those which match the given criteria. This should be more performant\n> than `pg_waldump | grep` as well as more reliable given specific variations in output style\n> depending on how the blocks are specified.\n\nSounds useful to me.\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Thu, 24 Feb 2022 11:26:09 -0800","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"Added to commitfest as:\n\nhttps://commitfest.postgresql.org/37/3565/\n\nAdded to commitfest as:https://commitfest.postgresql.org/37/3565/","msg_date":"Thu, 24 Feb 2022 13:28:30 -0600","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"\nOn Fri, 25 Feb 2022 at 03:02, David Christensen <david.christensen@crunchydata.com> wrote:\n> Greetings,\n>\n> This patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\n> pg_waldump records to only those which match the given criteria. This should be more performant\n> than `pg_waldump | grep` as well as more reliable given specific variations in output style\n> depending on how the blocks are specified.\n>\n> This currently affects only the main fork, but we could presumably add the option to filter by fork\n> as well, if that is considered useful.\n>\n\nCool. I think we can report an error instead of reading wal files,\nif the tablespace, database, or relation is invalid. Does there any\nWAL record that has invalid tablespace, database, or relation OID?\n\n-- \nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Fri, 25 Feb 2022 14:32:22 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"> Cool. I think we can report an error instead of reading wal files,\n> if the tablespace, database, or relation is invalid. Does there any\n> WAL record that has invalid tablespace, database, or relation OID?\n\nThe only sort of validity check we could do here is range checking for the underlying data types (which we certainly could/should add if it’s known to never be valid for the underlying types); non-existence of objects is a no-go, since that depends purely on the WAL range you are looking at and you’d have to, you know, scan it to see if it existed before marking as invalid. :)\n\nThanks,\n\nDavid\n\n\n\n","msg_date":"Fri, 25 Feb 2022 06:48:55 -0600","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"\nOn Fri, 25 Feb 2022 at 20:48, David Christensen <david.christensen@crunchydata.com> wrote:\n>> Cool. I think we can report an error instead of reading wal files,\n>> if the tablespace, database, or relation is invalid. Does there any\n>> WAL record that has invalid tablespace, database, or relation OID?\n>\n> The only sort of validity check we could do here is range checking for the underlying data types\n> (which we certainly could/should add if it’s known to never be valid for the underlying types);\n\nThe invalid OID I said here is such as negative number and zero, for those\nparameters, we do not need to read the WAL files, since it always invalid.\n\n> non-existence of objects is a no-go, since that depends purely on the WAL range you are\n> looking at and you’d have to, you know, scan it to see if it existed before marking as invalid. :)\n>\n\nAgreed.\n\n-- \nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Fri, 25 Feb 2022 21:08:20 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Fri, Feb 25, 2022 at 12:36 AM David Christensen\n<david.christensen@crunchydata.com> wrote:\n>\n> Greetings,\n>\n> This patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\n> pg_waldump records to only those which match the given criteria. This should be more performant\n> than `pg_waldump | grep` as well as more reliable given specific variations in output style\n> depending on how the blocks are specified.\n>\n> This currently affects only the main fork, but we could presumably add the option to filter by fork\n> as well, if that is considered useful.\n\nThanks for the patch. This is not adding something that users can't do\nright now, but definitely improves the usability of the pg_waldump as\nit avoids external filterings. Also, it can give the stats/info at\ntable and block level. So, +1 from my side.\n\nI have some comments on the patch:\n\n1) Let's use capitalized \"OID\" as is the case elsewhere in the documentation.\n+ specified via tablespace oid, database oid, and relfilenode separated\n\n2) Good idea to specify an example:\n+ by slashes.\nSomething like, \"by slashes, for instance, XXXX/XXXX/XXXX\n\n3) Crossing 80 char limit\n+/*\n+ * Boolean to return whether the given WAL record matches a specific\nrelation and optional block\n+ */\n+static bool\n+XLogRecordMatchesRelationBlock(XLogReaderState *record, RelFileNode\nmatchRnode, BlockNumber matchBlock\n)\n\n+ pg_log_error(\"could not parse block number \\\"%s\\\"\", optarg);\n+ pg_log_error(\"could not parse relation from \\\"%s\\\" (expecting\n\\\"spc/dat/rel\\\")\", optarg);\n\n4) How about (expecting \\\"tablespace OID/database OID/relation OID\\\")?\nLet's be clear.\n+ pg_log_error(\"could not parse relation from \\\"%s\\\" (expecting\n\\\"spc/dat/rel\\\")\", optarg);\n\n5) I would also see a need for \"filter by FPW\" i.e. list all WAL\nrecords with \"FPW\".\n\n6) How about \"--block option requires --relation option\" or some other\nbetter phrasing?\n+ pg_log_error(\"cannot filter by --block without also filtering --relation\");\n\n7) Extra new line between } and return false;\n+ return true;\n+ }\n+ return false;\n+}\n\n8) Can we have this for-loop local variables instead of function level\nvariables?\n+ RelFileNode rnode;\n+ ForkNumber forknum;\n+ BlockNumber blk;\n\nRegards,\nBharath Rupireddy.\n\n\n","msg_date":"Fri, 25 Feb 2022 19:03:38 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Fri, Feb 25, 2022 at 7:33 AM Bharath Rupireddy <\nbharath.rupireddyforpostgres@gmail.com> wrote:\n\n> Thanks for the patch. This is not adding something that users can't do\n> right now, but definitely improves the usability of the pg_waldump as\n> it avoids external filterings. Also, it can give the stats/info at\n> table and block level. So, +1 from my side.\n>\n\nThanks for the feedback; I will be incorporating most of this into a new\nversion, with a couple of responses below.\n\n\n> 3) Crossing 80 char limit\n>\n\nThis is neither here nor there, but have we as a project considered\nincreasing that to something more modern? I know a lot of current projects\nconsider 132 to be a more reasonable limit. (Will reduce it down to that\nfor now, but consider this a vote towards increasing that limit.)\n\n\n> 5) I would also see a need for \"filter by FPW\" i.e. list all WAL\n> records with \"FPW\".\n>\n\nYes, that wouldn't be too hard to add to this, can add to the next\nversion. We probably ought to also add the fork number as specifiable as\nwell. Other thoughts on could be some wildcard value in the relation part,\nso `1234/23456/*` could filter WAL to a specific database only, say, or\nsome other multiple specifier, like `--block=1234,123456,121234`. (I\nhonestly consider this to be more advanced than we'd need to support in\nthis patch, but if probably wouldn't be too hard to add to it.)\n\nThanks,\n\nDavid\n\nOn Fri, Feb 25, 2022 at 7:33 AM Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote:\nThanks for the patch. This is not adding something that users can't do\nright now, but definitely improves the usability of the pg_waldump as\nit avoids external filterings. Also, it can give the stats/info at\ntable and block level. So, +1 from my side.Thanks for the feedback; I will be incorporating most of this into a new version, with a couple of responses below. \n3) Crossing 80 char limitThis is neither here nor there, but have we as a project considered increasing that to something more modern? I know a lot of current projects consider 132 to be a more reasonable limit. (Will reduce it down to that for now, but consider this a vote towards increasing that limit.) 5) I would also see a need for \"filter by FPW\" i.e. list all WAL\nrecords with \"FPW\".Yes, that wouldn't be too hard to add to this, can add to the next version. We probably ought to also add the fork number as specifiable as well. Other thoughts on could be some wildcard value in the relation part, so `1234/23456/*` could filter WAL to a specific database only, say, or some other multiple specifier, like `--block=1234,123456,121234`. (I honestly consider this to be more advanced than we'd need to support in this patch, but if probably wouldn't be too hard to add to it.)Thanks,David","msg_date":"Fri, 25 Feb 2022 09:33:59 -0600","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Fri, Feb 25, 2022 at 7:08 AM Japin Li <japinli@hotmail.com> wrote:\n\n>\n> On Fri, 25 Feb 2022 at 20:48, David Christensen <\n> david.christensen@crunchydata.com> wrote:\n> >> Cool. I think we can report an error instead of reading wal files,\n> >> if the tablespace, database, or relation is invalid. Does there any\n> >> WAL record that has invalid tablespace, database, or relation OID?\n> >\n> > The only sort of validity check we could do here is range checking for\n> the underlying data types\n> > (which we certainly could/should add if it’s known to never be valid for\n> the underlying types);\n>\n> The invalid OID I said here is such as negative number and zero, for those\n> parameters, we do not need to read the WAL files, since it always invalid.\n>\n\nAgreed. Can add some additional range validation to the parsed values.\n\nDavid\n\nOn Fri, Feb 25, 2022 at 7:08 AM Japin Li <japinli@hotmail.com> wrote:\nOn Fri, 25 Feb 2022 at 20:48, David Christensen <david.christensen@crunchydata.com> wrote:\n>> Cool. I think we can report an error instead of reading wal files,\n>> if the tablespace, database, or relation is invalid. Does there any\n>> WAL record that has invalid tablespace, database, or relation OID?\n>\n> The only sort of validity check we could do here is range checking for the underlying data types\n> (which we certainly could/should add if it’s known to never be valid for the underlying types);\n\nThe invalid OID I said here is such as negative number and zero, for those\nparameters, we do not need to read the WAL files, since it always invalid.Agreed. Can add some additional range validation to the parsed values.David","msg_date":"Fri, 25 Feb 2022 09:35:03 -0600","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> writes:\n\n> On Fri, Feb 25, 2022 at 12:36 AM David Christensen\n> <david.christensen@crunchydata.com> wrote:\n>>\n>> Greetings,\n>>\n>> This patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\n>> pg_waldump records to only those which match the given criteria. This should be more performant\n>> than `pg_waldump | grep` as well as more reliable given specific variations in output style\n>> depending on how the blocks are specified.\n>>\n>> This currently affects only the main fork, but we could presumably add the option to filter by fork\n>> as well, if that is considered useful.\n>\n> Thanks for the patch. This is not adding something that users can't do\n> right now, but definitely improves the usability of the pg_waldump as\n> it avoids external filterings. Also, it can give the stats/info at\n> table and block level. So, +1 from my side.\n\nAttached is V2 with additional feedback from this email, as well as the specification of the\nForkNumber and FPW as specifiable options.\n\nBest,\n\nDavid\n\n\n\n\n--","msg_date":"Fri, 25 Feb 2022 12:56:16 -0600","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"The following review has been posted through the commitfest application:\nmake installcheck-world: tested, passed\nImplements feature: tested, passed\nSpec compliant: tested, passed\nDocumentation: tested, passed\n\nHi\r\n\r\nI am glad to find this patch here because it helps with my current development work, which involves a lot of debugging with the WAL records and this patch definitely make this much easier rather than using grep externally. \r\n\r\nI have tried all of the new options added by the patch and every combination seems to result correctly. \r\n\r\nThe only comment I would have is in the documentation, where I would replace:\r\n\"Display only records touching the given block\" with \"Display only records associated with the given block\" \r\n\"Display only records touching the given relation\" with \" Display only records associated with the given relation\"\r\n\r\njust to make it sound more formal. :)\r\n\r\nbest\r\n\r\nCary Huang\r\n------------------\r\nHighGo Software Canada\r\nwww.highgo.ca","msg_date":"Mon, 28 Feb 2022 22:56:31 +0000","msg_from":"Cary Huang <cary.huang@highgo.ca>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Sat, Feb 26, 2022 at 7:58 AM David Christensen\n<david.christensen@crunchydata.com> wrote:\n> Attached is V2 with additional feedback from this email, as well as the specification of the\n> ForkNumber and FPW as specifiable options.\n\nTrivial fixup needed after commit 3f1ce973.","msg_date":"Mon, 21 Mar 2022 16:36:40 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Mon, Mar 21, 2022 at 4:36 PM Thomas Munro <thomas.munro@gmail.com> wrote:\n> On Sat, Feb 26, 2022 at 7:58 AM David Christensen\n> <david.christensen@crunchydata.com> wrote:\n> > Attached is V2 with additional feedback from this email, as well as the specification of the\n> > ForkNumber and FPW as specifiable options.\n>\n> Trivial fixup needed after commit 3f1ce973.\n\n[04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\nargument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\n*’ [-Werror=format=]\n[04:30:50.630] 963 | if (sscanf(optarg, \"%u\",\n&config.filter_by_relation_forknum) != 1 ||\n[04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n[04:30:50.630] | | |\n[04:30:50.630] | | ForkNumber *\n[04:30:50.630] | unsigned int *\n\nAnd now that this gets to the CompilerWarnings CI task, it looks like\nGCC doesn't like an enum as a scanf %u destination (I didn't see that\nwarning locally when I compiled the above fixup because clearly Clang\nis cool with it...). Probably needs a temporary unsigned int to\nsscanf into first.\n\n\n","msg_date":"Mon, 21 Mar 2022 17:55:42 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Sun, Mar 20, 2022 at 11:56 PM Thomas Munro <thomas.munro@gmail.com>\nwrote:\n\n> On Mon, Mar 21, 2022 at 4:36 PM Thomas Munro <thomas.munro@gmail.com>\n> wrote:\n> > On Sat, Feb 26, 2022 at 7:58 AM David Christensen\n> > <david.christensen@crunchydata.com> wrote:\n> > > Attached is V2 with additional feedback from this email, as well as\n> the specification of the\n> > > ForkNumber and FPW as specifiable options.\n> >\n> > Trivial fixup needed after commit 3f1ce973.\n>\n> [04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\n> argument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\n> *’ [-Werror=format=]\n> [04:30:50.630] 963 | if (sscanf(optarg, \"%u\",\n> &config.filter_by_relation_forknum) != 1 ||\n> [04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n> [04:30:50.630] | | |\n> [04:30:50.630] | | ForkNumber *\n> [04:30:50.630] | unsigned int *\n>\n> And now that this gets to the CompilerWarnings CI task, it looks like\n> GCC doesn't like an enum as a scanf %u destination (I didn't see that\n> warning locally when I compiled the above fixup because clearly Clang\n> is cool with it...). Probably needs a temporary unsigned int to\n> sscanf into first.\n>\n\nDo you need me to fix this, or are you incorporating that into a V4 of this\npatch? (Similar to your fixup prior in this thread?)\n\nOn Sun, Mar 20, 2022 at 11:56 PM Thomas Munro <thomas.munro@gmail.com> wrote:On Mon, Mar 21, 2022 at 4:36 PM Thomas Munro <thomas.munro@gmail.com> wrote:\n> On Sat, Feb 26, 2022 at 7:58 AM David Christensen\n> <david.christensen@crunchydata.com> wrote:\n> > Attached is V2 with additional feedback from this email, as well as the specification of the\n> > ForkNumber and FPW as specifiable options.\n>\n> Trivial fixup needed after commit 3f1ce973.\n\n[04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\nargument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\n*’ [-Werror=format=]\n[04:30:50.630] 963 | if (sscanf(optarg, \"%u\",\n&config.filter_by_relation_forknum) != 1 ||\n[04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n[04:30:50.630] | | |\n[04:30:50.630] | | ForkNumber *\n[04:30:50.630] | unsigned int *\n\nAnd now that this gets to the CompilerWarnings CI task, it looks like\nGCC doesn't like an enum as a scanf %u destination (I didn't see that\nwarning locally when I compiled the above fixup because clearly Clang\nis cool with it...). Probably needs a temporary unsigned int to\nsscanf into first.Do you need me to fix this, or are you incorporating that into a V4 of this patch? (Similar to your fixup prior in this thread?)","msg_date":"Mon, 21 Mar 2022 11:10:23 -0500","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"Updated to include the V3 fixes as well as the unsigned int/enum fix.\n\n>","msg_date":"Mon, 21 Mar 2022 12:14:19 -0500","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Tue, Mar 22, 2022 at 6:14 AM David Christensen\n<david.christensen@crunchydata.com> wrote:\n> Updated to include the V3 fixes as well as the unsigned int/enum fix.\n\nHi David,\n\nI ran this though pg_indent and adjusted some remaining\nnon-project-style whitespace, and took it for a spin. Very minor\ncomments:\n\npg_waldump: error: could not parse valid relation from \"\"/ (expecting\n\"tablespace OID/database OID/relation filenode\")\n-> There was a stray \"/\" in that message, which I've removed in the attached.\n\npg_waldump: error: could not parse valid relation from \"1664/0/1262\"\n(expecting \"tablespace OID/database OID/relation filenode\")\n-> Why not? Shared relations like pg_database have invalid database\nOID, so I think it should be legal to write --relation=1664/0/1262. I\ntook out that restriction.\n\n+ if (sscanf(optarg, \"%u\",\n&forknum) != 1 ||\n+ forknum >= MAX_FORKNUM)\n+ {\n+ pg_log_error(\"could\nnot parse valid fork number (0..%d) \\\"%s\\\"\",\n+\n MAX_FORKNUM - 1, optarg);\n+ goto bad_argument;\n+ }\n\nI guess you did this because init fork references aren't really\nexpected in the WAL, but I think it's more consistent to allow up to\nMAX_FORKNUM, not least because your documentation mentions 3 as a\nvalid value. So I adjust this to allow MAX_FORKNUM. Make sense?\n\nHere are some more details I noticed, as a likely future user of this\nvery handy feature, which I haven't changed, because they seem more\ndebatable and you might disagree...\n\n1. I think it'd be less surprising if the default value for --fork\nwasn't 0... why not show all forks?\n2. I think it'd be less surprising if --fork without --relation\neither raised an error (like --block without --relation), or were\nallowed, with the meaning \"show me this fork of all relations\".\n3. It seems funny to have no short switch for --fork when everything\nelse has one... what about -F?","msg_date":"Tue, 22 Mar 2022 10:38:32 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Mon, Mar 21, 2022 at 4:39 PM Thomas Munro <thomas.munro@gmail.com> wrote:\n\n[snip]\n\nI guess you did this because init fork references aren't really\n> expected in the WAL, but I think it's more consistent to allow up to\n> MAX_FORKNUM, not least because your documentation mentions 3 as a\n> valid value. So I adjust this to allow MAX_FORKNUM. Make sense?\n>\n\nMakes sense, but I think I'd actually thought it was +1 of the max forks,\nso you give me more credit than I deserve in this case... :-)\n\n\n> Here are some more details I noticed, as a likely future user of this\n> very handy feature, which I haven't changed, because they seem more\n> debatable and you might disagree...\n>\n> 1. I think it'd be less surprising if the default value for --fork\n> wasn't 0... why not show all forks?\n>\n\nAgreed; made it default to all, with the ability to filter down if desired.\n\n\n> 2. I think it'd be less surprising if --fork without --relation\n> either raised an error (like --block without --relation), or were\n> allowed, with the meaning \"show me this fork of all relations\".\n>\n\nAgreed; reworked to support the use case of only showing target forks.\n\n\n> 3. It seems funny to have no short switch for --fork when everything\n> else has one... what about -F?\n>\n\nGood idea; I'd hadn't seen capitals in the getopt list so didn't consider\nthem, but I like this.\n\nEnclosed is v6, incorporating these fixes and docs tweaks.\n\nBest,\n\nDavid","msg_date":"Mon, 21 Mar 2022 18:01:02 -0500","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On 21.03.22 05:55, Thomas Munro wrote:\n> [04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\n> argument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\n> *’ [-Werror=format=]\n> [04:30:50.630] 963 | if (sscanf(optarg, \"%u\",\n> &config.filter_by_relation_forknum) != 1 ||\n> [04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n> [04:30:50.630] | | |\n> [04:30:50.630] | | ForkNumber *\n> [04:30:50.630] | unsigned int *\n> \n> And now that this gets to the CompilerWarnings CI task, it looks like\n> GCC doesn't like an enum as a scanf %u destination (I didn't see that\n> warning locally when I compiled the above fixup because clearly Clang\n> is cool with it...). Probably needs a temporary unsigned int to\n> sscanf into first.\n\nThat's because ForkNum is a signed type. You will probably succeed if \nyou use \"%d\" instead.\n\n\n","msg_date":"Wed, 23 Mar 2022 21:53:21 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Thu, Mar 24, 2022 at 9:53 AM Peter Eisentraut\n<peter.eisentraut@enterprisedb.com> wrote:\n> On 21.03.22 05:55, Thomas Munro wrote:\n> > [04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\n> > argument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\n> > *’ [-Werror=format=]\n> > [04:30:50.630] 963 | if (sscanf(optarg, \"%u\",\n> > &config.filter_by_relation_forknum) != 1 ||\n> > [04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n> > [04:30:50.630] | | |\n> > [04:30:50.630] | | ForkNumber *\n> > [04:30:50.630] | unsigned int *\n> >\n> > And now that this gets to the CompilerWarnings CI task, it looks like\n> > GCC doesn't like an enum as a scanf %u destination (I didn't see that\n> > warning locally when I compiled the above fixup because clearly Clang\n> > is cool with it...). Probably needs a temporary unsigned int to\n> > sscanf into first.\n>\n> That's because ForkNum is a signed type. You will probably succeed if\n> you use \"%d\" instead.\n\nErm, is that really OK? C says \"Each enumerated type shall be\ncompatible with char, a signed integer type, or an\nunsigned integer type. The choice of type is implementation-defined,\nbut shall be capable of representing the values of all the members of\nthe enumeration.\" It could even legally vary from enum to enum,\nthough in practice most compilers probably just use ints all the time\nunless you use weird pragma pack incantation. Therefore I think you\nneed an intermediate variable with the size and signedness matching the\nformat string, if you're going to scanf directly into it, which\nDavid's V6 did.\n\n\n","msg_date":"Thu, 24 Mar 2022 11:54:15 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On 2022-03-24 11:54:15 +1300, Thomas Munro wrote:\n> Erm, is that really OK? C says \"Each enumerated type shall be\n> compatible with char, a signed integer type, or an\n> unsigned integer type. The choice of type is implementation-defined,\n> but shall be capable of representing the values of all the members of\n> the enumeration.\" It could even legally vary from enum to enum,\n> though in practice most compilers probably just use ints all the time\n> unless you use weird pragma pack incantation. Therefore I think you\n> need an intermediate variable with the size and signedness matching the\n> format string, if you're going to scanf directly into it, which\n> David's V6 did.\n\n/me yearns for the perfectly reasonable C++ 11 feature of defining the base\ntype for enums (enum name : basetype { }). One of those features C should have\nadopted long ago. Not that we could use it yet, given we insist that C\nstandards have reached at least european drinking age before relying on them.\n\n\n","msg_date":"Wed, 23 Mar 2022 18:06:36 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Tue, Mar 22, 2022 at 12:01 PM David Christensen\n<david.christensen@crunchydata.com> wrote:\n> Enclosed is v6, incorporating these fixes and docs tweaks.\n\nThanks!\n\nI made a couple of minor changes in the docs, to wit: fixed\ncopy/paste-o \"-F block\" -> \"-F fork\", fork names didn't have initial\ncaps elsewhere, tablespace is better represented by\n<replaceable>tblspc</replaceable> than <replaceable>tbl</replaceable>,\nsome minor wording changes to avoid constructions with \"filter\" where\nit seemed to me a little ambiguous whether that means something is\nincluded or excluded, and some other wording changes for consistency\nwith nearby paragraphs.\n\nAnd... pushed.\n\n\n","msg_date":"Thu, 24 Mar 2022 21:50:34 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On 23.03.22 23:54, Thomas Munro wrote:\n>> That's because ForkNum is a signed type. You will probably succeed if\n>> you use \"%d\" instead.\n> \n> Erm, is that really OK? C says \"Each enumerated type shall be\n> compatible with char, a signed integer type, or an\n> unsigned integer type. The choice of type is implementation-defined,\n> but shall be capable of representing the values of all the members of\n> the enumeration.\" It could even legally vary from enum to enum,\n> though in practice most compilers probably just use ints all the time\n> unless you use weird pragma pack incantation. Therefore I think you\n> need an intermediate variable with the size and signedness matching the\n> format string, if you're going to scanf directly into it, which\n> David's V6 did.\n\nAn intermediate variable is probably the best way to avoid thinking \nabout this much more. ;-) But note that the committed patch uses a %u \nformat whereas the ForkNum enum is signed.\n\nBtw., why the sscanf() instead of just strtol/stroul?\n\n\n","msg_date":"Thu, 24 Mar 2022 11:57:56 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On 24.03.22 11:57, Peter Eisentraut wrote:\n> On 23.03.22 23:54, Thomas Munro wrote:\n>>> That's because ForkNum is a signed type. You will probably succeed if\n>>> you use \"%d\" instead.\n>>\n>> Erm, is that really OK? C says \"Each enumerated type shall be\n>> compatible with char, a signed integer type, or an\n>> unsigned integer type. The choice of type is implementation-defined,\n>> but shall be capable of representing the values of all the members of\n>> the enumeration.\" It could even legally vary from enum to enum,\n>> though in practice most compilers probably just use ints all the time\n>> unless you use weird pragma pack incantation. Therefore I think you\n>> need an intermediate variable with the size and signedness matching the\n>> format string, if you're going to scanf directly into it, which\n>> David's V6 did.\n> \n> An intermediate variable is probably the best way to avoid thinking \n> about this much more. ;-) But note that the committed patch uses a %u \n> format whereas the ForkNum enum is signed.\n> \n> Btw., why the sscanf() instead of just strtol/stroul?\n\nOr even: Why are we exposing fork *numbers* in the user interface? \nEven low-level tools such as pageinspect use fork *names* in their \ninterface.\n\n\n","msg_date":"Thu, 24 Mar 2022 12:01:47 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\n<peter.eisentraut@enterprisedb.com> wrote:\n> Or even: Why are we exposing fork *numbers* in the user interface?\n> Even low-level tools such as pageinspect use fork *names* in their\n> interface.\n\nI wondered about that but thought it seemed OK for such a low level\ntool. It's a fair point though, especially if other low level tools\nare doing that. Here's a patch to change it.","msg_date":"Fri, 25 Mar 2022 00:26:43 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Fri, Mar 25, 2022 at 12:26 AM Thomas Munro <thomas.munro@gmail.com> wrote:\n> On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\n> <peter.eisentraut@enterprisedb.com> wrote:\n> > Or even: Why are we exposing fork *numbers* in the user interface?\n> > Even low-level tools such as pageinspect use fork *names* in their\n> > interface.\n>\n> I wondered about that but thought it seemed OK for such a low level\n> tool. It's a fair point though, especially if other low level tools\n> are doing that. Here's a patch to change it.\n\nOh, and there's already a name lookup function to use for this.","msg_date":"Fri, 25 Mar 2022 00:42:50 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"\n> On Mar 24, 2022, at 6:43 AM, Thomas Munro <thomas.munro@gmail.com> wrote:\n> \n> On Fri, Mar 25, 2022 at 12:26 AM Thomas Munro <thomas.munro@gmail.com> wrote:\n>>> On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\n>>> <peter.eisentraut@enterprisedb.com> wrote:\n>>> Or even: Why are we exposing fork *numbers* in the user interface?\n>>> Even low-level tools such as pageinspect use fork *names* in their\n>>> interface.\n>> \n>> I wondered about that but thought it seemed OK for such a low level\n>> tool. It's a fair point though, especially if other low level tools\n>> are doing that. Here's a patch to change it.\n> \n> Oh, and there's already a name lookup function to use for this.\n\n+1 on the semantic names. \n\nDavid\n\n","msg_date":"Thu, 24 Mar 2022 07:43:02 -0500","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Fri, Mar 25, 2022 at 1:43 AM David Christensen\n<david.christensen@crunchydata.com> wrote:\n> > On Mar 24, 2022, at 6:43 AM, Thomas Munro <thomas.munro@gmail.com> wrote:\n> > On Fri, Mar 25, 2022 at 12:26 AM Thomas Munro <thomas.munro@gmail.com> wrote:\n> >>> On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\n> >>> <peter.eisentraut@enterprisedb.com> wrote:\n> >>> Or even: Why are we exposing fork *numbers* in the user interface?\n> >>> Even low-level tools such as pageinspect use fork *names* in their\n> >>> interface.\n> >>\n> >> I wondered about that but thought it seemed OK for such a low level\n> >> tool. It's a fair point though, especially if other low level tools\n> >> are doing that. Here's a patch to change it.\n> >\n> > Oh, and there's already a name lookup function to use for this.\n>\n> +1 on the semantic names.\n\nCool.\n\nI had another thought while changing that (and also re-alphabetising):\n Why don't we switch to -B for --block and -R for --relation? I\ngather you used -k and -l because -b and -r were already taken, but\nsince we already started using upper case for -F, it seems consistent\nthis way. Or were they chosen for consistency with something else?\n\nIt's also slightly more helpful to a user if the help says\n--relation=T/D/R instead of N/N/N (TS/DB/REL would be nicer but\ndoesn't fit in the space).","msg_date":"Fri, 25 Mar 2022 10:11:32 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"> On Mar 24, 2022, at 4:12 PM, Thomas Munro <thomas.munro@gmail.com> wrote:\n> \n> On Fri, Mar 25, 2022 at 1:43 AM David Christensen\n> <david.christensen@crunchydata.com> wrote:\n>>>> On Mar 24, 2022, at 6:43 AM, Thomas Munro <thomas.munro@gmail.com> wrote:\n>>> On Fri, Mar 25, 2022 at 12:26 AM Thomas Munro <thomas.munro@gmail.com> wrote:\n>>>>> On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\n>>>>> <peter.eisentraut@enterprisedb.com> wrote:\n>>>>> Or even: Why are we exposing fork *numbers* in the user interface?\n>>>>> Even low-level tools such as pageinspect use fork *names* in their\n>>>>> interface.\n>>>> \n>>>> I wondered about that but thought it seemed OK for such a low level\n>>>> tool. It's a fair point though, especially if other low level tools\n>>>> are doing that. Here's a patch to change it.\n>>> \n>>> Oh, and there's already a name lookup function to use for this.\n>> \n>> +1 on the semantic names.\n> \n> Cool.\n> \n> I had another thought while changing that (and also re-alphabetising):\n> Why don't we switch to -B for --block and -R for --relation? I\n> gather you used -k and -l because -b and -r were already taken, but\n> since we already started using upper case for -F, it seems consistent\n> this way. Or were they chosen for consistency with something else?\n\nWorks here; was just trying to get semi-memorable ones from the available lowercase ones, but I like your idea here, and it kind of puts them in the same mental space for remembering. \n\n> It's also slightly more helpful to a user if the help says\n> --relation=T/D/R instead of N/N/N (TS/DB/REL would be nicer but\n> doesn't fit in the space).","msg_date":"Thu, 24 Mar 2022 17:40:30 -0500","msg_from":"David Christensen <david.christensen@crunchydata.com>","msg_from_op":true,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"\nOn Fri, 25 Mar 2022 at 05:11, Thomas Munro <thomas.munro@gmail.com> wrote:\n> Cool.\n>\n> I had another thought while changing that (and also re-alphabetising):\n> Why don't we switch to -B for --block and -R for --relation? I\n> gather you used -k and -l because -b and -r were already taken, but\n> since we already started using upper case for -F, it seems consistent\n> this way. Or were they chosen for consistency with something else?\n>\n> It's also slightly more helpful to a user if the help says\n> --relation=T/D/R instead of N/N/N (TS/DB/REL would be nicer but\n> doesn't fit in the space).\n\nThanks for updating the patch!\n\n+ printf(_(\" -x, --xid=XID only show records with transaction ID XID\\n\"));\n\nI think the description of transaction ID is enough, IIUC, XID is use in core,\nwhich means transaction ID.\n\nSee: src/bin/pg_resetwal/pg_resetwal.c\n\n1239 printf(_(\" -V, --version output version information, then exit\\n\"));\n1240 printf(_(\" -x, --next-transaction-id=XID set next transaction ID\\n\"));\n\n\n+ if (sscanf(optarg, \"%u/%u/%u\",\n+ &config.filter_by_relation.spcNode,\n+ &config.filter_by_relation.dbNode,\n+ &config.filter_by_relation.relNode) != 3 ||\n+ !OidIsValid(config.filter_by_relation.spcNode) ||\n+ !OidIsValid(config.filter_by_relation.relNode))\n\nIt seems we should also check the dbNode.\n\n--\nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Fri, 25 Mar 2022 08:43:30 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"On Fri, Mar 25, 2022 at 1:43 PM Japin Li <japinli@hotmail.com> wrote:\n> + printf(_(\" -x, --xid=XID only show records with transaction ID XID\\n\"));\n>\n> I think the description of transaction ID is enough, IIUC, XID is use in core,\n> which means transaction ID.\n\nThe mention of \"XID\" corresponds to XID on the left, like a sort of\nvariable. That text is not changed by this patch.\n\n> See: src/bin/pg_resetwal/pg_resetwal.c\n>\n> 1239 printf(_(\" -V, --version output version information, then exit\\n\"));\n> 1240 printf(_(\" -x, --next-transaction-id=XID set next transaction ID\\n\"));\n\nHmm, yeah that is inconsistent, but it seems like it is pg_resetwal.c\nthat is not following the notational convention there. Other things\nin pg_resetwal's --help use that 'variable' style.\n\n> + if (sscanf(optarg, \"%u/%u/%u\",\n> + &config.filter_by_relation.spcNode,\n> + &config.filter_by_relation.dbNode,\n> + &config.filter_by_relation.relNode) != 3 ||\n> + !OidIsValid(config.filter_by_relation.spcNode) ||\n> + !OidIsValid(config.filter_by_relation.relNode))\n>\n> It seems we should also check the dbNode.\n\nThis was discussed earlier: it's OK for the dbNode to be invalid (0),\nbecause that's how shared relations like pg_database are referenced.\nLike this:\n\n$ pg_waldump pgdata/pg_wal/000000010000000000000001 --relation\n1664/0/1262 --fork vm --limit 1\nrmgr: Heap2 len (rec/tot): 64/ 8256, tx: 0, lsn:\n0/01491F20, prev 0/01491EC0, desc: VISIBLE cutoff xid 1 flags 0x03,\nblkref #0: rel 1664/0/1262 fork vm blk 0 FPW, blkref #1: rel\n1664/0/1262 blk 0\n\nThanks for looking! I've now pushed the improvements discussed so far.\n\n\n","msg_date":"Fri, 25 Mar 2022 13:55:04 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"},{"msg_contents":"\nOn Fri, 25 Mar 2022 at 08:55, Thomas Munro <thomas.munro@gmail.com> wrote:\n> On Fri, Mar 25, 2022 at 1:43 PM Japin Li <japinli@hotmail.com> wrote:\n>> + printf(_(\" -x, --xid=XID only show records with transaction ID XID\\n\"));\n>>\n>> I think the description of transaction ID is enough, IIUC, XID is use in core,\n>> which means transaction ID.\n>\n> The mention of \"XID\" corresponds to XID on the left, like a sort of\n> variable. That text is not changed by this patch.\n>\n>> See: src/bin/pg_resetwal/pg_resetwal.c\n>>\n>> 1239 printf(_(\" -V, --version output version information, then exit\\n\"));\n>> 1240 printf(_(\" -x, --next-transaction-id=XID set next transaction ID\\n\"));\n>\n> Hmm, yeah that is inconsistent, but it seems like it is pg_resetwal.c\n> that is not following the notational convention there. Other things\n> in pg_resetwal's --help use that 'variable' style.\n>\n\nThanks for your explanation!\n\n>> + if (sscanf(optarg, \"%u/%u/%u\",\n>> + &config.filter_by_relation.spcNode,\n>> + &config.filter_by_relation.dbNode,\n>> + &config.filter_by_relation.relNode) != 3 ||\n>> + !OidIsValid(config.filter_by_relation.spcNode) ||\n>> + !OidIsValid(config.filter_by_relation.relNode))\n>>\n>> It seems we should also check the dbNode.\n>\n> This was discussed earlier: it's OK for the dbNode to be invalid (0),\n> because that's how shared relations like pg_database are referenced.\n> Like this:\n>\n> $ pg_waldump pgdata/pg_wal/000000010000000000000001 --relation\n> 1664/0/1262 --fork vm --limit 1\n> rmgr: Heap2 len (rec/tot): 64/ 8256, tx: 0, lsn:\n> 0/01491F20, prev 0/01491EC0, desc: VISIBLE cutoff xid 1 flags 0x03,\n> blkref #0: rel 1664/0/1262 fork vm blk 0 FPW, blkref #1: rel\n> 1664/0/1262 blk 0\n>\n\nOh, my bad, I missed the discussion email. Sorry for the noise.\n\n\n--\nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Fri, 25 Mar 2022 09:32:49 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":false,"msg_subject":"Re: [PATCH] add relation and block-level filtering to pg_waldump"}],"string":"[\n {\n \"msg_contents\": \"Greetings,\\n\\nThis patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\\npg_waldump records to only those which match the given criteria. This should be more performant\\nthan `pg_waldump | grep` as well as more reliable given specific variations in output style\\ndepending on how the blocks are specified.\\n\\nThis currently affects only the main fork, but we could presumably add the option to filter by fork\\nas well, if that is considered useful.\\n\\nBest,\\n\\nDavid\\n\\n\\n\\n\\n--\",\n \"msg_date\": \"Thu, 24 Feb 2022 13:02:33 -0600\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"[PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 24, 2022 at 11:06 AM David Christensen\\n<david.christensen@crunchydata.com> wrote:\\n> This patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\\n> pg_waldump records to only those which match the given criteria. This should be more performant\\n> than `pg_waldump | grep` as well as more reliable given specific variations in output style\\n> depending on how the blocks are specified.\\n\\nSounds useful to me.\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 11:26:09 -0800\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"Added to commitfest as:\\n\\nhttps://commitfest.postgresql.org/37/3565/\\n\\nAdded to commitfest as:https://commitfest.postgresql.org/37/3565/\",\n \"msg_date\": \"Thu, 24 Feb 2022 13:28:30 -0600\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"\\nOn Fri, 25 Feb 2022 at 03:02, David Christensen <david.christensen@crunchydata.com> wrote:\\n> Greetings,\\n>\\n> This patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\\n> pg_waldump records to only those which match the given criteria. This should be more performant\\n> than `pg_waldump | grep` as well as more reliable given specific variations in output style\\n> depending on how the blocks are specified.\\n>\\n> This currently affects only the main fork, but we could presumably add the option to filter by fork\\n> as well, if that is considered useful.\\n>\\n\\nCool. I think we can report an error instead of reading wal files,\\nif the tablespace, database, or relation is invalid. Does there any\\nWAL record that has invalid tablespace, database, or relation OID?\\n\\n-- \\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 14:32:22 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"> Cool. I think we can report an error instead of reading wal files,\\n> if the tablespace, database, or relation is invalid. Does there any\\n> WAL record that has invalid tablespace, database, or relation OID?\\n\\nThe only sort of validity check we could do here is range checking for the underlying data types (which we certainly could/should add if it’s known to never be valid for the underlying types); non-existence of objects is a no-go, since that depends purely on the WAL range you are looking at and you’d have to, you know, scan it to see if it existed before marking as invalid. :)\\n\\nThanks,\\n\\nDavid\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 06:48:55 -0600\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"\\nOn Fri, 25 Feb 2022 at 20:48, David Christensen <david.christensen@crunchydata.com> wrote:\\n>> Cool. I think we can report an error instead of reading wal files,\\n>> if the tablespace, database, or relation is invalid. Does there any\\n>> WAL record that has invalid tablespace, database, or relation OID?\\n>\\n> The only sort of validity check we could do here is range checking for the underlying data types\\n> (which we certainly could/should add if it’s known to never be valid for the underlying types);\\n\\nThe invalid OID I said here is such as negative number and zero, for those\\nparameters, we do not need to read the WAL files, since it always invalid.\\n\\n> non-existence of objects is a no-go, since that depends purely on the WAL range you are\\n> looking at and you’d have to, you know, scan it to see if it existed before marking as invalid. :)\\n>\\n\\nAgreed.\\n\\n-- \\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 21:08:20 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 12:36 AM David Christensen\\n<david.christensen@crunchydata.com> wrote:\\n>\\n> Greetings,\\n>\\n> This patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\\n> pg_waldump records to only those which match the given criteria. This should be more performant\\n> than `pg_waldump | grep` as well as more reliable given specific variations in output style\\n> depending on how the blocks are specified.\\n>\\n> This currently affects only the main fork, but we could presumably add the option to filter by fork\\n> as well, if that is considered useful.\\n\\nThanks for the patch. This is not adding something that users can't do\\nright now, but definitely improves the usability of the pg_waldump as\\nit avoids external filterings. Also, it can give the stats/info at\\ntable and block level. So, +1 from my side.\\n\\nI have some comments on the patch:\\n\\n1) Let's use capitalized \\\"OID\\\" as is the case elsewhere in the documentation.\\n+ specified via tablespace oid, database oid, and relfilenode separated\\n\\n2) Good idea to specify an example:\\n+ by slashes.\\nSomething like, \\\"by slashes, for instance, XXXX/XXXX/XXXX\\n\\n3) Crossing 80 char limit\\n+/*\\n+ * Boolean to return whether the given WAL record matches a specific\\nrelation and optional block\\n+ */\\n+static bool\\n+XLogRecordMatchesRelationBlock(XLogReaderState *record, RelFileNode\\nmatchRnode, BlockNumber matchBlock\\n)\\n\\n+ pg_log_error(\\\"could not parse block number \\\\\\\"%s\\\\\\\"\\\", optarg);\\n+ pg_log_error(\\\"could not parse relation from \\\\\\\"%s\\\\\\\" (expecting\\n\\\\\\\"spc/dat/rel\\\\\\\")\\\", optarg);\\n\\n4) How about (expecting \\\\\\\"tablespace OID/database OID/relation OID\\\\\\\")?\\nLet's be clear.\\n+ pg_log_error(\\\"could not parse relation from \\\\\\\"%s\\\\\\\" (expecting\\n\\\\\\\"spc/dat/rel\\\\\\\")\\\", optarg);\\n\\n5) I would also see a need for \\\"filter by FPW\\\" i.e. list all WAL\\nrecords with \\\"FPW\\\".\\n\\n6) How about \\\"--block option requires --relation option\\\" or some other\\nbetter phrasing?\\n+ pg_log_error(\\\"cannot filter by --block without also filtering --relation\\\");\\n\\n7) Extra new line between } and return false;\\n+ return true;\\n+ }\\n+ return false;\\n+}\\n\\n8) Can we have this for-loop local variables instead of function level\\nvariables?\\n+ RelFileNode rnode;\\n+ ForkNumber forknum;\\n+ BlockNumber blk;\\n\\nRegards,\\nBharath Rupireddy.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 19:03:38 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 7:33 AM Bharath Rupireddy <\\nbharath.rupireddyforpostgres@gmail.com> wrote:\\n\\n> Thanks for the patch. This is not adding something that users can't do\\n> right now, but definitely improves the usability of the pg_waldump as\\n> it avoids external filterings. Also, it can give the stats/info at\\n> table and block level. So, +1 from my side.\\n>\\n\\nThanks for the feedback; I will be incorporating most of this into a new\\nversion, with a couple of responses below.\\n\\n\\n> 3) Crossing 80 char limit\\n>\\n\\nThis is neither here nor there, but have we as a project considered\\nincreasing that to something more modern? I know a lot of current projects\\nconsider 132 to be a more reasonable limit. (Will reduce it down to that\\nfor now, but consider this a vote towards increasing that limit.)\\n\\n\\n> 5) I would also see a need for \\\"filter by FPW\\\" i.e. list all WAL\\n> records with \\\"FPW\\\".\\n>\\n\\nYes, that wouldn't be too hard to add to this, can add to the next\\nversion. We probably ought to also add the fork number as specifiable as\\nwell. Other thoughts on could be some wildcard value in the relation part,\\nso `1234/23456/*` could filter WAL to a specific database only, say, or\\nsome other multiple specifier, like `--block=1234,123456,121234`. (I\\nhonestly consider this to be more advanced than we'd need to support in\\nthis patch, but if probably wouldn't be too hard to add to it.)\\n\\nThanks,\\n\\nDavid\\n\\nOn Fri, Feb 25, 2022 at 7:33 AM Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote:\\nThanks for the patch. This is not adding something that users can't do\\nright now, but definitely improves the usability of the pg_waldump as\\nit avoids external filterings. Also, it can give the stats/info at\\ntable and block level. So, +1 from my side.Thanks for the feedback; I will be incorporating most of this into a new version, with a couple of responses below. \\n3) Crossing 80 char limitThis is neither here nor there, but have we as a project considered increasing that to something more modern? I know a lot of current projects consider 132 to be a more reasonable limit. (Will reduce it down to that for now, but consider this a vote towards increasing that limit.) 5) I would also see a need for \\\"filter by FPW\\\" i.e. list all WAL\\nrecords with \\\"FPW\\\".Yes, that wouldn't be too hard to add to this, can add to the next version. We probably ought to also add the fork number as specifiable as well. Other thoughts on could be some wildcard value in the relation part, so `1234/23456/*` could filter WAL to a specific database only, say, or some other multiple specifier, like `--block=1234,123456,121234`. (I honestly consider this to be more advanced than we'd need to support in this patch, but if probably wouldn't be too hard to add to it.)Thanks,David\",\n \"msg_date\": \"Fri, 25 Feb 2022 09:33:59 -0600\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 7:08 AM Japin Li <japinli@hotmail.com> wrote:\\n\\n>\\n> On Fri, 25 Feb 2022 at 20:48, David Christensen <\\n> david.christensen@crunchydata.com> wrote:\\n> >> Cool. I think we can report an error instead of reading wal files,\\n> >> if the tablespace, database, or relation is invalid. Does there any\\n> >> WAL record that has invalid tablespace, database, or relation OID?\\n> >\\n> > The only sort of validity check we could do here is range checking for\\n> the underlying data types\\n> > (which we certainly could/should add if it’s known to never be valid for\\n> the underlying types);\\n>\\n> The invalid OID I said here is such as negative number and zero, for those\\n> parameters, we do not need to read the WAL files, since it always invalid.\\n>\\n\\nAgreed. Can add some additional range validation to the parsed values.\\n\\nDavid\\n\\nOn Fri, Feb 25, 2022 at 7:08 AM Japin Li <japinli@hotmail.com> wrote:\\nOn Fri, 25 Feb 2022 at 20:48, David Christensen <david.christensen@crunchydata.com> wrote:\\n>> Cool. I think we can report an error instead of reading wal files,\\n>> if the tablespace, database, or relation is invalid. Does there any\\n>> WAL record that has invalid tablespace, database, or relation OID?\\n>\\n> The only sort of validity check we could do here is range checking for the underlying data types\\n> (which we certainly could/should add if it’s known to never be valid for the underlying types);\\n\\nThe invalid OID I said here is such as negative number and zero, for those\\nparameters, we do not need to read the WAL files, since it always invalid.Agreed. Can add some additional range validation to the parsed values.David\",\n \"msg_date\": \"Fri, 25 Feb 2022 09:35:03 -0600\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> writes:\\n\\n> On Fri, Feb 25, 2022 at 12:36 AM David Christensen\\n> <david.christensen@crunchydata.com> wrote:\\n>>\\n>> Greetings,\\n>>\\n>> This patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\\n>> pg_waldump records to only those which match the given criteria. This should be more performant\\n>> than `pg_waldump | grep` as well as more reliable given specific variations in output style\\n>> depending on how the blocks are specified.\\n>>\\n>> This currently affects only the main fork, but we could presumably add the option to filter by fork\\n>> as well, if that is considered useful.\\n>\\n> Thanks for the patch. This is not adding something that users can't do\\n> right now, but definitely improves the usability of the pg_waldump as\\n> it avoids external filterings. Also, it can give the stats/info at\\n> table and block level. So, +1 from my side.\\n\\nAttached is V2 with additional feedback from this email, as well as the specification of the\\nForkNumber and FPW as specifiable options.\\n\\nBest,\\n\\nDavid\\n\\n\\n\\n\\n--\",\n \"msg_date\": \"Fri, 25 Feb 2022 12:56:16 -0600\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"The following review has been posted through the commitfest application:\\nmake installcheck-world: tested, passed\\nImplements feature: tested, passed\\nSpec compliant: tested, passed\\nDocumentation: tested, passed\\n\\nHi\\r\\n\\r\\nI am glad to find this patch here because it helps with my current development work, which involves a lot of debugging with the WAL records and this patch definitely make this much easier rather than using grep externally. \\r\\n\\r\\nI have tried all of the new options added by the patch and every combination seems to result correctly. \\r\\n\\r\\nThe only comment I would have is in the documentation, where I would replace:\\r\\n\\\"Display only records touching the given block\\\" with \\\"Display only records associated with the given block\\\" \\r\\n\\\"Display only records touching the given relation\\\" with \\\" Display only records associated with the given relation\\\"\\r\\n\\r\\njust to make it sound more formal. :)\\r\\n\\r\\nbest\\r\\n\\r\\nCary Huang\\r\\n------------------\\r\\nHighGo Software Canada\\r\\nwww.highgo.ca\",\n \"msg_date\": \"Mon, 28 Feb 2022 22:56:31 +0000\",\n \"msg_from\": \"Cary Huang <cary.huang@highgo.ca>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 7:58 AM David Christensen\\n<david.christensen@crunchydata.com> wrote:\\n> Attached is V2 with additional feedback from this email, as well as the specification of the\\n> ForkNumber and FPW as specifiable options.\\n\\nTrivial fixup needed after commit 3f1ce973.\",\n \"msg_date\": \"Mon, 21 Mar 2022 16:36:40 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 21, 2022 at 4:36 PM Thomas Munro <thomas.munro@gmail.com> wrote:\\n> On Sat, Feb 26, 2022 at 7:58 AM David Christensen\\n> <david.christensen@crunchydata.com> wrote:\\n> > Attached is V2 with additional feedback from this email, as well as the specification of the\\n> > ForkNumber and FPW as specifiable options.\\n>\\n> Trivial fixup needed after commit 3f1ce973.\\n\\n[04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\\nargument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\\n*’ [-Werror=format=]\\n[04:30:50.630] 963 | if (sscanf(optarg, \\\"%u\\\",\\n&config.filter_by_relation_forknum) != 1 ||\\n[04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\n[04:30:50.630] | | |\\n[04:30:50.630] | | ForkNumber *\\n[04:30:50.630] | unsigned int *\\n\\nAnd now that this gets to the CompilerWarnings CI task, it looks like\\nGCC doesn't like an enum as a scanf %u destination (I didn't see that\\nwarning locally when I compiled the above fixup because clearly Clang\\nis cool with it...). Probably needs a temporary unsigned int to\\nsscanf into first.\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Mar 2022 17:55:42 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 20, 2022 at 11:56 PM Thomas Munro <thomas.munro@gmail.com>\\nwrote:\\n\\n> On Mon, Mar 21, 2022 at 4:36 PM Thomas Munro <thomas.munro@gmail.com>\\n> wrote:\\n> > On Sat, Feb 26, 2022 at 7:58 AM David Christensen\\n> > <david.christensen@crunchydata.com> wrote:\\n> > > Attached is V2 with additional feedback from this email, as well as\\n> the specification of the\\n> > > ForkNumber and FPW as specifiable options.\\n> >\\n> > Trivial fixup needed after commit 3f1ce973.\\n>\\n> [04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\\n> argument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\\n> *’ [-Werror=format=]\\n> [04:30:50.630] 963 | if (sscanf(optarg, \\\"%u\\\",\\n> &config.filter_by_relation_forknum) != 1 ||\\n> [04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\n> [04:30:50.630] | | |\\n> [04:30:50.630] | | ForkNumber *\\n> [04:30:50.630] | unsigned int *\\n>\\n> And now that this gets to the CompilerWarnings CI task, it looks like\\n> GCC doesn't like an enum as a scanf %u destination (I didn't see that\\n> warning locally when I compiled the above fixup because clearly Clang\\n> is cool with it...). Probably needs a temporary unsigned int to\\n> sscanf into first.\\n>\\n\\nDo you need me to fix this, or are you incorporating that into a V4 of this\\npatch? (Similar to your fixup prior in this thread?)\\n\\nOn Sun, Mar 20, 2022 at 11:56 PM Thomas Munro <thomas.munro@gmail.com> wrote:On Mon, Mar 21, 2022 at 4:36 PM Thomas Munro <thomas.munro@gmail.com> wrote:\\n> On Sat, Feb 26, 2022 at 7:58 AM David Christensen\\n> <david.christensen@crunchydata.com> wrote:\\n> > Attached is V2 with additional feedback from this email, as well as the specification of the\\n> > ForkNumber and FPW as specifiable options.\\n>\\n> Trivial fixup needed after commit 3f1ce973.\\n\\n[04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\\nargument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\\n*’ [-Werror=format=]\\n[04:30:50.630] 963 | if (sscanf(optarg, \\\"%u\\\",\\n&config.filter_by_relation_forknum) != 1 ||\\n[04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\n[04:30:50.630] | | |\\n[04:30:50.630] | | ForkNumber *\\n[04:30:50.630] | unsigned int *\\n\\nAnd now that this gets to the CompilerWarnings CI task, it looks like\\nGCC doesn't like an enum as a scanf %u destination (I didn't see that\\nwarning locally when I compiled the above fixup because clearly Clang\\nis cool with it...). Probably needs a temporary unsigned int to\\nsscanf into first.Do you need me to fix this, or are you incorporating that into a V4 of this patch? (Similar to your fixup prior in this thread?)\",\n \"msg_date\": \"Mon, 21 Mar 2022 11:10:23 -0500\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"Updated to include the V3 fixes as well as the unsigned int/enum fix.\\n\\n>\",\n \"msg_date\": \"Mon, 21 Mar 2022 12:14:19 -0500\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 22, 2022 at 6:14 AM David Christensen\\n<david.christensen@crunchydata.com> wrote:\\n> Updated to include the V3 fixes as well as the unsigned int/enum fix.\\n\\nHi David,\\n\\nI ran this though pg_indent and adjusted some remaining\\nnon-project-style whitespace, and took it for a spin. Very minor\\ncomments:\\n\\npg_waldump: error: could not parse valid relation from \\\"\\\"/ (expecting\\n\\\"tablespace OID/database OID/relation filenode\\\")\\n-> There was a stray \\\"/\\\" in that message, which I've removed in the attached.\\n\\npg_waldump: error: could not parse valid relation from \\\"1664/0/1262\\\"\\n(expecting \\\"tablespace OID/database OID/relation filenode\\\")\\n-> Why not? Shared relations like pg_database have invalid database\\nOID, so I think it should be legal to write --relation=1664/0/1262. I\\ntook out that restriction.\\n\\n+ if (sscanf(optarg, \\\"%u\\\",\\n&forknum) != 1 ||\\n+ forknum >= MAX_FORKNUM)\\n+ {\\n+ pg_log_error(\\\"could\\nnot parse valid fork number (0..%d) \\\\\\\"%s\\\\\\\"\\\",\\n+\\n MAX_FORKNUM - 1, optarg);\\n+ goto bad_argument;\\n+ }\\n\\nI guess you did this because init fork references aren't really\\nexpected in the WAL, but I think it's more consistent to allow up to\\nMAX_FORKNUM, not least because your documentation mentions 3 as a\\nvalid value. So I adjust this to allow MAX_FORKNUM. Make sense?\\n\\nHere are some more details I noticed, as a likely future user of this\\nvery handy feature, which I haven't changed, because they seem more\\ndebatable and you might disagree...\\n\\n1. I think it'd be less surprising if the default value for --fork\\nwasn't 0... why not show all forks?\\n2. I think it'd be less surprising if --fork without --relation\\neither raised an error (like --block without --relation), or were\\nallowed, with the meaning \\\"show me this fork of all relations\\\".\\n3. It seems funny to have no short switch for --fork when everything\\nelse has one... what about -F?\",\n \"msg_date\": \"Tue, 22 Mar 2022 10:38:32 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 21, 2022 at 4:39 PM Thomas Munro <thomas.munro@gmail.com> wrote:\\n\\n[snip]\\n\\nI guess you did this because init fork references aren't really\\n> expected in the WAL, but I think it's more consistent to allow up to\\n> MAX_FORKNUM, not least because your documentation mentions 3 as a\\n> valid value. So I adjust this to allow MAX_FORKNUM. Make sense?\\n>\\n\\nMakes sense, but I think I'd actually thought it was +1 of the max forks,\\nso you give me more credit than I deserve in this case... :-)\\n\\n\\n> Here are some more details I noticed, as a likely future user of this\\n> very handy feature, which I haven't changed, because they seem more\\n> debatable and you might disagree...\\n>\\n> 1. I think it'd be less surprising if the default value for --fork\\n> wasn't 0... why not show all forks?\\n>\\n\\nAgreed; made it default to all, with the ability to filter down if desired.\\n\\n\\n> 2. I think it'd be less surprising if --fork without --relation\\n> either raised an error (like --block without --relation), or were\\n> allowed, with the meaning \\\"show me this fork of all relations\\\".\\n>\\n\\nAgreed; reworked to support the use case of only showing target forks.\\n\\n\\n> 3. It seems funny to have no short switch for --fork when everything\\n> else has one... what about -F?\\n>\\n\\nGood idea; I'd hadn't seen capitals in the getopt list so didn't consider\\nthem, but I like this.\\n\\nEnclosed is v6, incorporating these fixes and docs tweaks.\\n\\nBest,\\n\\nDavid\",\n \"msg_date\": \"Mon, 21 Mar 2022 18:01:02 -0500\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On 21.03.22 05:55, Thomas Munro wrote:\\n> [04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\\n> argument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\\n> *’ [-Werror=format=]\\n> [04:30:50.630] 963 | if (sscanf(optarg, \\\"%u\\\",\\n> &config.filter_by_relation_forknum) != 1 ||\\n> [04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\n> [04:30:50.630] | | |\\n> [04:30:50.630] | | ForkNumber *\\n> [04:30:50.630] | unsigned int *\\n> \\n> And now that this gets to the CompilerWarnings CI task, it looks like\\n> GCC doesn't like an enum as a scanf %u destination (I didn't see that\\n> warning locally when I compiled the above fixup because clearly Clang\\n> is cool with it...). Probably needs a temporary unsigned int to\\n> sscanf into first.\\n\\nThat's because ForkNum is a signed type. You will probably succeed if \\nyou use \\\"%d\\\" instead.\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Mar 2022 21:53:21 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 24, 2022 at 9:53 AM Peter Eisentraut\\n<peter.eisentraut@enterprisedb.com> wrote:\\n> On 21.03.22 05:55, Thomas Munro wrote:\\n> > [04:30:50.630] pg_waldump.c:963:26: error: format ‘%u’ expects\\n> > argument of type ‘unsigned int *’, but argument 3 has type ‘ForkNumber\\n> > *’ [-Werror=format=]\\n> > [04:30:50.630] 963 | if (sscanf(optarg, \\\"%u\\\",\\n> > &config.filter_by_relation_forknum) != 1 ||\\n> > [04:30:50.630] | ~^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\n> > [04:30:50.630] | | |\\n> > [04:30:50.630] | | ForkNumber *\\n> > [04:30:50.630] | unsigned int *\\n> >\\n> > And now that this gets to the CompilerWarnings CI task, it looks like\\n> > GCC doesn't like an enum as a scanf %u destination (I didn't see that\\n> > warning locally when I compiled the above fixup because clearly Clang\\n> > is cool with it...). Probably needs a temporary unsigned int to\\n> > sscanf into first.\\n>\\n> That's because ForkNum is a signed type. You will probably succeed if\\n> you use \\\"%d\\\" instead.\\n\\nErm, is that really OK? C says \\\"Each enumerated type shall be\\ncompatible with char, a signed integer type, or an\\nunsigned integer type. The choice of type is implementation-defined,\\nbut shall be capable of representing the values of all the members of\\nthe enumeration.\\\" It could even legally vary from enum to enum,\\nthough in practice most compilers probably just use ints all the time\\nunless you use weird pragma pack incantation. Therefore I think you\\nneed an intermediate variable with the size and signedness matching the\\nformat string, if you're going to scanf directly into it, which\\nDavid's V6 did.\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Mar 2022 11:54:15 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On 2022-03-24 11:54:15 +1300, Thomas Munro wrote:\\n> Erm, is that really OK? C says \\\"Each enumerated type shall be\\n> compatible with char, a signed integer type, or an\\n> unsigned integer type. The choice of type is implementation-defined,\\n> but shall be capable of representing the values of all the members of\\n> the enumeration.\\\" It could even legally vary from enum to enum,\\n> though in practice most compilers probably just use ints all the time\\n> unless you use weird pragma pack incantation. Therefore I think you\\n> need an intermediate variable with the size and signedness matching the\\n> format string, if you're going to scanf directly into it, which\\n> David's V6 did.\\n\\n/me yearns for the perfectly reasonable C++ 11 feature of defining the base\\ntype for enums (enum name : basetype { }). One of those features C should have\\nadopted long ago. Not that we could use it yet, given we insist that C\\nstandards have reached at least european drinking age before relying on them.\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Mar 2022 18:06:36 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 22, 2022 at 12:01 PM David Christensen\\n<david.christensen@crunchydata.com> wrote:\\n> Enclosed is v6, incorporating these fixes and docs tweaks.\\n\\nThanks!\\n\\nI made a couple of minor changes in the docs, to wit: fixed\\ncopy/paste-o \\\"-F block\\\" -> \\\"-F fork\\\", fork names didn't have initial\\ncaps elsewhere, tablespace is better represented by\\n<replaceable>tblspc</replaceable> than <replaceable>tbl</replaceable>,\\nsome minor wording changes to avoid constructions with \\\"filter\\\" where\\nit seemed to me a little ambiguous whether that means something is\\nincluded or excluded, and some other wording changes for consistency\\nwith nearby paragraphs.\\n\\nAnd... pushed.\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Mar 2022 21:50:34 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On 23.03.22 23:54, Thomas Munro wrote:\\n>> That's because ForkNum is a signed type. You will probably succeed if\\n>> you use \\\"%d\\\" instead.\\n> \\n> Erm, is that really OK? C says \\\"Each enumerated type shall be\\n> compatible with char, a signed integer type, or an\\n> unsigned integer type. The choice of type is implementation-defined,\\n> but shall be capable of representing the values of all the members of\\n> the enumeration.\\\" It could even legally vary from enum to enum,\\n> though in practice most compilers probably just use ints all the time\\n> unless you use weird pragma pack incantation. Therefore I think you\\n> need an intermediate variable with the size and signedness matching the\\n> format string, if you're going to scanf directly into it, which\\n> David's V6 did.\\n\\nAn intermediate variable is probably the best way to avoid thinking \\nabout this much more. ;-) But note that the committed patch uses a %u \\nformat whereas the ForkNum enum is signed.\\n\\nBtw., why the sscanf() instead of just strtol/stroul?\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Mar 2022 11:57:56 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On 24.03.22 11:57, Peter Eisentraut wrote:\\n> On 23.03.22 23:54, Thomas Munro wrote:\\n>>> That's because ForkNum is a signed type. You will probably succeed if\\n>>> you use \\\"%d\\\" instead.\\n>>\\n>> Erm, is that really OK? C says \\\"Each enumerated type shall be\\n>> compatible with char, a signed integer type, or an\\n>> unsigned integer type. The choice of type is implementation-defined,\\n>> but shall be capable of representing the values of all the members of\\n>> the enumeration.\\\" It could even legally vary from enum to enum,\\n>> though in practice most compilers probably just use ints all the time\\n>> unless you use weird pragma pack incantation. Therefore I think you\\n>> need an intermediate variable with the size and signedness matching the\\n>> format string, if you're going to scanf directly into it, which\\n>> David's V6 did.\\n> \\n> An intermediate variable is probably the best way to avoid thinking \\n> about this much more. ;-) But note that the committed patch uses a %u \\n> format whereas the ForkNum enum is signed.\\n> \\n> Btw., why the sscanf() instead of just strtol/stroul?\\n\\nOr even: Why are we exposing fork *numbers* in the user interface? \\nEven low-level tools such as pageinspect use fork *names* in their \\ninterface.\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Mar 2022 12:01:47 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\\n<peter.eisentraut@enterprisedb.com> wrote:\\n> Or even: Why are we exposing fork *numbers* in the user interface?\\n> Even low-level tools such as pageinspect use fork *names* in their\\n> interface.\\n\\nI wondered about that but thought it seemed OK for such a low level\\ntool. It's a fair point though, especially if other low level tools\\nare doing that. Here's a patch to change it.\",\n \"msg_date\": \"Fri, 25 Mar 2022 00:26:43 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 12:26 AM Thomas Munro <thomas.munro@gmail.com> wrote:\\n> On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\\n> <peter.eisentraut@enterprisedb.com> wrote:\\n> > Or even: Why are we exposing fork *numbers* in the user interface?\\n> > Even low-level tools such as pageinspect use fork *names* in their\\n> > interface.\\n>\\n> I wondered about that but thought it seemed OK for such a low level\\n> tool. It's a fair point though, especially if other low level tools\\n> are doing that. Here's a patch to change it.\\n\\nOh, and there's already a name lookup function to use for this.\",\n \"msg_date\": \"Fri, 25 Mar 2022 00:42:50 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"\\n> On Mar 24, 2022, at 6:43 AM, Thomas Munro <thomas.munro@gmail.com> wrote:\\n> \\n> On Fri, Mar 25, 2022 at 12:26 AM Thomas Munro <thomas.munro@gmail.com> wrote:\\n>>> On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\\n>>> <peter.eisentraut@enterprisedb.com> wrote:\\n>>> Or even: Why are we exposing fork *numbers* in the user interface?\\n>>> Even low-level tools such as pageinspect use fork *names* in their\\n>>> interface.\\n>> \\n>> I wondered about that but thought it seemed OK for such a low level\\n>> tool. It's a fair point though, especially if other low level tools\\n>> are doing that. Here's a patch to change it.\\n> \\n> Oh, and there's already a name lookup function to use for this.\\n\\n+1 on the semantic names. \\n\\nDavid\\n\\n\",\n \"msg_date\": \"Thu, 24 Mar 2022 07:43:02 -0500\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 1:43 AM David Christensen\\n<david.christensen@crunchydata.com> wrote:\\n> > On Mar 24, 2022, at 6:43 AM, Thomas Munro <thomas.munro@gmail.com> wrote:\\n> > On Fri, Mar 25, 2022 at 12:26 AM Thomas Munro <thomas.munro@gmail.com> wrote:\\n> >>> On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\\n> >>> <peter.eisentraut@enterprisedb.com> wrote:\\n> >>> Or even: Why are we exposing fork *numbers* in the user interface?\\n> >>> Even low-level tools such as pageinspect use fork *names* in their\\n> >>> interface.\\n> >>\\n> >> I wondered about that but thought it seemed OK for such a low level\\n> >> tool. It's a fair point though, especially if other low level tools\\n> >> are doing that. Here's a patch to change it.\\n> >\\n> > Oh, and there's already a name lookup function to use for this.\\n>\\n> +1 on the semantic names.\\n\\nCool.\\n\\nI had another thought while changing that (and also re-alphabetising):\\n Why don't we switch to -B for --block and -R for --relation? I\\ngather you used -k and -l because -b and -r were already taken, but\\nsince we already started using upper case for -F, it seems consistent\\nthis way. Or were they chosen for consistency with something else?\\n\\nIt's also slightly more helpful to a user if the help says\\n--relation=T/D/R instead of N/N/N (TS/DB/REL would be nicer but\\ndoesn't fit in the space).\",\n \"msg_date\": \"Fri, 25 Mar 2022 10:11:32 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"> On Mar 24, 2022, at 4:12 PM, Thomas Munro <thomas.munro@gmail.com> wrote:\\n> \\n> On Fri, Mar 25, 2022 at 1:43 AM David Christensen\\n> <david.christensen@crunchydata.com> wrote:\\n>>>> On Mar 24, 2022, at 6:43 AM, Thomas Munro <thomas.munro@gmail.com> wrote:\\n>>> On Fri, Mar 25, 2022 at 12:26 AM Thomas Munro <thomas.munro@gmail.com> wrote:\\n>>>>> On Fri, Mar 25, 2022 at 12:01 AM Peter Eisentraut\\n>>>>> <peter.eisentraut@enterprisedb.com> wrote:\\n>>>>> Or even: Why are we exposing fork *numbers* in the user interface?\\n>>>>> Even low-level tools such as pageinspect use fork *names* in their\\n>>>>> interface.\\n>>>> \\n>>>> I wondered about that but thought it seemed OK for such a low level\\n>>>> tool. It's a fair point though, especially if other low level tools\\n>>>> are doing that. Here's a patch to change it.\\n>>> \\n>>> Oh, and there's already a name lookup function to use for this.\\n>> \\n>> +1 on the semantic names.\\n> \\n> Cool.\\n> \\n> I had another thought while changing that (and also re-alphabetising):\\n> Why don't we switch to -B for --block and -R for --relation? I\\n> gather you used -k and -l because -b and -r were already taken, but\\n> since we already started using upper case for -F, it seems consistent\\n> this way. Or were they chosen for consistency with something else?\\n\\nWorks here; was just trying to get semi-memorable ones from the available lowercase ones, but I like your idea here, and it kind of puts them in the same mental space for remembering. \\n\\n> It's also slightly more helpful to a user if the help says\\n> --relation=T/D/R instead of N/N/N (TS/DB/REL would be nicer but\\n> doesn't fit in the space).\",\n \"msg_date\": \"Thu, 24 Mar 2022 17:40:30 -0500\",\n \"msg_from\": \"David Christensen <david.christensen@crunchydata.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"\\nOn Fri, 25 Mar 2022 at 05:11, Thomas Munro <thomas.munro@gmail.com> wrote:\\n> Cool.\\n>\\n> I had another thought while changing that (and also re-alphabetising):\\n> Why don't we switch to -B for --block and -R for --relation? I\\n> gather you used -k and -l because -b and -r were already taken, but\\n> since we already started using upper case for -F, it seems consistent\\n> this way. Or were they chosen for consistency with something else?\\n>\\n> It's also slightly more helpful to a user if the help says\\n> --relation=T/D/R instead of N/N/N (TS/DB/REL would be nicer but\\n> doesn't fit in the space).\\n\\nThanks for updating the patch!\\n\\n+ printf(_(\\\" -x, --xid=XID only show records with transaction ID XID\\\\n\\\"));\\n\\nI think the description of transaction ID is enough, IIUC, XID is use in core,\\nwhich means transaction ID.\\n\\nSee: src/bin/pg_resetwal/pg_resetwal.c\\n\\n1239 printf(_(\\\" -V, --version output version information, then exit\\\\n\\\"));\\n1240 printf(_(\\\" -x, --next-transaction-id=XID set next transaction ID\\\\n\\\"));\\n\\n\\n+ if (sscanf(optarg, \\\"%u/%u/%u\\\",\\n+ &config.filter_by_relation.spcNode,\\n+ &config.filter_by_relation.dbNode,\\n+ &config.filter_by_relation.relNode) != 3 ||\\n+ !OidIsValid(config.filter_by_relation.spcNode) ||\\n+ !OidIsValid(config.filter_by_relation.relNode))\\n\\nIt seems we should also check the dbNode.\\n\\n--\\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 08:43:30 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 1:43 PM Japin Li <japinli@hotmail.com> wrote:\\n> + printf(_(\\\" -x, --xid=XID only show records with transaction ID XID\\\\n\\\"));\\n>\\n> I think the description of transaction ID is enough, IIUC, XID is use in core,\\n> which means transaction ID.\\n\\nThe mention of \\\"XID\\\" corresponds to XID on the left, like a sort of\\nvariable. That text is not changed by this patch.\\n\\n> See: src/bin/pg_resetwal/pg_resetwal.c\\n>\\n> 1239 printf(_(\\\" -V, --version output version information, then exit\\\\n\\\"));\\n> 1240 printf(_(\\\" -x, --next-transaction-id=XID set next transaction ID\\\\n\\\"));\\n\\nHmm, yeah that is inconsistent, but it seems like it is pg_resetwal.c\\nthat is not following the notational convention there. Other things\\nin pg_resetwal's --help use that 'variable' style.\\n\\n> + if (sscanf(optarg, \\\"%u/%u/%u\\\",\\n> + &config.filter_by_relation.spcNode,\\n> + &config.filter_by_relation.dbNode,\\n> + &config.filter_by_relation.relNode) != 3 ||\\n> + !OidIsValid(config.filter_by_relation.spcNode) ||\\n> + !OidIsValid(config.filter_by_relation.relNode))\\n>\\n> It seems we should also check the dbNode.\\n\\nThis was discussed earlier: it's OK for the dbNode to be invalid (0),\\nbecause that's how shared relations like pg_database are referenced.\\nLike this:\\n\\n$ pg_waldump pgdata/pg_wal/000000010000000000000001 --relation\\n1664/0/1262 --fork vm --limit 1\\nrmgr: Heap2 len (rec/tot): 64/ 8256, tx: 0, lsn:\\n0/01491F20, prev 0/01491EC0, desc: VISIBLE cutoff xid 1 flags 0x03,\\nblkref #0: rel 1664/0/1262 fork vm blk 0 FPW, blkref #1: rel\\n1664/0/1262 blk 0\\n\\nThanks for looking! I've now pushed the improvements discussed so far.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 13:55:04 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n },\n {\n \"msg_contents\": \"\\nOn Fri, 25 Mar 2022 at 08:55, Thomas Munro <thomas.munro@gmail.com> wrote:\\n> On Fri, Mar 25, 2022 at 1:43 PM Japin Li <japinli@hotmail.com> wrote:\\n>> + printf(_(\\\" -x, --xid=XID only show records with transaction ID XID\\\\n\\\"));\\n>>\\n>> I think the description of transaction ID is enough, IIUC, XID is use in core,\\n>> which means transaction ID.\\n>\\n> The mention of \\\"XID\\\" corresponds to XID on the left, like a sort of\\n> variable. That text is not changed by this patch.\\n>\\n>> See: src/bin/pg_resetwal/pg_resetwal.c\\n>>\\n>> 1239 printf(_(\\\" -V, --version output version information, then exit\\\\n\\\"));\\n>> 1240 printf(_(\\\" -x, --next-transaction-id=XID set next transaction ID\\\\n\\\"));\\n>\\n> Hmm, yeah that is inconsistent, but it seems like it is pg_resetwal.c\\n> that is not following the notational convention there. Other things\\n> in pg_resetwal's --help use that 'variable' style.\\n>\\n\\nThanks for your explanation!\\n\\n>> + if (sscanf(optarg, \\\"%u/%u/%u\\\",\\n>> + &config.filter_by_relation.spcNode,\\n>> + &config.filter_by_relation.dbNode,\\n>> + &config.filter_by_relation.relNode) != 3 ||\\n>> + !OidIsValid(config.filter_by_relation.spcNode) ||\\n>> + !OidIsValid(config.filter_by_relation.relNode))\\n>>\\n>> It seems we should also check the dbNode.\\n>\\n> This was discussed earlier: it's OK for the dbNode to be invalid (0),\\n> because that's how shared relations like pg_database are referenced.\\n> Like this:\\n>\\n> $ pg_waldump pgdata/pg_wal/000000010000000000000001 --relation\\n> 1664/0/1262 --fork vm --limit 1\\n> rmgr: Heap2 len (rec/tot): 64/ 8256, tx: 0, lsn:\\n> 0/01491F20, prev 0/01491EC0, desc: VISIBLE cutoff xid 1 flags 0x03,\\n> blkref #0: rel 1664/0/1262 fork vm blk 0 FPW, blkref #1: rel\\n> 1664/0/1262 blk 0\\n>\\n\\nOh, my bad, I missed the discussion email. Sorry for the noise.\\n\\n\\n--\\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 09:32:49 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: [PATCH] add relation and block-level filtering to pg_waldump\"\n }\n]","isConversational":false}}},{"rowIdx":53576,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nVictor Yegorov reported a crash related to a GiST index on ltree [1],\nfollowing a pg_upgrade from 12.x to 14.x, with a data set reproducing\nthis. I spent some time investigating this, and it seems this is a silly\nbug in commit\n\n commit 911e70207703799605f5a0e8aad9f06cff067c63 (HEAD)\n Author: Alexander Korotkov <akorotkov(at)postgresql(dot)org>\n Date: Mon Mar 30 19:17:11 2020 +0300\n\n Implement operator class parameters\n ...\n\nin PG13, which modified ltree_gist so that siglen is opclass parameter\n(and not hard-coded). But the procedures use LTREE_GET_ASIGLEN macro to\nextract the value, which either extracts the value from the catalog (if\nthe index has opclass parameter) or uses a default value - but it always\nuses LTREE_ASIGLEN_DEFAULT, which belongs to _ltree_gist opclass (i.e.\nfor array of ltree). And that's 28 instead of the 8, as it should be.\n\nAttached is a simple patch tweaking the macros, which resolves the\nissue. Maybe there should be a separate macro though, because \"ASIGLEN\"\nprobably refers to \"array siglen\".\n\nBut the bigger issue is this fixes indexes created before the upgrade,\nbut it breaks indexes created after it. Because those were created with\nsiglen 28B, so reverting to 8B breaks them. A bit of a symmetry :-/\n\nI'm not sure what to do about this. There's no way to distinguish\nindexes, and I think an index may actually be broken both with and\nwithout the patch - imagine an index created before the upgrade (so\nusing siglen 8), but then receiving a couple new rows (siglen 28).\n\nAfter thinking about this I only see two options:\n\n1) Don't apply the patch and tell everyone using ltree_gist they need to\nrebuild the index after pg_upgrade from 12 to 13+. The downside of this\nis larger indexes (because some tuples are 20B larger).\n\n2) Apply the patch + tell those who already upgraded from 12 to rebuild\nltree_gist indexes, because those might be broken due to new inserts.\n\n\nMy opinion is to do (2), because at least those who'll upgrade later\n(which is a lot of people) will be fine without a rebuild. And it also\nmakes the indexes a bit smaller, thanks to saving 20B.\n\n\nregards\n\n\n[1]\nhttps://www.postgresql.org/message-id/17406-71e02820ae79bb40%40postgresql.org\n\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Thu, 24 Feb 2022 22:15:20 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":true,"msg_subject":"ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n> After thinking about this I only see two options:\n\n> 1) Don't apply the patch and tell everyone using ltree_gist they need to\n> rebuild the index after pg_upgrade from 12 to 13+. The downside of this\n> is larger indexes (because some tuples are 20B larger).\n\n> 2) Apply the patch + tell those who already upgraded from 12 to rebuild\n> ltree_gist indexes, because those might be broken due to new inserts.\n\nYeah, let's fix it and advise people to reindex those indexes.\nWon't be the first time release notes have contained such advice.\nPlus, it sounds like we'd have to advise people to reindex *anyway*,\njust in a slightly different set of cases.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 24 Feb 2022 16:51:36 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Hi,\n\nOn 2022-02-24 22:15:20 +0100, Tomas Vondra wrote:\n> After thinking about this I only see two options:\n> \n> 1) Don't apply the patch and tell everyone using ltree_gist they need to\n> rebuild the index after pg_upgrade from 12 to 13+. The downside of this\n> is larger indexes (because some tuples are 20B larger).\n> \n> 2) Apply the patch + tell those who already upgraded from 12 to rebuild\n> ltree_gist indexes, because those might be broken due to new inserts.\n> \n> \n> My opinion is to do (2), because at least those who'll upgrade later\n> (which is a lot of people) will be fine without a rebuild. And it also\n> makes the indexes a bit smaller, thanks to saving 20B.\n\nWon't 2) also break indexes created without a pg_upgrade? \"already upgraded\nfrom 12\" sounds like it wouldn't but I don't see why?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 24 Feb 2022 14:06:58 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"\n\nOn 2/24/22 23:06, Andres Freund wrote:\n> Hi,\n> \n> On 2022-02-24 22:15:20 +0100, Tomas Vondra wrote:\n>> After thinking about this I only see two options:\n>>\n>> 1) Don't apply the patch and tell everyone using ltree_gist they need to\n>> rebuild the index after pg_upgrade from 12 to 13+. The downside of this\n>> is larger indexes (because some tuples are 20B larger).\n>>\n>> 2) Apply the patch + tell those who already upgraded from 12 to rebuild\n>> ltree_gist indexes, because those might be broken due to new inserts.\n>>\n>>\n>> My opinion is to do (2), because at least those who'll upgrade later\n>> (which is a lot of people) will be fine without a rebuild. And it also\n>> makes the indexes a bit smaller, thanks to saving 20B.\n> \n> Won't 2) also break indexes created without a pg_upgrade? \"already upgraded\n> from 12\" sounds like it wouldn't but I don't see why?\n> \n\nIt will, unfortunately - that's why I wrote \"upgrade\" in that sentence.\nI should have been more explicit, sorry. But any new index tuples formed\nafter starting the 13+ cluster are/may be corrupted.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 24 Feb 2022 23:13:18 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"\n\nOn 2/24/22 23:13, Tomas Vondra wrote:\n> \n> \n> On 2/24/22 23:06, Andres Freund wrote:\n>> Hi,\n>>\n>> On 2022-02-24 22:15:20 +0100, Tomas Vondra wrote:\n>>> After thinking about this I only see two options:\n>>>\n>>> 1) Don't apply the patch and tell everyone using ltree_gist they need to\n>>> rebuild the index after pg_upgrade from 12 to 13+. The downside of this\n>>> is larger indexes (because some tuples are 20B larger).\n>>>\n>>> 2) Apply the patch + tell those who already upgraded from 12 to rebuild\n>>> ltree_gist indexes, because those might be broken due to new inserts.\n>>>\n>>>\n>>> My opinion is to do (2), because at least those who'll upgrade later\n>>> (which is a lot of people) will be fine without a rebuild. And it also\n>>> makes the indexes a bit smaller, thanks to saving 20B.\n>>\n>> Won't 2) also break indexes created without a pg_upgrade? \"already upgraded\n>> from 12\" sounds like it wouldn't but I don't see why?\n>>\n> \n> It will, unfortunately - that's why I wrote \"upgrade\" in that sentence.\n> I should have been more explicit, sorry. But any new index tuples formed\n> after starting the 13+ cluster are/may be corrupted.\n> \n\nI wonder if we could check the index tuple length, and adjust the siglen\nbased on that, somehow ...\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 24 Feb 2022 23:22:31 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n> I wonder if we could check the index tuple length, and adjust the siglen\n> based on that, somehow ...\n\nIn an already-corrupted index, there won't be a unique answer.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 24 Feb 2022 20:44:57 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Hi, \n\nOn February 24, 2022 5:44:57 PM PST, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n>> I wonder if we could check the index tuple length, and adjust the siglen\n>> based on that, somehow ...\n>\n>In an already-corrupted index, there won't be a unique answer.\n\nIf we're breaking every ltree gist index, can we detect that it's not an index generated by the new version? Most people don't read release notes, and this is just about guaranteed to break all. \n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Thu, 24 Feb 2022 17:51:41 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"On 25.02.2022 00:15, Tomas Vondra wrote:\n\n> Hi,\n>\n> Victor Yegorov reported a crash related to a GiST index on ltree [1],\n> following a pg_upgrade from 12.x to 14.x, with a data set reproducing\n> this. I spent some time investigating this, and it seems this is a silly\n> bug in commit\n>\n> commit 911e70207703799605f5a0e8aad9f06cff067c63 (HEAD)\n> Author: Alexander Korotkov <akorotkov(at)postgresql(dot)org>\n> Date: Mon Mar 30 19:17:11 2020 +0300\n>\n> Implement operator class parameters\n> ...\n>\n> in PG13, which modified ltree_gist so that siglen is opclass parameter\n> (and not hard-coded). But the procedures use LTREE_GET_ASIGLEN macro to\n> extract the value, which either extracts the value from the catalog (if\n> the index has opclass parameter) or uses a default value - but it always\n> uses LTREE_ASIGLEN_DEFAULT, which belongs to _ltree_gist opclass (i.e.\n> for array of ltree). And that's 28 instead of the 8, as it should be.\n\nIt seems that ltree extension simply was not updated to v1.2 after\npg_upgrade:\n ALTER EXTENSION ltree UPDATE TO '1.2'; -- is missing\n\nUpgrade script ltree--1.1--1.2.sql creates ltree_gist_options() and\nregisters it in the opclass. ltree_gist_options() initializes bytea\noptions using the correct SIGLEN_DEFAULT=8.\n\nIf ltree_gist_options() is absent, LTREE_GET_ASIGLEN() receives NULL\nand wrong LTREE_ASIGLEN_DEFAULT is used. But if ltree_gist_options()\nis registered, LTREE_GET_ASIGLEN() receives non-NULL bytea options\nwith the correct default value.\n\n\nSo, we probably have corrupted indexes that were updated since such\n\"incomplete\" upgrade of ltree.\n\n\n\nAlso I found that contrib/pg_trgm/trgm_gist.c uses wrongly named macro\nLTREE_GET_ASIGLEN().\n\n-- \nNikita Glukhov\nPostgres Professional:http://www.postgrespro.com\nThe Russian Postgres Company\n\n\n\n\n\n\nOn 25.02.2022 00:15, Tomas Vondra wrote:\n\n\nHi,\n\nVictor Yegorov reported a crash related to a GiST index on ltree [1],\nfollowing a pg_upgrade from 12.x to 14.x, with a data set reproducing\nthis. I spent some time investigating this, and it seems this is a silly\nbug in commit\n\n commit 911e70207703799605f5a0e8aad9f06cff067c63 (HEAD)\n Author: Alexander Korotkov <akorotkov(at)postgresql(dot)org>\n Date: Mon Mar 30 19:17:11 2020 +0300\n\n Implement operator class parameters\n ...\n\nin PG13, which modified ltree_gist so that siglen is opclass parameter\n(and not hard-coded). But the procedures use LTREE_GET_ASIGLEN macro to\nextract the value, which either extracts the value from the catalog (if\nthe index has opclass parameter) or uses a default value - but it always\nuses LTREE_ASIGLEN_DEFAULT, which belongs to _ltree_gist opclass (i.e.\nfor array of ltree). And that's 28 instead of the 8, as it should be.\n\n\nIt seems that ltree extension simply was not updated to v1.2 after \npg_upgrade:\n ALTER EXTENSION ltree UPDATE TO '1.2'; -- is missing\nUpgrade script ltree--1.1--1.2.sql creates ltree_gist_options() and \nregisters it in the opclass. ltree_gist_options() initializes bytea \noptions using the correct SIGLEN_DEFAULT=8.\n\nIf ltree_gist_options() is absent, LTREE_GET_ASIGLEN() receives NULL\nand wrong LTREE_ASIGLEN_DEFAULT is used. But if ltree_gist_options()\nis registered, LTREE_GET_ASIGLEN() receives non-NULL bytea options \nwith the correct default value.\n\n\nSo, we probably have corrupted indexes that were updated since such \n\"incomplete\" upgrade of ltree.\n\n\n\nAlso I found that contrib/pg_trgm/trgm_gist.c uses wrongly named macro\nLTREE_GET_ASIGLEN().\n\n-- \nNikita Glukhov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company","msg_date":"Fri, 4 Mar 2022 22:29:31 +0300","msg_from":"Nikita Glukhov <n.gluhov@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"\nOn 3/4/22 20:29, Nikita Glukhov wrote:\n> On 25.02.2022 00:15, Tomas Vondra wrote:\n> \n>> Hi,\n>>\n>> Victor Yegorov reported a crash related to a GiST index on ltree [1],\n>> following a pg_upgrade from 12.x to 14.x, with a data set reproducing\n>> this. I spent some time investigating this, and it seems this is a silly\n>> bug in commit\n>>\n>> commit 911e70207703799605f5a0e8aad9f06cff067c63 (HEAD)\n>> Author: Alexander Korotkov <akorotkov(at)postgresql(dot)org>\n>> Date: Mon Mar 30 19:17:11 2020 +0300\n>>\n>> Implement operator class parameters\n>> ...\n>>\n>> in PG13, which modified ltree_gist so that siglen is opclass parameter\n>> (and not hard-coded). But the procedures use LTREE_GET_ASIGLEN macro to\n>> extract the value, which either extracts the value from the catalog (if\n>> the index has opclass parameter) or uses a default value - but it always\n>> uses LTREE_ASIGLEN_DEFAULT, which belongs to _ltree_gist opclass (i.e.\n>> for array of ltree). And that's 28 instead of the 8, as it should be.\n> \n> It seems that ltree extension simply was not updated to v1.2 after \n> pg_upgrade:\n> ALTER EXTENSION ltree UPDATE TO '1.2'; -- is missing\n> \n> Upgrade script ltree--1.1--1.2.sql creates ltree_gist_options() and \n> registers it in the opclass. ltree_gist_options() initializes bytea \n> options using the correct SIGLEN_DEFAULT=8.\n> \n> If ltree_gist_options() is absent, LTREE_GET_ASIGLEN() receives NULL\n> and wrong LTREE_ASIGLEN_DEFAULT is used. But if ltree_gist_options()\n> is registered, LTREE_GET_ASIGLEN() receives non-NULL bytea options \n> with the correct default value.\n>\n> \n> So, we probably have corrupted indexes that were updated since such \n> \"incomplete\" upgrade of ltree.\n> \n\nIIRC pg_upgrade is not expected to upgrade extensions - it keeps the\ninstalled version of the extension, and that's intentional.\n\nMoreover, it's perfectly legal to install older version, so you can do\n\n CREATE EXTENSION ltree VERSION '1.1';\n\nSo I don't think we can call this \"incomplete upgrade\".\n\n> \n> Also I found that contrib/pg_trgm/trgm_gist.c uses wrongly named macro\n> LTREE_GET_ASIGLEN().\n\nDoes it? When I grep for LTREE_GET_ASIGLEN, I don't get any matches in\npg_trgm.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 4 Mar 2022 20:58:56 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n> On 3/4/22 20:29, Nikita Glukhov wrote:\n>> So, we probably have corrupted indexes that were updated since such \n>> \"incomplete\" upgrade of ltree.\n\n> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\n> installed version of the extension, and that's intentional.\n\nYeah, exactly. But this opens up an additional consideration we\nhave to account for: whatever we do needs to work with either 1.1\nor 1.2 SQL-level versions of the extension.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 04 Mar 2022 15:28:45 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"On 04.03.2022 23:28, Tom Lane wrote:\n\n> Tomas Vondra<tomas.vondra@enterprisedb.com> writes:\n>> On 3/4/22 20:29, Nikita Glukhov wrote:\n>>> So, we probably have corrupted indexes that were updated since such\n>>> \"incomplete\" upgrade of ltree.\n>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\n>> installed version of the extension, and that's intentional.\n> Yeah, exactly. But this opens up an additional consideration we\n> have to account for: whatever we do needs to work with either 1.1\n> or 1.2 SQL-level versions of the extension.\n>\n> \t\t\tregards, tom lane\n\nIt becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\nis not so much related to PG12 => PG13+ upgrades.\n\n\nThe following examples crashes PG13+:\n\nCREATE EXTENSION ltree VERSION \"1.1\";\n\nCREATE TABLE test AS\nSELECT i::text::ltree data\nFROM generate_series(1, 100000) i;\n\nCREATE INDEX ON test USING gist (data);\n\nALTER EXTENSION ltree UPDATE TO \"1.2\";\n\n-- works, cached bytea options is still NULL and siglen is 28\nSELECT * FROM test WHERE data = '12345';\n\n\\connect\n\n-- crash, siglen = 8\nSELECT * FROM test WHERE data = '12345';\n\n\n\nYou can see here another problem: installation of opclass options\nprocedure does not invalidate relation's opclass options cache.\n\n\n-- \nNikita Glukhov\nPostgres Professional:http://www.postgrespro.com\nThe Russian Postgres Company\n\n\n\n\n\n\nOn 04.03.2022 23:28, Tom Lane wrote:\n\n\nTomas Vondra <tomas.vondra@enterprisedb.com> writes:\n\n\nOn 3/4/22 20:29, Nikita Glukhov wrote:\n\n\nSo, we probably have corrupted indexes that were updated since such \n\"incomplete\" upgrade of ltree.\n\n\n\n\n\n\nIIRC pg_upgrade is not expected to upgrade extensions - it keeps the\ninstalled version of the extension, and that's intentional.\n\n\n\nYeah, exactly. But this opens up an additional consideration we\nhave to account for: whatever we do needs to work with either 1.1\nor 1.2 SQL-level versions of the extension.\n\n\t\t\tregards, tom lane\n\n\nIt becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem \nis not so much related to PG12 => PG13+ upgrades.\n\n\nThe following examples crashes PG13+:\n\nCREATE EXTENSION ltree VERSION \"1.1\";\n\nCREATE TABLE test AS \nSELECT i::text::ltree data \nFROM generate_series(1, 100000) i;\n\nCREATE INDEX ON test USING gist (data);\n\nALTER EXTENSION ltree UPDATE TO \"1.2\";\n\n-- works, cached bytea options is still NULL and siglen is 28\nSELECT * FROM test WHERE data = '12345';\n\n\\connect\n\n-- crash, siglen = 8\nSELECT * FROM test WHERE data = '12345'; \n\n\n\nYou can see here another problem: installation of opclass options \nprocedure does not invalidate relation's opclass options cache.\n\n\n-- \nNikita Glukhov\nPostgres Professional: http://www.postgrespro.com\nThe Russian Postgres Company","msg_date":"Sat, 5 Mar 2022 01:09:26 +0300","msg_from":"Nikita Glukhov <n.gluhov@postgrespro.ru>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"On 3/4/22 23:09, Nikita Glukhov wrote:\n> On 04.03.2022 23:28, Tom Lane wrote:\n> \n>> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n>>> On 3/4/22 20:29, Nikita Glukhov wrote:\n>>>> So, we probably have corrupted indexes that were updated since such \n>>>> \"incomplete\" upgrade of ltree.\n>>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\n>>> installed version of the extension, and that's intentional.\n>> Yeah, exactly. But this opens up an additional consideration we\n>> have to account for: whatever we do needs to work with either 1.1\n>> or 1.2 SQL-level versions of the extension.\n>>\n>> \t\t\tregards, tom lane\n> \n> It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem \n> is not so much related to PG12 => PG13+ upgrades.\n> \n\nWell, it's quite a mess :-(\n\nIt very clearly is not just 1.1 -> 1.2 upgrade issue, because you can\nget a crash with 1.1 after PG12 -> PG13 upgrade, as demonstrated\nearlier. So just \"fixing\" the extension upgrade is no enough.\n\nBut as you showed, 1.1 -> 1.2 upgrade is broken too.\n\n> \n> You can see here another problem: installation of opclass options \n> procedure does not invalidate relation's opclass options cache.\n>\n\nSeems like that.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Sat, 5 Mar 2022 08:13:00 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Hi!\n\nSorry for this terrible oversight by me.\n\nOn Sat, Mar 5, 2022 at 10:13 AM Tomas Vondra\n<tomas.vondra@enterprisedb.com> wrote:\n> On 3/4/22 23:09, Nikita Glukhov wrote:\n> > On 04.03.2022 23:28, Tom Lane wrote:\n> >\n> >> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n> >>> On 3/4/22 20:29, Nikita Glukhov wrote:\n> >>>> So, we probably have corrupted indexes that were updated since such\n> >>>> \"incomplete\" upgrade of ltree.\n> >>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\n> >>> installed version of the extension, and that's intentional.\n> >> Yeah, exactly. But this opens up an additional consideration we\n> >> have to account for: whatever we do needs to work with either 1.1\n> >> or 1.2 SQL-level versions of the extension.\n> >>\n> >> regards, tom lane\n> >\n> > It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\n> > is not so much related to PG12 => PG13+ upgrades.\n\nSo, it seems that ltree 1.1 in PG13+ is incompatible with ltree on\nPG12 and ltree 1.2 on PG13+. And there are many scenarios involving.\n\nIt seems too difficult to identify all the broken cases in the release\nnotes. What about applying a patch and asking all ltree users to\nreindex their indexes?\n\n> Well, it's quite a mess :-(\n>\n> It very clearly is not just 1.1 -> 1.2 upgrade issue, because you can\n> get a crash with 1.1 after PG12 -> PG13 upgrade, as demonstrated\n> earlier. So just \"fixing\" the extension upgrade is no enough.\n>\n> But as you showed, 1.1 -> 1.2 upgrade is broken too.\n>\n> >\n> > You can see here another problem: installation of opclass options\n> > procedure does not invalidate relation's opclass options cache.\n> >\n>\n> Seems like that.\n\nI think opclass options procedure shouldn't normally change opclass\noptions chache. Before installation of options procedure, there\nshould be no options. And options procedure shouldn't change the\ndefaults in this case.\n\n------\nRegards,\nAlexander Korotkov\n\n\n","msg_date":"Sun, 6 Mar 2022 10:09:43 +0300","msg_from":"Alexander Korotkov <aekorotkov@gmail.com>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"\nOn 3/4/22 15:28, Tom Lane wrote:\n> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n>> On 3/4/22 20:29, Nikita Glukhov wrote:\n>>> So, we probably have corrupted indexes that were updated since such \n>>> \"incomplete\" upgrade of ltree.\n>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\n>> installed version of the extension, and that's intentional.\n> Yeah, exactly. But this opens up an additional consideration we\n> have to account for: whatever we do needs to work with either 1.1\n> or 1.2 SQL-level versions of the extension.\n>\n> \t\t\t\n\n\nThis is an area not currently touched by the buildfarm's cross version\nupgrade testing, which basically compares a pre-upgrade and post-upgrade\ndump of the databases. The upgraded cluster does contain\ncontrib_regression_ltree.\n\nI'm open to suggestions on how we might improve the buildfarm's testing\nof upgraded indexes generally.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Sun, 6 Mar 2022 07:46:04 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"On 3/6/22 08:09, Alexander Korotkov wrote:\n> Hi!\n> \n> Sorry for this terrible oversight by me.\n> \n> On Sat, Mar 5, 2022 at 10:13 AM Tomas Vondra\n> <tomas.vondra@enterprisedb.com> wrote:\n>> On 3/4/22 23:09, Nikita Glukhov wrote:\n>>> On 04.03.2022 23:28, Tom Lane wrote:\n>>>\n>>>> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n>>>>> On 3/4/22 20:29, Nikita Glukhov wrote:\n>>>>>> So, we probably have corrupted indexes that were updated since such\n>>>>>> \"incomplete\" upgrade of ltree.\n>>>>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\n>>>>> installed version of the extension, and that's intentional.\n>>>> Yeah, exactly. But this opens up an additional consideration we\n>>>> have to account for: whatever we do needs to work with either 1.1\n>>>> or 1.2 SQL-level versions of the extension.\n>>>>\n>>>> regards, tom lane\n>>>\n>>> It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\n>>> is not so much related to PG12 => PG13+ upgrades.\n> \n> So, it seems that ltree 1.1 in PG13+ is incompatible with ltree on\n> PG12 and ltree 1.2 on PG13+. And there are many scenarios involving.\n> \n> It seems too difficult to identify all the broken cases in the release\n> notes. What about applying a patch and asking all ltree users to\n> reindex their indexes?\n> \n\nYeah. I think this is getting so complicated that there's little chance\nwe'd be able to clearly explain when to reindex.\n\n>> Well, it's quite a mess :-(\n>>\n>> It very clearly is not just 1.1 -> 1.2 upgrade issue, because you can\n>> get a crash with 1.1 after PG12 -> PG13 upgrade, as demonstrated\n>> earlier. So just \"fixing\" the extension upgrade is no enough.\n>>\n>> But as you showed, 1.1 -> 1.2 upgrade is broken too.\n>>\n>>>\n>>> You can see here another problem: installation of opclass options\n>>> procedure does not invalidate relation's opclass options cache.\n>>>\n>>\n>> Seems like that.\n> \n> I think opclass options procedure shouldn't normally change opclass\n> options chache. Before installation of options procedure, there\n> should be no options. And options procedure shouldn't change the\n> defaults in this case.\n> \n\nI should clarify I haven't looked into the details - I merely verified\naccessing the index in the same backend works, and after reconnection it\ncrashes (per Nikita's example). I don't know why that happans.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Sun, 6 Mar 2022 18:28:25 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Hi,\n\nOn 2022-03-06 07:46:04 -0500, Andrew Dunstan wrote:\n> This is an area not currently touched by the buildfarm's cross version\n> upgrade testing, which basically compares a pre-upgrade and post-upgrade\n> dump of the databases. The upgraded cluster does contain\n> contrib_regression_ltree.\n> \n> I'm open to suggestions on how we might improve the buildfarm's testing\n> of upgraded indexes generally.\n\nOne thing that's likely worth doing as part of the cross version upgrade test,\neven if it wouldn't even help in this case, is to run amcheck post\nupgrade. Just dumping data isn't going to touch indices at all.\n\nA sequence of\n pg_upgrade; amcheck; upgrade all extensions; amcheck;\nwould make sense.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sun, 6 Mar 2022 14:33:52 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"пн, 7 мар. 2022 г. в 00:34, Andres Freund <andres@anarazel.de>:\n\n> One thing that's likely worth doing as part of the cross version upgrade\n> test,\n> even if it wouldn't even help in this case, is to run amcheck post\n> upgrade. Just dumping data isn't going to touch indices at all.\n>\n> A sequence of\n> pg_upgrade; amcheck; upgrade all extensions; amcheck;\n> would make sense.\n>\n\nIs it possible to amcheck gist indexes?..\n\n\n-- \nVictor Yegorov\n\nпн, 7 мар. 2022 г. в 00:34, Andres Freund <andres@anarazel.de>:One thing that's likely worth doing as part of the cross version upgrade test,\neven if it wouldn't even help in this case, is to run amcheck post\nupgrade. Just dumping data isn't going to touch indices at all.\n\nA sequence of\n pg_upgrade; amcheck; upgrade all extensions; amcheck;\nwould make sense.Is it possible to amcheck gist indexes?.. -- Victor Yegorov","msg_date":"Mon, 7 Mar 2022 01:12:07 +0200","msg_from":"Victor Yegorov <vyegorov@gmail.com>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Hi,\n\nOn 2022-03-07 01:12:07 +0200, Victor Yegorov wrote:\n> пн, 7 мар. 2022 г. в 00:34, Andres Freund <andres@anarazel.de>:\n> \n> > One thing that's likely worth doing as part of the cross version upgrade\n> > test,\n> > even if it wouldn't even help in this case, is to run amcheck post\n> > upgrade. Just dumping data isn't going to touch indices at all.\n> >\n> > A sequence of\n> > pg_upgrade; amcheck; upgrade all extensions; amcheck;\n> > would make sense.\n> >\n> \n> Is it possible to amcheck gist indexes?..\n\nNo. That was what I was alluding to with \"even if it wouldn't even help in\nthis case\".\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sun, 6 Mar 2022 17:32:02 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"On Sun, Mar 6, 2022 at 8:28 PM Tomas Vondra\n<tomas.vondra@enterprisedb.com> wrote:\n> On 3/6/22 08:09, Alexander Korotkov wrote:\n> > Sorry for this terrible oversight by me.\n> >\n> > On Sat, Mar 5, 2022 at 10:13 AM Tomas Vondra\n> > <tomas.vondra@enterprisedb.com> wrote:\n> >> On 3/4/22 23:09, Nikita Glukhov wrote:\n> >>> On 04.03.2022 23:28, Tom Lane wrote:\n> >>>\n> >>>> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n> >>>>> On 3/4/22 20:29, Nikita Glukhov wrote:\n> >>>>>> So, we probably have corrupted indexes that were updated since such\n> >>>>>> \"incomplete\" upgrade of ltree.\n> >>>>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\n> >>>>> installed version of the extension, and that's intentional.\n> >>>> Yeah, exactly. But this opens up an additional consideration we\n> >>>> have to account for: whatever we do needs to work with either 1.1\n> >>>> or 1.2 SQL-level versions of the extension.\n> >>>>\n> >>>> regards, tom lane\n> >>>\n> >>> It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\n> >>> is not so much related to PG12 => PG13+ upgrades.\n> >\n> > So, it seems that ltree 1.1 in PG13+ is incompatible with ltree on\n> > PG12 and ltree 1.2 on PG13+. And there are many scenarios involving.\n> >\n> > It seems too difficult to identify all the broken cases in the release\n> > notes. What about applying a patch and asking all ltree users to\n> > reindex their indexes?\n> >\n>\n> Yeah. I think this is getting so complicated that there's little chance\n> we'd be able to clearly explain when to reindex.\n\nGood. The revised patch is attached. Instead of adding argument to\nLTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\nLTREE_GET_ASIGLEN() macros.\n\n------\nRegards,\nAlexander Korotkov","msg_date":"Tue, 8 Mar 2022 02:05:04 +0300","msg_from":"Alexander Korotkov <aekorotkov@gmail.com>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"On Tue, Mar 8, 2022 at 2:05 AM Alexander Korotkov <aekorotkov@gmail.com> wrote:\n> On Sun, Mar 6, 2022 at 8:28 PM Tomas Vondra\n> <tomas.vondra@enterprisedb.com> wrote:\n> > On 3/6/22 08:09, Alexander Korotkov wrote:\n> > > Sorry for this terrible oversight by me.\n> > >\n> > > On Sat, Mar 5, 2022 at 10:13 AM Tomas Vondra\n> > > <tomas.vondra@enterprisedb.com> wrote:\n> > >> On 3/4/22 23:09, Nikita Glukhov wrote:\n> > >>> On 04.03.2022 23:28, Tom Lane wrote:\n> > >>>\n> > >>>> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\n> > >>>>> On 3/4/22 20:29, Nikita Glukhov wrote:\n> > >>>>>> So, we probably have corrupted indexes that were updated since such\n> > >>>>>> \"incomplete\" upgrade of ltree.\n> > >>>>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\n> > >>>>> installed version of the extension, and that's intentional.\n> > >>>> Yeah, exactly. But this opens up an additional consideration we\n> > >>>> have to account for: whatever we do needs to work with either 1.1\n> > >>>> or 1.2 SQL-level versions of the extension.\n> > >>>>\n> > >>>> regards, tom lane\n> > >>>\n> > >>> It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\n> > >>> is not so much related to PG12 => PG13+ upgrades.\n> > >\n> > > So, it seems that ltree 1.1 in PG13+ is incompatible with ltree on\n> > > PG12 and ltree 1.2 on PG13+. And there are many scenarios involving.\n> > >\n> > > It seems too difficult to identify all the broken cases in the release\n> > > notes. What about applying a patch and asking all ltree users to\n> > > reindex their indexes?\n> > >\n> >\n> > Yeah. I think this is getting so complicated that there's little chance\n> > we'd be able to clearly explain when to reindex.\n>\n> Good. The revised patch is attached. Instead of adding argument to\n> LTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\n> LTREE_GET_ASIGLEN() macros.\n\nNo feedback yet. I'm going to push this if no objections.\n\n------\nRegards,\nAlexander Korotkov\n\n\n","msg_date":"Thu, 10 Mar 2022 10:09:24 +0300","msg_from":"Alexander Korotkov <aekorotkov@gmail.com>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"On 3/10/22 08:09, Alexander Korotkov wrote:\n> On Tue, Mar 8, 2022 at 2:05 AM Alexander Korotkov <aekorotkov@gmail.com> wrote:\n>> ...\n>>\n>> Good. The revised patch is attached. Instead of adding argument to\n>> LTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\n>> LTREE_GET_ASIGLEN() macros.\n> \n> No feedback yet. I'm going to push this if no objections.\n> \n\nWFM. I certainly don't have a better idea how to fix this.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 10 Mar 2022 13:09:12 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Alexander Korotkov <aekorotkov@gmail.com> writes:\n> Good. The revised patch is attached. Instead of adding argument to\n> LTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\n> LTREE_GET_ASIGLEN() macros.\n\nUm ... what I see after applying the patch is two redundant\ndefinitions of LTREE_GET_ASIGLEN, and no LTREE_GET_SIGLEN.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 10 Mar 2022 10:39:39 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"\nOn 3/6/22 17:33, Andres Freund wrote:\n> Hi,\n>\n> On 2022-03-06 07:46:04 -0500, Andrew Dunstan wrote:\n>> This is an area not currently touched by the buildfarm's cross version\n>> upgrade testing, which basically compares a pre-upgrade and post-upgrade\n>> dump of the databases. The upgraded cluster does contain\n>> contrib_regression_ltree.\n>>\n>> I'm open to suggestions on how we might improve the buildfarm's testing\n>> of upgraded indexes generally.\n> One thing that's likely worth doing as part of the cross version upgrade test,\n> even if it wouldn't even help in this case, is to run amcheck post\n> upgrade. Just dumping data isn't going to touch indices at all.\n>\n> A sequence of\n> pg_upgrade; amcheck; upgrade all extensions; amcheck;\n> would make sense.\n>\n\n\nSee\n<https://github.com/PGBuildFarm/client-code/commit/191df23bd25eb5546b0989d71ae92747151f9f39>\n\n\n\nThis will be in the next release\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 10 Mar 2022 10:43:06 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Andrew Dunstan <andrew@dunslane.net> writes:\n> On 3/6/22 17:33, Andres Freund wrote:\n>> A sequence of\n>> pg_upgrade; amcheck; upgrade all extensions; amcheck;\n>> would make sense.\n\n> See\n> <https://github.com/PGBuildFarm/client-code/commit/191df23bd25eb5546b0989d71ae92747151f9f39>\n\nDoes this detect the problem at hand?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 10 Mar 2022 10:53:49 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"\nOn 3/10/22 10:53, Tom Lane wrote:\n> Andrew Dunstan <andrew@dunslane.net> writes:\n>> On 3/6/22 17:33, Andres Freund wrote:\n>>> A sequence of\n>>> pg_upgrade; amcheck; upgrade all extensions; amcheck;\n>>> would make sense.\n>> See\n>> <https://github.com/PGBuildFarm/client-code/commit/191df23bd25eb5546b0989d71ae92747151f9f39>\n> Does this detect the problem at hand?\n>\n> \t\t\t\n\n\nAIUI, amcheck doesn't check gist indexes, hence Andres' \"even if it\nwouldn't even help in this case\". So no.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 10 Mar 2022 12:05:31 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"On 2022-03-10 10:43:06 -0500, Andrew Dunstan wrote:\n> On 3/6/22 17:33, Andres Freund wrote:\n> > One thing that's likely worth doing as part of the cross version upgrade test,\n> > even if it wouldn't even help in this case, is to run amcheck post\n> > upgrade. Just dumping data isn't going to touch indices at all.\n> >\n> > A sequence of\n> > pg_upgrade; amcheck; upgrade all extensions; amcheck;\n> > would make sense.\n\n> See\n> <https://github.com/PGBuildFarm/client-code/commit/191df23bd25eb5546b0989d71ae92747151f9f39>\n > This will be in the next release\n\nGreat, thanks for doing that!\n\n\n","msg_date":"Thu, 10 Mar 2022 09:24:19 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"On Thu, Mar 10, 2022 at 6:39 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Alexander Korotkov <aekorotkov@gmail.com> writes:\n> > Good. The revised patch is attached. Instead of adding argument to\n> > LTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\n> > LTREE_GET_ASIGLEN() macros.\n>\n> Um ... what I see after applying the patch is two redundant\n> definitions of LTREE_GET_ASIGLEN, and no LTREE_GET_SIGLEN.\n\nSorry for publishing a patch, which doesn't even compile.\nThe attached version should be good.\n\n------\nRegards,\nAlexander Korotkov","msg_date":"Thu, 10 Mar 2022 20:44:13 +0300","msg_from":"Alexander Korotkov <aekorotkov@gmail.com>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"},{"msg_contents":"Alexander Korotkov <aekorotkov@gmail.com> writes:\n> Sorry for publishing a patch, which doesn't even compile.\n> The attached version should be good.\n\nOK, better.\n\nSo to clarify: we are going to back-patch as far as v13, and\nthen in the minor releases we are going to tell people they\nneed to REINDEX all gist indexes on ltree? If so, please\nmake that clear in the commit message, so I don't forget it\ncome May ;-)\n\nAlso, it looks like reindex is *not* required for indexes on\nltree[] ? That is, gist_ltree_ops indexes are affected but\nnot gist__ltree_ops?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 10 Mar 2022 17:36:50 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nVictor Yegorov reported a crash related to a GiST index on ltree [1],\\nfollowing a pg_upgrade from 12.x to 14.x, with a data set reproducing\\nthis. I spent some time investigating this, and it seems this is a silly\\nbug in commit\\n\\n commit 911e70207703799605f5a0e8aad9f06cff067c63 (HEAD)\\n Author: Alexander Korotkov <akorotkov(at)postgresql(dot)org>\\n Date: Mon Mar 30 19:17:11 2020 +0300\\n\\n Implement operator class parameters\\n ...\\n\\nin PG13, which modified ltree_gist so that siglen is opclass parameter\\n(and not hard-coded). But the procedures use LTREE_GET_ASIGLEN macro to\\nextract the value, which either extracts the value from the catalog (if\\nthe index has opclass parameter) or uses a default value - but it always\\nuses LTREE_ASIGLEN_DEFAULT, which belongs to _ltree_gist opclass (i.e.\\nfor array of ltree). And that's 28 instead of the 8, as it should be.\\n\\nAttached is a simple patch tweaking the macros, which resolves the\\nissue. Maybe there should be a separate macro though, because \\\"ASIGLEN\\\"\\nprobably refers to \\\"array siglen\\\".\\n\\nBut the bigger issue is this fixes indexes created before the upgrade,\\nbut it breaks indexes created after it. Because those were created with\\nsiglen 28B, so reverting to 8B breaks them. A bit of a symmetry :-/\\n\\nI'm not sure what to do about this. There's no way to distinguish\\nindexes, and I think an index may actually be broken both with and\\nwithout the patch - imagine an index created before the upgrade (so\\nusing siglen 8), but then receiving a couple new rows (siglen 28).\\n\\nAfter thinking about this I only see two options:\\n\\n1) Don't apply the patch and tell everyone using ltree_gist they need to\\nrebuild the index after pg_upgrade from 12 to 13+. The downside of this\\nis larger indexes (because some tuples are 20B larger).\\n\\n2) Apply the patch + tell those who already upgraded from 12 to rebuild\\nltree_gist indexes, because those might be broken due to new inserts.\\n\\n\\nMy opinion is to do (2), because at least those who'll upgrade later\\n(which is a lot of people) will be fine without a rebuild. And it also\\nmakes the indexes a bit smaller, thanks to saving 20B.\\n\\n\\nregards\\n\\n\\n[1]\\nhttps://www.postgresql.org/message-id/17406-71e02820ae79bb40%40postgresql.org\\n\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Thu, 24 Feb 2022 22:15:20 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n> After thinking about this I only see two options:\\n\\n> 1) Don't apply the patch and tell everyone using ltree_gist they need to\\n> rebuild the index after pg_upgrade from 12 to 13+. The downside of this\\n> is larger indexes (because some tuples are 20B larger).\\n\\n> 2) Apply the patch + tell those who already upgraded from 12 to rebuild\\n> ltree_gist indexes, because those might be broken due to new inserts.\\n\\nYeah, let's fix it and advise people to reindex those indexes.\\nWon't be the first time release notes have contained such advice.\\nPlus, it sounds like we'd have to advise people to reindex *anyway*,\\njust in a slightly different set of cases.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 16:51:36 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-24 22:15:20 +0100, Tomas Vondra wrote:\\n> After thinking about this I only see two options:\\n> \\n> 1) Don't apply the patch and tell everyone using ltree_gist they need to\\n> rebuild the index after pg_upgrade from 12 to 13+. The downside of this\\n> is larger indexes (because some tuples are 20B larger).\\n> \\n> 2) Apply the patch + tell those who already upgraded from 12 to rebuild\\n> ltree_gist indexes, because those might be broken due to new inserts.\\n> \\n> \\n> My opinion is to do (2), because at least those who'll upgrade later\\n> (which is a lot of people) will be fine without a rebuild. And it also\\n> makes the indexes a bit smaller, thanks to saving 20B.\\n\\nWon't 2) also break indexes created without a pg_upgrade? \\\"already upgraded\\nfrom 12\\\" sounds like it wouldn't but I don't see why?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 14:06:58 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2/24/22 23:06, Andres Freund wrote:\\n> Hi,\\n> \\n> On 2022-02-24 22:15:20 +0100, Tomas Vondra wrote:\\n>> After thinking about this I only see two options:\\n>>\\n>> 1) Don't apply the patch and tell everyone using ltree_gist they need to\\n>> rebuild the index after pg_upgrade from 12 to 13+. The downside of this\\n>> is larger indexes (because some tuples are 20B larger).\\n>>\\n>> 2) Apply the patch + tell those who already upgraded from 12 to rebuild\\n>> ltree_gist indexes, because those might be broken due to new inserts.\\n>>\\n>>\\n>> My opinion is to do (2), because at least those who'll upgrade later\\n>> (which is a lot of people) will be fine without a rebuild. And it also\\n>> makes the indexes a bit smaller, thanks to saving 20B.\\n> \\n> Won't 2) also break indexes created without a pg_upgrade? \\\"already upgraded\\n> from 12\\\" sounds like it wouldn't but I don't see why?\\n> \\n\\nIt will, unfortunately - that's why I wrote \\\"upgrade\\\" in that sentence.\\nI should have been more explicit, sorry. But any new index tuples formed\\nafter starting the 13+ cluster are/may be corrupted.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 23:13:18 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2/24/22 23:13, Tomas Vondra wrote:\\n> \\n> \\n> On 2/24/22 23:06, Andres Freund wrote:\\n>> Hi,\\n>>\\n>> On 2022-02-24 22:15:20 +0100, Tomas Vondra wrote:\\n>>> After thinking about this I only see two options:\\n>>>\\n>>> 1) Don't apply the patch and tell everyone using ltree_gist they need to\\n>>> rebuild the index after pg_upgrade from 12 to 13+. The downside of this\\n>>> is larger indexes (because some tuples are 20B larger).\\n>>>\\n>>> 2) Apply the patch + tell those who already upgraded from 12 to rebuild\\n>>> ltree_gist indexes, because those might be broken due to new inserts.\\n>>>\\n>>>\\n>>> My opinion is to do (2), because at least those who'll upgrade later\\n>>> (which is a lot of people) will be fine without a rebuild. And it also\\n>>> makes the indexes a bit smaller, thanks to saving 20B.\\n>>\\n>> Won't 2) also break indexes created without a pg_upgrade? \\\"already upgraded\\n>> from 12\\\" sounds like it wouldn't but I don't see why?\\n>>\\n> \\n> It will, unfortunately - that's why I wrote \\\"upgrade\\\" in that sentence.\\n> I should have been more explicit, sorry. But any new index tuples formed\\n> after starting the 13+ cluster are/may be corrupted.\\n> \\n\\nI wonder if we could check the index tuple length, and adjust the siglen\\nbased on that, somehow ...\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 23:22:31 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n> I wonder if we could check the index tuple length, and adjust the siglen\\n> based on that, somehow ...\\n\\nIn an already-corrupted index, there won't be a unique answer.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 20:44:57 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Hi, \\n\\nOn February 24, 2022 5:44:57 PM PST, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n>> I wonder if we could check the index tuple length, and adjust the siglen\\n>> based on that, somehow ...\\n>\\n>In an already-corrupted index, there won't be a unique answer.\\n\\nIf we're breaking every ltree gist index, can we detect that it's not an index generated by the new version? Most people don't read release notes, and this is just about guaranteed to break all. \\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 17:51:41 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"On 25.02.2022 00:15, Tomas Vondra wrote:\\n\\n> Hi,\\n>\\n> Victor Yegorov reported a crash related to a GiST index on ltree [1],\\n> following a pg_upgrade from 12.x to 14.x, with a data set reproducing\\n> this. I spent some time investigating this, and it seems this is a silly\\n> bug in commit\\n>\\n> commit 911e70207703799605f5a0e8aad9f06cff067c63 (HEAD)\\n> Author: Alexander Korotkov <akorotkov(at)postgresql(dot)org>\\n> Date: Mon Mar 30 19:17:11 2020 +0300\\n>\\n> Implement operator class parameters\\n> ...\\n>\\n> in PG13, which modified ltree_gist so that siglen is opclass parameter\\n> (and not hard-coded). But the procedures use LTREE_GET_ASIGLEN macro to\\n> extract the value, which either extracts the value from the catalog (if\\n> the index has opclass parameter) or uses a default value - but it always\\n> uses LTREE_ASIGLEN_DEFAULT, which belongs to _ltree_gist opclass (i.e.\\n> for array of ltree). And that's 28 instead of the 8, as it should be.\\n\\nIt seems that ltree extension simply was not updated to v1.2 after\\npg_upgrade:\\n ALTER EXTENSION ltree UPDATE TO '1.2'; -- is missing\\n\\nUpgrade script ltree--1.1--1.2.sql creates ltree_gist_options() and\\nregisters it in the opclass. ltree_gist_options() initializes bytea\\noptions using the correct SIGLEN_DEFAULT=8.\\n\\nIf ltree_gist_options() is absent, LTREE_GET_ASIGLEN() receives NULL\\nand wrong LTREE_ASIGLEN_DEFAULT is used. But if ltree_gist_options()\\nis registered, LTREE_GET_ASIGLEN() receives non-NULL bytea options\\nwith the correct default value.\\n\\n\\nSo, we probably have corrupted indexes that were updated since such\\n\\\"incomplete\\\" upgrade of ltree.\\n\\n\\n\\nAlso I found that contrib/pg_trgm/trgm_gist.c uses wrongly named macro\\nLTREE_GET_ASIGLEN().\\n\\n-- \\nNikita Glukhov\\nPostgres Professional:http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\\n\\n\\n\\n\\nOn 25.02.2022 00:15, Tomas Vondra wrote:\\n\\n\\nHi,\\n\\nVictor Yegorov reported a crash related to a GiST index on ltree [1],\\nfollowing a pg_upgrade from 12.x to 14.x, with a data set reproducing\\nthis. I spent some time investigating this, and it seems this is a silly\\nbug in commit\\n\\n commit 911e70207703799605f5a0e8aad9f06cff067c63 (HEAD)\\n Author: Alexander Korotkov <akorotkov(at)postgresql(dot)org>\\n Date: Mon Mar 30 19:17:11 2020 +0300\\n\\n Implement operator class parameters\\n ...\\n\\nin PG13, which modified ltree_gist so that siglen is opclass parameter\\n(and not hard-coded). But the procedures use LTREE_GET_ASIGLEN macro to\\nextract the value, which either extracts the value from the catalog (if\\nthe index has opclass parameter) or uses a default value - but it always\\nuses LTREE_ASIGLEN_DEFAULT, which belongs to _ltree_gist opclass (i.e.\\nfor array of ltree). And that's 28 instead of the 8, as it should be.\\n\\n\\nIt seems that ltree extension simply was not updated to v1.2 after \\npg_upgrade:\\n ALTER EXTENSION ltree UPDATE TO '1.2'; -- is missing\\nUpgrade script ltree--1.1--1.2.sql creates ltree_gist_options() and \\nregisters it in the opclass. ltree_gist_options() initializes bytea \\noptions using the correct SIGLEN_DEFAULT=8.\\n\\nIf ltree_gist_options() is absent, LTREE_GET_ASIGLEN() receives NULL\\nand wrong LTREE_ASIGLEN_DEFAULT is used. But if ltree_gist_options()\\nis registered, LTREE_GET_ASIGLEN() receives non-NULL bytea options \\nwith the correct default value.\\n\\n\\nSo, we probably have corrupted indexes that were updated since such \\n\\\"incomplete\\\" upgrade of ltree.\\n\\n\\n\\nAlso I found that contrib/pg_trgm/trgm_gist.c uses wrongly named macro\\nLTREE_GET_ASIGLEN().\\n\\n-- \\nNikita Glukhov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\",\n \"msg_date\": \"Fri, 4 Mar 2022 22:29:31 +0300\",\n \"msg_from\": \"Nikita Glukhov <n.gluhov@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"\\nOn 3/4/22 20:29, Nikita Glukhov wrote:\\n> On 25.02.2022 00:15, Tomas Vondra wrote:\\n> \\n>> Hi,\\n>>\\n>> Victor Yegorov reported a crash related to a GiST index on ltree [1],\\n>> following a pg_upgrade from 12.x to 14.x, with a data set reproducing\\n>> this. I spent some time investigating this, and it seems this is a silly\\n>> bug in commit\\n>>\\n>> commit 911e70207703799605f5a0e8aad9f06cff067c63 (HEAD)\\n>> Author: Alexander Korotkov <akorotkov(at)postgresql(dot)org>\\n>> Date: Mon Mar 30 19:17:11 2020 +0300\\n>>\\n>> Implement operator class parameters\\n>> ...\\n>>\\n>> in PG13, which modified ltree_gist so that siglen is opclass parameter\\n>> (and not hard-coded). But the procedures use LTREE_GET_ASIGLEN macro to\\n>> extract the value, which either extracts the value from the catalog (if\\n>> the index has opclass parameter) or uses a default value - but it always\\n>> uses LTREE_ASIGLEN_DEFAULT, which belongs to _ltree_gist opclass (i.e.\\n>> for array of ltree). And that's 28 instead of the 8, as it should be.\\n> \\n> It seems that ltree extension simply was not updated to v1.2 after \\n> pg_upgrade:\\n> ALTER EXTENSION ltree UPDATE TO '1.2'; -- is missing\\n> \\n> Upgrade script ltree--1.1--1.2.sql creates ltree_gist_options() and \\n> registers it in the opclass. ltree_gist_options() initializes bytea \\n> options using the correct SIGLEN_DEFAULT=8.\\n> \\n> If ltree_gist_options() is absent, LTREE_GET_ASIGLEN() receives NULL\\n> and wrong LTREE_ASIGLEN_DEFAULT is used. But if ltree_gist_options()\\n> is registered, LTREE_GET_ASIGLEN() receives non-NULL bytea options \\n> with the correct default value.\\n>\\n> \\n> So, we probably have corrupted indexes that were updated since such \\n> \\\"incomplete\\\" upgrade of ltree.\\n> \\n\\nIIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\ninstalled version of the extension, and that's intentional.\\n\\nMoreover, it's perfectly legal to install older version, so you can do\\n\\n CREATE EXTENSION ltree VERSION '1.1';\\n\\nSo I don't think we can call this \\\"incomplete upgrade\\\".\\n\\n> \\n> Also I found that contrib/pg_trgm/trgm_gist.c uses wrongly named macro\\n> LTREE_GET_ASIGLEN().\\n\\nDoes it? When I grep for LTREE_GET_ASIGLEN, I don't get any matches in\\npg_trgm.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Mar 2022 20:58:56 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n> On 3/4/22 20:29, Nikita Glukhov wrote:\\n>> So, we probably have corrupted indexes that were updated since such \\n>> \\\"incomplete\\\" upgrade of ltree.\\n\\n> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\n> installed version of the extension, and that's intentional.\\n\\nYeah, exactly. But this opens up an additional consideration we\\nhave to account for: whatever we do needs to work with either 1.1\\nor 1.2 SQL-level versions of the extension.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 04 Mar 2022 15:28:45 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"On 04.03.2022 23:28, Tom Lane wrote:\\n\\n> Tomas Vondra<tomas.vondra@enterprisedb.com> writes:\\n>> On 3/4/22 20:29, Nikita Glukhov wrote:\\n>>> So, we probably have corrupted indexes that were updated since such\\n>>> \\\"incomplete\\\" upgrade of ltree.\\n>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\n>> installed version of the extension, and that's intentional.\\n> Yeah, exactly. But this opens up an additional consideration we\\n> have to account for: whatever we do needs to work with either 1.1\\n> or 1.2 SQL-level versions of the extension.\\n>\\n> \\t\\t\\tregards, tom lane\\n\\nIt becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\\nis not so much related to PG12 => PG13+ upgrades.\\n\\n\\nThe following examples crashes PG13+:\\n\\nCREATE EXTENSION ltree VERSION \\\"1.1\\\";\\n\\nCREATE TABLE test AS\\nSELECT i::text::ltree data\\nFROM generate_series(1, 100000) i;\\n\\nCREATE INDEX ON test USING gist (data);\\n\\nALTER EXTENSION ltree UPDATE TO \\\"1.2\\\";\\n\\n-- works, cached bytea options is still NULL and siglen is 28\\nSELECT * FROM test WHERE data = '12345';\\n\\n\\\\connect\\n\\n-- crash, siglen = 8\\nSELECT * FROM test WHERE data = '12345';\\n\\n\\n\\nYou can see here another problem: installation of opclass options\\nprocedure does not invalidate relation's opclass options cache.\\n\\n\\n-- \\nNikita Glukhov\\nPostgres Professional:http://www.postgrespro.com\\nThe Russian Postgres Company\\n\\n\\n\\n\\n\\n\\nOn 04.03.2022 23:28, Tom Lane wrote:\\n\\n\\nTomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n\\n\\nOn 3/4/22 20:29, Nikita Glukhov wrote:\\n\\n\\nSo, we probably have corrupted indexes that were updated since such \\n\\\"incomplete\\\" upgrade of ltree.\\n\\n\\n\\n\\n\\n\\nIIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\ninstalled version of the extension, and that's intentional.\\n\\n\\n\\nYeah, exactly. But this opens up an additional consideration we\\nhave to account for: whatever we do needs to work with either 1.1\\nor 1.2 SQL-level versions of the extension.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\nIt becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem \\nis not so much related to PG12 => PG13+ upgrades.\\n\\n\\nThe following examples crashes PG13+:\\n\\nCREATE EXTENSION ltree VERSION \\\"1.1\\\";\\n\\nCREATE TABLE test AS \\nSELECT i::text::ltree data \\nFROM generate_series(1, 100000) i;\\n\\nCREATE INDEX ON test USING gist (data);\\n\\nALTER EXTENSION ltree UPDATE TO \\\"1.2\\\";\\n\\n-- works, cached bytea options is still NULL and siglen is 28\\nSELECT * FROM test WHERE data = '12345';\\n\\n\\\\connect\\n\\n-- crash, siglen = 8\\nSELECT * FROM test WHERE data = '12345'; \\n\\n\\n\\nYou can see here another problem: installation of opclass options \\nprocedure does not invalidate relation's opclass options cache.\\n\\n\\n-- \\nNikita Glukhov\\nPostgres Professional: http://www.postgrespro.com\\nThe Russian Postgres Company\",\n \"msg_date\": \"Sat, 5 Mar 2022 01:09:26 +0300\",\n \"msg_from\": \"Nikita Glukhov <n.gluhov@postgrespro.ru>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"On 3/4/22 23:09, Nikita Glukhov wrote:\\n> On 04.03.2022 23:28, Tom Lane wrote:\\n> \\n>> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n>>> On 3/4/22 20:29, Nikita Glukhov wrote:\\n>>>> So, we probably have corrupted indexes that were updated since such \\n>>>> \\\"incomplete\\\" upgrade of ltree.\\n>>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\n>>> installed version of the extension, and that's intentional.\\n>> Yeah, exactly. But this opens up an additional consideration we\\n>> have to account for: whatever we do needs to work with either 1.1\\n>> or 1.2 SQL-level versions of the extension.\\n>>\\n>> \\t\\t\\tregards, tom lane\\n> \\n> It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem \\n> is not so much related to PG12 => PG13+ upgrades.\\n> \\n\\nWell, it's quite a mess :-(\\n\\nIt very clearly is not just 1.1 -> 1.2 upgrade issue, because you can\\nget a crash with 1.1 after PG12 -> PG13 upgrade, as demonstrated\\nearlier. So just \\\"fixing\\\" the extension upgrade is no enough.\\n\\nBut as you showed, 1.1 -> 1.2 upgrade is broken too.\\n\\n> \\n> You can see here another problem: installation of opclass options \\n> procedure does not invalidate relation's opclass options cache.\\n>\\n\\nSeems like that.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sat, 5 Mar 2022 08:13:00 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Hi!\\n\\nSorry for this terrible oversight by me.\\n\\nOn Sat, Mar 5, 2022 at 10:13 AM Tomas Vondra\\n<tomas.vondra@enterprisedb.com> wrote:\\n> On 3/4/22 23:09, Nikita Glukhov wrote:\\n> > On 04.03.2022 23:28, Tom Lane wrote:\\n> >\\n> >> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n> >>> On 3/4/22 20:29, Nikita Glukhov wrote:\\n> >>>> So, we probably have corrupted indexes that were updated since such\\n> >>>> \\\"incomplete\\\" upgrade of ltree.\\n> >>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\n> >>> installed version of the extension, and that's intentional.\\n> >> Yeah, exactly. But this opens up an additional consideration we\\n> >> have to account for: whatever we do needs to work with either 1.1\\n> >> or 1.2 SQL-level versions of the extension.\\n> >>\\n> >> regards, tom lane\\n> >\\n> > It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\\n> > is not so much related to PG12 => PG13+ upgrades.\\n\\nSo, it seems that ltree 1.1 in PG13+ is incompatible with ltree on\\nPG12 and ltree 1.2 on PG13+. And there are many scenarios involving.\\n\\nIt seems too difficult to identify all the broken cases in the release\\nnotes. What about applying a patch and asking all ltree users to\\nreindex their indexes?\\n\\n> Well, it's quite a mess :-(\\n>\\n> It very clearly is not just 1.1 -> 1.2 upgrade issue, because you can\\n> get a crash with 1.1 after PG12 -> PG13 upgrade, as demonstrated\\n> earlier. So just \\\"fixing\\\" the extension upgrade is no enough.\\n>\\n> But as you showed, 1.1 -> 1.2 upgrade is broken too.\\n>\\n> >\\n> > You can see here another problem: installation of opclass options\\n> > procedure does not invalidate relation's opclass options cache.\\n> >\\n>\\n> Seems like that.\\n\\nI think opclass options procedure shouldn't normally change opclass\\noptions chache. Before installation of options procedure, there\\nshould be no options. And options procedure shouldn't change the\\ndefaults in this case.\\n\\n------\\nRegards,\\nAlexander Korotkov\\n\\n\\n\",\n \"msg_date\": \"Sun, 6 Mar 2022 10:09:43 +0300\",\n \"msg_from\": \"Alexander Korotkov <aekorotkov@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"\\nOn 3/4/22 15:28, Tom Lane wrote:\\n> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n>> On 3/4/22 20:29, Nikita Glukhov wrote:\\n>>> So, we probably have corrupted indexes that were updated since such \\n>>> \\\"incomplete\\\" upgrade of ltree.\\n>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\n>> installed version of the extension, and that's intentional.\\n> Yeah, exactly. But this opens up an additional consideration we\\n> have to account for: whatever we do needs to work with either 1.1\\n> or 1.2 SQL-level versions of the extension.\\n>\\n> \\t\\t\\t\\n\\n\\nThis is an area not currently touched by the buildfarm's cross version\\nupgrade testing, which basically compares a pre-upgrade and post-upgrade\\ndump of the databases. The upgraded cluster does contain\\ncontrib_regression_ltree.\\n\\nI'm open to suggestions on how we might improve the buildfarm's testing\\nof upgraded indexes generally.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 6 Mar 2022 07:46:04 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"On 3/6/22 08:09, Alexander Korotkov wrote:\\n> Hi!\\n> \\n> Sorry for this terrible oversight by me.\\n> \\n> On Sat, Mar 5, 2022 at 10:13 AM Tomas Vondra\\n> <tomas.vondra@enterprisedb.com> wrote:\\n>> On 3/4/22 23:09, Nikita Glukhov wrote:\\n>>> On 04.03.2022 23:28, Tom Lane wrote:\\n>>>\\n>>>> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n>>>>> On 3/4/22 20:29, Nikita Glukhov wrote:\\n>>>>>> So, we probably have corrupted indexes that were updated since such\\n>>>>>> \\\"incomplete\\\" upgrade of ltree.\\n>>>>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\n>>>>> installed version of the extension, and that's intentional.\\n>>>> Yeah, exactly. But this opens up an additional consideration we\\n>>>> have to account for: whatever we do needs to work with either 1.1\\n>>>> or 1.2 SQL-level versions of the extension.\\n>>>>\\n>>>> regards, tom lane\\n>>>\\n>>> It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\\n>>> is not so much related to PG12 => PG13+ upgrades.\\n> \\n> So, it seems that ltree 1.1 in PG13+ is incompatible with ltree on\\n> PG12 and ltree 1.2 on PG13+. And there are many scenarios involving.\\n> \\n> It seems too difficult to identify all the broken cases in the release\\n> notes. What about applying a patch and asking all ltree users to\\n> reindex their indexes?\\n> \\n\\nYeah. I think this is getting so complicated that there's little chance\\nwe'd be able to clearly explain when to reindex.\\n\\n>> Well, it's quite a mess :-(\\n>>\\n>> It very clearly is not just 1.1 -> 1.2 upgrade issue, because you can\\n>> get a crash with 1.1 after PG12 -> PG13 upgrade, as demonstrated\\n>> earlier. So just \\\"fixing\\\" the extension upgrade is no enough.\\n>>\\n>> But as you showed, 1.1 -> 1.2 upgrade is broken too.\\n>>\\n>>>\\n>>> You can see here another problem: installation of opclass options\\n>>> procedure does not invalidate relation's opclass options cache.\\n>>>\\n>>\\n>> Seems like that.\\n> \\n> I think opclass options procedure shouldn't normally change opclass\\n> options chache. Before installation of options procedure, there\\n> should be no options. And options procedure shouldn't change the\\n> defaults in this case.\\n> \\n\\nI should clarify I haven't looked into the details - I merely verified\\naccessing the index in the same backend works, and after reconnection it\\ncrashes (per Nikita's example). I don't know why that happans.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sun, 6 Mar 2022 18:28:25 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-06 07:46:04 -0500, Andrew Dunstan wrote:\\n> This is an area not currently touched by the buildfarm's cross version\\n> upgrade testing, which basically compares a pre-upgrade and post-upgrade\\n> dump of the databases. The upgraded cluster does contain\\n> contrib_regression_ltree.\\n> \\n> I'm open to suggestions on how we might improve the buildfarm's testing\\n> of upgraded indexes generally.\\n\\nOne thing that's likely worth doing as part of the cross version upgrade test,\\neven if it wouldn't even help in this case, is to run amcheck post\\nupgrade. Just dumping data isn't going to touch indices at all.\\n\\nA sequence of\\n pg_upgrade; amcheck; upgrade all extensions; amcheck;\\nwould make sense.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sun, 6 Mar 2022 14:33:52 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"пн, 7 мар. 2022 г. в 00:34, Andres Freund <andres@anarazel.de>:\\n\\n> One thing that's likely worth doing as part of the cross version upgrade\\n> test,\\n> even if it wouldn't even help in this case, is to run amcheck post\\n> upgrade. Just dumping data isn't going to touch indices at all.\\n>\\n> A sequence of\\n> pg_upgrade; amcheck; upgrade all extensions; amcheck;\\n> would make sense.\\n>\\n\\nIs it possible to amcheck gist indexes?..\\n\\n\\n-- \\nVictor Yegorov\\n\\nпн, 7 мар. 2022 г. в 00:34, Andres Freund <andres@anarazel.de>:One thing that's likely worth doing as part of the cross version upgrade test,\\neven if it wouldn't even help in this case, is to run amcheck post\\nupgrade. Just dumping data isn't going to touch indices at all.\\n\\nA sequence of\\n pg_upgrade; amcheck; upgrade all extensions; amcheck;\\nwould make sense.Is it possible to amcheck gist indexes?.. -- Victor Yegorov\",\n \"msg_date\": \"Mon, 7 Mar 2022 01:12:07 +0200\",\n \"msg_from\": \"Victor Yegorov <vyegorov@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-07 01:12:07 +0200, Victor Yegorov wrote:\\n> пн, 7 мар. 2022 г. в 00:34, Andres Freund <andres@anarazel.de>:\\n> \\n> > One thing that's likely worth doing as part of the cross version upgrade\\n> > test,\\n> > even if it wouldn't even help in this case, is to run amcheck post\\n> > upgrade. Just dumping data isn't going to touch indices at all.\\n> >\\n> > A sequence of\\n> > pg_upgrade; amcheck; upgrade all extensions; amcheck;\\n> > would make sense.\\n> >\\n> \\n> Is it possible to amcheck gist indexes?..\\n\\nNo. That was what I was alluding to with \\\"even if it wouldn't even help in\\nthis case\\\".\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sun, 6 Mar 2022 17:32:02 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 6, 2022 at 8:28 PM Tomas Vondra\\n<tomas.vondra@enterprisedb.com> wrote:\\n> On 3/6/22 08:09, Alexander Korotkov wrote:\\n> > Sorry for this terrible oversight by me.\\n> >\\n> > On Sat, Mar 5, 2022 at 10:13 AM Tomas Vondra\\n> > <tomas.vondra@enterprisedb.com> wrote:\\n> >> On 3/4/22 23:09, Nikita Glukhov wrote:\\n> >>> On 04.03.2022 23:28, Tom Lane wrote:\\n> >>>\\n> >>>> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n> >>>>> On 3/4/22 20:29, Nikita Glukhov wrote:\\n> >>>>>> So, we probably have corrupted indexes that were updated since such\\n> >>>>>> \\\"incomplete\\\" upgrade of ltree.\\n> >>>>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\n> >>>>> installed version of the extension, and that's intentional.\\n> >>>> Yeah, exactly. But this opens up an additional consideration we\\n> >>>> have to account for: whatever we do needs to work with either 1.1\\n> >>>> or 1.2 SQL-level versions of the extension.\\n> >>>>\\n> >>>> regards, tom lane\\n> >>>\\n> >>> It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\\n> >>> is not so much related to PG12 => PG13+ upgrades.\\n> >\\n> > So, it seems that ltree 1.1 in PG13+ is incompatible with ltree on\\n> > PG12 and ltree 1.2 on PG13+. And there are many scenarios involving.\\n> >\\n> > It seems too difficult to identify all the broken cases in the release\\n> > notes. What about applying a patch and asking all ltree users to\\n> > reindex their indexes?\\n> >\\n>\\n> Yeah. I think this is getting so complicated that there's little chance\\n> we'd be able to clearly explain when to reindex.\\n\\nGood. The revised patch is attached. Instead of adding argument to\\nLTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\\nLTREE_GET_ASIGLEN() macros.\\n\\n------\\nRegards,\\nAlexander Korotkov\",\n \"msg_date\": \"Tue, 8 Mar 2022 02:05:04 +0300\",\n \"msg_from\": \"Alexander Korotkov <aekorotkov@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 8, 2022 at 2:05 AM Alexander Korotkov <aekorotkov@gmail.com> wrote:\\n> On Sun, Mar 6, 2022 at 8:28 PM Tomas Vondra\\n> <tomas.vondra@enterprisedb.com> wrote:\\n> > On 3/6/22 08:09, Alexander Korotkov wrote:\\n> > > Sorry for this terrible oversight by me.\\n> > >\\n> > > On Sat, Mar 5, 2022 at 10:13 AM Tomas Vondra\\n> > > <tomas.vondra@enterprisedb.com> wrote:\\n> > >> On 3/4/22 23:09, Nikita Glukhov wrote:\\n> > >>> On 04.03.2022 23:28, Tom Lane wrote:\\n> > >>>\\n> > >>>> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:\\n> > >>>>> On 3/4/22 20:29, Nikita Glukhov wrote:\\n> > >>>>>> So, we probably have corrupted indexes that were updated since such\\n> > >>>>>> \\\"incomplete\\\" upgrade of ltree.\\n> > >>>>> IIRC pg_upgrade is not expected to upgrade extensions - it keeps the\\n> > >>>>> installed version of the extension, and that's intentional.\\n> > >>>> Yeah, exactly. But this opens up an additional consideration we\\n> > >>>> have to account for: whatever we do needs to work with either 1.1\\n> > >>>> or 1.2 SQL-level versions of the extension.\\n> > >>>>\\n> > >>>> regards, tom lane\\n> > >>>\\n> > >>> It becomes clear that ltree upgrade 1.1 => 1.2 is broken, the problem\\n> > >>> is not so much related to PG12 => PG13+ upgrades.\\n> > >\\n> > > So, it seems that ltree 1.1 in PG13+ is incompatible with ltree on\\n> > > PG12 and ltree 1.2 on PG13+. And there are many scenarios involving.\\n> > >\\n> > > It seems too difficult to identify all the broken cases in the release\\n> > > notes. What about applying a patch and asking all ltree users to\\n> > > reindex their indexes?\\n> > >\\n> >\\n> > Yeah. I think this is getting so complicated that there's little chance\\n> > we'd be able to clearly explain when to reindex.\\n>\\n> Good. The revised patch is attached. Instead of adding argument to\\n> LTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\\n> LTREE_GET_ASIGLEN() macros.\\n\\nNo feedback yet. I'm going to push this if no objections.\\n\\n------\\nRegards,\\nAlexander Korotkov\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 10:09:24 +0300\",\n \"msg_from\": \"Alexander Korotkov <aekorotkov@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"On 3/10/22 08:09, Alexander Korotkov wrote:\\n> On Tue, Mar 8, 2022 at 2:05 AM Alexander Korotkov <aekorotkov@gmail.com> wrote:\\n>> ...\\n>>\\n>> Good. The revised patch is attached. Instead of adding argument to\\n>> LTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\\n>> LTREE_GET_ASIGLEN() macros.\\n> \\n> No feedback yet. I'm going to push this if no objections.\\n> \\n\\nWFM. I certainly don't have a better idea how to fix this.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 13:09:12 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Alexander Korotkov <aekorotkov@gmail.com> writes:\\n> Good. The revised patch is attached. Instead of adding argument to\\n> LTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\\n> LTREE_GET_ASIGLEN() macros.\\n\\nUm ... what I see after applying the patch is two redundant\\ndefinitions of LTREE_GET_ASIGLEN, and no LTREE_GET_SIGLEN.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 10:39:39 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"\\nOn 3/6/22 17:33, Andres Freund wrote:\\n> Hi,\\n>\\n> On 2022-03-06 07:46:04 -0500, Andrew Dunstan wrote:\\n>> This is an area not currently touched by the buildfarm's cross version\\n>> upgrade testing, which basically compares a pre-upgrade and post-upgrade\\n>> dump of the databases. The upgraded cluster does contain\\n>> contrib_regression_ltree.\\n>>\\n>> I'm open to suggestions on how we might improve the buildfarm's testing\\n>> of upgraded indexes generally.\\n> One thing that's likely worth doing as part of the cross version upgrade test,\\n> even if it wouldn't even help in this case, is to run amcheck post\\n> upgrade. Just dumping data isn't going to touch indices at all.\\n>\\n> A sequence of\\n> pg_upgrade; amcheck; upgrade all extensions; amcheck;\\n> would make sense.\\n>\\n\\n\\nSee\\n<https://github.com/PGBuildFarm/client-code/commit/191df23bd25eb5546b0989d71ae92747151f9f39>\\n\\n\\n\\nThis will be in the next release\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 10:43:06 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Andrew Dunstan <andrew@dunslane.net> writes:\\n> On 3/6/22 17:33, Andres Freund wrote:\\n>> A sequence of\\n>> pg_upgrade; amcheck; upgrade all extensions; amcheck;\\n>> would make sense.\\n\\n> See\\n> <https://github.com/PGBuildFarm/client-code/commit/191df23bd25eb5546b0989d71ae92747151f9f39>\\n\\nDoes this detect the problem at hand?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 10:53:49 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"\\nOn 3/10/22 10:53, Tom Lane wrote:\\n> Andrew Dunstan <andrew@dunslane.net> writes:\\n>> On 3/6/22 17:33, Andres Freund wrote:\\n>>> A sequence of\\n>>> pg_upgrade; amcheck; upgrade all extensions; amcheck;\\n>>> would make sense.\\n>> See\\n>> <https://github.com/PGBuildFarm/client-code/commit/191df23bd25eb5546b0989d71ae92747151f9f39>\\n> Does this detect the problem at hand?\\n>\\n> \\t\\t\\t\\n\\n\\nAIUI, amcheck doesn't check gist indexes, hence Andres' \\\"even if it\\nwouldn't even help in this case\\\". So no.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 12:05:31 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"On 2022-03-10 10:43:06 -0500, Andrew Dunstan wrote:\\n> On 3/6/22 17:33, Andres Freund wrote:\\n> > One thing that's likely worth doing as part of the cross version upgrade test,\\n> > even if it wouldn't even help in this case, is to run amcheck post\\n> > upgrade. Just dumping data isn't going to touch indices at all.\\n> >\\n> > A sequence of\\n> > pg_upgrade; amcheck; upgrade all extensions; amcheck;\\n> > would make sense.\\n\\n> See\\n> <https://github.com/PGBuildFarm/client-code/commit/191df23bd25eb5546b0989d71ae92747151f9f39>\\n > This will be in the next release\\n\\nGreat, thanks for doing that!\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 09:24:19 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 10, 2022 at 6:39 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Alexander Korotkov <aekorotkov@gmail.com> writes:\\n> > Good. The revised patch is attached. Instead of adding argument to\\n> > LTREE_GET_ASIGLEN(), it introduces separate LTREE_GET_SIGLEN() and\\n> > LTREE_GET_ASIGLEN() macros.\\n>\\n> Um ... what I see after applying the patch is two redundant\\n> definitions of LTREE_GET_ASIGLEN, and no LTREE_GET_SIGLEN.\\n\\nSorry for publishing a patch, which doesn't even compile.\\nThe attached version should be good.\\n\\n------\\nRegards,\\nAlexander Korotkov\",\n \"msg_date\": \"Thu, 10 Mar 2022 20:44:13 +0300\",\n \"msg_from\": \"Alexander Korotkov <aekorotkov@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n },\n {\n \"msg_contents\": \"Alexander Korotkov <aekorotkov@gmail.com> writes:\\n> Sorry for publishing a patch, which doesn't even compile.\\n> The attached version should be good.\\n\\nOK, better.\\n\\nSo to clarify: we are going to back-patch as far as v13, and\\nthen in the minor releases we are going to tell people they\\nneed to REINDEX all gist indexes on ltree? If so, please\\nmake that clear in the commit message, so I don't forget it\\ncome May ;-)\\n\\nAlso, it looks like reindex is *not* required for indexes on\\nltree[] ? That is, gist_ltree_ops indexes are affected but\\nnot gist__ltree_ops?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 17:36:50 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: ltree_gist indexes broken after pg_upgrade from 12 to 13\"\n }\n]","isConversational":false}}},{"rowIdx":53577,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi everyone,\n\nWould adding additional examples to testlibpq3.c on handling more data\ntypes be helpful? I've attached a patch adding how to handle a REAL to\ncurrent example.\n\nRegards,\nMark","msg_date":"Thu, 24 Feb 2022 13:42:40 -0800","msg_from":"Mark Wong <markwkm@gmail.com>","msg_from_op":true,"msg_subject":"real/float example for testlibpq3"},{"msg_contents":"Hi Mark,\nFetching a \"real\" type field from libpq doesn't look that intuitive.\nSo this example is super helpful. Thanks.\n\nI am wondering whether we should provide PQgetfloatvalue() which does\nwhat that example shows but transparently. It will return a float\nvalue of the given field instead of char *. That will standardize the\nway to fetch real typed values in libpq. That leads to the next\nquestion. Do we need to introduce different PQget*value() for standard\nC/SQL data types.\n\n-- \nBest Wishes,\nAshutosh Bapat\n\nOn Fri, Feb 25, 2022 at 3:12 AM Mark Wong <markwkm@gmail.com> wrote:\n>\n> Hi everyone,\n>\n> Would adding additional examples to testlibpq3.c on handling more data\n> types be helpful? I've attached a patch adding how to handle a REAL to\n> current example.\n>\n> Regards,\n> Mark\n\n\n","msg_date":"Mon, 28 Feb 2022 18:33:30 +0530","msg_from":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com> writes:\n> I am wondering whether we should provide PQgetfloatvalue() which does\n> what that example shows but transparently. It will return a float\n> value of the given field instead of char *.\n\nI'm against that, precisely because ...\n\n> That will standardize the\n> way to fetch real typed values in libpq. That leads to the next\n> question. Do we need to introduce different PQget*value() for standard\n> C/SQL data types.\n\n... I do not want to go here. Where would you stop? How would you\ndeal with cross-machine inconsistencies in integer widths?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Feb 2022 10:19:00 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On 02/28/22 10:19, Tom Lane wrote:\n>> That will standardize the\n>> way to fetch real typed values in libpq. That leads to the next\n>> question. Do we need to introduce different PQget*value() for standard\n>> C/SQL data types.\n> \n> ... I do not want to go here. Where would you stop? How would you\n> deal with cross-machine inconsistencies in integer widths?\n\nThis stimulates a question for me.\n\nNot just libpq, but all drivers implemented in whatever language, if they\nwish to support binary protocol, depend on knowing what the committed\nsend/recv wire formats are for whichever types they mean to support.\n\nIn the current state of affairs, what's considered the ur-source of that\ninformation?\n\nI have often seen those formats documented in code comments, usually above\nthe recv function in the .c file for a given adt.\n\nHave we got any more, well, machine-readable collection of that knowledge?\n\nRegards,\n-Chap\n\n\n","msg_date":"Mon, 28 Feb 2022 17:30:11 -0500","msg_from":"Chapman Flack <chap@anastigmatix.net>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"Chapman Flack <chap@anastigmatix.net> writes:\n> This stimulates a question for me.\n\n> Not just libpq, but all drivers implemented in whatever language, if they\n> wish to support binary protocol, depend on knowing what the committed\n> send/recv wire formats are for whichever types they mean to support.\n\n> In the current state of affairs, what's considered the ur-source of that\n> information?\n\nThe source code for the type's send/receive functions :-(. One could\nwish for something better, but no one has stepped up to produce such\ndocumentation.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Feb 2022 17:48:50 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Mon, 28 Feb 2022 at 17:50, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> Chapman Flack <chap@anastigmatix.net> writes:\n> > In the current state of affairs, what's considered the ur-source of that\n> > information?\n>\n> The source code for the type's send/receive functions :-(. One could\n> wish for something better, but no one has stepped up to produce such\n> documentation.\n\nFwiw the client library I heard of attempting to have good binary mode\nsupport was the Crystal language client\nhttps://github.com/will/crystal-pg. I think he was aiming for full\ncoverage of the built-in data types. That might make a good reference\nimplementation to write up documentation from. He probably uncovered\nsome corner cases in development that one might not find from just\ninspection of the server code.\n\n-- \ngreg\n\n\n","msg_date":"Wed, 30 Mar 2022 13:16:37 -0400","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Wed, Mar 30, 2022 at 01:16:37PM -0400, Greg Stark wrote:\n> On Mon, 28 Feb 2022 at 17:50, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> >\n> > Chapman Flack <chap@anastigmatix.net> writes:\n> > > In the current state of affairs, what's considered the ur-source of that\n> > > information?\n> >\n> > The source code for the type's send/receive functions :-(. One could\n> > wish for something better, but no one has stepped up to produce such\n> > documentation.\n> \n> Fwiw the client library I heard of attempting to have good binary mode\n> support was the Crystal language client\n> https://github.com/will/crystal-pg. I think he was aiming for full\n> coverage of the built-in data types. That might make a good reference\n> implementation to write up documentation from. He probably uncovered\n> some corner cases in development that one might not find from just\n> inspection of the server code.\n\nChecking in for quick feedback to see if this refactor makes sense.\n\nI've created a function for each data type with the idea that an example\nfor handling a specific data type can be more easily reviewed by looking\nin a single place.\n\nI've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\nto try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\nthis is a good way to go.\n\nRegards,\nMark","msg_date":"Tue, 14 Jun 2022 10:40:06 -0700","msg_from":"Mark Wong <markwkm@gmail.com>","msg_from_op":true,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Tue, Jun 14, 2022 at 1:40 PM Mark Wong <markwkm@gmail.com> wrote:\n> Checking in for quick feedback to see if this refactor makes sense.\n>\n> I've created a function for each data type with the idea that an example\n> for handling a specific data type can be more easily reviewed by looking\n> in a single place.\n>\n> I've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\n> to try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\n> this is a good way to go.\n\nI think this could be a reasonable kind of thing to do, but I think\nyou left the \"ts\" output out of the example output in the comments,\nand also your code's apparently not portable, because my compiler is\nOK with testlibpq3 right now, but with your patch it emits lengthy\nunhappy moaning, starting with:\n\ntestlibpq3.c:88:10: error: expected ')'\nuint64_t ntohll(uint64_t);\n ^\n/Library/Developer/CommandLineTools/SDKs/MacOSX10.15.sdk/usr/include/sys/_endian.h:140:25:\nnote: expanded from macro 'ntohll'\n#define ntohll(x) __DARWIN_OSSwapInt64(x)\n ^\n\nApparently macOS defines ntohll as a macro, and it's not amused by\nyour attempt to include a function prototype for it.\n\nI'm not sure that we want to let these test programs grow very large.\nIn particular, I don't think we should let ourselves get sucked into\nadding an example for every data type under the sun -- if that's\nwanted, the solution is perhaps to add documentation for the binary\nformats, not hide impromptu documentation inside a test program. But\ndoing this much seems OK to me.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 16 Jun 2022 15:06:01 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"Robert Haas <robertmhaas@gmail.com> writes:\n> On Tue, Jun 14, 2022 at 1:40 PM Mark Wong <markwkm@gmail.com> wrote:\n>> I've created a function for each data type with the idea that an example\n>> for handling a specific data type can be more easily reviewed by looking\n>> in a single place.\n>> I've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\n>> to try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\n>> this is a good way to go.\n\n> I'm not sure that we want to let these test programs grow very large.\n> In particular, I don't think we should let ourselves get sucked into\n> adding an example for every data type under the sun -- if that's\n> wanted, the solution is perhaps to add documentation for the binary\n> formats, not hide impromptu documentation inside a test program. But\n> doing this much seems OK to me.\n\nYeah, \"hiding impromptu documentation inside a test program\" is what\nthis looks like, and I'm not sure that's a reasonable way to go.\n\n(1) Who's going to think to look in src/test/examples/testlibpq3.c for\ndocumentation of binary formats?\n\n(2) The useful details are likely to get buried in notational and\nportability concerns, as I think your build failure illustrates.\n\n(3) I bet few if any packagers install these files, so that the new\ninfo would be unavailable to many people.\n\nI think some new appendix in the main SGML docs would be the appropriate\nplace if we want to provide real documentation.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 16 Jun 2022 15:41:50 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Thu, Jun 16, 2022 at 03:41:50PM -0400, Tom Lane wrote:\n> Robert Haas <robertmhaas@gmail.com> writes:\n> > On Tue, Jun 14, 2022 at 1:40 PM Mark Wong <markwkm@gmail.com> wrote:\n> >> I've created a function for each data type with the idea that an example\n> >> for handling a specific data type can be more easily reviewed by looking\n> >> in a single place.\n> >> I've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\n> >> to try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\n> >> this is a good way to go.\n> \n> > I'm not sure that we want to let these test programs grow very large.\n> > In particular, I don't think we should let ourselves get sucked into\n> > adding an example for every data type under the sun -- if that's\n> > wanted, the solution is perhaps to add documentation for the binary\n> > formats, not hide impromptu documentation inside a test program. But\n> > doing this much seems OK to me.\n> \n> Yeah, \"hiding impromptu documentation inside a test program\" is what\n> this looks like, and I'm not sure that's a reasonable way to go.\n> \n> (1) Who's going to think to look in src/test/examples/testlibpq3.c for\n> documentation of binary formats?\n> \n> (2) The useful details are likely to get buried in notational and\n> portability concerns, as I think your build failure illustrates.\n> \n> (3) I bet few if any packagers install these files, so that the new\n> info would be unavailable to many people.\n> \n> I think some new appendix in the main SGML docs would be the appropriate\n> place if we want to provide real documentation.\n\nOk, I'll leave the testlibpq3.c as it was then. If it's worth keeping\nany of those changes, then I can remove the timestamp example because of\nthe ntohll() portability since that is trivial.\n\nI'll start a new appendix and share again when I have something to show.\n\nRegards,\nMark\n\n--\nMark Wong\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 16 Jun 2022 15:56:50 -0700","msg_from":"Mark Wong <markwkm@gmail.com>","msg_from_op":true,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Thu, Jun 16, 2022 at 03:41:50PM -0400, Tom Lane wrote:\n> Robert Haas <robertmhaas@gmail.com> writes:\n> > On Tue, Jun 14, 2022 at 1:40 PM Mark Wong <markwkm@gmail.com> wrote:\n> >> I've created a function for each data type with the idea that an example\n> >> for handling a specific data type can be more easily reviewed by looking\n> >> in a single place.\n> >> I've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\n> >> to try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\n> >> this is a good way to go.\n> \n> > I'm not sure that we want to let these test programs grow very large.\n> > In particular, I don't think we should let ourselves get sucked into\n> > adding an example for every data type under the sun -- if that's\n> > wanted, the solution is perhaps to add documentation for the binary\n> > formats, not hide impromptu documentation inside a test program. But\n> > doing this much seems OK to me.\n> \n> Yeah, \"hiding impromptu documentation inside a test program\" is what\n> this looks like, and I'm not sure that's a reasonable way to go.\n> \n> (1) Who's going to think to look in src/test/examples/testlibpq3.c for\n> documentation of binary formats?\n> \n> (2) The useful details are likely to get buried in notational and\n> portability concerns, as I think your build failure illustrates.\n> \n> (3) I bet few if any packagers install these files, so that the new\n> info would be unavailable to many people.\n> \n> I think some new appendix in the main SGML docs would be the appropriate\n> place if we want to provide real documentation.\n\nJust wanted to do another quick check in before I go too far. How do\ndoes this start look? Extending the libpq section with a code snippet\nand description per data type?\n\nRegards,\nMark","msg_date":"Fri, 1 Jul 2022 10:18:20 -0700","msg_from":"Mark Wong <markwkm@gmail.com>","msg_from_op":true,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On 01.07.22 19:18, Mark Wong wrote:\n> Just wanted to do another quick check in before I go too far. How do\n> does this start look? Extending the libpq section with a code snippet\n> and description per data type?\n\nIf we want to start documenting the binary format for all data types, it \nshould go into the \"Data Types\" chapter. There, we already document the \ntext format in great detail, so putting the binary format in there as \nwell would make sense.\n\nThe libpq chapter could then contain some examples of applying that \ninformation in the context of libpq. But the libpq documentation \nshouldn't be the primary place where the formats are documented.\n\n\n\n","msg_date":"Tue, 1 Nov 2022 11:39:40 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> If we want to start documenting the binary format for all data types, it \n> should go into the \"Data Types\" chapter. There, we already document the \n> text format in great detail, so putting the binary format in there as \n> well would make sense.\n\nAgreed that the libpq manual is not the place for this, but I feel\nlike it will also be clutter in \"Data Types\". Perhaps we should\ninvent a new appendix or the like? Somewhere near the wire protocol\ndocs seems sensible.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 01 Nov 2022 09:15:46 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On 01.11.22 09:15, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n>> If we want to start documenting the binary format for all data types, it\n>> should go into the \"Data Types\" chapter. There, we already document the\n>> text format in great detail, so putting the binary format in there as\n>> well would make sense.\n> \n> Agreed that the libpq manual is not the place for this, but I feel\n> like it will also be clutter in \"Data Types\". Perhaps we should\n> invent a new appendix or the like? Somewhere near the wire protocol\n> docs seems sensible.\n\nWould that clutter the protocol docs? ;-)\n\nI suppose figuring out exactly where to put it and how to mark it up, \netc., in a repeatable fashion is part of the job here.\n\n\n\n","msg_date":"Thu, 3 Nov 2022 09:23:00 -0400","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> On 01.11.22 09:15, Tom Lane wrote:\n>> Agreed that the libpq manual is not the place for this, but I feel\n>> like it will also be clutter in \"Data Types\". Perhaps we should\n>> invent a new appendix or the like? Somewhere near the wire protocol\n>> docs seems sensible.\n\n> Would that clutter the protocol docs? ;-)\n\nI said \"near\", not \"in\". At the time I was thinking \"new appendix\",\nbut I now recall that the wire protocol docs are not an appendix\nbut a chapter in the Internals division. So that doesn't seem like\nquite the right place anyway.\n\nPerhaps a new chapter under \"IV. Client Interfaces\" is the right\nplace?\n\nIf we wanted to get aggressive, we could move most of the nitpicky details\nabout datatype text formatting (e.g., the array quoting rules) there too.\nI'm not set on that, but it'd make datatype.sgml smaller which could\nhardly be a bad thing.\n\n> I suppose figuring out exactly where to put it and how to mark it up, \n> etc., in a repeatable fashion is part of the job here.\n\nYup.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 03 Nov 2022 09:55:22 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Thu, Nov 03, 2022 at 09:55:22AM -0400, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> > On 01.11.22 09:15, Tom Lane wrote:\n> >> Agreed that the libpq manual is not the place for this, but I feel\n> >> like it will also be clutter in \"Data Types\". Perhaps we should\n> >> invent a new appendix or the like? Somewhere near the wire protocol\n> >> docs seems sensible.\n> \n> > Would that clutter the protocol docs? ;-)\n> \n> I said \"near\", not \"in\". At the time I was thinking \"new appendix\",\n> but I now recall that the wire protocol docs are not an appendix\n> but a chapter in the Internals division. So that doesn't seem like\n> quite the right place anyway.\n> \n> Perhaps a new chapter under \"IV. Client Interfaces\" is the right\n> place?\n> \n> If we wanted to get aggressive, we could move most of the nitpicky details\n> about datatype text formatting (e.g., the array quoting rules) there too.\n> I'm not set on that, but it'd make datatype.sgml smaller which could\n> hardly be a bad thing.\n> \n> > I suppose figuring out exactly where to put it and how to mark it up, \n> > etc., in a repeatable fashion is part of the job here.\n> \n> Yup.\n\nI'll take a stab at adding a new chapter and share how that looks.\n\nRegards,\nMark\n\n--\nMark Wong\nEDB https://enterprisedb.com\n\n\n","msg_date":"Wed, 9 Nov 2022 12:47:23 -0800","msg_from":"Mark Wong <markwkm@gmail.com>","msg_from_op":true,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Wed, Nov 09, 2022 at 12:47:23PM -0800, Mark Wong wrote:\n> I'll take a stab at adding a new chapter and share how that looks.\n\nThis thread has stalled for three weeks, so I have marked it as\nreturned with feedback. Once you have a new patch, please feel free\nto submit it.\n--\nMichael","msg_date":"Wed, 30 Nov 2022 15:59:12 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Thu, Nov 03, 2022 at 09:55:22AM -0400, Tom Lane wrote:\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\n> > On 01.11.22 09:15, Tom Lane wrote:\n> >> Agreed that the libpq manual is not the place for this, but I feel\n> >> like it will also be clutter in \"Data Types\". Perhaps we should\n> >> invent a new appendix or the like? Somewhere near the wire protocol\n> >> docs seems sensible.\n> \n> > Would that clutter the protocol docs? ;-)\n> \n> I said \"near\", not \"in\". At the time I was thinking \"new appendix\",\n> but I now recall that the wire protocol docs are not an appendix\n> but a chapter in the Internals division. So that doesn't seem like\n> quite the right place anyway.\n> \n> Perhaps a new chapter under \"IV. Client Interfaces\" is the right\n> place?\n> \n> If we wanted to get aggressive, we could move most of the nitpicky details\n> about datatype text formatting (e.g., the array quoting rules) there too.\n> I'm not set on that, but it'd make datatype.sgml smaller which could\n> hardly be a bad thing.\n> \n> > I suppose figuring out exactly where to put it and how to mark it up, \n> > etc., in a repeatable fashion is part of the job here.\n> \n> Yup.\n\nHow does this look?\n\nI've simply moved things around into a new \"Binary Format\" section with\nthe few parts that I've started for some quick feedback about whether\nthis is looking like the right landing place.\n\nRegards,\nMark","msg_date":"Fri, 13 Jan 2023 12:45:36 -0800","msg_from":"Mark Wong <markwkm@gmail.com>","msg_from_op":true,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"Mark Wong <markwkm@gmail.com> writes:\n> On Thu, Nov 03, 2022 at 09:55:22AM -0400, Tom Lane wrote:\n>> Perhaps a new chapter under \"IV. Client Interfaces\" is the right\n>> place?\n>> \n>> If we wanted to get aggressive, we could move most of the nitpicky details\n>> about datatype text formatting (e.g., the array quoting rules) there too.\n>> I'm not set on that, but it'd make datatype.sgml smaller which could\n>> hardly be a bad thing.\n\n> I suppose figuring out exactly where to put it and how to mark it up, \n> etc., in a repeatable fashion is part of the job here.\n\n> How does this look?\n\n> I've simply moved things around into a new \"Binary Format\" section with\n> the few parts that I've started for some quick feedback about whether\n> this is looking like the right landing place.\n\nI took a quick look at this. The patch cannot be applied as given\nbecause it seems to be editing binary-format.sgml from some prior\nstate, but that file doesn't exist at all yet. However, there's\nenough visible in the patch to comment on.\n\nPersonally I'm not excited about supplying fragments of C code\nlike this. Those seem quite useless to non-C users, and I'm not\nreally sure that they'd be very helpful even for C users, because\nyou have to make a whole lot of assumptions about what the user\nwants to do with the value. I think what *would* be helpful is\njust plain prose documentation of the on-the-wire binary format.\n\nI don't mind if you write something like\n\n A float4 value is a 4-byte IEEE single-precision floating point\n number. It is transmitted in network byte order, so you must\n convert to local byte order. (C users can do this portably\n using the standard ntohl() function.)\n\nbut I'm not sure an example is worth more than such a parenthetical\ncomment. Perhaps others disagree, though.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Fri, 20 Jan 2023 12:57:48 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Fri, Jan 20, 2023 at 12:58 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> I don't mind if you write something like\n>\n> A float4 value is a 4-byte IEEE single-precision floating point\n> number. It is transmitted in network byte order, so you must\n> convert to local byte order. (C users can do this portably\n> using the standard ntohl() function.)\n>\n> but I'm not sure an example is worth more than such a parenthetical\n> comment. Perhaps others disagree, though.\n\nI don't disagree with that.\n\nI do think that when you suggested documenting this rather than just\nadding some examples, you moved the goalposts a long way. If we're\ngoing to add this to the documentation, it probably ought to cover\nevery data type we ship. Overall, I think that would be a better\nresult than just adding a few examples for the most common data types\nto testlibpq*.c, but it's also substantially more work. I do agree\nwith you that if we're going to document this, as opposed to provide\nexamples, then a narrative style is more appropriate than a whole\nbunch of small sample programs; maintaining working code in the\ndocumentation seems like an annoying amount of maintenance and is\nprobably not the most efficient way to communicate useful information.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Fri, 20 Jan 2023 13:12:07 -0500","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Fri, Jan 20, 2023 at 01:12:07PM -0500, Robert Haas wrote:\n> On Fri, Jan 20, 2023 at 12:58 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> > I don't mind if you write something like\n> >\n> > A float4 value is a 4-byte IEEE single-precision floating point\n> > number. It is transmitted in network byte order, so you must\n> > convert to local byte order. (C users can do this portably\n> > using the standard ntohl() function.)\n> >\n> > but I'm not sure an example is worth more than such a parenthetical\n> > comment. Perhaps others disagree, though.\n> \n> I don't disagree with that.\n> \n> I do think that when you suggested documenting this rather than just\n> adding some examples, you moved the goalposts a long way. If we're\n> going to add this to the documentation, it probably ought to cover\n> every data type we ship. Overall, I think that would be a better\n> result than just adding a few examples for the most common data types\n> to testlibpq*.c, but it's also substantially more work. I do agree\n> with you that if we're going to document this, as opposed to provide\n> examples, then a narrative style is more appropriate than a whole\n> bunch of small sample programs; maintaining working code in the\n> documentation seems like an annoying amount of maintenance and is\n> probably not the most efficient way to communicate useful information.\n\nYeah, I will try to cover all the data types we ship. :) I'll keep at\nit and drop the code examples.\n\nRegards,\nMark\n-- \nMark Wong\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 23 Jan 2023 08:54:26 -0800","msg_from":"Mark Wong <markwkm@gmail.com>","msg_from_op":true,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Mon, 23 Jan 2023 at 11:54, Mark Wong <markwkm@gmail.com> wrote:\nfficient way to communicate useful information.\n>\n> Yeah, I will try to cover all the data types we ship. :) I'll keep at\n> it and drop the code examples.\n\nI assume this isn't going to happen for this commitfest? If you think\nit is then shout otherwise I'll mark it Returned with Feedback and\nmove it on to the next release.\n\n-- \nGregory Stark\nAs Commitfest Manager\n\n\n","msg_date":"Mon, 20 Mar 2023 13:37:57 -0400","msg_from":"\"Gregory Stark (as CFM)\" <stark.cfm@gmail.com>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"On Mon, Mar 20, 2023 at 01:37:57PM -0400, Gregory Stark (as CFM) wrote:\n> On Mon, 23 Jan 2023 at 11:54, Mark Wong <markwkm@gmail.com> wrote:\n> fficient way to communicate useful information.\n> >\n> > Yeah, I will try to cover all the data types we ship. :) I'll keep at\n> > it and drop the code examples.\n> \n> I assume this isn't going to happen for this commitfest? If you think\n> it is then shout otherwise I'll mark it Returned with Feedback and\n> move it on to the next release.\n\nSounds good. I actually thought I did that already, thanks for catching\nthat.\n\nRegards,\nMark\n--\nMark Wong\nEDB https://enterprisedb.com\n\n\n","msg_date":"Tue, 21 Mar 2023 06:44:08 -0700","msg_from":"Mark Wong <markwkm@gmail.com>","msg_from_op":true,"msg_subject":"Re: real/float example for testlibpq3"},{"msg_contents":"> On 21 Mar 2023, at 14:44, Mark Wong <markwkm@gmail.com> wrote:\n> \n> On Mon, Mar 20, 2023 at 01:37:57PM -0400, Gregory Stark (as CFM) wrote:\n>> On Mon, 23 Jan 2023 at 11:54, Mark Wong <markwkm@gmail.com> wrote:\n>> fficient way to communicate useful information.\n>>> \n>>> Yeah, I will try to cover all the data types we ship. :) I'll keep at\n>>> it and drop the code examples.\n>> \n>> I assume this isn't going to happen for this commitfest? If you think\n>> it is then shout otherwise I'll mark it Returned with Feedback and\n>> move it on to the next release.\n> \n> Sounds good. I actually thought I did that already, thanks for catching\n> that.\n\nThis has been waiting on author since January, unless there is a new patch\nready I will close this as Returned with Feedback and you can resubmit for\nanother CF when there is a new patch.\n\n--\nDaniel Gustafsson\n\n\n\n","msg_date":"Tue, 4 Jul 2023 09:29:01 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: real/float example for testlibpq3"}],"string":"[\n {\n \"msg_contents\": \"Hi everyone,\\n\\nWould adding additional examples to testlibpq3.c on handling more data\\ntypes be helpful? I've attached a patch adding how to handle a REAL to\\ncurrent example.\\n\\nRegards,\\nMark\",\n \"msg_date\": \"Thu, 24 Feb 2022 13:42:40 -0800\",\n \"msg_from\": \"Mark Wong <markwkm@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"Hi Mark,\\nFetching a \\\"real\\\" type field from libpq doesn't look that intuitive.\\nSo this example is super helpful. Thanks.\\n\\nI am wondering whether we should provide PQgetfloatvalue() which does\\nwhat that example shows but transparently. It will return a float\\nvalue of the given field instead of char *. That will standardize the\\nway to fetch real typed values in libpq. That leads to the next\\nquestion. Do we need to introduce different PQget*value() for standard\\nC/SQL data types.\\n\\n-- \\nBest Wishes,\\nAshutosh Bapat\\n\\nOn Fri, Feb 25, 2022 at 3:12 AM Mark Wong <markwkm@gmail.com> wrote:\\n>\\n> Hi everyone,\\n>\\n> Would adding additional examples to testlibpq3.c on handling more data\\n> types be helpful? I've attached a patch adding how to handle a REAL to\\n> current example.\\n>\\n> Regards,\\n> Mark\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 18:33:30 +0530\",\n \"msg_from\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"Ashutosh Bapat <ashutosh.bapat.oss@gmail.com> writes:\\n> I am wondering whether we should provide PQgetfloatvalue() which does\\n> what that example shows but transparently. It will return a float\\n> value of the given field instead of char *.\\n\\nI'm against that, precisely because ...\\n\\n> That will standardize the\\n> way to fetch real typed values in libpq. That leads to the next\\n> question. Do we need to introduce different PQget*value() for standard\\n> C/SQL data types.\\n\\n... I do not want to go here. Where would you stop? How would you\\ndeal with cross-machine inconsistencies in integer widths?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 10:19:00 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On 02/28/22 10:19, Tom Lane wrote:\\n>> That will standardize the\\n>> way to fetch real typed values in libpq. That leads to the next\\n>> question. Do we need to introduce different PQget*value() for standard\\n>> C/SQL data types.\\n> \\n> ... I do not want to go here. Where would you stop? How would you\\n> deal with cross-machine inconsistencies in integer widths?\\n\\nThis stimulates a question for me.\\n\\nNot just libpq, but all drivers implemented in whatever language, if they\\nwish to support binary protocol, depend on knowing what the committed\\nsend/recv wire formats are for whichever types they mean to support.\\n\\nIn the current state of affairs, what's considered the ur-source of that\\ninformation?\\n\\nI have often seen those formats documented in code comments, usually above\\nthe recv function in the .c file for a given adt.\\n\\nHave we got any more, well, machine-readable collection of that knowledge?\\n\\nRegards,\\n-Chap\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 17:30:11 -0500\",\n \"msg_from\": \"Chapman Flack <chap@anastigmatix.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"Chapman Flack <chap@anastigmatix.net> writes:\\n> This stimulates a question for me.\\n\\n> Not just libpq, but all drivers implemented in whatever language, if they\\n> wish to support binary protocol, depend on knowing what the committed\\n> send/recv wire formats are for whichever types they mean to support.\\n\\n> In the current state of affairs, what's considered the ur-source of that\\n> information?\\n\\nThe source code for the type's send/receive functions :-(. One could\\nwish for something better, but no one has stepped up to produce such\\ndocumentation.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 17:48:50 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Mon, 28 Feb 2022 at 17:50, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> Chapman Flack <chap@anastigmatix.net> writes:\\n> > In the current state of affairs, what's considered the ur-source of that\\n> > information?\\n>\\n> The source code for the type's send/receive functions :-(. One could\\n> wish for something better, but no one has stepped up to produce such\\n> documentation.\\n\\nFwiw the client library I heard of attempting to have good binary mode\\nsupport was the Crystal language client\\nhttps://github.com/will/crystal-pg. I think he was aiming for full\\ncoverage of the built-in data types. That might make a good reference\\nimplementation to write up documentation from. He probably uncovered\\nsome corner cases in development that one might not find from just\\ninspection of the server code.\\n\\n-- \\ngreg\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 13:16:37 -0400\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2022 at 01:16:37PM -0400, Greg Stark wrote:\\n> On Mon, 28 Feb 2022 at 17:50, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> >\\n> > Chapman Flack <chap@anastigmatix.net> writes:\\n> > > In the current state of affairs, what's considered the ur-source of that\\n> > > information?\\n> >\\n> > The source code for the type's send/receive functions :-(. One could\\n> > wish for something better, but no one has stepped up to produce such\\n> > documentation.\\n> \\n> Fwiw the client library I heard of attempting to have good binary mode\\n> support was the Crystal language client\\n> https://github.com/will/crystal-pg. I think he was aiming for full\\n> coverage of the built-in data types. That might make a good reference\\n> implementation to write up documentation from. He probably uncovered\\n> some corner cases in development that one might not find from just\\n> inspection of the server code.\\n\\nChecking in for quick feedback to see if this refactor makes sense.\\n\\nI've created a function for each data type with the idea that an example\\nfor handling a specific data type can be more easily reviewed by looking\\nin a single place.\\n\\nI've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\\nto try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\\nthis is a good way to go.\\n\\nRegards,\\nMark\",\n \"msg_date\": \"Tue, 14 Jun 2022 10:40:06 -0700\",\n \"msg_from\": \"Mark Wong <markwkm@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Tue, Jun 14, 2022 at 1:40 PM Mark Wong <markwkm@gmail.com> wrote:\\n> Checking in for quick feedback to see if this refactor makes sense.\\n>\\n> I've created a function for each data type with the idea that an example\\n> for handling a specific data type can be more easily reviewed by looking\\n> in a single place.\\n>\\n> I've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\\n> to try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\\n> this is a good way to go.\\n\\nI think this could be a reasonable kind of thing to do, but I think\\nyou left the \\\"ts\\\" output out of the example output in the comments,\\nand also your code's apparently not portable, because my compiler is\\nOK with testlibpq3 right now, but with your patch it emits lengthy\\nunhappy moaning, starting with:\\n\\ntestlibpq3.c:88:10: error: expected ')'\\nuint64_t ntohll(uint64_t);\\n ^\\n/Library/Developer/CommandLineTools/SDKs/MacOSX10.15.sdk/usr/include/sys/_endian.h:140:25:\\nnote: expanded from macro 'ntohll'\\n#define ntohll(x) __DARWIN_OSSwapInt64(x)\\n ^\\n\\nApparently macOS defines ntohll as a macro, and it's not amused by\\nyour attempt to include a function prototype for it.\\n\\nI'm not sure that we want to let these test programs grow very large.\\nIn particular, I don't think we should let ourselves get sucked into\\nadding an example for every data type under the sun -- if that's\\nwanted, the solution is perhaps to add documentation for the binary\\nformats, not hide impromptu documentation inside a test program. But\\ndoing this much seems OK to me.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Jun 2022 15:06:01 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"Robert Haas <robertmhaas@gmail.com> writes:\\n> On Tue, Jun 14, 2022 at 1:40 PM Mark Wong <markwkm@gmail.com> wrote:\\n>> I've created a function for each data type with the idea that an example\\n>> for handling a specific data type can be more easily reviewed by looking\\n>> in a single place.\\n>> I've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\\n>> to try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\\n>> this is a good way to go.\\n\\n> I'm not sure that we want to let these test programs grow very large.\\n> In particular, I don't think we should let ourselves get sucked into\\n> adding an example for every data type under the sun -- if that's\\n> wanted, the solution is perhaps to add documentation for the binary\\n> formats, not hide impromptu documentation inside a test program. But\\n> doing this much seems OK to me.\\n\\nYeah, \\\"hiding impromptu documentation inside a test program\\\" is what\\nthis looks like, and I'm not sure that's a reasonable way to go.\\n\\n(1) Who's going to think to look in src/test/examples/testlibpq3.c for\\ndocumentation of binary formats?\\n\\n(2) The useful details are likely to get buried in notational and\\nportability concerns, as I think your build failure illustrates.\\n\\n(3) I bet few if any packagers install these files, so that the new\\ninfo would be unavailable to many people.\\n\\nI think some new appendix in the main SGML docs would be the appropriate\\nplace if we want to provide real documentation.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Jun 2022 15:41:50 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 16, 2022 at 03:41:50PM -0400, Tom Lane wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n> > On Tue, Jun 14, 2022 at 1:40 PM Mark Wong <markwkm@gmail.com> wrote:\\n> >> I've created a function for each data type with the idea that an example\\n> >> for handling a specific data type can be more easily reviewed by looking\\n> >> in a single place.\\n> >> I've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\\n> >> to try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\\n> >> this is a good way to go.\\n> \\n> > I'm not sure that we want to let these test programs grow very large.\\n> > In particular, I don't think we should let ourselves get sucked into\\n> > adding an example for every data type under the sun -- if that's\\n> > wanted, the solution is perhaps to add documentation for the binary\\n> > formats, not hide impromptu documentation inside a test program. But\\n> > doing this much seems OK to me.\\n> \\n> Yeah, \\\"hiding impromptu documentation inside a test program\\\" is what\\n> this looks like, and I'm not sure that's a reasonable way to go.\\n> \\n> (1) Who's going to think to look in src/test/examples/testlibpq3.c for\\n> documentation of binary formats?\\n> \\n> (2) The useful details are likely to get buried in notational and\\n> portability concerns, as I think your build failure illustrates.\\n> \\n> (3) I bet few if any packagers install these files, so that the new\\n> info would be unavailable to many people.\\n> \\n> I think some new appendix in the main SGML docs would be the appropriate\\n> place if we want to provide real documentation.\\n\\nOk, I'll leave the testlibpq3.c as it was then. If it's worth keeping\\nany of those changes, then I can remove the timestamp example because of\\nthe ntohll() portability since that is trivial.\\n\\nI'll start a new appendix and share again when I have something to show.\\n\\nRegards,\\nMark\\n\\n--\\nMark Wong\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Jun 2022 15:56:50 -0700\",\n \"msg_from\": \"Mark Wong <markwkm@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Thu, Jun 16, 2022 at 03:41:50PM -0400, Tom Lane wrote:\\n> Robert Haas <robertmhaas@gmail.com> writes:\\n> > On Tue, Jun 14, 2022 at 1:40 PM Mark Wong <markwkm@gmail.com> wrote:\\n> >> I've created a function for each data type with the idea that an example\\n> >> for handling a specific data type can be more easily reviewed by looking\\n> >> in a single place.\\n> >> I've added examples for REAL, TIMESTAMP WITHOUT TIME ZONE, and BOOLEAN\\n> >> to try to illustrate how testlibpq3.sql and testlibpq3.c will grow if\\n> >> this is a good way to go.\\n> \\n> > I'm not sure that we want to let these test programs grow very large.\\n> > In particular, I don't think we should let ourselves get sucked into\\n> > adding an example for every data type under the sun -- if that's\\n> > wanted, the solution is perhaps to add documentation for the binary\\n> > formats, not hide impromptu documentation inside a test program. But\\n> > doing this much seems OK to me.\\n> \\n> Yeah, \\\"hiding impromptu documentation inside a test program\\\" is what\\n> this looks like, and I'm not sure that's a reasonable way to go.\\n> \\n> (1) Who's going to think to look in src/test/examples/testlibpq3.c for\\n> documentation of binary formats?\\n> \\n> (2) The useful details are likely to get buried in notational and\\n> portability concerns, as I think your build failure illustrates.\\n> \\n> (3) I bet few if any packagers install these files, so that the new\\n> info would be unavailable to many people.\\n> \\n> I think some new appendix in the main SGML docs would be the appropriate\\n> place if we want to provide real documentation.\\n\\nJust wanted to do another quick check in before I go too far. How do\\ndoes this start look? Extending the libpq section with a code snippet\\nand description per data type?\\n\\nRegards,\\nMark\",\n \"msg_date\": \"Fri, 1 Jul 2022 10:18:20 -0700\",\n \"msg_from\": \"Mark Wong <markwkm@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On 01.07.22 19:18, Mark Wong wrote:\\n> Just wanted to do another quick check in before I go too far. How do\\n> does this start look? Extending the libpq section with a code snippet\\n> and description per data type?\\n\\nIf we want to start documenting the binary format for all data types, it \\nshould go into the \\\"Data Types\\\" chapter. There, we already document the \\ntext format in great detail, so putting the binary format in there as \\nwell would make sense.\\n\\nThe libpq chapter could then contain some examples of applying that \\ninformation in the context of libpq. But the libpq documentation \\nshouldn't be the primary place where the formats are documented.\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Nov 2022 11:39:40 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> If we want to start documenting the binary format for all data types, it \\n> should go into the \\\"Data Types\\\" chapter. There, we already document the \\n> text format in great detail, so putting the binary format in there as \\n> well would make sense.\\n\\nAgreed that the libpq manual is not the place for this, but I feel\\nlike it will also be clutter in \\\"Data Types\\\". Perhaps we should\\ninvent a new appendix or the like? Somewhere near the wire protocol\\ndocs seems sensible.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 01 Nov 2022 09:15:46 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On 01.11.22 09:15, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n>> If we want to start documenting the binary format for all data types, it\\n>> should go into the \\\"Data Types\\\" chapter. There, we already document the\\n>> text format in great detail, so putting the binary format in there as\\n>> well would make sense.\\n> \\n> Agreed that the libpq manual is not the place for this, but I feel\\n> like it will also be clutter in \\\"Data Types\\\". Perhaps we should\\n> invent a new appendix or the like? Somewhere near the wire protocol\\n> docs seems sensible.\\n\\nWould that clutter the protocol docs? ;-)\\n\\nI suppose figuring out exactly where to put it and how to mark it up, \\netc., in a repeatable fashion is part of the job here.\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Nov 2022 09:23:00 -0400\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> On 01.11.22 09:15, Tom Lane wrote:\\n>> Agreed that the libpq manual is not the place for this, but I feel\\n>> like it will also be clutter in \\\"Data Types\\\". Perhaps we should\\n>> invent a new appendix or the like? Somewhere near the wire protocol\\n>> docs seems sensible.\\n\\n> Would that clutter the protocol docs? ;-)\\n\\nI said \\\"near\\\", not \\\"in\\\". At the time I was thinking \\\"new appendix\\\",\\nbut I now recall that the wire protocol docs are not an appendix\\nbut a chapter in the Internals division. So that doesn't seem like\\nquite the right place anyway.\\n\\nPerhaps a new chapter under \\\"IV. Client Interfaces\\\" is the right\\nplace?\\n\\nIf we wanted to get aggressive, we could move most of the nitpicky details\\nabout datatype text formatting (e.g., the array quoting rules) there too.\\nI'm not set on that, but it'd make datatype.sgml smaller which could\\nhardly be a bad thing.\\n\\n> I suppose figuring out exactly where to put it and how to mark it up, \\n> etc., in a repeatable fashion is part of the job here.\\n\\nYup.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 03 Nov 2022 09:55:22 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 03, 2022 at 09:55:22AM -0400, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> > On 01.11.22 09:15, Tom Lane wrote:\\n> >> Agreed that the libpq manual is not the place for this, but I feel\\n> >> like it will also be clutter in \\\"Data Types\\\". Perhaps we should\\n> >> invent a new appendix or the like? Somewhere near the wire protocol\\n> >> docs seems sensible.\\n> \\n> > Would that clutter the protocol docs? ;-)\\n> \\n> I said \\\"near\\\", not \\\"in\\\". At the time I was thinking \\\"new appendix\\\",\\n> but I now recall that the wire protocol docs are not an appendix\\n> but a chapter in the Internals division. So that doesn't seem like\\n> quite the right place anyway.\\n> \\n> Perhaps a new chapter under \\\"IV. Client Interfaces\\\" is the right\\n> place?\\n> \\n> If we wanted to get aggressive, we could move most of the nitpicky details\\n> about datatype text formatting (e.g., the array quoting rules) there too.\\n> I'm not set on that, but it'd make datatype.sgml smaller which could\\n> hardly be a bad thing.\\n> \\n> > I suppose figuring out exactly where to put it and how to mark it up, \\n> > etc., in a repeatable fashion is part of the job here.\\n> \\n> Yup.\\n\\nI'll take a stab at adding a new chapter and share how that looks.\\n\\nRegards,\\nMark\\n\\n--\\nMark Wong\\nEDB https://enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 9 Nov 2022 12:47:23 -0800\",\n \"msg_from\": \"Mark Wong <markwkm@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 09, 2022 at 12:47:23PM -0800, Mark Wong wrote:\\n> I'll take a stab at adding a new chapter and share how that looks.\\n\\nThis thread has stalled for three weeks, so I have marked it as\\nreturned with feedback. Once you have a new patch, please feel free\\nto submit it.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 30 Nov 2022 15:59:12 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Thu, Nov 03, 2022 at 09:55:22AM -0400, Tom Lane wrote:\\n> Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:\\n> > On 01.11.22 09:15, Tom Lane wrote:\\n> >> Agreed that the libpq manual is not the place for this, but I feel\\n> >> like it will also be clutter in \\\"Data Types\\\". Perhaps we should\\n> >> invent a new appendix or the like? Somewhere near the wire protocol\\n> >> docs seems sensible.\\n> \\n> > Would that clutter the protocol docs? ;-)\\n> \\n> I said \\\"near\\\", not \\\"in\\\". At the time I was thinking \\\"new appendix\\\",\\n> but I now recall that the wire protocol docs are not an appendix\\n> but a chapter in the Internals division. So that doesn't seem like\\n> quite the right place anyway.\\n> \\n> Perhaps a new chapter under \\\"IV. Client Interfaces\\\" is the right\\n> place?\\n> \\n> If we wanted to get aggressive, we could move most of the nitpicky details\\n> about datatype text formatting (e.g., the array quoting rules) there too.\\n> I'm not set on that, but it'd make datatype.sgml smaller which could\\n> hardly be a bad thing.\\n> \\n> > I suppose figuring out exactly where to put it and how to mark it up, \\n> > etc., in a repeatable fashion is part of the job here.\\n> \\n> Yup.\\n\\nHow does this look?\\n\\nI've simply moved things around into a new \\\"Binary Format\\\" section with\\nthe few parts that I've started for some quick feedback about whether\\nthis is looking like the right landing place.\\n\\nRegards,\\nMark\",\n \"msg_date\": \"Fri, 13 Jan 2023 12:45:36 -0800\",\n \"msg_from\": \"Mark Wong <markwkm@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"Mark Wong <markwkm@gmail.com> writes:\\n> On Thu, Nov 03, 2022 at 09:55:22AM -0400, Tom Lane wrote:\\n>> Perhaps a new chapter under \\\"IV. Client Interfaces\\\" is the right\\n>> place?\\n>> \\n>> If we wanted to get aggressive, we could move most of the nitpicky details\\n>> about datatype text formatting (e.g., the array quoting rules) there too.\\n>> I'm not set on that, but it'd make datatype.sgml smaller which could\\n>> hardly be a bad thing.\\n\\n> I suppose figuring out exactly where to put it and how to mark it up, \\n> etc., in a repeatable fashion is part of the job here.\\n\\n> How does this look?\\n\\n> I've simply moved things around into a new \\\"Binary Format\\\" section with\\n> the few parts that I've started for some quick feedback about whether\\n> this is looking like the right landing place.\\n\\nI took a quick look at this. The patch cannot be applied as given\\nbecause it seems to be editing binary-format.sgml from some prior\\nstate, but that file doesn't exist at all yet. However, there's\\nenough visible in the patch to comment on.\\n\\nPersonally I'm not excited about supplying fragments of C code\\nlike this. Those seem quite useless to non-C users, and I'm not\\nreally sure that they'd be very helpful even for C users, because\\nyou have to make a whole lot of assumptions about what the user\\nwants to do with the value. I think what *would* be helpful is\\njust plain prose documentation of the on-the-wire binary format.\\n\\nI don't mind if you write something like\\n\\n A float4 value is a 4-byte IEEE single-precision floating point\\n number. It is transmitted in network byte order, so you must\\n convert to local byte order. (C users can do this portably\\n using the standard ntohl() function.)\\n\\nbut I'm not sure an example is worth more than such a parenthetical\\ncomment. Perhaps others disagree, though.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Fri, 20 Jan 2023 12:57:48 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 20, 2023 at 12:58 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> I don't mind if you write something like\\n>\\n> A float4 value is a 4-byte IEEE single-precision floating point\\n> number. It is transmitted in network byte order, so you must\\n> convert to local byte order. (C users can do this portably\\n> using the standard ntohl() function.)\\n>\\n> but I'm not sure an example is worth more than such a parenthetical\\n> comment. Perhaps others disagree, though.\\n\\nI don't disagree with that.\\n\\nI do think that when you suggested documenting this rather than just\\nadding some examples, you moved the goalposts a long way. If we're\\ngoing to add this to the documentation, it probably ought to cover\\nevery data type we ship. Overall, I think that would be a better\\nresult than just adding a few examples for the most common data types\\nto testlibpq*.c, but it's also substantially more work. I do agree\\nwith you that if we're going to document this, as opposed to provide\\nexamples, then a narrative style is more appropriate than a whole\\nbunch of small sample programs; maintaining working code in the\\ndocumentation seems like an annoying amount of maintenance and is\\nprobably not the most efficient way to communicate useful information.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 20 Jan 2023 13:12:07 -0500\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 20, 2023 at 01:12:07PM -0500, Robert Haas wrote:\\n> On Fri, Jan 20, 2023 at 12:58 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> > I don't mind if you write something like\\n> >\\n> > A float4 value is a 4-byte IEEE single-precision floating point\\n> > number. It is transmitted in network byte order, so you must\\n> > convert to local byte order. (C users can do this portably\\n> > using the standard ntohl() function.)\\n> >\\n> > but I'm not sure an example is worth more than such a parenthetical\\n> > comment. Perhaps others disagree, though.\\n> \\n> I don't disagree with that.\\n> \\n> I do think that when you suggested documenting this rather than just\\n> adding some examples, you moved the goalposts a long way. If we're\\n> going to add this to the documentation, it probably ought to cover\\n> every data type we ship. Overall, I think that would be a better\\n> result than just adding a few examples for the most common data types\\n> to testlibpq*.c, but it's also substantially more work. I do agree\\n> with you that if we're going to document this, as opposed to provide\\n> examples, then a narrative style is more appropriate than a whole\\n> bunch of small sample programs; maintaining working code in the\\n> documentation seems like an annoying amount of maintenance and is\\n> probably not the most efficient way to communicate useful information.\\n\\nYeah, I will try to cover all the data types we ship. :) I'll keep at\\nit and drop the code examples.\\n\\nRegards,\\nMark\\n-- \\nMark Wong\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 23 Jan 2023 08:54:26 -0800\",\n \"msg_from\": \"Mark Wong <markwkm@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Mon, 23 Jan 2023 at 11:54, Mark Wong <markwkm@gmail.com> wrote:\\nfficient way to communicate useful information.\\n>\\n> Yeah, I will try to cover all the data types we ship. :) I'll keep at\\n> it and drop the code examples.\\n\\nI assume this isn't going to happen for this commitfest? If you think\\nit is then shout otherwise I'll mark it Returned with Feedback and\\nmove it on to the next release.\\n\\n-- \\nGregory Stark\\nAs Commitfest Manager\\n\\n\\n\",\n \"msg_date\": \"Mon, 20 Mar 2023 13:37:57 -0400\",\n \"msg_from\": \"\\\"Gregory Stark (as CFM)\\\" <stark.cfm@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 20, 2023 at 01:37:57PM -0400, Gregory Stark (as CFM) wrote:\\n> On Mon, 23 Jan 2023 at 11:54, Mark Wong <markwkm@gmail.com> wrote:\\n> fficient way to communicate useful information.\\n> >\\n> > Yeah, I will try to cover all the data types we ship. :) I'll keep at\\n> > it and drop the code examples.\\n> \\n> I assume this isn't going to happen for this commitfest? If you think\\n> it is then shout otherwise I'll mark it Returned with Feedback and\\n> move it on to the next release.\\n\\nSounds good. I actually thought I did that already, thanks for catching\\nthat.\\n\\nRegards,\\nMark\\n--\\nMark Wong\\nEDB https://enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 21 Mar 2023 06:44:08 -0700\",\n \"msg_from\": \"Mark Wong <markwkm@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n },\n {\n \"msg_contents\": \"> On 21 Mar 2023, at 14:44, Mark Wong <markwkm@gmail.com> wrote:\\n> \\n> On Mon, Mar 20, 2023 at 01:37:57PM -0400, Gregory Stark (as CFM) wrote:\\n>> On Mon, 23 Jan 2023 at 11:54, Mark Wong <markwkm@gmail.com> wrote:\\n>> fficient way to communicate useful information.\\n>>> \\n>>> Yeah, I will try to cover all the data types we ship. :) I'll keep at\\n>>> it and drop the code examples.\\n>> \\n>> I assume this isn't going to happen for this commitfest? If you think\\n>> it is then shout otherwise I'll mark it Returned with Feedback and\\n>> move it on to the next release.\\n> \\n> Sounds good. I actually thought I did that already, thanks for catching\\n> that.\\n\\nThis has been waiting on author since January, unless there is a new patch\\nready I will close this as Returned with Feedback and you can resubmit for\\nanother CF when there is a new patch.\\n\\n--\\nDaniel Gustafsson\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 4 Jul 2023 09:29:01 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: real/float example for testlibpq3\"\n }\n]","isConversational":false}}},{"rowIdx":53578,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"I'm trying to understand why hash indexes are built primarily in shared\nbuffers except when allocating a new splitpoint's worth of bucket pages\n-- which is done with smgrextend() directly in _hash_alloc_buckets().\n\nIs this just so that the value returned by smgrnblocks() includes the\nnew splitpoint's worth of bucket pages?\n\nAll writes of tuple data to pages in this new splitpoint will go\nthrough shared buffers (via hash_getnewbuf()).\n\nI asked this and got some thoughts from Robert in [1], but I still don't\nreally get it.\n\nWhen a new page is needed during the hash index build, why can't\n_hash_expandtable() just call ReadBufferExtended() with P_NEW instead of\n_hash_getnewbuf()? Does it have to do with the BUCKET_TO_BLKNO mapping?\n\n- Melanie\n\n[1] https://www.postgresql.org/message-id/CA%2BTgmoa%2BQFFhkHgPxyxv6t8aVU0r7GZmu7z8io4vGG7RHPpGzA%40mail.gmail.com\n\n\n","msg_date":"Thu, 24 Feb 2022 18:10:59 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"why do hash index builds use smgrextend() for new splitpoint pages"},{"msg_contents":"On Fri, Feb 25, 2022 at 4:41 AM Melanie Plageman\n<melanieplageman@gmail.com> wrote:\n>\n> I'm trying to understand why hash indexes are built primarily in shared\n> buffers except when allocating a new splitpoint's worth of bucket pages\n> -- which is done with smgrextend() directly in _hash_alloc_buckets().\n>\n> Is this just so that the value returned by smgrnblocks() includes the\n> new splitpoint's worth of bucket pages?\n>\n> All writes of tuple data to pages in this new splitpoint will go\n> through shared buffers (via hash_getnewbuf()).\n>\n> I asked this and got some thoughts from Robert in [1], but I still don't\n> really get it.\n>\n> When a new page is needed during the hash index build, why can't\n> _hash_expandtable() just call ReadBufferExtended() with P_NEW instead of\n> _hash_getnewbuf()? Does it have to do with the BUCKET_TO_BLKNO mapping?\n>\n\nWe allocate the chunk of pages (power-of-2 groups) at the time of\nsplit which allows them to appear consecutively in an index. This\nhelps us to compute the physical block number from bucket number\neasily (BUCKET_TO_BLKNO mapping) with some minimal control\ninformation.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 25 Feb 2022 08:54:29 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: why do hash index builds use smgrextend() for new splitpoint\n pages"},{"msg_contents":"On Fri, Feb 25, 2022 at 8:54 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, Feb 25, 2022 at 4:41 AM Melanie Plageman\n> <melanieplageman@gmail.com> wrote:\n> >\n> > I'm trying to understand why hash indexes are built primarily in shared\n> > buffers except when allocating a new splitpoint's worth of bucket pages\n> > -- which is done with smgrextend() directly in _hash_alloc_buckets().\n> >\n> > Is this just so that the value returned by smgrnblocks() includes the\n> > new splitpoint's worth of bucket pages?\n> >\n> > All writes of tuple data to pages in this new splitpoint will go\n> > through shared buffers (via hash_getnewbuf()).\n> >\n> > I asked this and got some thoughts from Robert in [1], but I still don't\n> > really get it.\n> >\n> > When a new page is needed during the hash index build, why can't\n> > _hash_expandtable() just call ReadBufferExtended() with P_NEW instead of\n> > _hash_getnewbuf()? Does it have to do with the BUCKET_TO_BLKNO mapping?\n> >\n>\n> We allocate the chunk of pages (power-of-2 groups) at the time of\n> split which allows them to appear consecutively in an index.\n>\n\nI think allocating chunks of pages via \"ReadBufferExtended() with\nP_NEW\" will be time-consuming as compared to what we do now.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 25 Feb 2022 09:01:15 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: why do hash index builds use smgrextend() for new splitpoint\n pages"},{"msg_contents":"On Thu, Feb 24, 2022 at 10:24 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, Feb 25, 2022 at 4:41 AM Melanie Plageman\n> <melanieplageman@gmail.com> wrote:\n> >\n> > I'm trying to understand why hash indexes are built primarily in shared\n> > buffers except when allocating a new splitpoint's worth of bucket pages\n> > -- which is done with smgrextend() directly in _hash_alloc_buckets().\n> >\n> > Is this just so that the value returned by smgrnblocks() includes the\n> > new splitpoint's worth of bucket pages?\n> >\n> > All writes of tuple data to pages in this new splitpoint will go\n> > through shared buffers (via hash_getnewbuf()).\n> >\n> > I asked this and got some thoughts from Robert in [1], but I still don't\n> > really get it.\n> >\n> > When a new page is needed during the hash index build, why can't\n> > _hash_expandtable() just call ReadBufferExtended() with P_NEW instead of\n> > _hash_getnewbuf()? Does it have to do with the BUCKET_TO_BLKNO mapping?\n> >\n>\n> We allocate the chunk of pages (power-of-2 groups) at the time of\n> split which allows them to appear consecutively in an index. This\n> helps us to compute the physical block number from bucket number\n> easily (BUCKET_TO_BLKNO mapping) with some minimal control\n> information.\n\ngot it, thanks.\n\nSince _hash_alloc_buckets() WAL-logs the last page of the\nsplitpoint, is it safe to skip the smgrimmedsync()? What if the last\npage of the splitpoint doesn't end up having any tuples added to it\nduring the index build and the redo pointer is moved past the WAL for\nthis page and then later there is a crash sometime before this page\nmakes it to permanent storage. Does it matter that this page is lost? If\nnot, then why bother WAL-logging it?\n\n- Melanie\n\n\n","msg_date":"Fri, 25 Feb 2022 16:31:23 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"Re: why do hash index builds use smgrextend() for new splitpoint\n pages"},{"msg_contents":"On Sat, Feb 26, 2022 at 3:01 AM Melanie Plageman\n<melanieplageman@gmail.com> wrote:\n>\n> Since _hash_alloc_buckets() WAL-logs the last page of the\n> splitpoint, is it safe to skip the smgrimmedsync()? What if the last\n> page of the splitpoint doesn't end up having any tuples added to it\n> during the index build and the redo pointer is moved past the WAL for\n> this page and then later there is a crash sometime before this page\n> makes it to permanent storage. Does it matter that this page is lost? If\n> not, then why bother WAL-logging it?\n>\n\nI think we don't care if the page is lost before we update the\nmeta-page in the caller because we will try to reallocate in that\ncase. But we do care after meta page update (having the updated\ninformation about this extension via different masks) in which case we\nwon't lose this last page because it would have registered the sync\nrequest for it via sgmrextend before meta page update.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Sat, 26 Feb 2022 09:47:21 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: why do hash index builds use smgrextend() for new splitpoint\n pages"},{"msg_contents":"On Fri, Feb 25, 2022 at 11:17 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Sat, Feb 26, 2022 at 3:01 AM Melanie Plageman\n> <melanieplageman@gmail.com> wrote:\n> >\n> > Since _hash_alloc_buckets() WAL-logs the last page of the\n> > splitpoint, is it safe to skip the smgrimmedsync()? What if the last\n> > page of the splitpoint doesn't end up having any tuples added to it\n> > during the index build and the redo pointer is moved past the WAL for\n> > this page and then later there is a crash sometime before this page\n> > makes it to permanent storage. Does it matter that this page is lost? If\n> > not, then why bother WAL-logging it?\n> >\n>\n> I think we don't care if the page is lost before we update the\n> meta-page in the caller because we will try to reallocate in that\n> case. But we do care after meta page update (having the updated\n> information about this extension via different masks) in which case we\n> won't lose this last page because it would have registered the sync\n> request for it via sgmrextend before meta page update.\n\nand could it happen that during smgrextend() for the last page, a\ncheckpoint starts and finishes between FileWrite() and\nregister_dirty_segment(), then index build finishes, and then a crash\noccurs before another checkpoint completes the pending fsync for that\nlast page?\n\n\n","msg_date":"Sat, 26 Feb 2022 10:47:34 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":true,"msg_subject":"Re: why do hash index builds use smgrextend() for new splitpoint\n pages"},{"msg_contents":"On Sat, Feb 26, 2022 at 9:17 PM Melanie Plageman\n<melanieplageman@gmail.com> wrote:\n>\n> On Fri, Feb 25, 2022 at 11:17 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Sat, Feb 26, 2022 at 3:01 AM Melanie Plageman\n> > <melanieplageman@gmail.com> wrote:\n> > >\n> > > Since _hash_alloc_buckets() WAL-logs the last page of the\n> > > splitpoint, is it safe to skip the smgrimmedsync()? What if the last\n> > > page of the splitpoint doesn't end up having any tuples added to it\n> > > during the index build and the redo pointer is moved past the WAL for\n> > > this page and then later there is a crash sometime before this page\n> > > makes it to permanent storage. Does it matter that this page is lost? If\n> > > not, then why bother WAL-logging it?\n> > >\n> >\n> > I think we don't care if the page is lost before we update the\n> > meta-page in the caller because we will try to reallocate in that\n> > case. But we do care after meta page update (having the updated\n> > information about this extension via different masks) in which case we\n> > won't lose this last page because it would have registered the sync\n> > request for it via sgmrextend before meta page update.\n>\n> and could it happen that during smgrextend() for the last page, a\n> checkpoint starts and finishes between FileWrite() and\n> register_dirty_segment(), then index build finishes, and then a crash\n> occurs before another checkpoint completes the pending fsync for that\n> last page?\n>\n\nYeah, this seems to be possible and then the problem could be that\nindex's idea and smgr's idea for EOF could be different which could\nlead to a problem when we try to get a new page via _hash_getnewbuf().\nIf this theory turns out to be true then probably, we can get an error\neither because of disk full or the index might request a block that is\nbeyond EOF as determined by RelationGetNumberOfBlocksInFork() in\n_hash_getnewbuf().\n\nCan we try to reproduce this scenario with the help of a debugger to\nsee if we are missing something?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 28 Feb 2022 11:29:32 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: why do hash index builds use smgrextend() for new splitpoint\n pages"}],"string":"[\n {\n \"msg_contents\": \"I'm trying to understand why hash indexes are built primarily in shared\\nbuffers except when allocating a new splitpoint's worth of bucket pages\\n-- which is done with smgrextend() directly in _hash_alloc_buckets().\\n\\nIs this just so that the value returned by smgrnblocks() includes the\\nnew splitpoint's worth of bucket pages?\\n\\nAll writes of tuple data to pages in this new splitpoint will go\\nthrough shared buffers (via hash_getnewbuf()).\\n\\nI asked this and got some thoughts from Robert in [1], but I still don't\\nreally get it.\\n\\nWhen a new page is needed during the hash index build, why can't\\n_hash_expandtable() just call ReadBufferExtended() with P_NEW instead of\\n_hash_getnewbuf()? Does it have to do with the BUCKET_TO_BLKNO mapping?\\n\\n- Melanie\\n\\n[1] https://www.postgresql.org/message-id/CA%2BTgmoa%2BQFFhkHgPxyxv6t8aVU0r7GZmu7z8io4vGG7RHPpGzA%40mail.gmail.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 24 Feb 2022 18:10:59 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"why do hash index builds use smgrextend() for new splitpoint pages\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 4:41 AM Melanie Plageman\\n<melanieplageman@gmail.com> wrote:\\n>\\n> I'm trying to understand why hash indexes are built primarily in shared\\n> buffers except when allocating a new splitpoint's worth of bucket pages\\n> -- which is done with smgrextend() directly in _hash_alloc_buckets().\\n>\\n> Is this just so that the value returned by smgrnblocks() includes the\\n> new splitpoint's worth of bucket pages?\\n>\\n> All writes of tuple data to pages in this new splitpoint will go\\n> through shared buffers (via hash_getnewbuf()).\\n>\\n> I asked this and got some thoughts from Robert in [1], but I still don't\\n> really get it.\\n>\\n> When a new page is needed during the hash index build, why can't\\n> _hash_expandtable() just call ReadBufferExtended() with P_NEW instead of\\n> _hash_getnewbuf()? Does it have to do with the BUCKET_TO_BLKNO mapping?\\n>\\n\\nWe allocate the chunk of pages (power-of-2 groups) at the time of\\nsplit which allows them to appear consecutively in an index. This\\nhelps us to compute the physical block number from bucket number\\neasily (BUCKET_TO_BLKNO mapping) with some minimal control\\ninformation.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 08:54:29 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: why do hash index builds use smgrextend() for new splitpoint\\n pages\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 8:54 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, Feb 25, 2022 at 4:41 AM Melanie Plageman\\n> <melanieplageman@gmail.com> wrote:\\n> >\\n> > I'm trying to understand why hash indexes are built primarily in shared\\n> > buffers except when allocating a new splitpoint's worth of bucket pages\\n> > -- which is done with smgrextend() directly in _hash_alloc_buckets().\\n> >\\n> > Is this just so that the value returned by smgrnblocks() includes the\\n> > new splitpoint's worth of bucket pages?\\n> >\\n> > All writes of tuple data to pages in this new splitpoint will go\\n> > through shared buffers (via hash_getnewbuf()).\\n> >\\n> > I asked this and got some thoughts from Robert in [1], but I still don't\\n> > really get it.\\n> >\\n> > When a new page is needed during the hash index build, why can't\\n> > _hash_expandtable() just call ReadBufferExtended() with P_NEW instead of\\n> > _hash_getnewbuf()? Does it have to do with the BUCKET_TO_BLKNO mapping?\\n> >\\n>\\n> We allocate the chunk of pages (power-of-2 groups) at the time of\\n> split which allows them to appear consecutively in an index.\\n>\\n\\nI think allocating chunks of pages via \\\"ReadBufferExtended() with\\nP_NEW\\\" will be time-consuming as compared to what we do now.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 09:01:15 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: why do hash index builds use smgrextend() for new splitpoint\\n pages\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 24, 2022 at 10:24 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, Feb 25, 2022 at 4:41 AM Melanie Plageman\\n> <melanieplageman@gmail.com> wrote:\\n> >\\n> > I'm trying to understand why hash indexes are built primarily in shared\\n> > buffers except when allocating a new splitpoint's worth of bucket pages\\n> > -- which is done with smgrextend() directly in _hash_alloc_buckets().\\n> >\\n> > Is this just so that the value returned by smgrnblocks() includes the\\n> > new splitpoint's worth of bucket pages?\\n> >\\n> > All writes of tuple data to pages in this new splitpoint will go\\n> > through shared buffers (via hash_getnewbuf()).\\n> >\\n> > I asked this and got some thoughts from Robert in [1], but I still don't\\n> > really get it.\\n> >\\n> > When a new page is needed during the hash index build, why can't\\n> > _hash_expandtable() just call ReadBufferExtended() with P_NEW instead of\\n> > _hash_getnewbuf()? Does it have to do with the BUCKET_TO_BLKNO mapping?\\n> >\\n>\\n> We allocate the chunk of pages (power-of-2 groups) at the time of\\n> split which allows them to appear consecutively in an index. This\\n> helps us to compute the physical block number from bucket number\\n> easily (BUCKET_TO_BLKNO mapping) with some minimal control\\n> information.\\n\\ngot it, thanks.\\n\\nSince _hash_alloc_buckets() WAL-logs the last page of the\\nsplitpoint, is it safe to skip the smgrimmedsync()? What if the last\\npage of the splitpoint doesn't end up having any tuples added to it\\nduring the index build and the redo pointer is moved past the WAL for\\nthis page and then later there is a crash sometime before this page\\nmakes it to permanent storage. Does it matter that this page is lost? If\\nnot, then why bother WAL-logging it?\\n\\n- Melanie\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 16:31:23 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: why do hash index builds use smgrextend() for new splitpoint\\n pages\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 3:01 AM Melanie Plageman\\n<melanieplageman@gmail.com> wrote:\\n>\\n> Since _hash_alloc_buckets() WAL-logs the last page of the\\n> splitpoint, is it safe to skip the smgrimmedsync()? What if the last\\n> page of the splitpoint doesn't end up having any tuples added to it\\n> during the index build and the redo pointer is moved past the WAL for\\n> this page and then later there is a crash sometime before this page\\n> makes it to permanent storage. Does it matter that this page is lost? If\\n> not, then why bother WAL-logging it?\\n>\\n\\nI think we don't care if the page is lost before we update the\\nmeta-page in the caller because we will try to reallocate in that\\ncase. But we do care after meta page update (having the updated\\ninformation about this extension via different masks) in which case we\\nwon't lose this last page because it would have registered the sync\\nrequest for it via sgmrextend before meta page update.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 09:47:21 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: why do hash index builds use smgrextend() for new splitpoint\\n pages\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 11:17 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Sat, Feb 26, 2022 at 3:01 AM Melanie Plageman\\n> <melanieplageman@gmail.com> wrote:\\n> >\\n> > Since _hash_alloc_buckets() WAL-logs the last page of the\\n> > splitpoint, is it safe to skip the smgrimmedsync()? What if the last\\n> > page of the splitpoint doesn't end up having any tuples added to it\\n> > during the index build and the redo pointer is moved past the WAL for\\n> > this page and then later there is a crash sometime before this page\\n> > makes it to permanent storage. Does it matter that this page is lost? If\\n> > not, then why bother WAL-logging it?\\n> >\\n>\\n> I think we don't care if the page is lost before we update the\\n> meta-page in the caller because we will try to reallocate in that\\n> case. But we do care after meta page update (having the updated\\n> information about this extension via different masks) in which case we\\n> won't lose this last page because it would have registered the sync\\n> request for it via sgmrextend before meta page update.\\n\\nand could it happen that during smgrextend() for the last page, a\\ncheckpoint starts and finishes between FileWrite() and\\nregister_dirty_segment(), then index build finishes, and then a crash\\noccurs before another checkpoint completes the pending fsync for that\\nlast page?\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 10:47:34 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: why do hash index builds use smgrextend() for new splitpoint\\n pages\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 9:17 PM Melanie Plageman\\n<melanieplageman@gmail.com> wrote:\\n>\\n> On Fri, Feb 25, 2022 at 11:17 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Sat, Feb 26, 2022 at 3:01 AM Melanie Plageman\\n> > <melanieplageman@gmail.com> wrote:\\n> > >\\n> > > Since _hash_alloc_buckets() WAL-logs the last page of the\\n> > > splitpoint, is it safe to skip the smgrimmedsync()? What if the last\\n> > > page of the splitpoint doesn't end up having any tuples added to it\\n> > > during the index build and the redo pointer is moved past the WAL for\\n> > > this page and then later there is a crash sometime before this page\\n> > > makes it to permanent storage. Does it matter that this page is lost? If\\n> > > not, then why bother WAL-logging it?\\n> > >\\n> >\\n> > I think we don't care if the page is lost before we update the\\n> > meta-page in the caller because we will try to reallocate in that\\n> > case. But we do care after meta page update (having the updated\\n> > information about this extension via different masks) in which case we\\n> > won't lose this last page because it would have registered the sync\\n> > request for it via sgmrextend before meta page update.\\n>\\n> and could it happen that during smgrextend() for the last page, a\\n> checkpoint starts and finishes between FileWrite() and\\n> register_dirty_segment(), then index build finishes, and then a crash\\n> occurs before another checkpoint completes the pending fsync for that\\n> last page?\\n>\\n\\nYeah, this seems to be possible and then the problem could be that\\nindex's idea and smgr's idea for EOF could be different which could\\nlead to a problem when we try to get a new page via _hash_getnewbuf().\\nIf this theory turns out to be true then probably, we can get an error\\neither because of disk full or the index might request a block that is\\nbeyond EOF as determined by RelationGetNumberOfBlocksInFork() in\\n_hash_getnewbuf().\\n\\nCan we try to reproduce this scenario with the help of a debugger to\\nsee if we are missing something?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 11:29:32 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: why do hash index builds use smgrextend() for new splitpoint\\n pages\"\n }\n]","isConversational":false}}},{"rowIdx":53579,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nThe final commitfest for pg15 will start in a few days, and I didn't see any\ndiscussion on it or anyone volunteering to be a CFM.\n\nI thought it would be a good idea to send this reminder now and avoid the same\nsituation as the last commitfest, to avoid unnecessary pain for the CFM(s).\n\nIs there any volunteer?\n\nFor the record there are already 246 active patches registered.\n\n\n","msg_date":"Fri, 25 Feb 2022 18:30:46 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":true,"msg_subject":"Commitfest manager for 2022-03"},{"msg_contents":"I would like to volunteer.\n\nOn Fri, 25 Feb 2022 at 05:31, Julien Rouhaud <rjuju123@gmail.com> wrote:\n>\n> Hi,\n>\n> The final commitfest for pg15 will start in a few days, and I didn't see any\n> discussion on it or anyone volunteering to be a CFM.\n>\n> I thought it would be a good idea to send this reminder now and avoid the same\n> situation as the last commitfest, to avoid unnecessary pain for the CFM(s).\n>\n> Is there any volunteer?\n>\n> For the record there are already 246 active patches registered.\n>\n>\n\n\n-- \ngreg\n\n\n","msg_date":"Fri, 25 Feb 2022 13:39:25 -0500","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"Hi Greg,\n\nOn 2/25/22 12:39, Greg Stark wrote:\n> I would like to volunteer.\n> \n> On Fri, 25 Feb 2022 at 05:31, Julien Rouhaud <rjuju123@gmail.com> wrote:\n>>\n>> Is there any volunteer?\n\nSince I've done the March CF for the last seven years, I thought it \nwould be good if I chimed in. I've been agonizing a bit because I have \ntravel planned during March and while I *could* do it I don't think I'd \nbe able to do a good job.\n\nI've been hoping somebody would volunteer, so I'm all in favor of you \nbeing CF.\n\nRegards,\n-David\n\n\n","msg_date":"Fri, 25 Feb 2022 13:58:55 -0600","msg_from":"David Steele <david@pgmasters.net>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"On Fri, Feb 25, 2022 at 01:58:55PM -0600, David Steele wrote:\n> On 2/25/22 12:39, Greg Stark wrote:\n>> I would like to volunteer.\n> \n> Since I've done the March CF for the last seven years, I thought it would be\n> good if I chimed in. I've been agonizing a bit because I have travel planned\n> during March and while I *could* do it I don't think I'd be able to do a\n> good job.\n\nThat's a time-consuming task, and there have been little occasions to\nenjoy travelling lately, so have a good time :)\n\n> I've been hoping somebody would volunteer, so I'm all in favor of you being\n> CF.\n\nGreg as CFM would be fine. It's nice to see someone volunteer.\n--\nMichael","msg_date":"Sat, 26 Feb 2022 14:42:33 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"On Sat, Feb 26, 2022 at 02:42:33PM +0900, Michael Paquier wrote:\n> On Fri, Feb 25, 2022 at 01:58:55PM -0600, David Steele wrote:\n> > On 2/25/22 12:39, Greg Stark wrote:\n> >> I would like to volunteer.\n> \n> > I've been hoping somebody would volunteer, so I'm all in favor of you being\n> > CF.\n> \n> Greg as CFM would be fine. It's nice to see someone volunteer.\n\n+1, thanks a lot Greg!\n\nNote that the last CF of a major version is probably even more work than\n\"regular\" CF, so if other people are interested there's probably room for\nmultiple CFMs.\n\n\n","msg_date":"Sat, 26 Feb 2022 14:32:43 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":true,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"On Sat, 26 Feb 2022 at 01:33, Julien Rouhaud <rjuju123@gmail.com> wrote:\n>\n> On Sat, Feb 26, 2022 at 02:42:33PM +0900, Michael Paquier wrote:\n> > On Fri, Feb 25, 2022 at 01:58:55PM -0600, David Steele wrote:\n> > > On 2/25/22 12:39, Greg Stark wrote:\n> > >> I would like to volunteer.\n> >\n> > > I've been hoping somebody would volunteer, so I'm all in favor of you being\n> > > CF.\n> >\n> > Greg as CFM would be fine. It's nice to see someone volunteer.\n>\n> +1, thanks a lot Greg!\n>\n> Note that the last CF of a major version is probably even more work than\n> \"regular\" CF, so if other people are interested there's probably room for\n> multiple CFMs.\n\nI do have the time available. What I don't necessarily have is insight\ninto everything that needs to be done, especially behind the scenes.\nSo if I could have someone to give me tips about things I'm missing or\nreview the things I am doing to see that I've covered everything that\nwould be great.\n\n\n-- \ngreg\n\n\n","msg_date":"Sat, 26 Feb 2022 16:12:27 -0500","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"Can I suggest to update the CF APP to allow:\n| Target version: 16\n\nI also suggest to update patches to indicate which are (not) being considered\nfor v15.\n\nA few specific ones from myself:\n\n|https://commitfest.postgresql.org/37/2573/\n|pg_dump - read data for some options from external file \tReady for Committer \t15\nMarked RFC by Daniel Gustafsson since 2021-10-01.\nIs it more likely than not to be included in v15, or should the \"v15\" be\nremoved ?\n\n|https://commitfest.postgresql.org/37/3499/\n|libpq compression \tWaiting on Author \t15\nI re-raised the same concerns made ~12 months ago but haven't heard back.\nWOA since 35 days. Unlikely to be considered/included in v15.\n\n|https://commitfest.postgresql.org/37/3571/\n|Add LZ4 compression in pg_dump \tNeeds review\nThis patch was just submitted on 2022-02-25.\nI did a lot of the same things this patch does for a previous patch submission\n(32/2888) for pg_dump/ZSTD, so I could review this, if there were interest to\ninclude it in v15.\n\n|https://commitfest.postgresql.org/37/2349/\n|Global temporary table \tNeeds review \t15\nThe handling/hijacking of pg_class and pg_statistic is not likely to be\nacceptable. I doubt this will be progressed for v15.\n\n|https://commitfest.postgresql.org/37/3448/\n|reduce impact of lengthy startup and checkpoint tasks \tNeeds review \t15\nI think this is in the design/concept phase, and not likely for v15, except\nmaybe in reduced scope.\n\n|https://commitfest.postgresql.org/37/2906/\n|Row filtering for logical replication\nIf I'm not wrong, this is merged and should be closed?\n\nFor my own patches, it'd be helpful if someone would suggest if any are (not)\nlikely to be progressed for v15.\n\nThese are marked RFC, but have no \"committer\" set. Are they all targetting\nv15?\n\n\thttps://commitfest.postgresql.org/37/2863/\n\tFunction to log backtrace of postgres processes \tReady for Committer\n\tShould target v15 ?\n\n\thttps://commitfest.postgresql.org/37/2897/\n\tFaster pglz compression \tReady for Committer \t15\n\tTomas, are you still planning to merge this one ?\n\n\thttps://commitfest.postgresql.org/37/2992/\n\tAllow batched insert during cross-partition updates \tReady for Committer\n\n\thttps://commitfest.postgresql.org/37/3073/\n\tAdd callback table access method to reset filenode when dropping relation \tReady for Committer\n\n\thttps://commitfest.postgresql.org/37/3384/\n\tuse has_privs_for_role for predefined roles \tReady for Committer \t15\n\n\tDo all of the RFC patches target v15 (except bugfixes) ?\n\thttps://commitfest.postgresql.org/37/?status=3\n\nThese are not marked as targetting any version .. are any of them \nbeing considered for v15 ?\n\n\thttps://commitfest.postgresql.org/37/2903/\n\tParallel Hash Full Join \tNeeds review \t\n\n\thttps://commitfest.postgresql.org/37/3508/\n\tAvoid smgrimmedsync() during index build and add unbuffered IO API \tNeeds review\n\n\thttps://commitfest.postgresql.org/37/3182/\n\tautomatically generating node support functions \tNeeds review \t\n\n\thttps://commitfest.postgresql.org/37/3183/\n\tDetectable crashes and unlogged table resets \tNeeds review \t\tJeff Davis (jdavis)\n\nThese are set as \"targetting v15\", but have no committer set. If someone\nthinks it's unrealistic they'll be included in v15, I suggest to update not to\nsay so.\n\n\thttps://commitfest.postgresql.org/37/3272/\n\tAdd system view tracking shared buffer actions \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3224/\n\tFix ExecRTCheckPerms() inefficiency with many prunable partitions \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3270/\n\tCache tuple routing info during bulk loads into partitioned tables \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3461/\n\tIn-place persistence change of a relation (fast ALTER TABLE ... SET LOGGED with wal_level=minimal) \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3539/\n\tAllow parallel plan for referential integrity checks \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3397/\n\tPrefetching in recovery, take II \tNeeds review \t15 \tThomas Munro (macdice)\n\n\thttps://commitfest.postgresql.org/37/2482/\n\tjsonpath syntax extensions \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/2901/\n\tSQL/JSON: functions \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/2902/\n\tSQL/JSON: JSON_TABLE \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3099/\n\tAsymmetric partition-wise JOIN \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3293/\n\tTags in errordata \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3358/\n\tUpdate relfrozenxmin when truncating temp tables \tNeeds review \t15 \tGreg Stark (stark)\n\n\thttps://commitfest.postgresql.org/37/3490/\n\tPluggable toaster \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3500/\n\tCollecting statistics about contents of JSONB columns \tNeeds review \t15 \tTomas Vondra (fuzzycz),\n\n\thttps://commitfest.postgresql.org/37/2138/\n\tIncremental Materialized View Maintenance \tNeeds review \t15\n\n\thttps://commitfest.postgresql.org/37/3565/\n\tAdd relation and block-level filtering to pg_waldump \tNeeds review \t15 \tDavid Christensen (dwc-pgguru)\n\n\thttps://commitfest.postgresql.org/37/3436/\n\tremove exclusive backup mode \tNeeds review \t15 \tDavid Steele (dsteele)\n\n\thttps://commitfest.postgresql.org/37/3457/\n\tpg_stat_toast \tNeeds review \t15\n\n\n","msg_date":"Sat, 26 Feb 2022 18:37:21 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"On Sat, Feb 26, 2022 at 06:37:21PM -0600, Justin Pryzby wrote:\n> Can I suggest to update the CF APP to allow:\n> | Target version: 16\n> \n> I also suggest to update patches to indicate which are (not) being considered\n> for v15.\n\nI don't really understand what that field is supposed to mean. But now that\nwe're in the final pg15 commit fest, wouldn't it be simpler to actually move\nthe patches for which there's a agreement that they can't make it to pg15?\nTagging them and letting them rot for a month isn't helpful for the authors or\nthe CFMs, especially when there are already 250 patches that needs to be\nhandled.\n\nThere's already an on-going a discussion for the PGTT patchset, maybe it should\nalso happen for some of the thread you mentioned.\n\n\n","msg_date":"Sun, 27 Feb 2022 16:51:16 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":true,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"On Sun, Feb 27, 2022 at 04:51:16PM +0800, Julien Rouhaud wrote:\n> I don't really understand what that field is supposed to mean. But now that\n> we're in the final pg15 commit fest, wouldn't it be simpler to actually move\n> the patches for which there's a agreement that they can't make it to pg15?\n> Tagging them and letting them rot for a month isn't helpful for the authors or\n> the CFMs, especially when there are already 250 patches that needs to be\n> handled.\n\nYes, I don't see any point in having a new tag just to mark patches\nthat will have to be moved to the next CF anyway. These should just\nbe moved to the July one rather than stay in the March one.\n--\nMichael","msg_date":"Sun, 27 Feb 2022 18:02:01 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"On Sat, Feb 26, 2022, at 9:37 PM, Justin Pryzby wrote:\n> |https://commitfest.postgresql.org/37/2906/\n> |Row filtering for logical replication\n> If I'm not wrong, this is merged and should be closed?\nI think Amit forgot to mark it as committed. Done.\n\n\n--\nEuler Taveira\nEDB https://www.enterprisedb.com/\n\nOn Sat, Feb 26, 2022, at 9:37 PM, Justin Pryzby wrote:|https://commitfest.postgresql.org/37/2906/|Row filtering for logical replicationIf I'm not wrong, this is merged and should be closed?I think Amit forgot to mark it as committed. Done.--Euler TaveiraEDB https://www.enterprisedb.com/","msg_date":"Sun, 27 Feb 2022 15:07:16 -0300","msg_from":"\"Euler Taveira\" <euler@eulerto.com>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"On Sun, Feb 27, 2022 at 11:37 PM Euler Taveira <euler@eulerto.com> wrote:\n>\n> On Sat, Feb 26, 2022, at 9:37 PM, Justin Pryzby wrote:\n>\n> |https://commitfest.postgresql.org/37/2906/\n> |Row filtering for logical replication\n> If I'm not wrong, this is merged and should be closed?\n>\n> I think Amit forgot to mark it as committed. Done.\n>\n\nI was waiting for some build farm cycles to complete but it is fine to\nmark it now.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 28 Feb 2022 09:16:59 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"On Sun, Feb 27, 2022 at 04:51:16PM +0800, Julien Rouhaud wrote:\n> On Sat, Feb 26, 2022 at 06:37:21PM -0600, Justin Pryzby wrote:\n> > Can I suggest to update the CF APP to allow:\n> > | Target version: 16\n> > \n> > I also suggest to update patches to indicate which are (not) being considered\n> > for v15.\n> \n> I don't really understand what that field is supposed to mean.\n\nIt can mean whatever we decide it's convenient for it to mean. I suppose its\nprimary purpose may have been to indicate which backbranch a bugfix applies to\n(?)\n\nI find it useful to indicate whether or not a patch is intended to be\nconsidered for the next release. If a patch is targetting v15, it's more\ninteresting to review.\n\n@Thomas Munro: I think it'd be useful if cfbot would also show 1) the \"target\nversion\" (if any); and, 2) whether a patch has a committer set.\n\n-- \nJustin\n\n\n","msg_date":"Sat, 19 Mar 2022 15:35:05 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"Hi,\n\nOn 2022-02-26 16:12:27 -0500, Greg Stark wrote:\n> I do have the time available. What I don't necessarily have is insight\n> into everything that needs to be done, especially behind the scenes.\n\nOne thing that desperately needs doing, particularly during the last\ncommitfest, is looking through CF entries and pruning stuff that shouldn't be\nthere anymore or that are in the wrong state.\n\nI just went through the list of patches that are failing on\nhttp://cfbot.cputube.org/index.html\n\nThere were several CF entries that haven't made progress in months marked as\n\"needs review\", despite the last things on the thread being asks of the author(s).\n\nOne can't just automatically mark all failing runs as \"waiting on author\"\nbecause sometimes there are issues with somebody else posting an incremental\ndiff confusing cfbot or spurious test failures...\n\nIf a patch fails to apply and it looks to be a \"real patch\" clearly some\naction has to be taken by the author, so marking entries as \"waiting on\nauthor\" is good.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 21 Mar 2022 18:48:21 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"},{"msg_contents":"On Mon, 21 Mar 2022 at 21:48, Andres Freund <andres@anarazel.de> wrote:\n>\n> Hi,\n>\n> On 2022-02-26 16:12:27 -0500, Greg Stark wrote:\n> > I do have the time available. What I don't necessarily have is insight\n> > into everything that needs to be done, especially behind the scenes.\n>\n> One thing that desperately needs doing, particularly during the last\n> commitfest, is looking through CF entries and pruning stuff that shouldn't be\n> there anymore or that are in the wrong state.\n\nThanks\n\n> One can't just automatically mark all failing runs as \"waiting on author\"\n> because sometimes there are issues with somebody else posting an incremental\n> diff confusing cfbot or spurious test failures...\n\nWhat I'm seeing is patches that are failing with either the\n027_stream_regress.pl failure that I see is being actively worked on\nin another thread or simply a timeout which I'm not sure but may be\nthe same issue?\n\nBut I'll do a pass and then do another pass later in the week when\nthose problems may have been ironed out.\n\n\n--\ngreg\n\n\n","msg_date":"Tue, 22 Mar 2022 00:01:33 -0400","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Commitfest manager for 2022-03"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nThe final commitfest for pg15 will start in a few days, and I didn't see any\\ndiscussion on it or anyone volunteering to be a CFM.\\n\\nI thought it would be a good idea to send this reminder now and avoid the same\\nsituation as the last commitfest, to avoid unnecessary pain for the CFM(s).\\n\\nIs there any volunteer?\\n\\nFor the record there are already 246 active patches registered.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 18:30:46 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"I would like to volunteer.\\n\\nOn Fri, 25 Feb 2022 at 05:31, Julien Rouhaud <rjuju123@gmail.com> wrote:\\n>\\n> Hi,\\n>\\n> The final commitfest for pg15 will start in a few days, and I didn't see any\\n> discussion on it or anyone volunteering to be a CFM.\\n>\\n> I thought it would be a good idea to send this reminder now and avoid the same\\n> situation as the last commitfest, to avoid unnecessary pain for the CFM(s).\\n>\\n> Is there any volunteer?\\n>\\n> For the record there are already 246 active patches registered.\\n>\\n>\\n\\n\\n-- \\ngreg\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 13:39:25 -0500\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"Hi Greg,\\n\\nOn 2/25/22 12:39, Greg Stark wrote:\\n> I would like to volunteer.\\n> \\n> On Fri, 25 Feb 2022 at 05:31, Julien Rouhaud <rjuju123@gmail.com> wrote:\\n>>\\n>> Is there any volunteer?\\n\\nSince I've done the March CF for the last seven years, I thought it \\nwould be good if I chimed in. I've been agonizing a bit because I have \\ntravel planned during March and while I *could* do it I don't think I'd \\nbe able to do a good job.\\n\\nI've been hoping somebody would volunteer, so I'm all in favor of you \\nbeing CF.\\n\\nRegards,\\n-David\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 13:58:55 -0600\",\n \"msg_from\": \"David Steele <david@pgmasters.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 01:58:55PM -0600, David Steele wrote:\\n> On 2/25/22 12:39, Greg Stark wrote:\\n>> I would like to volunteer.\\n> \\n> Since I've done the March CF for the last seven years, I thought it would be\\n> good if I chimed in. I've been agonizing a bit because I have travel planned\\n> during March and while I *could* do it I don't think I'd be able to do a\\n> good job.\\n\\nThat's a time-consuming task, and there have been little occasions to\\nenjoy travelling lately, so have a good time :)\\n\\n> I've been hoping somebody would volunteer, so I'm all in favor of you being\\n> CF.\\n\\nGreg as CFM would be fine. It's nice to see someone volunteer.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 26 Feb 2022 14:42:33 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 02:42:33PM +0900, Michael Paquier wrote:\\n> On Fri, Feb 25, 2022 at 01:58:55PM -0600, David Steele wrote:\\n> > On 2/25/22 12:39, Greg Stark wrote:\\n> >> I would like to volunteer.\\n> \\n> > I've been hoping somebody would volunteer, so I'm all in favor of you being\\n> > CF.\\n> \\n> Greg as CFM would be fine. It's nice to see someone volunteer.\\n\\n+1, thanks a lot Greg!\\n\\nNote that the last CF of a major version is probably even more work than\\n\\\"regular\\\" CF, so if other people are interested there's probably room for\\nmultiple CFMs.\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 14:32:43 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"On Sat, 26 Feb 2022 at 01:33, Julien Rouhaud <rjuju123@gmail.com> wrote:\\n>\\n> On Sat, Feb 26, 2022 at 02:42:33PM +0900, Michael Paquier wrote:\\n> > On Fri, Feb 25, 2022 at 01:58:55PM -0600, David Steele wrote:\\n> > > On 2/25/22 12:39, Greg Stark wrote:\\n> > >> I would like to volunteer.\\n> >\\n> > > I've been hoping somebody would volunteer, so I'm all in favor of you being\\n> > > CF.\\n> >\\n> > Greg as CFM would be fine. It's nice to see someone volunteer.\\n>\\n> +1, thanks a lot Greg!\\n>\\n> Note that the last CF of a major version is probably even more work than\\n> \\\"regular\\\" CF, so if other people are interested there's probably room for\\n> multiple CFMs.\\n\\nI do have the time available. What I don't necessarily have is insight\\ninto everything that needs to be done, especially behind the scenes.\\nSo if I could have someone to give me tips about things I'm missing or\\nreview the things I am doing to see that I've covered everything that\\nwould be great.\\n\\n\\n-- \\ngreg\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 16:12:27 -0500\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"Can I suggest to update the CF APP to allow:\\n| Target version: 16\\n\\nI also suggest to update patches to indicate which are (not) being considered\\nfor v15.\\n\\nA few specific ones from myself:\\n\\n|https://commitfest.postgresql.org/37/2573/\\n|pg_dump - read data for some options from external file \\tReady for Committer \\t15\\nMarked RFC by Daniel Gustafsson since 2021-10-01.\\nIs it more likely than not to be included in v15, or should the \\\"v15\\\" be\\nremoved ?\\n\\n|https://commitfest.postgresql.org/37/3499/\\n|libpq compression \\tWaiting on Author \\t15\\nI re-raised the same concerns made ~12 months ago but haven't heard back.\\nWOA since 35 days. Unlikely to be considered/included in v15.\\n\\n|https://commitfest.postgresql.org/37/3571/\\n|Add LZ4 compression in pg_dump \\tNeeds review\\nThis patch was just submitted on 2022-02-25.\\nI did a lot of the same things this patch does for a previous patch submission\\n(32/2888) for pg_dump/ZSTD, so I could review this, if there were interest to\\ninclude it in v15.\\n\\n|https://commitfest.postgresql.org/37/2349/\\n|Global temporary table \\tNeeds review \\t15\\nThe handling/hijacking of pg_class and pg_statistic is not likely to be\\nacceptable. I doubt this will be progressed for v15.\\n\\n|https://commitfest.postgresql.org/37/3448/\\n|reduce impact of lengthy startup and checkpoint tasks \\tNeeds review \\t15\\nI think this is in the design/concept phase, and not likely for v15, except\\nmaybe in reduced scope.\\n\\n|https://commitfest.postgresql.org/37/2906/\\n|Row filtering for logical replication\\nIf I'm not wrong, this is merged and should be closed?\\n\\nFor my own patches, it'd be helpful if someone would suggest if any are (not)\\nlikely to be progressed for v15.\\n\\nThese are marked RFC, but have no \\\"committer\\\" set. Are they all targetting\\nv15?\\n\\n\\thttps://commitfest.postgresql.org/37/2863/\\n\\tFunction to log backtrace of postgres processes \\tReady for Committer\\n\\tShould target v15 ?\\n\\n\\thttps://commitfest.postgresql.org/37/2897/\\n\\tFaster pglz compression \\tReady for Committer \\t15\\n\\tTomas, are you still planning to merge this one ?\\n\\n\\thttps://commitfest.postgresql.org/37/2992/\\n\\tAllow batched insert during cross-partition updates \\tReady for Committer\\n\\n\\thttps://commitfest.postgresql.org/37/3073/\\n\\tAdd callback table access method to reset filenode when dropping relation \\tReady for Committer\\n\\n\\thttps://commitfest.postgresql.org/37/3384/\\n\\tuse has_privs_for_role for predefined roles \\tReady for Committer \\t15\\n\\n\\tDo all of the RFC patches target v15 (except bugfixes) ?\\n\\thttps://commitfest.postgresql.org/37/?status=3\\n\\nThese are not marked as targetting any version .. are any of them \\nbeing considered for v15 ?\\n\\n\\thttps://commitfest.postgresql.org/37/2903/\\n\\tParallel Hash Full Join \\tNeeds review \\t\\n\\n\\thttps://commitfest.postgresql.org/37/3508/\\n\\tAvoid smgrimmedsync() during index build and add unbuffered IO API \\tNeeds review\\n\\n\\thttps://commitfest.postgresql.org/37/3182/\\n\\tautomatically generating node support functions \\tNeeds review \\t\\n\\n\\thttps://commitfest.postgresql.org/37/3183/\\n\\tDetectable crashes and unlogged table resets \\tNeeds review \\t\\tJeff Davis (jdavis)\\n\\nThese are set as \\\"targetting v15\\\", but have no committer set. If someone\\nthinks it's unrealistic they'll be included in v15, I suggest to update not to\\nsay so.\\n\\n\\thttps://commitfest.postgresql.org/37/3272/\\n\\tAdd system view tracking shared buffer actions \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3224/\\n\\tFix ExecRTCheckPerms() inefficiency with many prunable partitions \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3270/\\n\\tCache tuple routing info during bulk loads into partitioned tables \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3461/\\n\\tIn-place persistence change of a relation (fast ALTER TABLE ... SET LOGGED with wal_level=minimal) \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3539/\\n\\tAllow parallel plan for referential integrity checks \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3397/\\n\\tPrefetching in recovery, take II \\tNeeds review \\t15 \\tThomas Munro (macdice)\\n\\n\\thttps://commitfest.postgresql.org/37/2482/\\n\\tjsonpath syntax extensions \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/2901/\\n\\tSQL/JSON: functions \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/2902/\\n\\tSQL/JSON: JSON_TABLE \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3099/\\n\\tAsymmetric partition-wise JOIN \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3293/\\n\\tTags in errordata \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3358/\\n\\tUpdate relfrozenxmin when truncating temp tables \\tNeeds review \\t15 \\tGreg Stark (stark)\\n\\n\\thttps://commitfest.postgresql.org/37/3490/\\n\\tPluggable toaster \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3500/\\n\\tCollecting statistics about contents of JSONB columns \\tNeeds review \\t15 \\tTomas Vondra (fuzzycz),\\n\\n\\thttps://commitfest.postgresql.org/37/2138/\\n\\tIncremental Materialized View Maintenance \\tNeeds review \\t15\\n\\n\\thttps://commitfest.postgresql.org/37/3565/\\n\\tAdd relation and block-level filtering to pg_waldump \\tNeeds review \\t15 \\tDavid Christensen (dwc-pgguru)\\n\\n\\thttps://commitfest.postgresql.org/37/3436/\\n\\tremove exclusive backup mode \\tNeeds review \\t15 \\tDavid Steele (dsteele)\\n\\n\\thttps://commitfest.postgresql.org/37/3457/\\n\\tpg_stat_toast \\tNeeds review \\t15\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 18:37:21 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 06:37:21PM -0600, Justin Pryzby wrote:\\n> Can I suggest to update the CF APP to allow:\\n> | Target version: 16\\n> \\n> I also suggest to update patches to indicate which are (not) being considered\\n> for v15.\\n\\nI don't really understand what that field is supposed to mean. But now that\\nwe're in the final pg15 commit fest, wouldn't it be simpler to actually move\\nthe patches for which there's a agreement that they can't make it to pg15?\\nTagging them and letting them rot for a month isn't helpful for the authors or\\nthe CFMs, especially when there are already 250 patches that needs to be\\nhandled.\\n\\nThere's already an on-going a discussion for the PGTT patchset, maybe it should\\nalso happen for some of the thread you mentioned.\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Feb 2022 16:51:16 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 04:51:16PM +0800, Julien Rouhaud wrote:\\n> I don't really understand what that field is supposed to mean. But now that\\n> we're in the final pg15 commit fest, wouldn't it be simpler to actually move\\n> the patches for which there's a agreement that they can't make it to pg15?\\n> Tagging them and letting them rot for a month isn't helpful for the authors or\\n> the CFMs, especially when there are already 250 patches that needs to be\\n> handled.\\n\\nYes, I don't see any point in having a new tag just to mark patches\\nthat will have to be moved to the next CF anyway. These should just\\nbe moved to the July one rather than stay in the March one.\\n--\\nMichael\",\n \"msg_date\": \"Sun, 27 Feb 2022 18:02:01 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022, at 9:37 PM, Justin Pryzby wrote:\\n> |https://commitfest.postgresql.org/37/2906/\\n> |Row filtering for logical replication\\n> If I'm not wrong, this is merged and should be closed?\\nI think Amit forgot to mark it as committed. Done.\\n\\n\\n--\\nEuler Taveira\\nEDB https://www.enterprisedb.com/\\n\\nOn Sat, Feb 26, 2022, at 9:37 PM, Justin Pryzby wrote:|https://commitfest.postgresql.org/37/2906/|Row filtering for logical replicationIf I'm not wrong, this is merged and should be closed?I think Amit forgot to mark it as committed. Done.--Euler TaveiraEDB https://www.enterprisedb.com/\",\n \"msg_date\": \"Sun, 27 Feb 2022 15:07:16 -0300\",\n \"msg_from\": \"\\\"Euler Taveira\\\" <euler@eulerto.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 11:37 PM Euler Taveira <euler@eulerto.com> wrote:\\n>\\n> On Sat, Feb 26, 2022, at 9:37 PM, Justin Pryzby wrote:\\n>\\n> |https://commitfest.postgresql.org/37/2906/\\n> |Row filtering for logical replication\\n> If I'm not wrong, this is merged and should be closed?\\n>\\n> I think Amit forgot to mark it as committed. Done.\\n>\\n\\nI was waiting for some build farm cycles to complete but it is fine to\\nmark it now.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 09:16:59 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 04:51:16PM +0800, Julien Rouhaud wrote:\\n> On Sat, Feb 26, 2022 at 06:37:21PM -0600, Justin Pryzby wrote:\\n> > Can I suggest to update the CF APP to allow:\\n> > | Target version: 16\\n> > \\n> > I also suggest to update patches to indicate which are (not) being considered\\n> > for v15.\\n> \\n> I don't really understand what that field is supposed to mean.\\n\\nIt can mean whatever we decide it's convenient for it to mean. I suppose its\\nprimary purpose may have been to indicate which backbranch a bugfix applies to\\n(?)\\n\\nI find it useful to indicate whether or not a patch is intended to be\\nconsidered for the next release. If a patch is targetting v15, it's more\\ninteresting to review.\\n\\n@Thomas Munro: I think it'd be useful if cfbot would also show 1) the \\\"target\\nversion\\\" (if any); and, 2) whether a patch has a committer set.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sat, 19 Mar 2022 15:35:05 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-26 16:12:27 -0500, Greg Stark wrote:\\n> I do have the time available. What I don't necessarily have is insight\\n> into everything that needs to be done, especially behind the scenes.\\n\\nOne thing that desperately needs doing, particularly during the last\\ncommitfest, is looking through CF entries and pruning stuff that shouldn't be\\nthere anymore or that are in the wrong state.\\n\\nI just went through the list of patches that are failing on\\nhttp://cfbot.cputube.org/index.html\\n\\nThere were several CF entries that haven't made progress in months marked as\\n\\\"needs review\\\", despite the last things on the thread being asks of the author(s).\\n\\nOne can't just automatically mark all failing runs as \\\"waiting on author\\\"\\nbecause sometimes there are issues with somebody else posting an incremental\\ndiff confusing cfbot or spurious test failures...\\n\\nIf a patch fails to apply and it looks to be a \\\"real patch\\\" clearly some\\naction has to be taken by the author, so marking entries as \\\"waiting on\\nauthor\\\" is good.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Mar 2022 18:48:21 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n },\n {\n \"msg_contents\": \"On Mon, 21 Mar 2022 at 21:48, Andres Freund <andres@anarazel.de> wrote:\\n>\\n> Hi,\\n>\\n> On 2022-02-26 16:12:27 -0500, Greg Stark wrote:\\n> > I do have the time available. What I don't necessarily have is insight\\n> > into everything that needs to be done, especially behind the scenes.\\n>\\n> One thing that desperately needs doing, particularly during the last\\n> commitfest, is looking through CF entries and pruning stuff that shouldn't be\\n> there anymore or that are in the wrong state.\\n\\nThanks\\n\\n> One can't just automatically mark all failing runs as \\\"waiting on author\\\"\\n> because sometimes there are issues with somebody else posting an incremental\\n> diff confusing cfbot or spurious test failures...\\n\\nWhat I'm seeing is patches that are failing with either the\\n027_stream_regress.pl failure that I see is being actively worked on\\nin another thread or simply a timeout which I'm not sure but may be\\nthe same issue?\\n\\nBut I'll do a pass and then do another pass later in the week when\\nthose problems may have been ironed out.\\n\\n\\n--\\ngreg\\n\\n\\n\",\n \"msg_date\": \"Tue, 22 Mar 2022 00:01:33 -0400\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Commitfest manager for 2022-03\"\n }\n]","isConversational":false}}},{"rowIdx":53580,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nplease find attached a patchset which adds lz4 compression in pg_dump.\n\nThe first commit does the heavy lifting required for additional compression methods.\nIt expands testing coverage for the already supported gzip compression. Commit\nbf9aa490db introduced cfp in compress_io.{c,h} with the intent of unifying\ncompression related code and allow for the introduction of additional archive\nformats. However, pg_backup_archiver.c was not using that API. This commit\nteaches pg_backup_archiver.c about cfp and is using it through out.\n\nFurthermore, compression was chosen based on the value of the level passed\nas an argument during the invocation of pg_dump or some hardcoded defaults. This\ndoes not scale for more than one compression methods. Now the method used for\ncompression can be explicitly requested during command invocation, or set during\nhardcoded defaults. Then it is stored in the relevant structs and passed in the\nrelevant functions, along side compression level which has lost it's special\nmeaning. The method for compression is not yet stored in the actual archive.\nThis is done in the next commit which does introduce a new method.\n\nThe previously named CompressionAlgorithm enum is changed for\nCompressionMethod so that it matches better similar variables found through out\nthe code base.\n\nIn a fashion similar to the binary for pg_basebackup, the method for compression\nis passed using the already existing -Z/--compress parameter of pg_dump. The\nlegacy format and behaviour is maintained. Additionally, the user can explicitly\npass a requested method and optionaly the level to be used after a semicolon,e.g. --compress=gzip:6\n\nThe second commit adds LZ4 compression in pg_dump and pg_restore.\n\nWithin compress_io.{c,h} there are two distinct APIs exposed, the streaming API\nand a file API. The first one, is aimed at inlined use cases and thus simple\nlz4.h calls can be used directly. The second one is generating output, or is\nparsing input, which can be read/generated via the lz4 utility.\n\nIn the later case, the API is using an opaque wrapper around a file stream,\nwhich aquired via fopen() or gzopen() respectively. It would then provide\nwrappers around fread(), fwrite(), fgets(), fgetc(), feof(), and fclose(); or\ntheir gz equivallents. However the LZ4F api does not provide this functionality.\nSo this has been implemented localy.\n\nIn order to maintain the API compatibility a new structure LZ4File is\nintroduced. It is responsible for keeping state and any yet unused generated\ncontent. The later is required when the generated decompressed output, exceeds\nthe caller's buffer capacity.\n\nCustom compressed archives need to now store the compression method in their\nheader. This requires a bump in the version number. The level of compression is\nstill stored in the dump, though admittedly is of no apparent use.\n\nThe series is authored by me. Rachel Heaton helped out with the expansion\nof the testing coverage, testing in different platforms and providing debug information\non those, as well as native speaker wording.\n\nCheers,\n//Georgios","msg_date":"Fri, 25 Feb 2022 12:05:31 +0000","msg_from":"Georgios <gkokolatos@protonmail.com>","msg_from_op":true,"msg_subject":"Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Feb 25, 2022 at 12:05:31PM +0000, Georgios wrote:\n> The first commit does the heavy lifting required for additional compression methods.\n> It expands testing coverage for the already supported gzip compression. Commit\n> bf9aa490db introduced cfp in compress_io.{c,h} with the intent of unifying\n> compression related code and allow for the introduction of additional archive\n> formats. However, pg_backup_archiver.c was not using that API. This commit\n> teaches pg_backup_archiver.c about cfp and is using it through out.\n\nThanks for the patch. I have a few high-level comments.\n\n+ # Do not use --no-sync to give test coverage for data sync.\n+ compression_gzip_directory_format => {\n+ test_key => 'compression',\n\nThe tests for GZIP had better be split into their own commit, as\nthat's a coverage improvement for the existing code.\n\nI was assuming that this was going to be much larger :)\n\n+/* Routines that support LZ4 compressed data I/O */\n+#ifdef HAVE_LIBLZ4\n+static void InitCompressorLZ4(CompressorState *cs);\n+static void ReadDataFromArchiveLZ4(ArchiveHandle *AH, ReadFunc\nreadF);\n+static void WriteDataToArchiveLZ4(ArchiveHandle *AH, CompressorState *cs,\n+ const char *data, size_t dLen);\n+static void EndCompressorLZ4(ArchiveHandle *AH, CompressorState *cs);\n+#endif\n\nHmm. This is the same set of APIs as ZLIB and NONE to init, read,\nwrite and end, but for the LZ4 compressor (NONE has no init/end).\nWouldn't it be better to refactor the existing pg_dump code to have a\ncentral structure holding all the function definitions in a common\nstructure so as all those function signatures are set in stone in the\nshape of a catalog of callbacks, making the addition of more\ncompression formats easier? I would imagine that we'd split the code\nof each compression method into their own file with their own context\ndata. This would lead to a removal of compress_io.c, with its entry\npoints ReadDataFromArchive(), WriteDataToArchive() & co replaced by\npointers to each per-compression callback.\n\n> Furthermore, compression was chosen based on the value of the level passed\n> as an argument during the invocation of pg_dump or some hardcoded defaults. This\n> does not scale for more than one compression methods. Now the method used for\n> compression can be explicitly requested during command invocation, or set during\n> hardcoded defaults. Then it is stored in the relevant structs and passed in the\n> relevant functions, along side compression level which has lost it's special\n> meaning. The method for compression is not yet stored in the actual archive.\n> This is done in the next commit which does introduce a new method.\n\nThat's one thing Robert was arguing about with pg_basebackup, so that\nwould be consistent, and the option set is backward-compatible as far\nas I get it by reading the code.\n--\nMichael","msg_date":"Tue, 1 Mar 2022 17:04:35 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"The patch is failing on cfbot/freebsd.\nhttp://cfbot.cputube.org/georgios-kokolatos.html\n\nAlso, I wondered if you'd looked at the \"high compression\" interfaces in\nlz4hc.h ? Should pg_dump use that ?\n\nOn Fri, Feb 25, 2022 at 08:03:40AM -0600, Justin Pryzby wrote:\n> Thanks for working on this. Your 0001 looks similar to what I did for zstd 1-2\n> years ago.\n> https://commitfest.postgresql.org/32/2888/\n> \n> I rebased and attached the latest patches I had in case they're useful to you.\n> I'd like to see zstd included in pg_dump eventually, but it was too much work\n> to shepherd the patches. Now that seems reasonable for pg16.\n> \n> With the other compression patches I've worked on, we've used an extra patch\n> with changes the default to the new compression algorithm, to force cfbot to\n> exercize the new code.\n> \n> Do you know the process with commitfests and cfbot ?\n> There's also this, which allows running the tests on cirrus before mailing the\n> patch to the hackers list.\n> ./src/tools/ci/README\n\n\n","msg_date":"Fri, 4 Mar 2022 10:10:39 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"It seems development on this has stalled. If there's no further work\nhappening I guess I'll mark the patch returned with feedback. Feel\nfree to resubmit it to the next CF when there's progress.\n\n\n","msg_date":"Fri, 25 Mar 2022 01:20:47 -0400","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Mar 25, 2022 at 01:20:47AM -0400, Greg Stark wrote:\n> It seems development on this has stalled. If there's no further work\n> happening I guess I'll mark the patch returned with feedback. Feel\n> free to resubmit it to the next CF when there's progress.\n\nSince it's a reasonably large patch (and one that I had myself started before)\nand it's only been 20some days since (minor) review comments, and since the\nfocus right now is on committing features, and not reviewing new patches, and\nthis patch is new one month ago, and its 0002 not intended for pg15, therefor\nI'm moving it to the next CF, where I hope to work with its authors to progress\nit.\n\n-- \nJustin\n\n\n","msg_date":"Fri, 25 Mar 2022 08:22:53 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Mar 25, 2022 at 6:22 AM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> On Fri, Mar 25, 2022 at 01:20:47AM -0400, Greg Stark wrote:\n> > It seems development on this has stalled. If there's no further work\n> > happening I guess I'll mark the patch returned with feedback. Feel\n> > free to resubmit it to the next CF when there's progress.\n>\n> Since it's a reasonably large patch (and one that I had myself started before)\n> and it's only been 20some days since (minor) review comments, and since the\n> focus right now is on committing features, and not reviewing new patches, and\n> this patch is new one month ago, and its 0002 not intended for pg15, therefor\n> I'm moving it to the next CF, where I hope to work with its authors to progress\n> it.\n>\nHi Folks,\n\nHere is an updated patchset from Georgios, with minor assistance from myself.\nThe comments above should be addressed, but please let us know if\nthere are other things to go over. A functional change in this\npatchset is when `--compress=none` is passed to pg_dump, it will not\ncompress for directory type (previously, it would use gzip if\npresent). The previous default behavior is retained.\n\n- Rachel","msg_date":"Fri, 25 Mar 2022 16:13:08 -0700","msg_from":"Rachel Heaton <rachelmheaton@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n------- Original Message -------\n\nOn Saturday, March 26th, 2022 at 12:13 AM, Rachel Heaton <rachelmheaton@gmail.com> wrote:\n\n> On Fri, Mar 25, 2022 at 6:22 AM Justin Pryzby pryzby@telsasoft.com wrote:\n>\n> > On Fri, Mar 25, 2022 at 01:20:47AM -0400, Greg Stark wrote:\n> >\n> > > It seems development on this has stalled. If there's no further work\n> > > happening I guess I'll mark the patch returned with feedback. Feel\n> > > free to resubmit it to the next CF when there's progress.\n\nWe had some progress yet we didn't want to distract the list with too many\nemails. Of course, it seemed stalled to the outside observer, yet I simply\nwanted to set the record straight and say that we are actively working on it.\n\n> >\n> > Since it's a reasonably large patch (and one that I had myself started before)\n> > and it's only been 20some days since (minor) review comments, and since the\n> > focus right now is on committing features, and not reviewing new patches, and\n> > this patch is new one month ago, and its 0002 not intended for pg15, therefor\n> > I'm moving it to the next CF, where I hope to work with its authors to progress\n> > it.\n\nThank you. It is much appreciated. We will sent updates when the next commitfest\nstarts in July as to not distract from the 15 work. Then, we can take it from there.\n\n>\n> Hi Folks,\n>\n> Here is an updated patchset from Georgios, with minor assistance from myself.\n> The comments above should be addressed, but please let us know if\n\nA small amendment to the above statement. This patchset does not include the\nrefactoring of compress_io suggested by Mr Paquier in the same thread, as it is\nmissing documentation. An updated version will be sent to include those changes\non the next commitfest.\n\n> there are other things to go over. A functional change in this\n> patchset is when `--compress=none` is passed to pg_dump, it will not\n> compress for directory type (previously, it would use gzip if\n> present). The previous default behavior is retained.\n>\n> - Rachel\n\n\n","msg_date":"Fri, 25 Mar 2022 23:43:17 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Mar 25, 2022 at 11:43:17PM +0000, gkokolatos@pm.me wrote:\n> On Saturday, March 26th, 2022 at 12:13 AM, Rachel Heaton <rachelmheaton@gmail.com> wrote:\n>> Here is an updated patchset from Georgios, with minor assistance from myself.\n>> The comments above should be addressed, but please let us know if\n> \n> A small amendment to the above statement. This patchset does not include the\n> refactoring of compress_io suggested by Mr Paquier in the same thread, as it is\n> missing documentation. An updated version will be sent to include those changes\n> on the next commitfest.\n\nThe refactoring using callbacks would make the code much cleaner IMO\nin the long term, with zstd waiting in the queue. Now, I see some\npieces of the patch set that could be merged now without waiting for\nthe development cycle of 16 to begin, as of 0001 to add more tests and\n0002.\n\nI have a question about 0002, actually. What has led you to the\nconclusion that this code is dead and could be removed?\n--\nMichael","msg_date":"Sat, 26 Mar 2022 14:57:50 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, Mar 26, 2022 at 02:57:50PM +0900, Michael Paquier wrote:\n> I have a question about 0002, actually. What has led you to the\n> conclusion that this code is dead and could be removed?\n\nSee 0001 and the manpage.\n\n+ 'pg_dump: compression is not supported by tar archive format');\n\nWhen I submitted a patch to support zstd, I spent awhile trying to make\ncompression work with tar, but it's a significant effort and better done\nseparately.\n\n\n","msg_date":"Sat, 26 Mar 2022 01:14:41 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\n\nI ran into that on an ubuntu LTS, so I don't think it's so old that it\nshouldn't be handled more gracefully. LZ4 should either have an explicit\nversion check, or else shouldn't depend on that feature (or should define a\nsafe fallback version if the library header doesn't define it).\n\nhttps://packages.ubuntu.com/liblz4-1\n\n0003: typo: of legacy => or legacy\n\nThere are a large number of ifdefs being added here - it'd be nice to minimize\nthat. basebackup was organized to use separate files, which is one way.\n\n$ git grep -c 'ifdef .*LZ4' src/bin/pg_dump/compress_io.c\nsrc/bin/pg_dump/compress_io.c:19\n\nIn last year's CF entry, I had made a union within CompressorState. LZ4\ndoesn't need z_streamp (and ztsd will need ZSTD_outBuffer, ZSTD_inBuffer,\nZSTD_CStream).\n\n0002: I wonder if you're able to re-use any of the basebackup parsing stuff\nfrom commit ffd53659c. You're passing both the compression method *and* level.\nI think there should be a structure which includes both. In the future, that\ncan also handle additional options. I hope to re-use these same things for\nwal_compression=method:level.\n\nYou renamed this:\n\n|- COMPR_ALG_LIBZ\n|-} CompressionAlgorithm;\n|+ COMPRESSION_GZIP,\n|+} CompressionMethod;\n\n..But I don't think that's an improvement. If you were to change it, it should\nsay something like PGDUMP_COMPRESS_ZLIB, since there are other compression\nstructs and typedefs. zlib is not idential to gzip, which uses a different\nheader, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\n\nThe cf* changes in pg_backup_archiver could be split out into a separate\ncommit. It's strictly a code simplification - not just preparation for more\ncompression algorithms. The commit message should \"See also:\nbf9aa490db24b2334b3595ee33653bf2fe39208c\".\n\nThe changes in 0002 for cfopen_write seem insufficient:\n|+ if (compressionMethod == COMPRESSION_NONE)\n|+ fp = cfopen(path, mode, compressionMethod, 0);\n| else\n| {\n| #ifdef HAVE_LIBZ\n| char *fname;\n| \n| fname = psprintf(\"%s.gz\", path);\n|- fp = cfopen(fname, mode, compression);\n|+ fp = cfopen(fname, mode, compressionMethod, compressionLevel);\n| free_keep_errno(fname);\n| #else\n\nThe only difference between the LIBZ and uncompressed case is the file\nextension, and it'll be the only difference with LZ4 too. So I suggest to\nfirst handle the file extension, and the rest of the code path is not\nconditional on the compression method. I don't think cfopen_write even needs\nHAVE_LIBZ - can't you handle that in cfopen_internal() ?\n\nThis patch rejects -Z0, which ought to be accepted:\n./src/bin/pg_dump/pg_dump -h /tmp regression -Fc -Z0 |wc\npg_dump: error: can only specify -Z/--compress [LEVEL] when method is set\n\nYour 0003 patch shouldn't reference LZ4:\n+#ifndef HAVE_LIBLZ4\n+ if (*compressionMethod == COMPRESSION_LZ4)\n+ supports_compression = false;\n+#endif\n\nThe 0004 patch renames zlibOutSize to outsize - I think the patch series should\nbe constructed such as to minimize the size of the method-specific patches. I\nsay this anticipating also adding support for zstd. The preliminary patches\nshould have all the boring stuff. It would help for reviewing to keep the \npatches split up, or to enumerate all the boring things that are being renamed\n(like change OutputContext to cfp, rename zlibOutSize, ...).\n\n0004: The include should use <lz4.h> and not \"lz4.h\"\n\nfreebsd/cfbot is failing.\n\nI suggested off-list to add an 0099 patch to change LZ4 to the default, to\nexercise it more on CI.\n\n-- \nJustin\n\n\n","msg_date":"Sat, 26 Mar 2022 11:21:56 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\n> You're passing both the compression method *and* level. I think there should\n> be a structure which includes both. In the future, that can also handle\n> additional options.\n\nI'm not sure if there's anything worth saving, but I did that last year with\n0003-Support-multiple-compression-algs-levels-opts.patch\nI sent a rebased copy off-list.\nhttps://www.postgresql.org/message-id/flat/20210104025321.GA9712@telsasoft.com#ca1b9f9d3552c87fa874731cad9d8391\n\n|\tfatal(\"not built with LZ4 support\");\n|\tfatal(\"not built with lz4 support\");\n\nPlease use consistent capitalization of \"lz4\" - then the compiler can optimize\naway duplicate strings.\n\n> 0004: The include should use <lz4.h> and not \"lz4.h\"\n\nAlso, use USE_LZ4 rather than HAVE_LIBLZ4, per 75eae0908.\n\n\n","msg_date":"Sat, 26 Mar 2022 13:33:36 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"> On 26 Mar 2022, at 17:21, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n> I suggested off-list to add an 0099 patch to change LZ4 to the default, to\n> exercise it more on CI.\n\nNo need to change the defaults in autoconf for that. The CFBot uses the cirrus\nfile in the tree so changing what the job includes can be easily done (assuming\nthe CFBot hasn't changed this recently which I think it hasn't). I used that\ntrick in the NSS patchset to add a completely new job for --with-ssl=nss beside\nthe --with-ssl=openssl job.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Sun, 27 Mar 2022 00:37:27 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sun, Mar 27, 2022 at 12:37:27AM +0100, Daniel Gustafsson wrote:\n> > On 26 Mar 2022, at 17:21, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> \n> > I suggested off-list to add an 0099 patch to change LZ4 to the default, to\n> > exercise it more on CI.\n> \n> No need to change the defaults in autoconf for that. The CFBot uses the cirrus\n> file in the tree so changing what the job includes can be easily done (assuming\n> the CFBot hasn't changed this recently which I think it hasn't). I used that\n> trick in the NSS patchset to add a completely new job for --with-ssl=nss beside\n> the --with-ssl=openssl job.\n\nI think you misunderstood - I'm suggesting not only to use with-lz4 (which was\nalways true since 93d973494), but to change pg_dump -Fc and -Fd to use LZ4 by\ndefault (the same as I suggested for toast_compression, wal_compression, and\nagain in last year's patch to add zstd compression to pg_dump, for which\npostgres was not ready).\n\n@@ -781,6 +807,11 @@ main(int argc, char **argv)\n compress.alg = COMPR_ALG_LIBZ;\n compress.level = Z_DEFAULT_COMPRESSION;\n #endif\n+\n+#ifdef USE_ZSTD\n+ compress.alg = COMPR_ALG_ZSTD; // Set default for testing purposes\n+ compress.level = ZSTD_CLEVEL_DEFAULT;\n+#endif\n\n\n\n","msg_date":"Sat, 26 Mar 2022 18:51:34 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, Mar 26, 2022 at 12:22 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\n> from commit ffd53659c. You're passing both the compression method *and* level.\n> I think there should be a structure which includes both. In the future, that\n> can also handle additional options. I hope to re-use these same things for\n> wal_compression=method:level.\n\nYeah, we should really try to use that infrastructure instead of\ninventing a bunch of different ways to do it. It might require some\nrenaming here and there, and I'm not sure whether we really want to\ntry to rush all this into the current release, but I think we should\nfind a way to get it done.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Sun, 27 Mar 2022 10:13:00 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sun, Mar 27, 2022 at 10:13:00AM -0400, Robert Haas wrote:\n> On Sat, Mar 26, 2022 at 12:22 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\n> > from commit ffd53659c. You're passing both the compression method *and* level.\n> > I think there should be a structure which includes both. In the future, that\n> > can also handle additional options. I hope to re-use these same things for\n> > wal_compression=method:level.\n> \n> Yeah, we should really try to use that infrastructure instead of\n> inventing a bunch of different ways to do it. It might require some\n> renaming here and there, and I'm not sure whether we really want to\n> try to rush all this into the current release, but I think we should\n> find a way to get it done.\n\nIt seems like something a whole lot like parse_compress_options() should be in\ncommon/. Nobody wants to write it again, and I couldn't convince myself to\ncopy it when I looked at using it for wal_compression.\n\nMaybe it should take an argument which specifies the default algorithm to use\nfor input of a numeric \"level\". And reject such input if not specified, since\nwal_compression has never taken a \"level\", so it's not useful or desirable to\nhave that default to some new algorithm.\n\nI could write this down if you want, although I'm not sure how/if you intend\nother people to use bc_algorithm and bc_algorithm. I don't think it's\nimportant to do for v15, but it seems like it could be done after featue\nfreeze. pg_dump+lz4 is targetting v16, although there's a cleanup patch that\ncould also go in before branching.\n\n-- \nJustin\n\n\n","msg_date":"Sun, 27 Mar 2022 11:06:35 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"> On 27 Mar 2022, at 00:51, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> \n> On Sun, Mar 27, 2022 at 12:37:27AM +0100, Daniel Gustafsson wrote:\n>>> On 26 Mar 2022, at 17:21, Justin Pryzby <pryzby@telsasoft.com> wrote:\n>> \n>>> I suggested off-list to add an 0099 patch to change LZ4 to the default, to\n>>> exercise it more on CI.\n>> \n>> No need to change the defaults in autoconf for that. The CFBot uses the cirrus\n>> file in the tree so changing what the job includes can be easily done (assuming\n>> the CFBot hasn't changed this recently which I think it hasn't). I used that\n>> trick in the NSS patchset to add a completely new job for --with-ssl=nss beside\n>> the --with-ssl=openssl job.\n> \n> I think you misunderstood - I'm suggesting not only to use with-lz4 (which was\n> always true since 93d973494), but to change pg_dump -Fc and -Fd to use LZ4 by\n> default (the same as I suggested for toast_compression, wal_compression, and\n> again in last year's patch to add zstd compression to pg_dump, for which\n> postgres was not ready).\n\nRight, I clearly misunderstood, thanks for the clarification.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Sun, 27 Mar 2022 21:55:52 +0200","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sun, Mar 27, 2022 at 12:06 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> Maybe it should take an argument which specifies the default algorithm to use\n> for input of a numeric \"level\". And reject such input if not specified, since\n> wal_compression has never taken a \"level\", so it's not useful or desirable to\n> have that default to some new algorithm.\n\nThat sounds odd to me. Wouldn't it be rather confusing if a bare\ninteger meant gzip for one case and lz4 for another?\n\n> I could write this down if you want, although I'm not sure how/if you intend\n> other people to use bc_algorithm and bc_algorithm. I don't think it's\n> important to do for v15, but it seems like it could be done after featue\n> freeze. pg_dump+lz4 is targetting v16, although there's a cleanup patch that\n> could also go in before branching.\n\nWell, I think the first thing we should do is get rid of enum\nWalCompressionMethod and use enum WalCompression instead. They've got\nthe same elements and very similar names, but the WalCompressionMethod\nones just have names like COMPRESSION_NONE, which is too generic,\nwhereas WalCompressionMethod uses WAL_COMPRESSION_NONE, which is\nbetter. Then I think we should also rename the COMPR_ALG_* constants\nin pg_dump.h to names like DUMP_COMPRESSION_*. Once we do that we've\ngot rid of all the unprefixed things that purport to be a list of\ncompression algorithms.\n\nThen, if people are willing to adopt the syntax that the\nbackup_compression.c/h stuff supports as a project standard (+1 from\nme) we can go the other way and rename that stuff to be more generic,\ntaking backup out of the name.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Mon, 28 Mar 2022 08:36:15 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Mar 28, 2022 at 08:36:15AM -0400, Robert Haas wrote:\n> Well, I think the first thing we should do is get rid of enum\n> WalCompressionMethod and use enum WalCompression instead. They've got\n> the same elements and very similar names, but the WalCompressionMethod\n> ones just have names like COMPRESSION_NONE, which is too generic,\n> whereas WalCompressionMethod uses WAL_COMPRESSION_NONE, which is\n> better. Then I think we should also rename the COMPR_ALG_* constants\n> in pg_dump.h to names like DUMP_COMPRESSION_*. Once we do that we've\n> got rid of all the unprefixed things that purport to be a list of\n> compression algorithms.\n\nYes, having a centralized enum for the compression method would make\nsense, along with the routines to parse and get the compression method\nnames. At least that would be one step towards more unity in\nsrc/common/.\n\n> Then, if people are willing to adopt the syntax that the\n> backup_compression.c/h stuff supports as a project standard (+1 from\n> me) we can go the other way and rename that stuff to be more generic,\n> taking backup out of the name.\n\nI am not sure about the specification part which is only used by base\nbackups that has no client-server requirements, so option values would\nstill require their own grammar.\n--\nMichael","msg_date":"Tue, 29 Mar 2022 14:03:12 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, Mar 26, 2022 at 01:14:41AM -0500, Justin Pryzby wrote:\n> See 0001 and the manpage.\n> \n> + 'pg_dump: compression is not supported by tar archive format');\n> \n> When I submitted a patch to support zstd, I spent awhile trying to make\n> compression work with tar, but it's a significant effort and better done\n> separately.\n\nWow. This stuff is old enough to vote (c3e18804), dead since its\nintroduction. There is indeed an argument for removing that, it is\nnot good to keep around that that has never been stressed and/or\nused. Upon review, the cleanup done looks correct, as we have never\nbeen able to generate .dat.gz files in for a dump in the tar format.\n\n+ command_fails_like(\n+ [ 'pg_dump', '--compress', '1', '--format', 'tar' ],\nThis addition depending on HAVE_LIBZ is a good thing as a reminder of\nany work that could be done in 0002. Now that's waiting for 20 years\nso I would not hold my breath on this support. I think that this\ncould be just applied first, with 0002 on top of it, as a first\nimprovement.\n\n+ compress_cmd => [\n+ $ENV{'GZIP_PROGRAM'},\nPatch 0001 is missing and update of pg_dump's Makefile to pass down\nthis environment variable to the test scripts, no?\n\n+ compress_cmd => [\n+ $ENV{'GZIP_PROGRAM'},\n+ '-f',\n[...]\n+ $ENV{'GZIP_PROGRAM'},\n+ '-k', '-d',\n-f and -d are available everywhere I looked at, but is -k/--keep a\nportable choice with a gzip command? I don't see this option in\nOpenBSD, for one. So this test is going to cause problems on those\nbuildfarm machines, at least. Couldn't this part be replaced by a\nsimple --test to check that what has been compressed is in correct\nshape? We know that this works, based on our recent experiences with\nthe other tests.\n--\nMichael","msg_date":"Tue, 29 Mar 2022 16:27:32 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\n\nOn Tuesday, March 29th, 2022 at 9:27 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Sat, Mar 26, 2022 at 01:14:41AM -0500, Justin Pryzby wrote:\n> > See 0001 and the manpage.\n> > + 'pg_dump: compression is not supported by tar archive format');\n> > When I submitted a patch to support zstd, I spent awhile trying to make\n> > compression work with tar, but it's a significant effort and better done\n> > separately.\n>\n> Wow. This stuff is old enough to vote (c3e18804), dead since its\n> introduction. There is indeed an argument for removing that, it is\n> not good to keep around that that has never been stressed and/or\n> used. Upon review, the cleanup done looks correct, as we have never\n> been able to generate .dat.gz files in for a dump in the tar format.\n\nCorrect. My driving force behind it was to ease up the cleanup/refactoring\nwork that follows, by eliminating the callers of the GZ*() macros.\n\n> + command_fails_like(\n>\n> + [ 'pg_dump', '--compress', '1', '--format', 'tar' ],\n> This addition depending on HAVE_LIBZ is a good thing as a reminder of\n> any work that could be done in 0002. Now that's waiting for 20 years\n> so I would not hold my breath on this support. I think that this\n> could be just applied first, with 0002 on top of it, as a first\n> improvement.\n\nExcellent, thank you.\n\n> + compress_cmd => [\n> + $ENV{'GZIP_PROGRAM'},\n> Patch 0001 is missing and update of pg_dump's Makefile to pass down\n> this environment variable to the test scripts, no?\n\nAgreed. It was not properly moved forward. Fixed.\n\n> + compress_cmd => [\n> + $ENV{'GZIP_PROGRAM'},\n> + '-f',\n> [...]\n> + $ENV{'GZIP_PROGRAM'},\n> + '-k', '-d',\n> -f and -d are available everywhere I looked at, but is -k/--keep a\n> portable choice with a gzip command? I don't see this option in\n> OpenBSD, for one. So this test is going to cause problems on those\n> buildfarm machines, at least. Couldn't this part be replaced by a\n> simple --test to check that what has been compressed is in correct\n> shape? We know that this works, based on our recent experiences with\n> the other tests.\n\nI would argue that the simple '--test' will not do in this case, as the\nTAP tests do need a file named <test>.sql to compare the contents with.\nThis file is generated either directly by pg_dump itself, or by running\npg_restore on pg_dump's output. In the case of compression pg_dump will\ngenerate a <test>.sql.<compression program suffix> which can not be\nused in the comparison tests. So the intention of this block, is not to\nsimply test for validity, but to also decompress pg_dump's output for it\nto be able to be used.\n\nI updated the patch to simply remove the '-k' flag.\n\nPlease find v3 attached. (only 0001 and 0002 are relevant, 0003 and\n0004 are only for reference and are currently under active modification).\n\nCheers,\n//Georgios","msg_date":"Tue, 29 Mar 2022 09:46:27 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Mar 29, 2022 at 1:03 AM Michael Paquier <michael@paquier.xyz> wrote:\n> > Then, if people are willing to adopt the syntax that the\n> > backup_compression.c/h stuff supports as a project standard (+1 from\n> > me) we can go the other way and rename that stuff to be more generic,\n> > taking backup out of the name.\n>\n> I am not sure about the specification part which is only used by base\n> backups that has no client-server requirements, so option values would\n> still require their own grammar.\n\nI don't know what you mean by this. I think the specification stuff\ncould be reused in a lot of places. If you can ask for a base backup\nwith zstd:level=3,long=1,fancystuff=yes or whatever we end up with,\nwhy not enable exactly the same for every other place that uses\ncompression? I don't know what \"client-server requirements\" is or what\nthat has to do with this.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 29 Mar 2022 09:14:03 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Mar 29, 2022 at 09:46:27AM +0000, gkokolatos@pm.me wrote:\n> On Tuesday, March 29th, 2022 at 9:27 AM, Michael Paquier <michael@paquier.xyz> wrote:\n>> On Sat, Mar 26, 2022 at 01:14:41AM -0500, Justin Pryzby wrote:\n>> Wow. This stuff is old enough to vote (c3e18804), dead since its\n>> introduction. There is indeed an argument for removing that, it is\n>> not good to keep around that that has never been stressed and/or\n>> used. Upon review, the cleanup done looks correct, as we have never\n>> been able to generate .dat.gz files in for a dump in the tar format.\n> \n> Correct. My driving force behind it was to ease up the cleanup/refactoring\n> work that follows, by eliminating the callers of the GZ*() macros.\n\nMakes sense to me.\n\n>> + command_fails_like(\n>>\n>> + [ 'pg_dump', '--compress', '1', '--format', 'tar' ],\n>> This addition depending on HAVE_LIBZ is a good thing as a reminder of\n>> any work that could be done in 0002. Now that's waiting for 20 years\n>> so I would not hold my breath on this support. I think that this\n>> could be just applied first, with 0002 on top of it, as a first\n>> improvement.\n> \n> Excellent, thank you.\n\nI have applied the test for --compress and --format=tar, separating it\nfrom the rest.\n\nWhile moving on with 0002, I have noticed the following in\n_StartBlob():\n if (AH->compression != 0)\n sfx = \".gz\";\n else\n sfx = \"\";\n\t\nShouldn't this bit also be simplified, adding a fatal() like the other\ncode paths, for safety?\n\n>> + compress_cmd => [\n>> + $ENV{'GZIP_PROGRAM'},\n>> + '-f',\n>> [...]\n>> + $ENV{'GZIP_PROGRAM'},\n>> + '-k', '-d',\n>> -f and -d are available everywhere I looked at, but is -k/--keep a\n>> portable choice with a gzip command? I don't see this option in\n>> OpenBSD, for one. So this test is going to cause problems on those\n>> buildfarm machines, at least. Couldn't this part be replaced by a\n>> simple --test to check that what has been compressed is in correct\n>> shape? We know that this works, based on our recent experiences with\n>> the other tests.\n> \n> I would argue that the simple '--test' will not do in this case, as the\n> TAP tests do need a file named <test>.sql to compare the contents with.\n> This file is generated either directly by pg_dump itself, or by running\n> pg_restore on pg_dump's output. In the case of compression pg_dump will\n> generate a <test>.sql.<compression program suffix> which can not be\n> used in the comparison tests. So the intention of this block, is not to\n> simply test for validity, but to also decompress pg_dump's output for it\n> to be able to be used.\n\nAhh, I see, thanks. I would add a comment about that in the area of\ncompression_gzip_plain_format.\n\n+ my $supports_compression = check_pg_config(\"#define HAVE_LIBZ 1\");\n\nThis part could be moved within the if block a couple of lines down.\n\n+ my $compress_program = $ENV{GZIP_PROGRAM};\n\nIt seems to me that it is enough to rely on {compress_cmd}, hence\nthere should be no need for $compress_program, no?\n\nIt seems to me that we should have a description for compress_cmd at\nthe top of 002_pg_dump.pl (close to \"Definition of the pg_dump runs to\nmake\"). There is an order dependency with restore_cmd.\n\n> I updated the patch to simply remove the '-k' flag.\n\nOkay.\n--\nMichael","msg_date":"Wed, 30 Mar 2022 14:54:34 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Mar 29, 2022 at 09:14:03AM -0400, Robert Haas wrote:\n> I don't know what you mean by this. I think the specification stuff\n> could be reused in a lot of places. If you can ask for a base backup\n> with zstd:level=3,long=1,fancystuff=yes or whatever we end up with,\n> why not enable exactly the same for every other place that uses\n> compression? I don't know what \"client-server requirements\" is or what\n> that has to do with this.\n\nOh. I think that I got confused here. I saw the backup component in\nthe file name and this has been associated with the client/server\nchoice that can be done in the options of pg_basebackup. But\nparse_bc_specification() does not include any knowledge about that:\npg_basebackup does this job in parse_compress_options(). I agree that\nit looks possible to reuse that stuff in more places than just base\nbackups.\n--\nMichael","msg_date":"Wed, 30 Mar 2022 15:36:02 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\n\nOn Wednesday, March 30th, 2022 at 7:54 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Tue, Mar 29, 2022 at 09:46:27AM +0000, gkokolatos@pm.me wrote:\n> > On Tuesday, March 29th, 2022 at 9:27 AM, Michael Paquier michael@paquier.xyz wrote:\n> > > On Sat, Mar 26, 2022 at 01:14:41AM -0500, Justin Pryzby wrote:\n> > > + command_fails_like(\n> > > + [ 'pg_dump', '--compress', '1', '--format', 'tar' ],\n> > > This addition depending on HAVE_LIBZ is a good thing as a reminder of\n> > > any work that could be done in 0002. Now that's waiting for 20 years\n> > > so I would not hold my breath on this support. I think that this\n> > > could be just applied first, with 0002 on top of it, as a first\n> > > improvement.\n> >\n> > Excellent, thank you.\n>\n> I have applied the test for --compress and --format=tar, separating it\n> from the rest.\n\nThank you.\n\n> While moving on with 0002, I have noticed the following in\n>\n> _StartBlob():\n> if (AH->compression != 0)\n> sfx = \".gz\";\n> else\n> sfx = \"\";\n>\n> Shouldn't this bit also be simplified, adding a fatal() like the other\n> code paths, for safety?\n\nAgreed. Fixed.\n\n> > > + compress_cmd => [\n> > > + $ENV{'GZIP_PROGRAM'},\n> > > + '-f',\n> > > [...]\n> > > + $ENV{'GZIP_PROGRAM'},\n> > > + '-k', '-d',\n> > > -f and -d are available everywhere I looked at, but is -k/--keep a\n> > > portable choice with a gzip command? I don't see this option in\n> > > OpenBSD, for one. So this test is going to cause problems on those\n> > > buildfarm machines, at least. Couldn't this part be replaced by a\n> > > simple --test to check that what has been compressed is in correct\n> > > shape? We know that this works, based on our recent experiences with\n> > > the other tests.\n> >\n> > I would argue that the simple '--test' will not do in this case, as the\n> > TAP tests do need a file named <test>.sql to compare the contents with.\n> > This file is generated either directly by pg_dump itself, or by running\n> > pg_restore on pg_dump's output. In the case of compression pg_dump will\n> > generate a <test>.sql.<compression program suffix> which can not be\n> > used in the comparison tests. So the intention of this block, is not to\n> > simply test for validity, but to also decompress pg_dump's output for it\n> > to be able to be used.\n>\n> Ahh, I see, thanks. I would add a comment about that in the area of\n> compression_gzip_plain_format.\n\nAgreed. Comment added.\n\n> + my $supports_compression = check_pg_config(\"#define HAVE_LIBZ 1\");\n>\n> This part could be moved within the if block a couple of lines down.\n\nI moved it instead out of the for loop above to not have to call it on\neach iteration.\n\n> + my $compress_program = $ENV{GZIP_PROGRAM};\n>\n> It seems to me that it is enough to rely on {compress_cmd}, hence\n> there should be no need for $compress_program, no?\n\nMaybe not. We don't want to the tests to fail if the utility is not\ninstalled. That becomes even more evident as more methods are added.\nHowever I realized that the presence of the environmental variable does\nnot guarrantee that the utility is actually installed. In the attached,\nthe existance of the utility is based on the return value of system_log().\n\n> It seems to me that we should have a description for compress_cmd at\n> the top of 002_pg_dump.pl (close to \"Definition of the pg_dump runs to\n> make\"). There is an order dependency with restore_cmd.\n\nAgreed. Comment added.\n\nCheers,\n//Georgios","msg_date":"Wed, 30 Mar 2022 15:32:55 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Mar 30, 2022 at 03:32:55PM +0000, gkokolatos@pm.me wrote:\n> On Wednesday, March 30th, 2022 at 7:54 AM, Michael Paquier <michael@paquier.xyz> wrote:\n>> While moving on with 0002, I have noticed the following in\n>>\n>> _StartBlob():\n>> if (AH->compression != 0)\n>> sfx = \".gz\";\n>> else\n>> sfx = \"\";\n>>\n>> Shouldn't this bit also be simplified, adding a fatal() like the other\n>> code paths, for safety?\n> \n> Agreed. Fixed.\n\nOkay. 0002 looks fine as-is, and I don't mind the extra fatal()\ncalls. These could be asserts but that's not a big deal one way or\nthe other. And the cleanup is now applied.\n\n>> + my $compress_program = $ENV{GZIP_PROGRAM};\n>>\n>> It seems to me that it is enough to rely on {compress_cmd}, hence\n>> there should be no need for $compress_program, no?\n> \n> Maybe not. We don't want to the tests to fail if the utility is not\n> installed. That becomes even more evident as more methods are added.\n> However I realized that the presence of the environmental variable does\n> not guarrantee that the utility is actually installed. In the attached,\n> the existance of the utility is based on the return value of system_log().\n\nHmm. [.. thinks ..] The thing that's itching me here is that you\nalign the concept of compression with gzip, but that's not going to be\ntrue once more compression options are added to pg_dump, and that\nwould make $supports_compression and $compress_program_exists\nincorrect. Perhaps the right answer would be to rename all that with\na suffix like \"_gzip\" to make a difference? Or would there be enough\ncontrol with a value of \"compression_gzip\" instead of \"compression\" in\ntest_key?\n\n+my $compress_program_exists = (system_log(\"$ENV{GZIP_PROGRAM}\", '-h',\n+ '>', '/dev/null') == 0);\nDo we need this command execution at all? In all the other tests, we\nrely on a simple \"if (!defined $gzip || $gzip eq '');\", so we could do\nthe same here.\n\nA last thing is that we should perhaps make a clear difference between\nthe check that looks at if the code has been built with zlib and the\ncheck for the presence of GZIP_PROGRAM, as it can be useful in some\nenvironments to be able to run pg_dump built with zlib, even if the\nGZIP_PROGRAM command does not exist (I don't think this can be the\ncase, but other tests are flexible). As of now, the patch relies on\npg_dump enforcing uncompression if building under --without-zlib even\nif --compress/-Z is used, but that also means that those compression\ntests overlap with the existing tests in this case. Wouldn't it be\nmore consistent to check after $supports_compression when executing\nthe dump command for test_key = \"compression[_gzip]\"? This would mean\nkeeping GZIP_PROGRAM as sole check when executing the compression\ncommand.\n--\nMichael","msg_date":"Thu, 31 Mar 2022 11:34:52 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thursday, March 31st, 2022 at 4:34 AM, Michael Paquier <michael@paquier.xyz> wrote:\n> On Wed, Mar 30, 2022 at 03:32:55PM +0000, gkokolatos@pm.me wrote:\n> > On Wednesday, March 30th, 2022 at 7:54 AM, Michael Paquier michael@paquier.xyz wrote:\n>\n> Okay. 0002 looks fine as-is, and I don't mind the extra fatal()\n> calls. These could be asserts but that's not a big deal one way or\n> the other. And the cleanup is now applied.\n\nThank you very much.\n\n> > > + my $compress_program = $ENV{GZIP_PROGRAM};\n> > > It seems to me that it is enough to rely on {compress_cmd}, hence\n> > > there should be no need for $compress_program, no?\n> >\n> > Maybe not. We don't want to the tests to fail if the utility is not\n> > installed. That becomes even more evident as more methods are added.\n> > However I realized that the presence of the environmental variable does\n> > not guarrantee that the utility is actually installed. In the attached,\n> > the existance of the utility is based on the return value of system_log().\n>\n> Hmm. [.. thinks ..] The thing that's itching me here is that you\n> align the concept of compression with gzip, but that's not going to be\n> true once more compression options are added to pg_dump, and that\n> would make $supports_compression and $compress_program_exists\n> incorrect. Perhaps the right answer would be to rename all that with\n> a suffix like \"_gzip\" to make a difference? Or would there be enough\n> control with a value of \"compression_gzip\" instead of \"compression\" in\n> test_key?\n\nI understand the itch. Indeed when LZ4 is added as compression method, this\nblock changes slightly. I went with the minimum amount changed. Please find\nin 0001 of the attached this variable renamed as $gzip_program_exist. I thought\nthat as prefix it will match better the already used $ENV{GZIP_PROGRAM}.\n\n> +my $compress_program_exists = (system_log(\"$ENV{GZIP_PROGRAM}\", '-h',\n> + '>', '/dev/null') == 0);\n>\n> Do we need this command execution at all? In all the other tests, we\n> rely on a simple \"if (!defined $gzip || $gzip eq '');\", so we could do\n> the same here.\n\nYou are very correct that we are using the simple version, and that is what\nit was included in the previous versions of the current patch. However, I\ndid notice that the variable is hard-coded in Makefile.global.in and it does\nnot go through configure. By now, gzip is considered an essential package\nin most installations, and this hard-code makes sense. Though I did remove\nthe utility from my system, (apt remove gzip) and tried the test with the\nsimple \"if (!defined $gzip || $gzip eq '');\", which predictably failed. For\nthis, I went with the system call, it is not too expensive and is rather\nreliable.\n\nIt is true that the rest of the TAP tests that use this, e.g. in pg_basebackup,\nalso failed. There is an argument to go simple and I will be happy to revert\nto the previous version.\n\n> A last thing is that we should perhaps make a clear difference between\n> the check that looks at if the code has been built with zlib and the\n> check for the presence of GZIP_PROGRAM, as it can be useful in some\n> environments to be able to run pg_dump built with zlib, even if the\n> GZIP_PROGRAM command does not exist (I don't think this can be the\n> case, but other tests are flexible).\n\nYou are very correct. We do that already in the current patch. Note that we skip\nthe test only when we specifically have to execute a compression command. Not\nall compression tests define such command, exactly so that we can test those\ncases as well. The point of using an external utility program is in order to\nextend the coverage in previously untested yet supported scenarios, e.g. manual\ncompression of the *.toc files.\n\nAlso in the case where it will actually skip the compression command because the\ngzip program is not present, it will execute the pg_dump command first.\n\n> As of now, the patch relies on\n> pg_dump enforcing uncompression if building under --without-zlib even\n> if --compress/-Z is used, but that also means that those compression\n> tests overlap with the existing tests in this case. Wouldn't it be\n> more consistent to check after $supports_compression when executing\n> the dump command for test_key = \"compression[_gzip]\"? This would mean\n> keeping GZIP_PROGRAM as sole check when executing the compression\n> command.\n\nI can see the overlap case. Yet, I understand the test_key as serving different\npurpose, as it is a key of %tests and %full_runs. I do not expect the database\ncontent of the generated dump to change based on which compression method is used.\n\nIn the next round, I can see one explitcly requesting --compress=none to override\ndefaults. There is a benefit to group the tests for this scenario under the same\ntest_key, i.e. compression.\n\nAlso there will be cases where if the program exists, yet the codebase is compiled\nwithout support for the method. Then compress_cmd or the restore_cmd that follows\nwill fail. For example, in the plain output, if we try to uncompress the generated\nthe test will fail with 'gzip: <filename> not in gzip format'. In the directory\nformat the compress_cmd will compress the *.toc files, but the restore_cmd will\nfail because it does not build with support for them.\n\nIn the attached version, I propose that the compression_cmd is converted into\na hash. It contains two keys, the program and the arguments. Maybe it is easier\nto read than before or than simply grabbing the first element of the array.\n\nCheers,\n//Georgios","msg_date":"Fri, 01 Apr 2022 15:06:40 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Apr 01, 2022 at 03:06:40PM +0000, gkokolatos@pm.me wrote:\n> I understand the itch. Indeed when LZ4 is added as compression method, this\n> block changes slightly. I went with the minimum amount changed. Please find\n> in 0001 of the attached this variable renamed as $gzip_program_exist. I thought\n> that as prefix it will match better the already used $ENV{GZIP_PROGRAM}.\n\nHmm. I have spent some time on that, and upon review I really think\nthat we should skip the tests marked as dedicated to the gzip\ncompression entirely if the build is not compiled with this option,\nrather than letting the code run a dump for nothing in some cases,\nrelying on the default to uncompress the contents in others. In the\nlatter case, it happens that we have already some checks like\ndefaults_custom_format, but you already mentioned that.\n\nWe should also skip the later parts of the tests if the compression\nprogram does not exist as we rely on it, but only if the command does\nnot exist. This will count for LZ4.\n\n> I can see the overlap case. Yet, I understand the test_key as serving different\n> purpose, as it is a key of %tests and %full_runs. I do not expect the database\n> content of the generated dump to change based on which compression method is used.\n\nContrary to the current LZ4 tests in pg_dump, what we have here is a\ncheck for a command-level run and not a data-level check. So what's\nintroduced is a new concept, and we need a new way to control if the\ntests should be entirely skipped or not, particularly if we finish by\nnot using test_key to make the difference. Perhaps the best way to\naddress that is to have a new keyword in the $runs structure. The\nattached defines a new compile_option, that can be completed later for\nnew compression methods introduced in the tests. So the idea is to\nmark all the tests related to compression with the same test_key, and\nthe tests can be skipped depending on what compile_option requires.\n\n> In the attached version, I propose that the compression_cmd is converted into\n> a hash. It contains two keys, the program and the arguments. Maybe it is easier\n> to read than before or than simply grabbing the first element of the array.\n\nSplitting the program and its arguments makes sense.\n\nAt the end I am finishing with the attached. I also saw an overlap\nwith the addition of --jobs for the directory format vs not using the\noption, so I have removed the case where --jobs was not used in the\ndirectory format.\n--\nMichael","msg_date":"Tue, 5 Apr 2022 10:34:49 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n------- Original Message -------\n\nOn Tuesday, April 5th, 2022 at 3:34 AM, Michael Paquier <michael@paquier.xyz> wrote:\n> On Fri, Apr 01, 2022 at 03:06:40PM +0000, gkokolatos@pm.me wrote:\n\n> Splitting the program and its arguments makes sense.\n\nGreat.\n\n> At the end I am finishing with the attached. I also saw an overlap\n> with the addition of --jobs for the directory format vs not using the\n> option, so I have removed the case where --jobs was not used in the\n> directory format.\n\nThank you. I agree with the attached and I will carry it forward to the\nrest of the patchset.\n\nCheers,\n//Georgios\n\n> --\n> Michael\n\n\n","msg_date":"Tue, 05 Apr 2022 07:13:35 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Apr 05, 2022 at 07:13:35AM +0000, gkokolatos@pm.me wrote:\n> Thank you. I agree with the attached and I will carry it forward to the\n> rest of the patchset.\n\nNo need to carry it forward anymore, I think ;)\n--\nMichael","msg_date":"Tue, 5 Apr 2022 19:55:05 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n------- Original Message -------\n\nOn Tuesday, April 5th, 2022 at 12:55 PM, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On Tue, Apr 05, 2022 at 07:13:35AM +0000, gkokolatos@pm.me wrote:\n> No need to carry it forward anymore, I think ;)\n\nThank you for committing!\n\nCheers,\n//Georgios\n\n> --\n> Michael\n\n\n","msg_date":"Tue, 05 Apr 2022 11:12:28 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Hi,\n\nWill you be able to send a rebased patch for the next CF ?\n\nIf you update for the review comments I sent in March, I'll plan to do another\nround of review.\n\nOn Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\n> LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\n> \n> I ran into that on an ubuntu LTS, so I don't think it's so old that it\n> shouldn't be handled more gracefully. LZ4 should either have an explicit\n> version check, or else shouldn't depend on that feature (or should define a\n> safe fallback version if the library header doesn't define it).\n> \n> https://packages.ubuntu.com/liblz4-1\n> \n> 0003: typo: of legacy => or legacy\n> \n> There are a large number of ifdefs being added here - it'd be nice to minimize\n> that. basebackup was organized to use separate files, which is one way.\n> \n> $ git grep -c 'ifdef .*LZ4' src/bin/pg_dump/compress_io.c\n> src/bin/pg_dump/compress_io.c:19\n> \n> In last year's CF entry, I had made a union within CompressorState. LZ4\n> doesn't need z_streamp (and ztsd will need ZSTD_outBuffer, ZSTD_inBuffer,\n> ZSTD_CStream).\n> \n> 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\n> from commit ffd53659c. You're passing both the compression method *and* level.\n> I think there should be a structure which includes both. In the future, that\n> can also handle additional options. I hope to re-use these same things for\n> wal_compression=method:level.\n> \n> You renamed this:\n> \n> |- COMPR_ALG_LIBZ\n> |-} CompressionAlgorithm;\n> |+ COMPRESSION_GZIP,\n> |+} CompressionMethod;\n> \n> ..But I don't think that's an improvement. If you were to change it, it should\n> say something like PGDUMP_COMPRESS_ZLIB, since there are other compression\n> structs and typedefs. zlib is not idential to gzip, which uses a different\n> header, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\n> \n> The cf* changes in pg_backup_archiver could be split out into a separate\n> commit. It's strictly a code simplification - not just preparation for more\n> compression algorithms. The commit message should \"See also:\n> bf9aa490db24b2334b3595ee33653bf2fe39208c\".\n> \n> The changes in 0002 for cfopen_write seem insufficient:\n> |+ if (compressionMethod == COMPRESSION_NONE)\n> |+ fp = cfopen(path, mode, compressionMethod, 0);\n> | else\n> | {\n> | #ifdef HAVE_LIBZ\n> | char *fname;\n> | \n> | fname = psprintf(\"%s.gz\", path);\n> |- fp = cfopen(fname, mode, compression);\n> |+ fp = cfopen(fname, mode, compressionMethod, compressionLevel);\n> | free_keep_errno(fname);\n> | #else\n> \n> The only difference between the LIBZ and uncompressed case is the file\n> extension, and it'll be the only difference with LZ4 too. So I suggest to\n> first handle the file extension, and the rest of the code path is not\n> conditional on the compression method. I don't think cfopen_write even needs\n> HAVE_LIBZ - can't you handle that in cfopen_internal() ?\n> \n> This patch rejects -Z0, which ought to be accepted:\n> ./src/bin/pg_dump/pg_dump -h /tmp regression -Fc -Z0 |wc\n> pg_dump: error: can only specify -Z/--compress [LEVEL] when method is set\n> \n> Your 0003 patch shouldn't reference LZ4:\n> +#ifndef HAVE_LIBLZ4\n> + if (*compressionMethod == COMPRESSION_LZ4)\n> + supports_compression = false;\n> +#endif\n> \n> The 0004 patch renames zlibOutSize to outsize - I think the patch series should\n> be constructed such as to minimize the size of the method-specific patches. I\n> say this anticipating also adding support for zstd. The preliminary patches\n> should have all the boring stuff. It would help for reviewing to keep the \n> patches split up, or to enumerate all the boring things that are being renamed\n> (like change OutputContext to cfp, rename zlibOutSize, ...).\n> \n> 0004: The include should use <lz4.h> and not \"lz4.h\"\n> \n> freebsd/cfbot is failing.\n> \n> I suggested off-list to add an 0099 patch to change LZ4 to the default, to\n> exercise it more on CI.\n\nOn Sat, Mar 26, 2022 at 01:33:36PM -0500, Justin Pryzby wrote:\n> On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\n> > You're passing both the compression method *and* level. I think there should\n> > be a structure which includes both. In the future, that can also handle\n> > additional options.\n> \n> I'm not sure if there's anything worth saving, but I did that last year with\n> 0003-Support-multiple-compression-algs-levels-opts.patch\n> I sent a rebased copy off-list.\n> https://www.postgresql.org/message-id/flat/20210104025321.GA9712@telsasoft.com#ca1b9f9d3552c87fa874731cad9d8391\n> \n> |\tfatal(\"not built with LZ4 support\");\n> |\tfatal(\"not built with lz4 support\");\n> \n> Please use consistent capitalization of \"lz4\" - then the compiler can optimize\n> away duplicate strings.\n> \n> > 0004: The include should use <lz4.h> and not \"lz4.h\"\n> \n> Also, use USE_LZ4 rather than HAVE_LIBLZ4, per 75eae0908.\n\n\n","msg_date":"Sun, 26 Jun 2022 10:55:32 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Sunday, June 26th, 2022 at 5:55 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n>\n>\n> Hi,\n>\n> Will you be able to send a rebased patch for the next CF ?\n\nThank you for taking an interest in the PR. The plan is indeed to sent\na new version.\n\n> If you update for the review comments I sent in March, I'll plan to do another\n> round of review.\n\nThank you.\n\n>\n> On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\n>\n> > LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\n> >\n> > I ran into that on an ubuntu LTS, so I don't think it's so old that it\n> > shouldn't be handled more gracefully. LZ4 should either have an explicit\n> > version check, or else shouldn't depend on that feature (or should define a\n> > safe fallback version if the library header doesn't define it).\n> >\n> > https://packages.ubuntu.com/liblz4-1\n> >\n> > 0003: typo: of legacy => or legacy\n> >\n> > There are a large number of ifdefs being added here - it'd be nice to minimize\n> > that. basebackup was organized to use separate files, which is one way.\n> >\n> > $ git grep -c 'ifdef .*LZ4' src/bin/pg_dump/compress_io.c\n> > src/bin/pg_dump/compress_io.c:19\n> >\n> > In last year's CF entry, I had made a union within CompressorState. LZ4\n> > doesn't need z_streamp (and ztsd will need ZSTD_outBuffer, ZSTD_inBuffer,\n> > ZSTD_CStream).\n> >\n> > 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\n> > from commit ffd53659c. You're passing both the compression method and level.\n> > I think there should be a structure which includes both. In the future, that\n> > can also handle additional options. I hope to re-use these same things for\n> > wal_compression=method:level.\n> >\n> > You renamed this:\n> >\n> > |- COMPR_ALG_LIBZ\n> > |-} CompressionAlgorithm;\n> > |+ COMPRESSION_GZIP,\n> > |+} CompressionMethod;\n> >\n> > ..But I don't think that's an improvement. If you were to change it, it should\n> > say something like PGDUMP_COMPRESS_ZLIB, since there are other compression\n> > structs and typedefs. zlib is not idential to gzip, which uses a different\n> > header, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\n> >\n> > The cf* changes in pg_backup_archiver could be split out into a separate\n> > commit. It's strictly a code simplification - not just preparation for more\n> > compression algorithms. The commit message should \"See also:\n> > bf9aa490db24b2334b3595ee33653bf2fe39208c\".\n> >\n> > The changes in 0002 for cfopen_write seem insufficient:\n> > |+ if (compressionMethod == COMPRESSION_NONE)\n> > |+ fp = cfopen(path, mode, compressionMethod, 0);\n> > | else\n> > | {\n> > | #ifdef HAVE_LIBZ\n> > | char *fname;\n> > |\n> > | fname = psprintf(\"%s.gz\", path);\n> > |- fp = cfopen(fname, mode, compression);\n> > |+ fp = cfopen(fname, mode, compressionMethod, compressionLevel);\n> > | free_keep_errno(fname);\n> > | #else\n> >\n> > The only difference between the LIBZ and uncompressed case is the file\n> > extension, and it'll be the only difference with LZ4 too. So I suggest to\n> > first handle the file extension, and the rest of the code path is not\n> > conditional on the compression method. I don't think cfopen_write even needs\n> > HAVE_LIBZ - can't you handle that in cfopen_internal() ?\n> >\n> > This patch rejects -Z0, which ought to be accepted:\n> > ./src/bin/pg_dump/pg_dump -h /tmp regression -Fc -Z0 |wc\n> > pg_dump: error: can only specify -Z/--compress [LEVEL] when method is set\n> >\n> > Your 0003 patch shouldn't reference LZ4:\n> > +#ifndef HAVE_LIBLZ4\n> > + if (*compressionMethod == COMPRESSION_LZ4)\n> > + supports_compression = false;\n> > +#endif\n> >\n> > The 0004 patch renames zlibOutSize to outsize - I think the patch series should\n> > be constructed such as to minimize the size of the method-specific patches. I\n> > say this anticipating also adding support for zstd. The preliminary patches\n> > should have all the boring stuff. It would help for reviewing to keep the\n> > patches split up, or to enumerate all the boring things that are being renamed\n> > (like change OutputContext to cfp, rename zlibOutSize, ...).\n> >\n> > 0004: The include should use <lz4.h> and not \"lz4.h\"\n> >\n> > freebsd/cfbot is failing.\n> >\n> > I suggested off-list to add an 0099 patch to change LZ4 to the default, to\n> > exercise it more on CI.\n>\n>\n> On Sat, Mar 26, 2022 at 01:33:36PM -0500, Justin Pryzby wrote:\n>\n> > On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\n> >\n> > > You're passing both the compression method and level. I think there should\n> > > be a structure which includes both. In the future, that can also handle\n> > > additional options.\n> >\n> > I'm not sure if there's anything worth saving, but I did that last year with\n> > 0003-Support-multiple-compression-algs-levels-opts.patch\n> > I sent a rebased copy off-list.\n> > https://www.postgresql.org/message-id/flat/20210104025321.GA9712@telsasoft.com#ca1b9f9d3552c87fa874731cad9d8391\n> >\n> > | fatal(\"not built with LZ4 support\");\n> > | fatal(\"not built with lz4 support\");\n> >\n> > Please use consistent capitalization of \"lz4\" - then the compiler can optimize\n> > away duplicate strings.\n> >\n> > > 0004: The include should use <lz4.h> and not \"lz4.h\"\n> >\n> > Also, use USE_LZ4 rather than HAVE_LIBLZ4, per 75eae0908.\n>\n>\n\n\n","msg_date":"Tue, 28 Jun 2022 08:14:25 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Sunday, June 26th, 2022 at 5:55 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n>\n> Hi,\n>\n> Will you be able to send a rebased patch for the next CF ?\n\nPlease find a rebased and heavily refactored patchset. Since parts of this\npatchset were already committed, I restarted numbering. I am not certain if\nthis is the preferred way. This makes alignment with previous comments a bit\nharder\n\n> If you update for the review comments I sent in March, I'll plan to do another\n> round of review.\n\nI have updated for \"some\" of the comments. This is not an unwillingness to\nincorporate those specific comments. Simply this patchset had started to divert\nheavily already based on comments from Mr. Paquier who had already requested for\nthe APIs to be refactored to use function pointers. This is happening in 0002 of\nthe patchset. 0001 of the patchset is using the new compression.h under common.\n\nThis patchset should be considered a late draft, as commentary, documentation,\nand some finer details are not yet finalized; because I am expecting the proposed\nrefactor to receive a wealth of comments. It would be helpful to understand if\nthe proposed direction is something worth to be worked upon, before moving to the\nfiner details.\n\nFor what is worth, I am the sole author of the current patchset.\n\nCheers,\n//Georgios\n\n\n> On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\n>\n> > LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\n> >\n> > I ran into that on an ubuntu LTS, so I don't think it's so old that it\n> > shouldn't be handled more gracefully. LZ4 should either have an explicit\n> > version check, or else shouldn't depend on that feature (or should define a\n> > safe fallback version if the library header doesn't define it).\n> >\n> > https://packages.ubuntu.com/liblz4-1\n> >\n> > 0003: typo: of legacy => or legacy\n> >\n> > There are a large number of ifdefs being added here - it'd be nice to minimize\n> > that. basebackup was organized to use separate files, which is one way.\n> >\n> > $ git grep -c 'ifdef .*LZ4' src/bin/pg_dump/compress_io.c\n> > src/bin/pg_dump/compress_io.c:19\n> >\n> > In last year's CF entry, I had made a union within CompressorState. LZ4\n> > doesn't need z_streamp (and ztsd will need ZSTD_outBuffer, ZSTD_inBuffer,\n> > ZSTD_CStream).\n> >\n> > 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\n> > from commit ffd53659c. You're passing both the compression method and level.\n> > I think there should be a structure which includes both. In the future, that\n> > can also handle additional options. I hope to re-use these same things for\n> > wal_compression=method:level.\n> >\n> > You renamed this:\n> >\n> > |- COMPR_ALG_LIBZ\n> > |-} CompressionAlgorithm;\n> > |+ COMPRESSION_GZIP,\n> > |+} CompressionMethod;\n> >\n> > ..But I don't think that's an improvement. If you were to change it, it should\n> > say something like PGDUMP_COMPRESS_ZLIB, since there are other compression\n> > structs and typedefs. zlib is not idential to gzip, which uses a different\n> > header, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\n> >\n> > The cf* changes in pg_backup_archiver could be split out into a separate\n> > commit. It's strictly a code simplification - not just preparation for more\n> > compression algorithms. The commit message should \"See also:\n> > bf9aa490db24b2334b3595ee33653bf2fe39208c\".\n> >\n> > The changes in 0002 for cfopen_write seem insufficient:\n> > |+ if (compressionMethod == COMPRESSION_NONE)\n> > |+ fp = cfopen(path, mode, compressionMethod, 0);\n> > | else\n> > | {\n> > | #ifdef HAVE_LIBZ\n> > | char *fname;\n> > |\n> > | fname = psprintf(\"%s.gz\", path);\n> > |- fp = cfopen(fname, mode, compression);\n> > |+ fp = cfopen(fname, mode, compressionMethod, compressionLevel);\n> > | free_keep_errno(fname);\n> > | #else\n> >\n> > The only difference between the LIBZ and uncompressed case is the file\n> > extension, and it'll be the only difference with LZ4 too. So I suggest to\n> > first handle the file extension, and the rest of the code path is not\n> > conditional on the compression method. I don't think cfopen_write even needs\n> > HAVE_LIBZ - can't you handle that in cfopen_internal() ?\n> >\n> > This patch rejects -Z0, which ought to be accepted:\n> > ./src/bin/pg_dump/pg_dump -h /tmp regression -Fc -Z0 |wc\n> > pg_dump: error: can only specify -Z/--compress [LEVEL] when method is set\n> >\n> > Your 0003 patch shouldn't reference LZ4:\n> > +#ifndef HAVE_LIBLZ4\n> > + if (*compressionMethod == COMPRESSION_LZ4)\n> > + supports_compression = false;\n> > +#endif\n> >\n> > The 0004 patch renames zlibOutSize to outsize - I think the patch series should\n> > be constructed such as to minimize the size of the method-specific patches. I\n> > say this anticipating also adding support for zstd. The preliminary patches\n> > should have all the boring stuff. It would help for reviewing to keep the\n> > patches split up, or to enumerate all the boring things that are being renamed\n> > (like change OutputContext to cfp, rename zlibOutSize, ...).\n> >\n> > 0004: The include should use <lz4.h> and not \"lz4.h\"\n> >\n> > freebsd/cfbot is failing.\n> >\n> > I suggested off-list to add an 0099 patch to change LZ4 to the default, to\n> > exercise it more on CI.\n>\n>\n> On Sat, Mar 26, 2022 at 01:33:36PM -0500, Justin Pryzby wrote:\n>\n> > On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\n> >\n> > > You're passing both the compression method and level. I think there should\n> > > be a structure which includes both. In the future, that can also handle\n> > > additional options.\n> >\n> > I'm not sure if there's anything worth saving, but I did that last year with\n> > 0003-Support-multiple-compression-algs-levels-opts.patch\n> > I sent a rebased copy off-list.\n> > https://www.postgresql.org/message-id/flat/20210104025321.GA9712@telsasoft.com#ca1b9f9d3552c87fa874731cad9d8391\n> >\n> > | fatal(\"not built with LZ4 support\");\n> > | fatal(\"not built with lz4 support\");\n> >\n> > Please use consistent capitalization of \"lz4\" - then the compiler can optimize\n> > away duplicate strings.\n> >\n> > > 0004: The include should use <lz4.h> and not \"lz4.h\"\n> >\n> > Also, use USE_LZ4 rather than HAVE_LIBLZ4, per 75eae0908.\n>\n>","msg_date":"Tue, 05 Jul 2022 13:22:47 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"This is a review of 0001.\n\nOn Tue, Jul 05, 2022 at 01:22:47PM +0000, gkokolatos@pm.me wrote:\n> Simply this patchset had started to divert\n> heavily already based on comments from Mr. Paquier who had already requested for\n> the APIs to be refactored to use function pointers. This is happening in 0002 of\n> the patchset.\n\nI said something about reducing ifdefs, but I'm having trouble finding what\nMichael said about this ?\n\n> > On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\n> >\n> > > LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\n> > >\n> > > I ran into that on an ubuntu LTS, so I don't think it's so old that it\n> > > shouldn't be handled more gracefully. LZ4 should either have an explicit\n> > > version check, or else shouldn't depend on that feature (or should define a\n> > > safe fallback version if the library header doesn't define it).\n> > > https://packages.ubuntu.com/liblz4-1\n\nThe constant still seems to be used without defining a fallback or a minimum version.\n\n> > > 0003: typo: of legacy => or legacy\n\nThis is still there\n\n> > > You renamed this:\n> > >\n> > > |- COMPR_ALG_LIBZ\n> > > |-} CompressionAlgorithm;\n> > > |+ COMPRESSION_GZIP,\n> > > |+} CompressionMethod;\n> > >\n> > > ..But I don't think that's an improvement. If you were to change it, it should\n> > > say something like PGDUMP_COMPRESS_ZLIB, since there are other compression\n> > > structs and typedefs. zlib is not idential to gzip, which uses a different\n> > > header, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\n\nThis comment still applies - zlib's gz* functions are \"gzip\" but the others are\n\"zlib\". https://zlib.net/manual.html\n\nThat affects both the 0001 and 0002 patches.\n\nActually, I think that \"gzip\" should not be the name of the user-facing option,\nsince (except for \"plain\" format) it isn't using gzip.\n\n+Robert, since this suggests amending parse_compress_algorithm(). Maybe \"zlib\"\nshould be parsed the same way as \"gzip\" - I don't think we ever expose both to\na user, but in some cases (basebackup and pg_dump -Fp -Z1) the output is \"gzip\"\nand in some cases NO it's zlib (pg_dump -Fc -Z1).\n\n> > > The cf* changes in pg_backup_archiver could be split out into a separate\n> > > commit. It's strictly a code simplification - not just preparation for more\n> > > compression algorithms. The commit message should \"See also:\n> > > bf9aa490db24b2334b3595ee33653bf2fe39208c\".\n\nI still think this could be an early, 0000 patch.\n\n> > > freebsd/cfbot is failing.\n\nThis is still failing for bsd, windows and compiler warnings.\nWindows also has compiler warnings.\nhttp://cfbot.cputube.org/georgios-kokolatos.html\n\nPlease see: src/tools/ci/README, which you can use to run check-world on 4 OS\nby pushing a branch to github.\n\n> > > I suggested off-list to add an 0099 patch to change LZ4 to the default, to\n> > > exercise it more on CI.\n\nWhat about this ? I think the patch needs to pass CI on all 4 OS with\ndefault=zlib and default=lz4.\n\n> > On Sat, Mar 26, 2022 at 01:33:36PM -0500, Justin Pryzby wrote:\n\n> @@ -254,7 +251,12 @@ CreateArchive(const char *FileSpec, const ArchiveFormat fmt,\n> Archive *\n> OpenArchive(const char *FileSpec, const ArchiveFormat fmt)\n> {\n> -\tArchiveHandle *AH = _allocAH(FileSpec, fmt, 0, true, archModeRead, setupRestoreWorker);\n> +\tArchiveHandle *AH;\n> +\tpg_compress_specification compress_spec;\n\nShould this be initialized to {0} ?\n\n> @@ -969,6 +969,8 @@ NewRestoreOptions(void)\n> \topts->format = archUnknown;\n> \topts->cparams.promptPassword = TRI_DEFAULT;\n> \topts->dumpSections = DUMP_UNSECTIONED;\n> +\topts->compress_spec.algorithm = PG_COMPRESSION_NONE;\n> +\topts->compress_spec.level = INT_MIN;\n\nWhy INT_MIN ?\n\n> @@ -1115,23 +1117,28 @@ PrintTOCSummary(Archive *AHX)\n> \tArchiveHandle *AH = (ArchiveHandle *) AHX;\n> \tRestoreOptions *ropt = AH->public.ropt;\n> \tTocEntry *te;\n> +\tpg_compress_specification out_compress_spec;\n\nShould have {0} ?\nI suggest to write it like my 2020 patch for this, which says:\nno_compression = {0};\n\n> \t/* Open stdout with no compression for AH output handle */\n> -\tAH->gzOut = 0;\n> -\tAH->OF = stdout;\n> +\tout_compress_spec.algorithm = PG_COMPRESSION_NONE;\n> +\tAH->OF = cfdopen(dup(fileno(stdout)), PG_BINARY_A, out_compress_spec);\n\nIdeally this should check the success of dup().\n\n> @@ -3776,21 +3746,25 @@ ReadHead(ArchiveHandle *AH)\n> +\tif (AH->compress_spec.level != INT_MIN)\n\nWhy is it testing the level and not the algorithm ?\n\n> --- a/src/bin/pg_dump/pg_backup_custom.c\n> +++ b/src/bin/pg_dump/pg_backup_custom.c\n> @@ -298,7 +298,7 @@ _StartData(ArchiveHandle *AH, TocEntry *te)\n> \t_WriteByte(AH, BLK_DATA);\t/* Block type */\n> \tWriteInt(AH, te->dumpId);\t/* For sanity check */\n> \n> -\tctx->cs = AllocateCompressor(AH->compression, _CustomWriteFunc);\n> +\tctx->cs = AllocateCompressor(AH->compress_spec, _CustomWriteFunc);\n\nIs it necessary to rename the data structure ?\nIf not, this file can remain unchanged.\n\n> --- a/src/bin/pg_dump/pg_backup_directory.c\n> +++ b/src/bin/pg_dump/pg_backup_directory.c\n> @@ -573,6 +574,7 @@ _CloseArchive(ArchiveHandle *AH)\n> \tif (AH->mode == archModeWrite)\n> \t{\n> \t\tcfp\t\t *tocFH;\n> +\t\tpg_compress_specification compress_spec;\n\nShould use {0} ?\n\n> @@ -639,12 +642,14 @@ static void\n> _StartBlobs(ArchiveHandle *AH, TocEntry *te)\n> {\n> \tlclContext *ctx = (lclContext *) AH->formatData;\n> +\tpg_compress_specification compress_spec;\n\nSame\n\n> +\t/*\n> +\t * Custom and directory formats are compressed by default (zlib), others\n> +\t * not\n> +\t */\n> +\tif (user_compression_defined == false)\n\nShould be: !user_compression_defined\n\nYour 0001+0002 patches (without 0003) fail to compile:\n\npg_backup_directory.c: In function ‘_ReadByte’:\npg_backup_directory.c:519:12: error: ‘CompressFileHandle’ {aka ‘struct CompressFileHandle’} has no member named ‘_IO_getc’\n 519 | return CFH->getc(CFH);\n | ^~\npg_backup_directory.c:520:1: warning: control reaches end of non-void function [-Wreturn-type]\n 520 | }\n\n-- \nJustin\n\n\n","msg_date":"Tue, 5 Jul 2022 10:13:28 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Jul 05, 2022 at 01:22:47PM +0000, gkokolatos@pm.me wrote:\n> I have updated for \"some\" of the comments. This is not an unwillingness to\n> incorporate those specific comments. Simply this patchset had started to divert\n> heavily already based on comments from Mr. Paquier who had already requested for\n> the APIs to be refactored to use function pointers. This is happening in 0002 of\n> the patchset. 0001 of the patchset is using the new compression.h under common.\n> \n> This patchset should be considered a late draft, as commentary, documentation,\n> and some finer details are not yet finalized; because I am expecting the proposed\n> refactor to receive a wealth of comments. It would be helpful to understand if\n> the proposed direction is something worth to be worked upon, before moving to the\n> finer details.\n\nI have read through the patch set, and I like a lot the separation you\nare doing here with CompressFileHandle where a compression method has\nto specify a full set of callbacks depending on the actions that need\nto be taken. One advantage, as you patch shows, is that you reduce\nthe dependency of each code path depending on the compression method,\nwith #ifdefs and such located mostly into their own file structure, so\nas adding a new compression method becomes really easier. These\ncallbacks are going to require much more documentation to describe\nwhat anybody using them should expect from them, and perhaps they\ncould be renamed in a more generic way as the currect names come from\nPOSIX (say read_char(), read_string()?), even if this patch has just\ninherited the names coming from pg_dump itself, but this can be tuned\nover and over.\n\nThe split into three parts as of 0001 to plug into pg_dump the new\ncompression option set, 0002 to introduce the callbacks and 0003 to\nadd LZ4, building on the two first parts, makes sense to me. 0001 and\n0002 could be done in a reversed order as they are mostly independent,\nthis order is fine as-is.\n\nIn short, I am fine with the proposed approach.\n\n+#define K_VERS_1_15 MAKE_ARCHIVE_VERSION(1, 15, 0) /* add compressionMethod\n+ * in header */\nIndeed, the dump format needs a version bump for this information.\n\n+static bool\n+parse_compression_option(const char *opt,\n+ pg_compress_specification *compress_spec)\nThis parsing logic in pg_dump.c looks a lot like what pg_receivewal.c\ndoes with its parse_compress_options() where, for compatibility:\n- If only a number is given:\n-- Assume no compression if level is 0.\n-- Assume gzip with given compression if level > 0.\n- If a string is found, assume a full spec, with optionally a level.\nSo some consolidation could be done between both.\n\nBy the way, I can see that GZCLOSE(), etc. are still defined in\ncompress_io.h but they are not used.\n--\nMichael","msg_date":"Thu, 21 Jul 2022 14:04:18 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"This entry has been waiting on author input for a while (our current\nthreshold is roughly two weeks), so I've marked it Returned with\nFeedback.\n\nOnce you think the patchset is ready for review again, you (or any\ninterested party) can resurrect the patch entry by visiting\n\n https://commitfest.postgresql.org/38/3571/\n\nand changing the status to \"Needs Review\", and then changing the\nstatus again to \"Move to next CF\". (Don't forget the second step;\nhopefully we will have streamlined this in the near future!)\n\nThanks,\n--Jacob\n\n\n","msg_date":"Tue, 2 Aug 2022 11:42:03 -0700","msg_from":"Jacob Champion <jchampion@timescale.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Thank you for your work during commitfest.\r\n\r\nThe patch is still in development. Given vacation status, expect the next patches to be ready for the November commitfest.\r\nFor now it has moved to the September one. Further action will be taken then as needed.\r\n\r\nEnjoy the rest of the summer!","msg_date":"Fri, 05 Aug 2022 14:23:45 +0000","msg_from":"Georgios Kokolatos <gkokolatos@protonmail.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Checking if you'll be able to submit new patches soon ?\n\n\n","msg_date":"Wed, 2 Nov 2022 08:28:03 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Nov 2, 2022 at 14:28, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n> Checking if you'll be able to submit new patches soon ?\n\nThank you for checking up. Expect new versions within this commitfest cycle.\n On Wed, Nov 2, 2022 at 14:28, Justin Pryzby <pryzby@telsasoft.com> wrote: Checking if you'll be able to submit new patches soon ?Thank you for checking up. Expect new versions within this commitfest cycle.","msg_date":"Sun, 06 Nov 2022 14:53:12 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Aug 05, 2022 at 02:23:45PM +0000, Georgios Kokolatos wrote:\n> Thank you for your work during commitfest.\n> \n> The patch is still in development. Given vacation status, expect the next patches to be ready for the November commitfest.\n> For now it has moved to the September one. Further action will be taken then as needed.\n\nOn Sun, Nov 06, 2022 at 02:53:12PM +0000, gkokolatos@pm.me wrote:\n> On Wed, Nov 2, 2022 at 14:28, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > Checking if you'll be able to submit new patches soon ?\n> \n> Thank you for checking up. Expect new versions within this commitfest cycle.\n\nHi,\n\nI think this patch record should be closed for now. You can re-open the\nexisting patch record once a patch is ready to be reviewed.\n\nThe commitfest is a time for committing/reviewing patches that were\npreviously submitted, but there's no new patch since July. Making a\npatch available for review at the start of the commitfest seems like a\nrequirement for current patch records (same as for new patch records).\n\nI wrote essentially the same patch as your early patches 2 years ago\n(before postgres was ready to consider new compression algorithms), so\nI'm happy to review a new patch when it's available, regardless of its\nstatus in the cfapp.\n\nBTW, some of my own review comments from March weren't addressed.\nPlease check. Also, in February, I asked if you knew how to use\ncirrusci to run checks on cirrusci, but the patches still had\ncompilation errors and warnings on various OS.\n\nhttps://cirrus-ci.com/github/postgresql-cfbot/postgresql/commitfest/39/3571\n\n-- \nJustin\n\n\n","msg_date":"Sun, 20 Nov 2022 11:26:11 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sun, Nov 20, 2022 at 11:26:11AM -0600, Justin Pryzby wrote:\n> I think this patch record should be closed for now. You can re-open the\n> existing patch record once a patch is ready to be reviewed.\n\nIndeed. As of things are, this is just a dead entry in the CF which\nwould be confusing. I have marked it as RwF.\n--\nMichael","msg_date":"Mon, 21 Nov 2022 08:13:16 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Monday, November 21st, 2022 at 12:13 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Sun, Nov 20, 2022 at 11:26:11AM -0600, Justin Pryzby wrote:\n> \n> > I think this patch record should be closed for now. You can re-open the\n> > existing patch record once a patch is ready to be reviewed.\n> \n> \n> Indeed. As of things are, this is just a dead entry in the CF which\n> would be confusing. I have marked it as RwF.\n\nThank you for closing it.\n\nFor the record I am currently working on it simply unsure if I should submit\nWIP patches and add noise to the list or wait until it is in a state that I\nfeel that the comments have been addressed.\n\nA new version that I feel that is in a decent enough state for review should\nbe ready within this week. I am happy to drop the patch if you think I should\nnot work on it though.\n\nCheers,\n//Georgios\n\n> --\n> Michael\n\n\n","msg_date":"Tue, 22 Nov 2022 10:00:47 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Nov 22, 2022 at 10:00:47AM +0000, gkokolatos@pm.me wrote:\n> A new version that I feel that is in a decent enough state for review should\n> be ready within this week. I am happy to drop the patch if you think I should\n> not work on it though.\n\nIf you can post a new version of the patch, that's fine, of course.\nI'll be happy to look over it more.\n--\nMichael","msg_date":"Tue, 22 Nov 2022 19:49:13 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Nov 22, 2022 at 10:00:47AM +0000, gkokolatos@pm.me wrote:\n> For the record I am currently working on it simply unsure if I should submit\n> WIP patches and add noise to the list or wait until it is in a state that I\n> feel that the comments have been addressed.\n> \n> A new version that I feel that is in a decent enough state for review should\n> be ready within this week. I am happy to drop the patch if you think I should\n> not work on it though.\n\nI hope you'll want to continue work on it. The patch record is like a\nrequest for review, so it's closed if there's nothing ready to review.\n\nI think you should re-send patches (and update the CF app) as often as\nthey're ready for more review. Your 001 commit (which is almost the\nsame as what I wrote 2 years ago) still needs to account for some review\ncomments, and the whole patch set ought to pass cirrusci tests. At that\npoint, you'll be ready for another round of review, even if there's\nknown TODO/FIXME items in later patches.\n\nBTW I saw that you updated your branch on github. You'll need to make\nthe corresponding changes to ./meson.build that you made to ./Makefile.\nhttps://wiki.postgresql.org/wiki/Meson_for_patch_authors\nhttps://wiki.postgresql.org/wiki/Meson \n\n-- \nJustin\n\n\n","msg_date":"Wed, 23 Nov 2022 14:52:22 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Tuesday, November 22nd, 2022 at 11:49 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Tue, Nov 22, 2022 at 10:00:47AM +0000, gkokolatos@pm.me wrote:\n> \n> > A new version that I feel that is in a decent enough state for review should\n> > be ready within this week. I am happy to drop the patch if you think I should\n> > not work on it though.\n> \n> \n> If you can post a new version of the patch, that's fine, of course.\n> I'll be happy to look over it more.\n\nThank you Michael (and Justin). Allow me to present v8.\n\nThe focus of this version of this series is 0001 and 0002.\n\nAdmittedly 0001 could be presented in a separate thread though given its size and\nproximity to the topic, I present it here.\n\nIn an earlier review you spotted the similarity between pg_dump's and pg_receivewal's\nparsing of compression options. However there exists a substantial difference in the\nbehaviour of the two programs; one treats the lack of support for the requested\nalgorithm as a fatal error, whereas the other does not. The existing functions in\ncommon/compression.c do not account for the later. 0002 proposes an implementation\nfor this. It's usefulness is shown in 0003.\n\nPlease consider 0003-0005 as work in progress. They are differences from v7 yet they\nmay contain unaddressed comments for now.\n\nA welcome feedback would be in splitting and/or reordering of 0003-0005. I think\nthat they now split in coherent units and are presented in a logical order. Let me\nknow if you disagree and where should the breakpoints be.\n\nCheers,\n//Georgios\n\n> --\n> Michael","msg_date":"Mon, 28 Nov 2022 16:32:43 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Nov 28, 2022 at 04:32:43PM +0000, gkokolatos@pm.me wrote:\n> The focus of this version of this series is 0001 and 0002.\n> \n> Admittedly 0001 could be presented in a separate thread though given its size and\n> proximity to the topic, I present it here.\n\nI don't mind. This was a hole in meson.build, so nice catch! I have\nnoticed a second defect with pg_verifybackup for all the commands, and\napplied both at the same time.\n\n> In an earlier review you spotted the similarity between pg_dump's and pg_receivewal's\n> parsing of compression options. However there exists a substantial difference in the\n> behaviour of the two programs; one treats the lack of support for the requested\n> algorithm as a fatal error, whereas the other does not. The existing functions in\n> common/compression.c do not account for the later. 0002 proposes an implementation\n> for this. It's usefulness is shown in 0003.\n\nIn what does it matter? The logic in compression.c provides an error\nwhen looking at a spec or validating it, but the caller is free to\nconsume it as it wants because this is shared between the frontend and\nthe backend, and that includes consuming it as a warning rather than a\nahrd failure. If we don't want to issue an error and force\nnon-compression if attempting to use a compression method not\nsupported in pg_dump, that's fine by me as a historical behavior, but\nI don't see why these routines have any need to be split more as\nproposed in 0002.\n\nSaying that, I do agree that it would be nice to remove the\nduplication between the option parsing of pg_basebackup and\npg_receivewal. Your patch is very close to that, actually, and it\noccured to me that if we move the check on \"server-\" and \"client-\" in\npg_basebackup to be just before the integer-only check then we can\nconsolidate the whole thing.\n\nAttached is an alternative that does not sacrifice the pluggability of\nthe existing routines while allowing 0003~ to still use them (I don't\nreally want to move around the checks on the supported build options\nnow in parse_compress_specification(), that was hard enough to settle\non this location). On top of that, pg_basebackup is able to cope with\nthe case of --compress=0 already, enforcing \"none\" (BaseBackup could\nbe simplified a bit more before StartLogStreamer). This refactoring\nshaves a little bit of code.\n\n> Please consider 0003-0005 as work in progress. They are differences from v7 yet they\n> may contain unaddressed comments for now.\n\nOkay.\n--\nMichael","msg_date":"Tue, 29 Nov 2022 15:19:17 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Nov 29, 2022 at 03:19:17PM +0900, Michael Paquier wrote:\n> Attached is an alternative that does not sacrifice the pluggability of\n> the existing routines while allowing 0003~ to still use them (I don't\n> really want to move around the checks on the supported build options\n> now in parse_compress_specification(), that was hard enough to settle\n> on this location). On top of that, pg_basebackup is able to cope with\n> the case of --compress=0 already, enforcing \"none\" (BaseBackup could\n> be simplified a bit more before StartLogStreamer). This refactoring\n> shaves a little bit of code.\n\nOne thing that I forgot to mention is that this refactoring would\ntreat things like server-N, client-N as valid grammars (in this case N\ntakes precedence over an optional detail string), implying that N = 0\nis \"none\" and N > 0 is gzip, so that makes for an extra grammar flavor\nwithout impacting the existing ones. I am not sure that it is worth\ndocumenting, still worth mentioning.\n--\nMichael","msg_date":"Tue, 29 Nov 2022 16:08:27 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Tuesday, November 29th, 2022 at 7:19 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Mon, Nov 28, 2022 at 04:32:43PM +0000, gkokolatos@pm.me wrote:\n> \n> > The focus of this version of this series is 0001 and 0002.\n> > \n> > Admittedly 0001 could be presented in a separate thread though given its size and\n> > proximity to the topic, I present it here.\n> \n> \n> I don't mind. This was a hole in meson.build, so nice catch! I have\n> noticed a second defect with pg_verifybackup for all the commands, and\n> applied both at the same time.\n\nThank you.\n\n> \n> > In an earlier review you spotted the similarity between pg_dump's and pg_receivewal's\n> > parsing of compression options. However there exists a substantial difference in the\n> > behaviour of the two programs; one treats the lack of support for the requested\n> > algorithm as a fatal error, whereas the other does not. The existing functions in\n> > common/compression.c do not account for the later. 0002 proposes an implementation\n> > for this. It's usefulness is shown in 0003.\n> \n> \n> In what does it matter? The logic in compression.c provides an error\n> when looking at a spec or validating it, but the caller is free to\n> consume it as it wants because this is shared between the frontend and\n> the backend, and that includes consuming it as a warning rather than a\n> ahrd failure. If we don't want to issue an error and force\n> non-compression if attempting to use a compression method not\n> supported in pg_dump, that's fine by me as a historical behavior, but\n> I don't see why these routines have any need to be split more as\n> proposed in 0002.\n\nI understand. The reason for the change in the routines was because it was\nimpossible to distinguish a genuine parse error from a missing library in\nparse_compress_specification(). If the zlib library is missing, then both\n'--compress=gzip:garbage' and '--compress=gzip:7' would populate the\nparse_error member of the struct and subsequent calls to\nvalidate_compress_specification() would error out, although only one of\nthe two options is truly an error. Historically the code would fail on\ninvalid input regardless of whether the library was present or not.\n\n> Saying that, I do agree that it would be nice to remove the\n> duplication between the option parsing of pg_basebackup and\n> pg_receivewal. Your patch is very close to that, actually, and it\n> occured to me that if we move the check on \"server-\" and \"client-\" in\n> pg_basebackup to be just before the integer-only check then we can\n> consolidate the whole thing.\n\nGreat. I did notice the possible benefit but chose to not tread too far\noff the necessary in my patch.\n \n> Attached is an alternative that does not sacrifice the pluggability of\n> the existing routines while allowing 0003~ to still use them (I don't\n> really want to move around the checks on the supported build options\n> now in parse_compress_specification(), that was hard enough to settle\n> on this location).\n\nYeah, I thought that it would be a hard sell, hence an \"earlier\"\nversion.\n\nThe attached version 10, contains verbatim your proposed v9 as 0001.\nThen 0002 is switching a bit the parsing order in pg_dump and will not\nfail as described above on missing libraries. Now, it will first parse\nthe algorithm, discard it when unsupported, and only parse the rest of\nthe option if the algorithm is supported. Granted it is a bit 'uglier'\nwith the preprocessing blocks, yet it maintains most of the historic\nbehaviour without altering the common compression interfaces. Now, as\nshown in 001_basic.pl, invalid detail will fail only if the algorithm\nis supported.\n\n> On top of that, pg_basebackup is able to cope with\n> the case of --compress=0 already, enforcing \"none\" (BaseBackup could\n> be simplified a bit more before StartLogStreamer). This refactoring\n> shaves a little bit of code.\n> \n> > Please consider 0003-0005 as work in progress. They are differences from v7 yet they\n> > may contain unaddressed comments for now.\n> \n> \n> Okay.\n\nThank you. Please advice if is preferable to split 0002 in two parts.\nI think not but I will happily do so if you think otherwise.\n\nCheers,\n//Georgios\n\n> --\n> Michael","msg_date":"Tue, 29 Nov 2022 12:10:46 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Nov 29, 2022 at 12:10:46PM +0000, gkokolatos@pm.me wrote:\n> Thank you. Please advice if is preferable to split 0002 in two parts.\n> I think not but I will happily do so if you think otherwise.\n\nThis one makes me curious. What kind of split are you talking about?\nIf it makes the code review and the git history cleaner and easier, I\nam usually a lot in favor of such incremental changes. As far as I\ncan see, there is the switch from the compression integer to\ncompression specification as one thing. The second thing is the\nrefactoring of cfclose() and these routines, paving the way for 0003.\nHmm, it may be cleaner to move the switch to the compression spec in\none patch, and move the logic around cfclose() to its own, paving the\nway to 0003.\n\nBy the way, I think that this 0002 should drop all the default clauses\nin the switches for the compression method so as we'd catch any\nmissing code paths with compiler warnings if a new compression method\nis added in the future.\n\nAnyway, I have applied 0001, adding you as a primary author because\nyou did most of it with only tweaks from me for pg_basebackup. The\ndocs of pg_basebackup have been amended to mention the slight change\nin grammar, affecting the case where we do not have a detail string.\n--\nMichael","msg_date":"Wed, 30 Nov 2022 09:50:34 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Wednesday, November 30th, 2022 at 1:50 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Tue, Nov 29, 2022 at 12:10:46PM +0000, gkokolatos@pm.me wrote:\n> \n> > Thank you. Please advice if is preferable to split 0002 in two parts.\n> > I think not but I will happily do so if you think otherwise.\n> \n> \n> This one makes me curious. What kind of split are you talking about?\n> If it makes the code review and the git history cleaner and easier, I\n> am usually a lot in favor of such incremental changes. As far as I\n> can see, there is the switch from the compression integer to\n> compression specification as one thing. The second thing is the\n> refactoring of cfclose() and these routines, paving the way for 0003.\n> Hmm, it may be cleaner to move the switch to the compression spec in\n> one patch, and move the logic around cfclose() to its own, paving the\n> way to 0003.\n\nFair enough. The atteched v11 does that. 0001 introduces compression\nspecification and is using it throughout. 0002 paves the way to the\nnew interface by homogenizing the use of cfp. 0003 introduces the new\nAPI and stores the compression algorithm in the custom format header\ninstead of the compression level integer. Finally 0004 adds support for\nLZ4.\n\nBesides the version bump in 0003 which can possibly be split out and\nas an independent and earlier step, I think that the patchset consists\nof coherent units.\n\n> By the way, I think that this 0002 should drop all the default clauses\n> in the switches for the compression method so as we'd catch any\n> missing code paths with compiler warnings if a new compression method\n> is added in the future.\n\nSure. \n\n> Anyway, I have applied 0001, adding you as a primary author because\n> you did most of it with only tweaks from me for pg_basebackup. The\n> docs of pg_basebackup have been amended to mention the slight change\n> in grammar, affecting the case where we do not have a detail string.\n\nVery kind of you, thank you.\n\nCheers,\n//Georgios\n\n> --\n> Michael","msg_date":"Wed, 30 Nov 2022 17:11:44 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Nov 30, 2022 at 05:11:44PM +0000, gkokolatos@pm.me wrote:\n> Fair enough. The atteched v11 does that. 0001 introduces compression\n> specification and is using it throughout. 0002 paves the way to the\n> new interface by homogenizing the use of cfp. 0003 introduces the new\n> API and stores the compression algorithm in the custom format header\n> instead of the compression level integer. Finally 0004 adds support for\n> LZ4.\n\nI have been looking at 0001, and.. Hmm. I am really wondering\nwhether it would not be better to just nuke this warning into orbit.\nThis stuff enforces non-compression even if -Z has been used to a\nnon-default value. This has been moved to its current location by\ncae2bb1 as of this thread:\nhttps://www.postgresql.org/message-id/20160526.185551.242041780.horiguchi.kyotaro%40lab.ntt.co.jp\n\nHowever, this is only active if -Z is used when not building with\nzlib. At the end, it comes down to whether we want to prioritize the\nportability of pg_dump commands specifying a -Z/--compress across\nenvironments knowing that these may or may not be built with zlib,\nvs the amount of simplification/uniformity we would get across the\nbinaries in the tree once we switch everything to use the compression\nspecifications. Now that pg_basebackup and pg_receivewal are managed\nby compression specifications, and that we'd want more compression\noptions for pg_dump, I would tend to do the latter and from now on\ncomplain if attempting to do a pg_dump -Z under --without-zlib with a\ncompression level > 0. zlib is also widely available, and we don't\ndocument the fact that non-compression is enforced in this case,\neither. (Two TAP tests with the custom format had to be tweaked.)\n\nAs per the patch, it is true that we do not need to bump the format of\nthe dump archives, as we can still store only the compression level\nand guess the method from it. I have added some notes about that in\nReadHead and WriteHead to not forget.\n\nMost of the changes are really-straight forward, and it has resisted\nmy tests, so I think that this is in a rather-commitable shape as-is.\n--\nMichael","msg_date":"Thu, 1 Dec 2022 11:05:54 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Thursday, December 1st, 2022 at 3:05 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Wed, Nov 30, 2022 at 05:11:44PM +0000, gkokolatos@pm.me wrote:\n> \n> > Fair enough. The atteched v11 does that. 0001 introduces compression\n> > specification and is using it throughout. 0002 paves the way to the\n> > new interface by homogenizing the use of cfp. 0003 introduces the new\n> > API and stores the compression algorithm in the custom format header\n> > instead of the compression level integer. Finally 0004 adds support for\n> > LZ4.\n> \n> \n> I have been looking at 0001, and.. Hmm. I am really wondering\n> whether it would not be better to just nuke this warning into orbit.\n> This stuff enforces non-compression even if -Z has been used to a\n> non-default value. This has been moved to its current location by\n> cae2bb1 as of this thread:\n> https://www.postgresql.org/message-id/20160526.185551.242041780.horiguchi.kyotaro%40lab.ntt.co.jp\n> \n> However, this is only active if -Z is used when not building with\n> zlib. At the end, it comes down to whether we want to prioritize the\n> portability of pg_dump commands specifying a -Z/--compress across\n> environments knowing that these may or may not be built with zlib,\n> vs the amount of simplification/uniformity we would get across the\n> binaries in the tree once we switch everything to use the compression\n> specifications. Now that pg_basebackup and pg_receivewal are managed\n> by compression specifications, and that we'd want more compression\n> options for pg_dump, I would tend to do the latter and from now on\n> complain if attempting to do a pg_dump -Z under --without-zlib with a\n> compression level > 0. zlib is also widely available, and we don't\n> document the fact that non-compression is enforced in this case,\n> either. (Two TAP tests with the custom format had to be tweaked.)\n\nFair enough. Thank you for looking. However I have a small comment\non your new patch.\n\n- /* Custom and directory formats are compressed by default, others not */\n- if (compressLevel == -1)\n- {\n-#ifdef HAVE_LIBZ\n- if (archiveFormat == archCustom || archiveFormat == archDirectory)\n- compressLevel = Z_DEFAULT_COMPRESSION;\n- else\n-#endif\n- compressLevel = 0;\n- }\n\n\nNuking the warning from orbit and changing the behaviour around disabling\nthe requested compression when the libraries are not present, should not\nmean that we need to change the behaviour of default values for different\nformats. Please find v13 attached which reinstates it.\n\nWhich in itself it got me looking and wondering why the tests succeeded.\nThe only existing test covering that path is `defaults_dir_format` in\n`002_pg_dump.pl`. However as the test is currently written it does not\ncheck whether the output was compressed. The restore command would succeed\nin either case. A simple `gzip -t -r` against the directory will not\nsuffice to test it, because there exist files which are never compressed\nin this format (.toc). A little bit more involved test case would need\nto be written, yet before I embark to this journey, I would like to know\nif you would agree to reinstate the defaults for those formats.\n\n> \n> As per the patch, it is true that we do not need to bump the format of\n> the dump archives, as we can still store only the compression level\n> and guess the method from it. I have added some notes about that in\n> ReadHead and WriteHead to not forget.\n\nAgreed. A minor suggestion if you may.\n\n #ifndef HAVE_LIBZ\n- if (AH->compression != 0)\n+ if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n #endif\n\nIt would seem a more consistent to error out in this case. We do error\nin all other cases where the compression is not available.\n\n> \n> Most of the changes are really-straight forward, and it has resisted\n> my tests, so I think that this is in a rather-commitable shape as-is.\n\nThank you.\n\nCheers,\n//Georgios\n\n> --\n> Michael","msg_date":"Thu, 01 Dec 2022 14:58:35 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Dec 01, 2022 at 02:58:35PM +0000, gkokolatos@pm.me wrote:\n> Nuking the warning from orbit and changing the behaviour around disabling\n> the requested compression when the libraries are not present, should not\n> mean that we need to change the behaviour of default values for different\n> formats. Please find v13 attached which reinstates it.\n\nGah, thanks! And this default behavior is documented as dependent on\nthe compilation as well.\n\n> Which in itself it got me looking and wondering why the tests succeeded.\n> The only existing test covering that path is `defaults_dir_format` in\n> `002_pg_dump.pl`. However as the test is currently written it does not\n> check whether the output was compressed. The restore command would succeed\n> in either case. A simple `gzip -t -r` against the directory will not\n> suffice to test it, because there exist files which are never compressed\n> in this format (.toc). A little bit more involved test case would need\n> to be written, yet before I embark to this journey, I would like to know\n> if you would agree to reinstate the defaults for those formats.\n\nOn top of my mind, I briefly recall that -r is not that portable. And\nthe toc format makes the files generated non-deterministic as these\nuse OIDs..\n\n[.. thinks ..]\n\nWe are going to need a new thing here, as compress_cmd cannot be\ndirectly used. What if we used only an array of glob()-able elements?\nLet's say \"expected_contents\" that could include a \"dir_path/*.gz\"\nconditional on $supports_gzip? glob() can only be calculated when the\ntest is run as the file names cannot be known beforehand :/\n\n>> As per the patch, it is true that we do not need to bump the format of\n>> the dump archives, as we can still store only the compression level\n>> and guess the method from it. I have added some notes about that in\n>> ReadHead and WriteHead to not forget.\n> \n> Agreed. A minor suggestion if you may.\n> \n> #ifndef HAVE_LIBZ\n> - if (AH->compression != 0)\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n> #endif\n> \n> It would seem a more consistent to error out in this case. We do error\n> in all other cases where the compression is not available.\n\nMakes sense.\n\nI have gone through the patch again, and applied it. Thanks!\n--\nMichael","msg_date":"Fri, 2 Dec 2022 10:56:26 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Friday, December 2nd, 2022 at 2:56 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n> On top of my mind, I briefly recall that -r is not that portable. And\n> the toc format makes the files generated non-deterministic as these\n> use OIDs..\n> \n> [.. thinks ..]\n> \n> We are going to need a new thing here, as compress_cmd cannot be\n> directly used. What if we used only an array of glob()-able elements?\n> Let's say \"expected_contents\" that could include a \"dir_path/*.gz\"\n> conditional on $supports_gzip? glob() can only be calculated when the\n> test is run as the file names cannot be known beforehand :/\n\nYou are very correct. However one can glob after the fact. Please find\n0001 of the attached v14 which attempts to implement it.\n\n> I have gone through the patch again, and applied it. Thanks!\n\nThank you. Please find the rest of of the patchset series rebased on top\nof it. I dare to say that 0002 is in a state worth of your consideration.\n\nCheers,\n//Georgios\n\n> --\n> Michael","msg_date":"Fri, 02 Dec 2022 16:15:10 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Dec 02, 2022 at 04:15:10PM +0000, gkokolatos@pm.me wrote:\n> You are very correct. However one can glob after the fact. Please find\n> 0001 of the attached v14 which attempts to implement it.\n\n+ if ($pgdump_runs{$run}->{glob_pattern})\n+ {\n+ my $glob_pattern = $pgdump_runs{$run}->{glob_pattern};\n+ my @glob_output = glob($glob_pattern);\n+ is(scalar(@glob_output) > 0, 1, \"glob pattern matched\")\n+ }\n\nWhile this is correct in checking that the contents are compressed\nunder --with-zlib, this also removes the coverage where we make sure\nthat this command is able to complete under --without-zlib without\ncompressing any of the table data files. Hence my point from\nupthread: this test had better not use compile_option, but change\nglob_pattern depending on if the build uses zlib or not.\n\nIn order to check this behavior with defaults_custom_format, perhaps\nwe could just remove the -Z6 from it or add an extra command for its\ndefault behavior?\n--\nMichael","msg_date":"Sat, 3 Dec 2022 11:45:30 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, Dec 03, 2022 at 11:45:30AM +0900, Michael Paquier wrote:\n> While this is correct in checking that the contents are compressed\n> under --with-zlib, this also removes the coverage where we make sure\n> that this command is able to complete under --without-zlib without\n> compressing any of the table data files. Hence my point from\n> upthread: this test had better not use compile_option, but change\n> glob_pattern depending on if the build uses zlib or not.\n\nIn short, I mean something like the attached. I have named the flag\ncontent_patterns, and switched it to an array so as we can check that\ntoc.dat is always uncompression and that the other data files are\nalways uncompressed.\n\n> In order to check this behavior with defaults_custom_format, perhaps\n> we could just remove the -Z6 from it or add an extra command for its\n> default behavior?\n\nThis is slightly more complicated as there is just one file generated\nfor the compression and non-compression cases, so I have let that as\nit is now.\n--\nMichael","msg_date":"Mon, 5 Dec 2022 16:05:23 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Monday, December 5th, 2022 at 8:05 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Sat, Dec 03, 2022 at 11:45:30AM +0900, Michael Paquier wrote:\n> \n> > While this is correct in checking that the contents are compressed\n> > under --with-zlib, this also removes the coverage where we make sure\n> > that this command is able to complete under --without-zlib without\n> > compressing any of the table data files. Hence my point from\n> > upthread: this test had better not use compile_option, but change\n> > glob_pattern depending on if the build uses zlib or not.\n> \n> In short, I mean something like the attached. I have named the flag\n> content_patterns, and switched it to an array so as we can check that\n> toc.dat is always uncompression and that the other data files are\n> always uncompressed.\n\nI see. This approach is much better than my proposal, thanks. If you\nallow me, I find 'content_patterns' to be slightly ambiguous. While is\ntrue that it refers to the contents of a directory, it is not the\ncontents of the dump that it is examining. I took the liberty of proposing\nan alternative name in the attached v16.\n\nI also took the liberty of applying the test pattern when it the dump\nis explicitly compressed.\n\n> > In order to check this behavior with defaults_custom_format, perhaps\n> > we could just remove the -Z6 from it or add an extra command for its\n> > default behavior?\n> \n> This is slightly more complicated as there is just one file generated\n> for the compression and non-compression cases, so I have let that as\n> it is now.\n\nI was thinking a bit more about this. I think that we can use the list\nTOC option of pg_restore. This option will first print out the header\ninfo which contains the compression. Perl utils already support to \nparse the generated output of a command. Please find an attempt to do\nso in the attached. The benefits of having some testing for this case\nbecome a bit more obvious in 0004 of the patchset, when lz4 is\nintroduced.\n\nCheers,\n//Georgios\n\n> --\n> Michael","msg_date":"Mon, 05 Dec 2022 12:48:28 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Dec 05, 2022 at 12:48:28PM +0000, gkokolatos@pm.me wrote:\n> I also took the liberty of applying the test pattern when it the dump\n> is explicitly compressed.\n\nSticking with glob_patterns is fine by me.\n\n> I was thinking a bit more about this. I think that we can use the list\n> TOC option of pg_restore. This option will first print out the header\n> info which contains the compression. Perl utils already support to \n> parse the generated output of a command. Please find an attempt to do\n> so in the attached. The benefits of having some testing for this case\n> become a bit more obvious in 0004 of the patchset, when lz4 is\n> introduced.\n\nThis is where the fun is. What you are doing here is more complete,\nand we would make sure that the custom and data directory would always\nsee their contents compressed by default. And it would have caught\nthe bug you mentioned upthread for the custom format.\n\nI have kept things as you proposed at the end, added a few comments,\ndocumented the new command_like and an extra command_like for\ndefaults_dir_format. Glad to see this addressed, thanks!\n--\nMichael","msg_date":"Tue, 6 Dec 2022 09:22:57 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Tuesday, December 6th, 2022 at 1:22 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Mon, Dec 05, 2022 at 12:48:28PM +0000, gkokolatos@pm.me wrote:\n> \n> This is where the fun is. What you are doing here is more complete,\n> and we would make sure that the custom and data directory would always\n> see their contents compressed by default. And it would have caught\n> the bug you mentioned upthread for the custom format.\n\nThank you very much Michael.\n\n> I have kept things as you proposed at the end, added a few comments,\n> documented the new command_like and an extra command_like for\n> defaults_dir_format. Glad to see this addressed, thanks!\n\nPlease find attached v17, which builds on top of what is already\ncommitted. I dare to think 0001 as ready to be reviewed. 0002 is\nalso complete albeit with some documentation gaps.\n\nCheers,\n//Georgios\n\n> --\n> Michael","msg_date":"Tue, 06 Dec 2022 15:52:01 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"001: still refers to \"gzip\", which is correct for -Fp and -Fd but not\nfor -Fc, for which it's more correct to say \"zlib\". That affects the\nname of the function, structures, comments, etc. I'm not sure if it's\nan issue to re-use the basebackup compression routines here. Maybe we\nshould accept \"-Zn\" for zlib output (-Fc), but reject \"gzip:9\", which\nI'm sure some will find confusing, as it does not output. Maybe 001\nshould be split into a patch to re-use the existing \"cfp\" interface\n(which is a clear win), and 002 to re-use the basebackup interfaces for\nuser input and constants, etc.\n\n001 still doesn't compile on freebsd, and 002 doesn't compile on\nwindows. Have you checked test results from cirrusci on your private\ngithub account ?\n\n002 says:\n+ save_errno = errno; \n+ errno = save_errno; \n\nI suppose that's intended to wrap the preceding library call.\n\n002 breaks \"pg_dump -Fc -Z2\" because (I think) AllocateCompressor()\ndoesn't store the passed-in compression_spec.\n\n003 still uses <lz4.h> and not \"lz4.h\".\n\nEarlier this year I also suggested to include an 999 patch to change to\nuse LZ4 as the default compression, to exercise the new code under CI.\nI suggest to re-open the cf patch entry after that passes tests on all\nplatforms and when it's ready for more review.\n\nBTW, some of these review comments are the same as what I sent earlier\nthis year.\n\nhttps://www.postgresql.org/message-id/20220326162156.GI28503%40telsasoft.com\nhttps://www.postgresql.org/message-id/20220705151328.GQ13040%40telsasoft.com\n\n-- \nJustin\n\n\n","msg_date":"Sat, 17 Dec 2022 17:26:15 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, Dec 17, 2022 at 05:26:15PM -0600, Justin Pryzby wrote:\n> 001: still refers to \"gzip\", which is correct for -Fp and -Fd but not\n> for -Fc, for which it's more correct to say \"zlib\".\n\nOr should we begin by changing all these existing \"not built with zlib \nsupport\" error strings to the more generic \"this build does not\nsupport compression with %s\" to reduce the number of messages to\ntranslate? That would bring consistency with the other tools dealing\nwith compression.\n\n> That affects the\n> name of the function, structures, comments, etc. I'm not sure if it's\n> an issue to re-use the basebackup compression routines here. Maybe we\n> should accept \"-Zn\" for zlib output (-Fc), but reject \"gzip:9\", which\n> I'm sure some will find confusing, as it does not output. Maybe 001\n> should be split into a patch to re-use the existing \"cfp\" interface\n> (which is a clear win), and 002 to re-use the basebackup interfaces for\n> user input and constants, etc.\n> \n> 001 still doesn't compile on freebsd, and 002 doesn't compile on\n> windows. Have you checked test results from cirrusci on your private\n> github account ?\n\nFYI, I have re-added an entry to the CF app to get some automated\ncoverage:\nhttps://commitfest.postgresql.org/41/3571/\n\nOn MinGW, a complain about the open() callback, which I guess ought to\nbe avoided with a rename:\n[00:16:37.254] compress_gzip.c:356:38: error: macro \"open\" passed 4 arguments, but takes just 3\n[00:16:37.254] 356 | ret = CFH->open(fname, -1, mode, CFH);\n[00:16:37.254] | ^\n[00:16:37.254] In file included from ../../../src/include/c.h:1309,\n[00:16:37.254] from ../../../src/include/postgres_fe.h:25,\n[00:16:37.254] from compress_gzip.c:15:\n\nOn MSVC, some declaration conflicts, for a similar issue:\n[00:12:31.966] ../src/bin/pg_dump/compress_io.c(193): error C2371: '_read': redefinition; different basic types\n[00:12:31.966] C:\\Program Files (x86)\\Windows Kits\\10\\include\\10.0.20348.0\\ucrt\\corecrt_io.h(252): note: see declaration of '_read'\n[00:12:31.966] ../src/bin/pg_dump/compress_io.c(210): error C2371: '_write': redefinition; different basic types\n[00:12:31.966] C:\\Program Files (x86)\\Windows Kits\\10\\include\\10.0.20348.0\\ucrt\\corecrt_io.h(294): note: see declaration of '_write'\n\n> 002 breaks \"pg_dump -Fc -Z2\" because (I think) AllocateCompressor()\n> doesn't store the passed-in compression_spec.\n\nHmm. This looks like a gap in the existing tests that we'd better fix\nfirst. This CI is green on Linux.\n\n> 003 still uses <lz4.h> and not \"lz4.h\".\n\nThis should be <lz4.h>, not \"lz4.h\".\n--\nMichael","msg_date":"Mon, 19 Dec 2022 13:06:00 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Monday, December 19th, 2022 at 5:06 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Sat, Dec 17, 2022 at 05:26:15PM -0600, Justin Pryzby wrote:\n> \n\nThank you for the comments, please find v18 attached.\n\n> > 001: still refers to \"gzip\", which is correct for -Fp and -Fd but not\n> > for -Fc, for which it's more correct to say \"zlib\".\n> \n> \n> Or should we begin by changing all these existing \"not built with zlib\n> support\" error strings to the more generic \"this build does not\n> support compression with %s\" to reduce the number of messages to\n> translate? That would bring consistency with the other tools dealing\n> with compression.\n\nThis has been the approach from 0002 on-wards. In the attached it is also\napplied on the remaining location in 0001. \n\n> \n> > That affects the\n> > name of the function, structures, comments, etc. I'm not sure if it's\n> > an issue to re-use the basebackup compression routines here. Maybe we\n> > should accept \"-Zn\" for zlib output (-Fc), but reject \"gzip:9\", which\n> > I'm sure some will find confusing, as it does not output. Maybe 001\n> > should be split into a patch to re-use the existing \"cfp\" interface\n> > (which is a clear win), and 002 to re-use the basebackup interfaces for\n> > user input and constants, etc.\n> > \n> > 001 still doesn't compile on freebsd, and 002 doesn't compile on\n> > windows. Have you checked test results from cirrusci on your private\n> > github account ?\n\nThere are still known gaps in 0002 and 0003, for example documentation,\nand I have not been focusing too much on those. You are right, it is helpful\nand kind to try to reduce the noise. The attached should have hopefully\ntackled the ci errors.\n\n> \n> FYI, I have re-added an entry to the CF app to get some automated\n> coverage:\n> https://commitfest.postgresql.org/41/3571/\n\nMuch obliged. Should I change the state to \"ready for review\" when post a\nnew version or should I leave that to the senior personnel? \n\n> \n> On MinGW, a complain about the open() callback, which I guess ought to\n> be avoided with a rename:\n> [00:16:37.254] compress_gzip.c:356:38: error: macro \"open\" passed 4 arguments, but takes just 3\n> [00:16:37.254] 356 | ret = CFH->open(fname, -1, mode, CFH);\n> \n> [00:16:37.254] | ^\n> [00:16:37.254] In file included from ../../../src/include/c.h:1309,\n> [00:16:37.254] from ../../../src/include/postgres_fe.h:25,\n> [00:16:37.254] from compress_gzip.c:15:\n> \n> On MSVC, some declaration conflicts, for a similar issue:\n> [00:12:31.966] ../src/bin/pg_dump/compress_io.c(193): error C2371: '_read': redefinition; different basic types\n> [00:12:31.966] C:\\Program Files (x86)\\Windows Kits\\10\\include\\10.0.20348.0\\ucrt\\corecrt_io.h(252): note: see declaration of '_read'\n> [00:12:31.966] ../src/bin/pg_dump/compress_io.c(210): error C2371: '_write': redefinition; different basic types\n> [00:12:31.966] C:\\Program Files (x86)\\Windows Kits\\10\\include\\10.0.20348.0\\ucrt\\corecrt_io.h(294): note: see declaration of '_write'\n> \n\nA rename was enough.\n\n> > 002 breaks \"pg_dump -Fc -Z2\" because (I think) AllocateCompressor()\n> > doesn't store the passed-in compression_spec.\n> \n\nI am afraid I have not been able to reproduce this error. I tried both\ndebian and freebsd after I addressed the compilation warnings. Which\nerror did you get? Is it still present in the attached?\n\n> Hmm. This looks like a gap in the existing tests that we'd better fix\n> first. This CI is green on Linux.\n\nAs the code stands, the compression level is not stored in the custom\nformat's header as it is no longer relevant information. We can decide\nto make it relevant for the tests only on the expense of increasing\ndump size by four bytes. In either case this is not applicable in\ncurrent head and can wait for 0002's turn.\n\nCheers,\n//Georgios\n\n> \n> > 003 still uses <lz4.h> and not \"lz4.h\".\n> \n> \n> This should be <lz4.h>, not \"lz4.h\".\n> \n> --\n> Michael","msg_date":"Mon, 19 Dec 2022 17:03:21 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\n> > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\n> > > windows. Have you checked test results from cirrusci on your private\n> > > github account ?\n> \n> There are still known gaps in 0002 and 0003, for example documentation,\n> and I have not been focusing too much on those. You are right, it is helpful\n> and kind to try to reduce the noise. The attached should have hopefully\n> tackled the ci errors.\n\nYep. Are you using cirrusci under your github account ?\n\n> > FYI, I have re-added an entry to the CF app to get some automated\n> > coverage:\n> > https://commitfest.postgresql.org/41/3571/\n> \n> Much obliged. Should I change the state to \"ready for review\" when post a\n> new version or should I leave that to the senior personnel? \n\nIt's better to update it to reflect what you think its current status\nis. If you think it's ready for review.\n\n> > > 002 breaks \"pg_dump -Fc -Z2\" because (I think) AllocateCompressor()\n> > > doesn't store the passed-in compression_spec.\n> \n> I am afraid I have not been able to reproduce this error. I tried both\n> debian and freebsd after I addressed the compilation warnings. Which\n> error did you get? Is it still present in the attached?\n\nIt's not that there's an error - it's that compression isn't working.\n\n$ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\n659956\n$ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\n637192\n\n$ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\n1954890\n$ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\n1954890\n\n-- \nJustin\n\n\n","msg_date":"Mon, 19 Dec 2022 11:27:06 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Dec 19, 2022 at 01:06:00PM +0900, Michael Paquier wrote:\n> On Sat, Dec 17, 2022 at 05:26:15PM -0600, Justin Pryzby wrote:\n> > 001: still refers to \"gzip\", which is correct for -Fp and -Fd but not\n> > for -Fc, for which it's more correct to say \"zlib\".\n> \n> Or should we begin by changing all these existing \"not built with zlib \n> support\" error strings to the more generic \"this build does not\n> support compression with %s\" to reduce the number of messages to\n> translate? That would bring consistency with the other tools dealing\n> with compression.\n\nThat's fine, but it doesn't touch on the issue I'm talking about, which\nis that zlib != gzip.\n\nBTW I noticed that that also affects the pg_dump file itself; 002\nchanges the file format to say \"gzip\", but that's wrong for -Fc, which\ndoes not use gzip headers, which could be surprising to someone who\nspecified \"gzip\".\n\n-- \nJustin\n\n\n","msg_date":"Mon, 19 Dec 2022 11:42:58 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Monday, December 19th, 2022 at 6:27 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n> \n> \n> On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\n> \n> > > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\n> > > > windows. Have you checked test results from cirrusci on your private\n> > > > github account ?\n> > \n> > There are still known gaps in 0002 and 0003, for example documentation,\n> > and I have not been focusing too much on those. You are right, it is helpful\n> > and kind to try to reduce the noise. The attached should have hopefully\n> > tackled the ci errors.\n> \n> \n> Yep. Are you using cirrusci under your github account ?\n\nThank you. To be very honest, I am not using github exclusively to post patches.\nSometimes I do, sometimes I do not. Is github a requirement?\n\nTo answer your question, some of my github accounts are integrated with cirrusci,\nothers are not.\n\nThe current cfbot build is green for what is worth.\nhttps://cirrus-ci.com/build/5934319840002048\n\n> \n> > > FYI, I have re-added an entry to the CF app to get some automated\n> > > coverage:\n> > > https://commitfest.postgresql.org/41/3571/\n> > \n> > Much obliged. Should I change the state to \"ready for review\" when post a\n> > new version or should I leave that to the senior personnel?\n> \n> \n> It's better to update it to reflect what you think its current status\n> is. If you think it's ready for review.\n\nThank you.\n\n> \n> > > > 002 breaks \"pg_dump -Fc -Z2\" because (I think) AllocateCompressor()\n> > > > doesn't store the passed-in compression_spec.\n> > \n> > I am afraid I have not been able to reproduce this error. I tried both\n> > debian and freebsd after I addressed the compilation warnings. Which\n> > error did you get? Is it still present in the attached?\n> \n> \n> It's not that there's an error - it's that compression isn't working.\n> \n> $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\n> 659956\n> $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\n> 637192\n> \n> $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\n> 1954890\n> $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\n> 1954890\n> \n\nThank you. Now I understand what you mean. Trying the same on top of v18-0003\non Ubuntu 22.04 yields:\n\n$ for compression in none gzip:1 gzip:6 gzip:9; do \\\n pg_dump --format=custom --compress=\"$compression\" -f regression.\"$compression\".dump -d regression; \\\n wc -c regression.\"$compression\".dump; \\\n done;\n14963753 regression.none.dump\n3600183 regression.gzip:1.dump\n3223755 regression.gzip:6.dump\n3196903 regression.gzip:9.dump\n\nand on FreeBSD 13.1\n\n$ for compression in none gzip:1 gzip:6 gzip:9; do \\\n pg_dump --format=custom --compress=\"$compression\" -f regression.\"$compression\".dump -d regression; \\\n wc -c regression.\"$compression\".dump; \\\n done;\n14828822 regression.none.dump\n3584304 regression.gzip:1.dump\n3208548 regression.gzip:6.dump\n3182044 regression.gzip:9.dump\n\nAlthough there are some variations between the installations, within the same\ninstallation the size of the dump file is shrinking as expected.\n\nInvestigating a bit further on the issue, you are correct in identifying an\nissue in v17. Up until v16, the compressor function looked like:\n\n+InitCompressorGzip(CompressorState *cs, int compressionLevel)\n+{\n+ GzipCompressorState *gzipcs;\n+\n+ cs->readData = ReadDataFromArchiveGzip;\n+ cs->writeData = WriteDataToArchiveGzip;\n+ cs->end = EndCompressorGzip;\n+\n+ gzipcs = (GzipCompressorState *) pg_malloc0(sizeof(GzipCompressorState));\n+ gzipcs->compressionLevel = compressionLevel;\n\nV17 considered that more options could become available in the future\nand changed the signature of the relevant Init functions to:\n\n+InitCompressorGzip(CompressorState *cs, const pg_compress_specification compression_spec)\n+{\n+ GzipCompressorState *gzipcs;\n+\n+ cs->readData = ReadDataFromArchiveGzip;\n+ cs->writeData = WriteDataToArchiveGzip;\n+ cs->end = EndCompressorGzip;\n+\n+ gzipcs = (GzipCompressorState *) pg_malloc0(sizeof(GzipCompressorState));\n+\n\nV18 reinstated the assignment in similar fashion to InitCompressorNone and \nInitCompressorLz4:\n\n+void\n+InitCompressorGzip(CompressorState *cs, const pg_compress_specification compression_spec)\n+{\n+ GzipCompressorState *gzipcs;\n+\n+ cs->readData = ReadDataFromArchiveGzip;\n+ cs->writeData = WriteDataToArchiveGzip;\n+ cs->end = EndCompressorGzip;\n+\n+ cs->compression_spec = compression_spec;\n+\n+ gzipcs = (GzipCompressorState *) pg_malloc0(sizeof(GzipCompressorState));\n\nA test case can be added which performs a check similar to the loop above.\nCreate a custom dump with the least and most compression for each method.\nThen verify that the output sizes differ as expected. This addition could\nbecome 0001 in the current series.\n\nThoughts?\n\nCheers,\n//Georgios\n\n> --\n> Justin\n\n\n","msg_date":"Tue, 20 Dec 2022 11:19:15 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Dec 20, 2022 at 11:19:15AM +0000, gkokolatos@pm.me wrote:\n> ------- Original Message -------\n> On Monday, December 19th, 2022 at 6:27 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\n> > \n> > > > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\n> > > > > windows. Have you checked test results from cirrusci on your private\n> > > > > github account ?\n> > > \n> > > There are still known gaps in 0002 and 0003, for example documentation,\n> > > and I have not been focusing too much on those. You are right, it is helpful\n> > > and kind to try to reduce the noise. The attached should have hopefully\n> > > tackled the ci errors.\n> > \n> > \n> > Yep. Are you using cirrusci under your github account ?\n> \n> Thank you. To be very honest, I am not using github exclusively to post patches.\n> Sometimes I do, sometimes I do not. Is github a requirement?\n\nGithub isn't a requirement for postgres (but cirrusci only supports\ngithub). I wasn't not trying to say that it's required, only trying to\nmake sure that you (and others) know that it's available, since our\ncirrus.yml is relatively new.\n\n> > > > > 002 breaks \"pg_dump -Fc -Z2\" because (I think) AllocateCompressor()\n> > > > > doesn't store the passed-in compression_spec.\n> > > \n> > > I am afraid I have not been able to reproduce this error. I tried both\n> > > debian and freebsd after I addressed the compilation warnings. Which\n> > > error did you get? Is it still present in the attached?\n> > \n> > \n> > It's not that there's an error - it's that compression isn't working.\n> > \n> > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\n> > 659956\n> > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\n> > 637192\n> > \n> > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\n> > 1954890\n> > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\n> > 1954890\n> > \n> \n> Thank you. Now I understand what you mean. Trying the same on top of v18-0003\n> on Ubuntu 22.04 yields:\n\nYou're right; this seems to be fixed in v18. Thanks.\n\nIt looks like I'd forgotten to run \"meson test tmp_install\", so had\nretested v17...\n\n-- \nJustin\n\n\n","msg_date":"Tue, 20 Dec 2022 09:26:10 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Tuesday, December 20th, 2022 at 4:26 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n> \n> \n> On Tue, Dec 20, 2022 at 11:19:15AM +0000, gkokolatos@pm.me wrote:\n> \n> > ------- Original Message -------\n> > On Monday, December 19th, 2022 at 6:27 PM, Justin Pryzby pryzby@telsasoft.com wrote:\n> > \n> > > On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\n> > > \n> > > > > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\n> > > > > > windows. Have you checked test results from cirrusci on your private\n> > > > > > github account ?\n> > > > \n> > > > There are still known gaps in 0002 and 0003, for example documentation,\n> > > > and I have not been focusing too much on those. You are right, it is helpful\n> > > > and kind to try to reduce the noise. The attached should have hopefully\n> > > > tackled the ci errors.\n> > > \n> > > Yep. Are you using cirrusci under your github account ?\n> > \n> > Thank you. To be very honest, I am not using github exclusively to post patches.\n> > Sometimes I do, sometimes I do not. Is github a requirement?\n> \n> \n> Github isn't a requirement for postgres (but cirrusci only supports\n> github). I wasn't not trying to say that it's required, only trying to\n> make sure that you (and others) know that it's available, since our\n> cirrus.yml is relatively new.\n\nGot it. Thank you very much for spreading the word. It is a useful feature which\nshould be known. \n\n> \n> > > > > > 002 breaks \"pg_dump -Fc -Z2\" because (I think) AllocateCompressor()\n> > > > > > doesn't store the passed-in compression_spec.\n> > > > \n> > > > I am afraid I have not been able to reproduce this error. I tried both\n> > > > debian and freebsd after I addressed the compilation warnings. Which\n> > > > error did you get? Is it still present in the attached?\n> > > \n> > > It's not that there's an error - it's that compression isn't working.\n> > > \n> > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\n> > > 659956\n> > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\n> > > 637192\n> > > \n> > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\n> > > 1954890\n> > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\n> > > 1954890\n> > \n> > Thank you. Now I understand what you mean. Trying the same on top of v18-0003\n> > on Ubuntu 22.04 yields:\n> \n> \n> You're right; this seems to be fixed in v18. Thanks.\n\nGreat. Still there was a bug in v17 which you discovered. Thank you for the review\neffort.\n\nPlease find in the attached v19 an extra check right before calling deflateInit().\nThis check will verify that only compressed output will be generated for this\nmethod.\n\nAlso v19 is rebased on top f450695e889 and applies cleanly.\n\nCheers.\n//Georgios\n\n> --\n> Justin","msg_date":"Wed, 21 Dec 2022 10:09:38 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"There's a couple of lz4 bits which shouldn't be present in 002: file\nextension and comments.\n\n\n","msg_date":"Thu, 22 Dec 2022 11:08:59 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Dec 22, 2022 at 11:08:59AM -0600, Justin Pryzby wrote:\n> There's a couple of lz4 bits which shouldn't be present in 002: file\n> extension and comments.\n\nThere were \"LZ4\" comments and file extension stuff in the preparatory\ncommit. But now it seems like you *removed* them in the LZ4 commit\n(where it actually belongs) rather than *moving* it from the\nprior/parent commit *to* the lz4 commit. I recommend to run something\nlike \"git diff @{1}\" whenever doing this kind of patch surgery.\n\n+ if (AH->compression_spec.algorithm != PG_COMPRESSION_NONE &&\n+ AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\n\nThis looks wrong/redundant. The gzip part should be removed, right ?\n\nMaybe other places that check if (compression==PG_COMPRESSION_GZIP)\nshould maybe change to say compression!=NONE?\n\n_PrepParallelRestore() references \".gz\", so I think it needs to be\nretrofitted to handle .lz4. Ideally, that's built into a struct or list\nof file extensions to try. Maybe compression.h should have a function\nto return the file extension of a given algorithm. I'm planning to send\na patch for zstd, and hoping its changes will be minimized by these\npreparatory commits.\n\n+ errno = errno ? : ENOSPC;\n\n\"?:\" is a GNU extension (not the ternary operator, but the ternary\noperator with only 2 args). It's not in use anywhere else in postgres.\nYou could instead write it with 3 \"errno\"s or as \"if (errno==0):\nerrno=ENOSPC\"\n\nYou wrote \"eol_flag == false\" and \"eol_flag == 0\" and true. But it's\ncleaner to test it as a boolean: if (eol_flag) / if (!eol_flag).\n\nBoth LZ4File_init() and its callers check \"inited\". Better to do it in\none place than 3. It's a static function, so I think there's no\nperformance concern.\n\nGzip_close() still has a useless save_errno (or rebase issue?).\n\nI think it's confusing to have two functions, one named\nInitCompressLZ4() and InitCompressorLZ4().\n\npg_compress_specification is being passed by value, but I think it\nshould be passed as a pointer, as is done everywhere else.\n\npg_compress_algorithm is being writen directly into the pg_dump header.\nCurrently, I think that's not an externally-visible value (it could be\nrenumbered, theoretically even in a minor release). Maybe there should\nbe a \"private\" enum for encoding the pg_dump header, similar to\nWAL_COMPRESSION_LZ4 vs BKPIMAGE_COMPRESS_LZ4 ? Or else a comment there\nshould warn that the values are encoded in pg_dump, and must never be\nchanged.\n\n+ Verify that data files where compressed\ntypo: s/where/were/\n\nAlso:\ns/occurance/occurrence/\ns/begining/beginning/\ns/Verfiy/Verify/\ns/nessary/necessary/\n\nBTW I noticed that cfdopen() was accidentally committed to compress_io.h\nin master without being defined anywhere.\n\n-- \nJustin\n\n\n","msg_date":"Sun, 8 Jan 2023 13:45:25 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, 21 Dec 2022 at 15:40, <gkokolatos@pm.me> wrote:\n>\n>\n>\n>\n>\n>\n> ------- Original Message -------\n> On Tuesday, December 20th, 2022 at 4:26 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n>\n> >\n> >\n> > On Tue, Dec 20, 2022 at 11:19:15AM +0000, gkokolatos@pm.me wrote:\n> >\n> > > ------- Original Message -------\n> > > On Monday, December 19th, 2022 at 6:27 PM, Justin Pryzby pryzby@telsasoft.com wrote:\n> > >\n> > > > On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\n> > > >\n> > > > > > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\n> > > > > > > windows. Have you checked test results from cirrusci on your private\n> > > > > > > github account ?\n> > > > >\n> > > > > There are still known gaps in 0002 and 0003, for example documentation,\n> > > > > and I have not been focusing too much on those. You are right, it is helpful\n> > > > > and kind to try to reduce the noise. The attached should have hopefully\n> > > > > tackled the ci errors.\n> > > >\n> > > > Yep. Are you using cirrusci under your github account ?\n> > >\n> > > Thank you. To be very honest, I am not using github exclusively to post patches.\n> > > Sometimes I do, sometimes I do not. Is github a requirement?\n> >\n> >\n> > Github isn't a requirement for postgres (but cirrusci only supports\n> > github). I wasn't not trying to say that it's required, only trying to\n> > make sure that you (and others) know that it's available, since our\n> > cirrus.yml is relatively new.\n>\n> Got it. Thank you very much for spreading the word. It is a useful feature which\n> should be known.\n>\n> >\n> > > > > > > 002 breaks \"pg_dump -Fc -Z2\" because (I think) AllocateCompressor()\n> > > > > > > doesn't store the passed-in compression_spec.\n> > > > >\n> > > > > I am afraid I have not been able to reproduce this error. I tried both\n> > > > > debian and freebsd after I addressed the compilation warnings. Which\n> > > > > error did you get? Is it still present in the attached?\n> > > >\n> > > > It's not that there's an error - it's that compression isn't working.\n> > > >\n> > > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\n> > > > 659956\n> > > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\n> > > > 637192\n> > > >\n> > > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\n> > > > 1954890\n> > > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\n> > > > 1954890\n> > >\n> > > Thank you. Now I understand what you mean. Trying the same on top of v18-0003\n> > > on Ubuntu 22.04 yields:\n> >\n> >\n> > You're right; this seems to be fixed in v18. Thanks.\n>\n> Great. Still there was a bug in v17 which you discovered. Thank you for the review\n> effort.\n>\n> Please find in the attached v19 an extra check right before calling deflateInit().\n> This check will verify that only compressed output will be generated for this\n> method.\n>\n> Also v19 is rebased on top f450695e889 and applies cleanly.\n\nThe patch does not apply on top of HEAD as in [1], please post a rebased patch:\n=== Applying patches on top of PostgreSQL commit ID\nff23b592ad6621563d3128b26860bcb41daf9542 ===\n=== applying patch ./v19-0002-Introduce-Compressor-API-in-pg_dump.patch\npatching file src/bin/pg_dump/compress_io.h\nHunk #1 FAILED at 37.\n1 out of 1 hunk FAILED -- saving rejects to file\nsrc/bin/pg_dump/compress_io.h.rej\n\n[1] - http://cfbot.cputube.org/patch_41_3571.log\n\nRegards,\nVignesh\n\n\n","msg_date":"Sat, 14 Jan 2023 12:21:31 +0530","msg_from":"vignesh C <vignesh21@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\n> On Thu, Dec 22, 2022 at 11:08:59AM -0600, Justin Pryzby wrote:\n> > There's a couple of lz4 bits which shouldn't be present in 002: file\n> > extension and comments.\n\n> BTW I noticed that cfdopen() was accidentally committed to compress_io.h\n> in master without being defined anywhere.\n\nThis was resolved in 69fb29d1a (so now needs to be re-added for this\npatch series).\n\n> pg_compress_specification is being passed by value, but I think it\n> should be passed as a pointer, as is done everywhere else.\n\nISTM that was an issue with 5e73a6048, affecting a few public and\nprivate functions. I wrote a pre-preparatory patch which changes to\npass by reference.\n\nAnd addressed a handful of other issues I reported as separate fixup\ncommits. And changed to use LZ4 by default for CI.\n\nI also rebased my 2 year old patch to support zstd in pg_dump. I hope\nit can finally added for v16. I'll send it for the next CF if these\npatches progress.\n\nOne more thing: some comments still refer to the cfopen API, which this\npatch removes.\n\n> There were \"LZ4\" comments and file extension stuff in the preparatory\n> commit. But now it seems like you *removed* them in the LZ4 commit\n> (where it actually belongs) rather than *moving* it from the\n> prior/parent commit *to* the lz4 commit. I recommend to run something\n> like \"git diff @{1}\" whenever doing this kind of patch surgery.\n\nTODO\n\n> Maybe other places that check if (compression==PG_COMPRESSION_GZIP)\n> should maybe change to say compression!=NONE?\n> \n> _PrepParallelRestore() references \".gz\", so I think it needs to be\n> retrofitted to handle .lz4. Ideally, that's built into a struct or list\n> of file extensions to try. Maybe compression.h should have a function\n> to return the file extension of a given algorithm. I'm planning to send\n> a patch for zstd, and hoping its changes will be minimized by these\n> preparatory commits.\n\nTODO\n\n> I think it's confusing to have two functions, one named\n> InitCompressLZ4() and InitCompressorLZ4().\n\nTODO\n\n> pg_compress_algorithm is being writen directly into the pg_dump header.\n> Currently, I think that's not an externally-visible value (it could be\n> renumbered, theoretically even in a minor release). Maybe there should\n> be a \"private\" enum for encoding the pg_dump header, similar to\n> WAL_COMPRESSION_LZ4 vs BKPIMAGE_COMPRESS_LZ4 ? Or else a comment there\n> should warn that the values are encoded in pg_dump, and must never be\n> changed.\n\nMichael, WDYT ?\n\n-- \nJustin","msg_date":"Sat, 14 Jan 2023 15:43:09 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, Jan 14, 2023 at 03:43:08PM -0600, Justin Pryzby wrote:\n> On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\n> > pg_compress_specification is being passed by value, but I think it\n> > should be passed as a pointer, as is done everywhere else.\n> \n> ISTM that was an issue with 5e73a6048, affecting a few public and\n> private functions. I wrote a pre-preparatory patch which changes to\n> pass by reference.\n\nI updated 001 to change SetOutput() to pass by reference, too (before,\nthat ended up in the 002 patch).\n\nI can't see any issue in 002 other than the == GZIP change (the fix for\nwhich I'd previously included in a later patch).\n\n> One more thing: some comments still refer to the cfopen API, which this\n> patch removes.\n> \n> > There were \"LZ4\" comments and file extension stuff in the preparatory\n> > commit. But now it seems like you *removed* them in the LZ4 commit\n> > (where it actually belongs) rather than *moving* it from the\n> > prior/parent commit *to* the lz4 commit. I recommend to run something\n> > like \"git diff @{1}\" whenever doing this kind of patch surgery.\n> \n> TODO\n\nI addressed that in the fixup commits 005 and 007.\n\n-- \nJustin","msg_date":"Sun, 15 Jan 2023 16:32:11 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, Jan 14, 2023 at 03:43:09PM -0600, Justin Pryzby wrote:\n> On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\n>> pg_compress_specification is being passed by value, but I think it\n>> should be passed as a pointer, as is done everywhere else.\n> \n> ISTM that was an issue with 5e73a6048, affecting a few public and\n> private functions. I wrote a pre-preparatory patch which changes to\n> pass by reference.\n\nThe functions changed by 0001 are cfopen[_write](),\nAllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\nto change these interfaces which basically exist to handle inputs? Is\nthere some benefit in changing compression_spec within the internals\nof these routines before going back one layer down to their callers?\nChanging the compression_spec on-the-fly in these internal paths could\nbe risky, actually, no?\n\n> And addressed a handful of other issues I reported as separate fixup\n> commits. And changed to use LZ4 by default for CI.\n\nAre your slight changes shaped as of 0003-f.patch, 0005-f.patch and\n0007-f.patch on top of the original patches sent by Georgios?\n\n> I also rebased my 2 year old patch to support zstd in pg_dump. I hope\n> it can finally added for v16. I'll send it for the next CF if these\n> patches progress.\n\nGood idea to see if what you have done for zstd fits with what's\npresented here.\n\n>> pg_compress_algorithm is being writen directly into the pg_dump header.\n\nDo you mean that this is what happens once the patch series 0001~0008\nsent upthread is applied on HEAD?\n\n>> Currently, I think that's not an externally-visible value (it could be\n>> renumbered, theoretically even in a minor release). Maybe there should\n>> be a \"private\" enum for encoding the pg_dump header, similar to\n>> WAL_COMPRESSION_LZ4 vs BKPIMAGE_COMPRESS_LZ4 ? Or else a comment there\n>> should warn that the values are encoded in pg_dump, and must never be\n>> changed.\n> \n> Michael, WDYT ?\n\nChanging the order of the members in an enum would cause an ABI\nbreakage, so that would not happen, and we tend to be very careful\nabout that. Appending new members would be fine, though. FWIW, I'd\nrather avoid adding more enums that would just be exact maps to\npg_compress_algorithm.\n\n- /*\n- * For now the compression type is implied by the level. This will need\n- * to change once support for more compression algorithms is added,\n- * requiring a format bump.\n- */\n- WriteInt(AH, AH->compression_spec.level);\n+ AH->WriteBytePtr(AH, AH->compression_spec.algorithm);\n\nI may be missing something here, but it seems to me that you ought to\nstore as well the level in the dump header, or it would not be\npossible to report in the dump's description what was used? Hence,\nK_VERS_1_15 should imply that we have both the method compression and\nthe compression level.\n--\nMichael","msg_date":"Mon, 16 Jan 2023 10:28:50 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\n> On Sat, Jan 14, 2023 at 03:43:09PM -0600, Justin Pryzby wrote:\n> > On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\n> >> pg_compress_specification is being passed by value, but I think it\n> >> should be passed as a pointer, as is done everywhere else.\n> > \n> > ISTM that was an issue with 5e73a6048, affecting a few public and\n> > private functions. I wrote a pre-preparatory patch which changes to\n> > pass by reference.\n> \n> The functions changed by 0001 are cfopen[_write](),\n> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\n> to change these interfaces which basically exist to handle inputs?\n\nI changed to pass pg_compress_specification as a pointer, since that's\nthe usual convention for structs, as followed by the existing uses of\npg_compress_specification.\n\n> Is there some benefit in changing compression_spec within the\n> internals of these routines before going back one layer down to their\n> callers? Changing the compression_spec on-the-fly in these internal\n> paths could be risky, actually, no?\n\nI think what you're saying is that if the spec is passed as a pointer,\nthen the called functions shouldn't set spec->algorithm=something.\n\nI agree that if they need to do that, they should use a local variable.\nWhich looks to be true for the functions that were changed in 001.\n\n> > And addressed a handful of other issues I reported as separate fixup\n> > commits. And changed to use LZ4 by default for CI.\n> \n> Are your slight changes shaped as of 0003-f.patch, 0005-f.patch and\n> 0007-f.patch on top of the original patches sent by Georgios?\n\nYes, the original patches, rebased as needed on top of HEAD and 001...\n\n> >> pg_compress_algorithm is being writen directly into the pg_dump header.\n> \n> Do you mean that this is what happens once the patch series 0001~0008\n> sent upthread is applied on HEAD?\n\nYes\n\n> - /*\n> - * For now the compression type is implied by the level. This will need\n> - * to change once support for more compression algorithms is added,\n> - * requiring a format bump.\n> - */\n> - WriteInt(AH, AH->compression_spec.level);\n> + AH->WriteBytePtr(AH, AH->compression_spec.algorithm);\n> \n> I may be missing something here, but it seems to me that you ought to\n> store as well the level in the dump header, or it would not be\n> possible to report in the dump's description what was used? Hence,\n> K_VERS_1_15 should imply that we have both the method compression and\n> the compression level.\n\nMaybe. But the \"level\" isn't needed for decompression for any case I'm\naware of.\n\nAlso, dumps with the default compression level currently say:\n\"Compression: -1\", which does't seem valuable.\n\n-- \nJustin\n\n\n","msg_date":"Sun, 15 Jan 2023 19:56:25 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Oh, I didn’t realize you took over Justin? Why? After almost a year of work?\n\nThis is rather disheartening.\n\nOn Mon, Jan 16, 2023 at 02:56, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n> On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\n>> On Sat, Jan 14, 2023 at 03:43:09PM -0600, Justin Pryzby wrote:\n>> > On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\n>> >> pg_compress_specification is being passed by value, but I think it\n>> >> should be passed as a pointer, as is done everywhere else.\n>> >\n>> > ISTM that was an issue with 5e73a6048, affecting a few public and\n>> > private functions. I wrote a pre-preparatory patch which changes to\n>> > pass by reference.\n>>\n>> The functions changed by 0001 are cfopen[_write](),\n>> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\n>> to change these interfaces which basically exist to handle inputs?\n>\n> I changed to pass pg_compress_specification as a pointer, since that's\n> the usual convention for structs, as followed by the existing uses of\n> pg_compress_specification.\n>\n>> Is there some benefit in changing compression_spec within the\n>> internals of these routines before going back one layer down to their\n>> callers? Changing the compression_spec on-the-fly in these internal\n>> paths could be risky, actually, no?\n>\n> I think what you're saying is that if the spec is passed as a pointer,\n> then the called functions shouldn't set spec->algorithm=something.\n>\n> I agree that if they need to do that, they should use a local variable.\n> Which looks to be true for the functions that were changed in 001.\n>\n>> > And addressed a handful of other issues I reported as separate fixup\n>> > commits. And changed to use LZ4 by default for CI.\n>>\n>> Are your slight changes shaped as of 0003-f.patch, 0005-f.patch and\n>> 0007-f.patch on top of the original patches sent by Georgios?\n>\n> Yes, the original patches, rebased as needed on top of HEAD and 001...\n>\n>> >> pg_compress_algorithm is being writen directly into the pg_dump header.\n>>\n>> Do you mean that this is what happens once the patch series 0001~0008\n>> sent upthread is applied on HEAD?\n>\n> Yes\n>\n>> - /*\n>> - * For now the compression type is implied by the level. This will need\n>> - * to change once support for more compression algorithms is added,\n>> - * requiring a format bump.\n>> - */\n>> - WriteInt(AH, AH->compression_spec.level);\n>> + AH->WriteBytePtr(AH, AH->compression_spec.algorithm);\n>>\n>> I may be missing something here, but it seems to me that you ought to\n>> store as well the level in the dump header, or it would not be\n>> possible to report in the dump's description what was used? Hence,\n>> K_VERS_1_15 should imply that we have both the method compression and\n>> the compression level.\n>\n> Maybe. But the \"level\" isn't needed for decompression for any case I'm\n> aware of.\n>\n> Also, dumps with the default compression level currently say:\n> \"Compression: -1\", which does't seem valuable.\n>\n> --\n> Justin\n Oh, I didn’t realize you took over Justin? Why? After almost a year of work?This is rather disheartening. On Mon, Jan 16, 2023 at 02:56, Justin Pryzby <pryzby@telsasoft.com> wrote: On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:> On Sat, Jan 14, 2023 at 03:43:09PM -0600, Justin Pryzby wrote:> > On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:> >> pg_compress_specification is being passed by value, but I think it> >> should be passed as a pointer, as is done everywhere else.> >> > ISTM that was an issue with 5e73a6048, affecting a few public and> > private functions. I wrote a pre-preparatory patch which changes to> > pass by reference.>> The functions changed by 0001 are cfopen[_write](),> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea> to change these interfaces which basically exist to handle inputs?I changed to pass pg_compress_specification as a pointer, since that'sthe usual convention for structs, as followed by the existing uses ofpg_compress_specification.> Is there some benefit in changing compression_spec within the> internals of these routines before going back one layer down to their> callers? Changing the compression_spec on-the-fly in these internal> paths could be risky, actually, no?I think what you're saying is that if the spec is passed as a pointer,then the called functions shouldn't set spec->algorithm=something.I agree that if they need to do that, they should use a local variable.Which looks to be true for the functions that were changed in 001.> > And addressed a handful of other issues I reported as separate fixup> > commits. And changed to use LZ4 by default for CI.>> Are your slight changes shaped as of 0003-f.patch, 0005-f.patch and> 0007-f.patch on top of the original patches sent by Georgios?Yes, the original patches, rebased as needed on top of HEAD and 001...> >> pg_compress_algorithm is being writen directly into the pg_dump header.>> Do you mean that this is what happens once the patch series 0001~0008> sent upthread is applied on HEAD?Yes> - /*> - * For now the compression type is implied by the level. This will need> - * to change once support for more compression algorithms is added,> - * requiring a format bump.> - */> - WriteInt(AH, AH->compression_spec.level);> + AH->WriteBytePtr(AH, AH->compression_spec.algorithm);>> I may be missing something here, but it seems to me that you ought to> store as well the level in the dump header, or it would not be> possible to report in the dump's description what was used? Hence,> K_VERS_1_15 should imply that we have both the method compression and> the compression level.Maybe. But the \"level\" isn't needed for decompression for any case I'maware of.Also, dumps with the default compression level currently say:\"Compression: -1\", which does't seem valuable.--Justin","msg_date":"Mon, 16 Jan 2023 02:27:56 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Jan 16, 2023 at 02:27:56AM +0000, gkokolatos@pm.me wrote:\n> Oh, I didn’t realize you took over Justin? Why? After almost a year of work?\n> \n> This is rather disheartening.\n\nI believe you've misunderstood my intent here. I sent rebased versions\nof your patches with fixup commits implementing fixes that I'd\npreviously sent. I don't think that's unusual. I hope your patches\nwill be included in v16, and I hope to facilitate that. I don't mean\nany offense. Actually, the fixups are provided as separate patches so\nyou can adopt the changes easily into your branch. \n\n-- \nJustin\n\n\n","msg_date":"Sun, 15 Jan 2023 20:47:13 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sun, Jan 15, 2023 at 07:56:25PM -0600, Justin Pryzby wrote:\n> On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\n>> The functions changed by 0001 are cfopen[_write](),\n>> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\n>> to change these interfaces which basically exist to handle inputs?\n> \n> I changed to pass pg_compress_specification as a pointer, since that's\n> the usual convention for structs, as followed by the existing uses of\n> pg_compress_specification.\n\nOkay, but what do we gain here? It seems to me that this introduces\nthe risk that a careless change in one of the internal routines if\nthey change slight;ly compress_spec, hence impacting any of their\ncallers? Or is that fixing an actual bug (except if I am missing your\npoint, that does not seem to be the case)? \n\n>> Is there some benefit in changing compression_spec within the\n>> internals of these routines before going back one layer down to their\n>> callers? Changing the compression_spec on-the-fly in these internal\n>> paths could be risky, actually, no?\n> \n> I think what you're saying is that if the spec is passed as a pointer,\n> then the called functions shouldn't set spec->algorithm=something.\n\nYes. HEAD makes sure of that, 0001 would not prevent that. So I am a\nbit confused in seeing how this is a benefit.\n--\nMichael","msg_date":"Mon, 16 Jan 2023 11:54:46 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Hi,\n\nI admit I am completely at lost as to what is expected from me anymore.\n\nI had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\nPlease find a rebased v20 attached.\n\nAlso please let me know if I should silently step away from it and let other people lead\nit. I would be glad to comply either way.\n\nCheers,\n//Georgios\n\n\n------- Original Message -------\nOn Monday, January 16th, 2023 at 3:54 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n>\n>\n> On Sun, Jan 15, 2023 at 07:56:25PM -0600, Justin Pryzby wrote:\n>\n> > On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\n> >\n> > > The functions changed by 0001 are cfopen_write,\n> > > AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\n> > > to change these interfaces which basically exist to handle inputs?\n> >\n> > I changed to pass pg_compress_specification as a pointer, since that's\n> > the usual convention for structs, as followed by the existing uses of\n> > pg_compress_specification.\n>\n>\n> Okay, but what do we gain here? It seems to me that this introduces\n> the risk that a careless change in one of the internal routines if\n> they change slight;ly compress_spec, hence impacting any of their\n> callers? Or is that fixing an actual bug (except if I am missing your\n> point, that does not seem to be the case)?\n>\n> > > Is there some benefit in changing compression_spec within the\n> > > internals of these routines before going back one layer down to their\n> > > callers? Changing the compression_spec on-the-fly in these internal\n> > > paths could be risky, actually, no?\n> >\n> > I think what you're saying is that if the spec is passed as a pointer,\n> > then the called functions shouldn't set spec->algorithm=something.\n>\n>\n> Yes. HEAD makes sure of that, 0001 would not prevent that. So I am a\n> bit confused in seeing how this is a benefit.\n> --\n> Michael","msg_date":"Mon, 16 Jan 2023 15:14:43 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Hi,\n\nOn 1/16/23 16:14, gkokolatos@pm.me wrote:\n> Hi,\n> \n> I admit I am completely at lost as to what is expected from me anymore.\n> \n\n:-(\n\nI understand it's frustrating not to know why a patch is not moving\nforward. Particularly when is seems fairly straightforward ...\n\nLet me briefly explain my personal (and admittedly very subjective) view\non picking what patches to review/commit. I'm sure other committers have\nother criteria, but maybe this will help.\n\nThere are always more patches than I can review/commit, so I have to\nprioritize, and pick which patches to look at. For me, it's mostly about\ncost/benefit of the patch. The cost is e.g. the amount of time I need to\nspend to review/commit the stuff, maybe read the thread, etc. Benefits\nis mainly the new features/improvements.\n\nIt's oversimplified, we could talk about various bits that contribute to\nthe costs and benefits, but this is what it boils down.\n\nThere's always the aspect of time - patches A and B have roughly the\nsame benefits, but with A we get it \"immediately\" while B requires\nadditional parts that we don't have ready yet (and if they don't make it\nwe get no benefit), I'll probably pick A.\n\nUnfortunately, this plays against this patch - I'm certainly in favor of\nadding lz4 (and other compression algos) into pg_dump, but if I commit\n0001 we get little benefit, and the other parts actually adding lz4/zstd\nare treated as \"WIP / for completeness\" so it's unclear when we'd get to\ncommit them.\n\nSo if I could recommend one thing, it'd be to get at least one of those\nWIP patches into a shape that's likely committable right after 0001.\n\n> I had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\n> Please find a rebased v20 attached.\n> \n\nI took a quick look at 0001, so a couple comments (sorry if some of this\nwas already discussed in the thread):\n\n1) I don't think a \"refactoring\" patch should reference particular\ncompression algorithms (lz4/zstd), and in particular I don't think we\nshould have \"not yet implemented\" messages. We only have a couple other\nplaces doing that, when we didn't have a better choice. But here we can\nsimply reject the algorithm when parsing the options, we don't need to\ndo that in a dozen other places.\n\n2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\n\"none\" at the end. It might make backpatches harder.\n\n3) While building, I get bunch of warnings about missing cfdopen()\nprototype and pg_backup_archiver.c not knowing about cfdopen() and\nadding an implicit prototype (so I doubt it actually works).\n\n4) \"cfp\" struct no longer wraps gzFile, but the comment was not updated.\nFWIW I'm not sure switching to \"void *\" is an improvement, maybe it'd be\nbetter to have a \"union\" of correct types?\n\n5) cfopen/cfdopen are missing comments. cfopen_internal has an updated\ncomment, but that's a static function while cfopen/cfdopen are the\nactual API.\n\n> Also please let me know if I should silently step away from it and let other people lead\n> it. I would be glad to comply either way.\n> \n\nPlease don't. I promise to take a look at this patch again.\n\nThanks for doing all the work.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Wed, 18 Jan 2023 15:00:34 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Wednesday, January 18th, 2023 at 3:00 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> Hi,\n> \n> On 1/16/23 16:14, gkokolatos@pm.me wrote:\n> \n> > Hi,\n> > \n> > I admit I am completely at lost as to what is expected from me anymore.\n> \n<snip>\n> \n> Unfortunately, this plays against this patch - I'm certainly in favor of\n> adding lz4 (and other compression algos) into pg_dump, but if I commit\n> 0001 we get little benefit, and the other parts actually adding lz4/zstd\n> are treated as \"WIP / for completeness\" so it's unclear when we'd get to\n> commit them.\n\nThank you for your kindness and for taking the time to explain.\n \n> So if I could recommend one thing, it'd be to get at least one of those\n> WIP patches into a shape that's likely committable right after 0001.\n\nThis was clearly my fault. I misunderstood a suggestion upthread to focus\non the first patch of the series and ignore documentation and comments on\nthe rest.\n\nPlease find v21 to contain 0002 and 0003 in a state which I no longer consider\nas WIP but worthy of proper consideration. Some guidance on where is best to add\ndocumentation in 0002 for the function pointers in CompressFileHandle will\nbe welcomed.\n\n> \n> > I had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\n> > Please find a rebased v20 attached.\n> \n> \n> I took a quick look at 0001, so a couple comments (sorry if some of this\n> was already discussed in the thread):\n\nMuch appreciated!\n\n> \n> 1) I don't think a \"refactoring\" patch should reference particular\n> compression algorithms (lz4/zstd), and in particular I don't think we\n> should have \"not yet implemented\" messages. We only have a couple other\n> places doing that, when we didn't have a better choice. But here we can\n> simply reject the algorithm when parsing the options, we don't need to\n> do that in a dozen other places.\n\nI have now removed lz4/zstd from where they were present with the exception\nof pg_dump.c which is responsible for parsing.\n\n> 2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\n> \"none\" at the end. It might make backpatches harder.\n\nAgreed. However a 'default' is needed in order to avoid compilation warnings.\nAlso note that 0002 completely does away with cases within WriteDataToArchive.\n\n> 3) While building, I get bunch of warnings about missing cfdopen()\n> prototype and pg_backup_archiver.c not knowing about cfdopen() and\n> adding an implicit prototype (so I doubt it actually works).\n\nFixed. cfdopen() got prematurely introduced in 5e73a6048 and then got removed\nin 69fb29d1af. v20 failed to properly take 69fb29d1af in consideration. Note\nthat cfdopen is removed in 0002 which explains why cfbot didn't complain.\n \n> 4) \"cfp\" struct no longer wraps gzFile, but the comment was not updated.\n> FWIW I'm not sure switching to \"void *\" is an improvement, maybe it'd be\n> better to have a \"union\" of correct types?\n\nPlease find and updated comment and a union in place of the void *. Also\nnote that 0002 completely does away with cfp in favour of a new struct\nCompressFileHandle. I maintained the void * there because it is used by\nprivate methods of the compressors. 0003 contains such an example with\nLZ4CompressorState.\n\n> 5) cfopen/cfdopen are missing comments. cfopen_internal has an updated\n> comment, but that's a static function while cfopen/cfdopen are the\n> actual API.\n\nAdded comments to cfopen/cfdopen.\n\n> \n> > Also please let me know if I should silently step away from it and let other people lead\n> > it. I would be glad to comply either way.\n> \n> \n> Please don't. I promise to take a look at this patch again.\n\nThank you very much.\n\n> Thanks for doing all the work.\n\nThank you.\n\nCheers,\n//Georgios\n\n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Wed, 18 Jan 2023 19:05:43 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 1/18/23 20:05, gkokolatos@pm.me wrote:\n> \n> \n> \n> \n> \n> ------- Original Message -------\n> On Wednesday, January 18th, 2023 at 3:00 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n> \n> \n>>\n>>\n>> Hi,\n>>\n>> On 1/16/23 16:14, gkokolatos@pm.me wrote:\n>>\n>>> Hi,\n>>>\n>>> I admit I am completely at lost as to what is expected from me anymore.\n>>\n> <snip>\n>>\n>> Unfortunately, this plays against this patch - I'm certainly in favor of\n>> adding lz4 (and other compression algos) into pg_dump, but if I commit\n>> 0001 we get little benefit, and the other parts actually adding lz4/zstd\n>> are treated as \"WIP / for completeness\" so it's unclear when we'd get to\n>> commit them.\n> \n> Thank you for your kindness and for taking the time to explain.\n> \n>> So if I could recommend one thing, it'd be to get at least one of those\n>> WIP patches into a shape that's likely committable right after 0001.\n> \n> This was clearly my fault. I misunderstood a suggestion upthread to focus\n> on the first patch of the series and ignore documentation and comments on\n> the rest.\n> \n> Please find v21 to contain 0002 and 0003 in a state which I no longer consider\n> as WIP but worthy of proper consideration. Some guidance on where is best to add\n> documentation in 0002 for the function pointers in CompressFileHandle will\n> be welcomed.\n> \n\nThis is internal-only API, not meant for use by regular users and/or\nextension authors, so I don't think we need sgml docs. I'd just add\nregular code-level documentation to compress_io.h.\n\nFor inspiration see docs for \"struct ReorderBuffer\" in reorderbuffer.h,\nor \"struct _archiveHandle\" in pg_backup_archiver.h.\n\nOr what other kind of documentation you had in mind?\n\n>>\n>>> I had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\n>>> Please find a rebased v20 attached.\n>>\n>>\n>> I took a quick look at 0001, so a couple comments (sorry if some of this\n>> was already discussed in the thread):\n> \n> Much appreciated!\n> \n>>\n>> 1) I don't think a \"refactoring\" patch should reference particular\n>> compression algorithms (lz4/zstd), and in particular I don't think we\n>> should have \"not yet implemented\" messages. We only have a couple other\n>> places doing that, when we didn't have a better choice. But here we can\n>> simply reject the algorithm when parsing the options, we don't need to\n>> do that in a dozen other places.\n> \n> I have now removed lz4/zstd from where they were present with the exception\n> of pg_dump.c which is responsible for parsing.\n> \n\nI'm not sure I understand why leave the lz4/zstd in this place?\n\n>> 2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\n>> \"none\" at the end. It might make backpatches harder.\n> \n> Agreed. However a 'default' is needed in order to avoid compilation warnings.\n> Also note that 0002 completely does away with cases within WriteDataToArchive.\n> \n\nOK, although that's also a consequence of using a \"switch\" instead of\nplan \"if\" branches.\n\nFurthermore, I'm not sure we really need the pg_fatal() about invalid\ncompression method in these default blocks. I mean, how could we even\nget to these places when the build does not support the algorithm? All\nof this (ReadDataFromArchive, WriteDataToArchive, EndCompressor, ...)\nhappens looong after the compressor was initialized and the method\nchecked, no? So maybe either this should simply do Assert(false) or use\na different error message.\n\n>> 3) While building, I get bunch of warnings about missing cfdopen()\n>> prototype and pg_backup_archiver.c not knowing about cfdopen() and\n>> adding an implicit prototype (so I doubt it actually works).\n> \n> Fixed. cfdopen() got prematurely introduced in 5e73a6048 and then got removed\n> in 69fb29d1af. v20 failed to properly take 69fb29d1af in consideration. Note\n> that cfdopen is removed in 0002 which explains why cfbot didn't complain.\n> \n\nOK.\n\n>> 4) \"cfp\" struct no longer wraps gzFile, but the comment was not updated.\n>> FWIW I'm not sure switching to \"void *\" is an improvement, maybe it'd be\n>> better to have a \"union\" of correct types?\n> \n> Please find and updated comment and a union in place of the void *. Also\n> note that 0002 completely does away with cfp in favour of a new struct\n> CompressFileHandle. I maintained the void * there because it is used by\n> private methods of the compressors. 0003 contains such an example with\n> LZ4CompressorState.\n> \n\nI wonder if this (and also the previous item) makes sense to keep 0001\nand 0002 or to combine them. The \"intermediate\" state is a bit annoying.\n\n>> 5) cfopen/cfdopen are missing comments. cfopen_internal has an updated\n>> comment, but that's a static function while cfopen/cfdopen are the\n>> actual API.\n> \n> Added comments to cfopen/cfdopen.\n> \n\nOK.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 19 Jan 2023 16:45:10 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Thursday, January 19th, 2023 at 4:45 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> On 1/18/23 20:05, gkokolatos@pm.me wrote:\n> \n> > ------- Original Message -------\n> > On Wednesday, January 18th, 2023 at 3:00 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\n> > \n> > > Hi,\n> > > \n> > > On 1/16/23 16:14, gkokolatos@pm.me wrote:\n> > > \n> > > > Hi,\n> > > > \n> > > > I admit I am completely at lost as to what is expected from me anymore.\n> > \n> > <snip>\n> > \n> > > Unfortunately, this plays against this patch - I'm certainly in favor of\n> > > adding lz4 (and other compression algos) into pg_dump, but if I commit\n> > > 0001 we get little benefit, and the other parts actually adding lz4/zstd\n> > > are treated as \"WIP / for completeness\" so it's unclear when we'd get to\n> > > commit them.\n> > \n> > Thank you for your kindness and for taking the time to explain.\n> > \n> > > So if I could recommend one thing, it'd be to get at least one of those\n> > > WIP patches into a shape that's likely committable right after 0001.\n> > \n> > This was clearly my fault. I misunderstood a suggestion upthread to focus\n> > on the first patch of the series and ignore documentation and comments on\n> > the rest.\n> > \n> > Please find v21 to contain 0002 and 0003 in a state which I no longer consider\n> > as WIP but worthy of proper consideration. Some guidance on where is best to add\n> > documentation in 0002 for the function pointers in CompressFileHandle will\n> > be welcomed.\n> \n> \n> This is internal-only API, not meant for use by regular users and/or\n> extension authors, so I don't think we need sgml docs. I'd just add\n> regular code-level documentation to compress_io.h.\n> \n> For inspiration see docs for \"struct ReorderBuffer\" in reorderbuffer.h,\n> or \"struct _archiveHandle\" in pg_backup_archiver.h.\n> \n> Or what other kind of documentation you had in mind?\n\nThis is exactly what I was after. I was between compress_io.c and compress_io.h.\nThank you.\n\n> > > > I had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\n> > > > Please find a rebased v20 attached.\n> > > \n> > > I took a quick look at 0001, so a couple comments (sorry if some of this\n> > > was already discussed in the thread):\n> > \n> > Much appreciated!\n> > \n> > > 1) I don't think a \"refactoring\" patch should reference particular\n> > > compression algorithms (lz4/zstd), and in particular I don't think we\n> > > should have \"not yet implemented\" messages. We only have a couple other\n> > > places doing that, when we didn't have a better choice. But here we can\n> > > simply reject the algorithm when parsing the options, we don't need to\n> > > do that in a dozen other places.\n> > \n> > I have now removed lz4/zstd from where they were present with the exception\n> > of pg_dump.c which is responsible for parsing.\n> \n> \n> I'm not sure I understand why leave the lz4/zstd in this place?\n\nYou are right, it is not obvious. Those were added in 5e73a60488 which is\nalready committed in master and I didn't want to backtrack. Of course, I am\nnot opposing in doing so if you wish.\n\n> \n> > > 2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\n> > > \"none\" at the end. It might make backpatches harder.\n> > \n> > Agreed. However a 'default' is needed in order to avoid compilation warnings.\n> > Also note that 0002 completely does away with cases within WriteDataToArchive.\n> \n> \n> OK, although that's also a consequence of using a \"switch\" instead of\n> plan \"if\" branches.\n> \n> Furthermore, I'm not sure we really need the pg_fatal() about invalid\n> compression method in these default blocks. I mean, how could we even\n> get to these places when the build does not support the algorithm? All\n> of this (ReadDataFromArchive, WriteDataToArchive, EndCompressor, ...)\n> happens looong after the compressor was initialized and the method\n> checked, no? So maybe either this should simply do Assert(false) or use\n> a different error message.\n\nI like Assert(false).\n\n> > > 3) While building, I get bunch of warnings about missing cfdopen()\n> > > prototype and pg_backup_archiver.c not knowing about cfdopen() and\n> > > adding an implicit prototype (so I doubt it actually works).\n> > \n> > Fixed. cfdopen() got prematurely introduced in 5e73a6048 and then got removed\n> > in 69fb29d1af. v20 failed to properly take 69fb29d1af in consideration. Note\n> > that cfdopen is removed in 0002 which explains why cfbot didn't complain.\n> \n> \n> OK.\n> \n> > > 4) \"cfp\" struct no longer wraps gzFile, but the comment was not updated.\n> > > FWIW I'm not sure switching to \"void *\" is an improvement, maybe it'd be\n> > > better to have a \"union\" of correct types?\n> > \n> > Please find and updated comment and a union in place of the void *. Also\n> > note that 0002 completely does away with cfp in favour of a new struct\n> > CompressFileHandle. I maintained the void * there because it is used by\n> > private methods of the compressors. 0003 contains such an example with\n> > LZ4CompressorState.\n> \n> \n> I wonder if this (and also the previous item) makes sense to keep 0001\n> and 0002 or to combine them. The \"intermediate\" state is a bit annoying.\n\nAgreed. It was initially submitted as one patch. Then it was requested to be\nsplit up in two parts, one to expand the use of the existing API and one to\nreplace with the new interface. Unfortunately the expansion of usage of the\nexisting API requires some tweaking, but that is not a very good reason for\nthe current patch set. I should have done a better job there.\n\nPlease find v22 attach which combines back 0001 and 0002. It is missing the\ndocumentation that was discussed above as I wanted to give a quick feedback.\nLet me know if you think that the combined version is the one to move forward\nwith.\n\nCheers,\n//Georgios\n\n> \n> > > 5) cfopen/cfdopen are missing comments. cfopen_internal has an updated\n> > > comment, but that's a static function while cfopen/cfdopen are the\n> > > actual API.\n> > \n> > Added comments to cfopen/cfdopen.\n> \n> \n> OK.\n> \n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Thu, 19 Jan 2023 16:42:06 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Hi,\n\nOn 1/19/23 17:42, gkokolatos@pm.me wrote:\n> \n> ------- Original Message -------\n> On Thursday, January 19th, 2023 at 4:45 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n>>\n>> On 1/18/23 20:05, gkokolatos@pm.me wrote:\n>>\n>>> ------- Original Message -------\n>>> On Wednesday, January 18th, 2023 at 3:00 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\n>>\n>> I'm not sure I understand why leave the lz4/zstd in this place?\n> \n> You are right, it is not obvious. Those were added in 5e73a60488 which is\n> already committed in master and I didn't want to backtrack. Of course, I am\n> not opposing in doing so if you wish.\n> \n\nAh, I didn't realize it was already added by earlier commit. In that\ncase let's not worry about it.\n\n>>\n>>>> 2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\n>>>> \"none\" at the end. It might make backpatches harder.\n>>>\n>>> Agreed. However a 'default' is needed in order to avoid compilation warnings.\n>>> Also note that 0002 completely does away with cases within WriteDataToArchive.\n>>\n>>\n>> OK, although that's also a consequence of using a \"switch\" instead of\n>> plan \"if\" branches.\n>>\n>> Furthermore, I'm not sure we really need the pg_fatal() about invalid\n>> compression method in these default blocks. I mean, how could we even\n>> get to these places when the build does not support the algorithm? All\n>> of this (ReadDataFromArchive, WriteDataToArchive, EndCompressor, ...)\n>> happens looong after the compressor was initialized and the method\n>> checked, no? So maybe either this should simply do Assert(false) or use\n>> a different error message.\n> \n> I like Assert(false).\n> \n\nOK, good. Do you agree we should never actually get there, if the\nearlier checks work correctly?\n\n>>\n>>>> 4) \"cfp\" struct no longer wraps gzFile, but the comment was not updated.\n>>>> FWIW I'm not sure switching to \"void *\" is an improvement, maybe it'd be\n>>>> better to have a \"union\" of correct types?\n>>>\n>>> Please find and updated comment and a union in place of the void *. Also\n>>> note that 0002 completely does away with cfp in favour of a new struct\n>>> CompressFileHandle. I maintained the void * there because it is used by\n>>> private methods of the compressors. 0003 contains such an example with\n>>> LZ4CompressorState.\n>>\n>>\n>> I wonder if this (and also the previous item) makes sense to keep 0001\n>> and 0002 or to combine them. The \"intermediate\" state is a bit annoying.\n> \n> Agreed. It was initially submitted as one patch. Then it was requested to be\n> split up in two parts, one to expand the use of the existing API and one to\n> replace with the new interface. Unfortunately the expansion of usage of the\n> existing API requires some tweaking, but that is not a very good reason for\n> the current patch set. I should have done a better job there.\n> \n> Please find v22 attach which combines back 0001 and 0002. It is missing the\n> documentation that was discussed above as I wanted to give a quick feedback.\n> Let me know if you think that the combined version is the one to move forward\n> with.\n> \n\nThanks, I'll take a look.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 19 Jan 2023 18:55:43 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 1/19/23 18:55, Tomas Vondra wrote:\n> Hi,\n> \n> On 1/19/23 17:42, gkokolatos@pm.me wrote:\n>>\n>> ...\n>>\n>> Agreed. It was initially submitted as one patch. Then it was requested to be\n>> split up in two parts, one to expand the use of the existing API and one to\n>> replace with the new interface. Unfortunately the expansion of usage of the\n>> existing API requires some tweaking, but that is not a very good reason for\n>> the current patch set. I should have done a better job there.\n>>\n>> Please find v22 attach which combines back 0001 and 0002. It is missing the\n>> documentation that was discussed above as I wanted to give a quick feedback.\n>> Let me know if you think that the combined version is the one to move forward\n>> with.\n>>\n> \n> Thanks, I'll take a look.\n> \n\nAfter taking a look and thinking about it a bit more, I think we should\nkeep the two parts separate. I think Michael (or whoever proposed) the\nsplit was right, it makes the patches easier to grok.\n\nSorry for the noise, hopefully we can just revert to the last version.\n\nWhile reading the thread, I also noticed this:\n\n> By the way, I think that this 0002 should drop all the default clauses\n> in the switches for the compression method so as we'd catch any\n> missing code paths with compiler warnings if a new compression method\n> is added in the future.\n\nNow I realize why there were \"not yet implemented\" errors for lz4/zstd\nin all the switches, and why after removing them you had to add a\ndefault branch.\n\nWe DON'T want a default branch, because the idea is that after adding a\nnew compression algorithm, we get warnings about switches not handling\nit correctly.\n\nSo I guess we should walk back this change too :-( It's probably easier\nto go back to v20 from January 16, and redo the couple remaining things\nI commented on.\n\n\nFWIW I think this is a hint that adding LZ4/ZSTD options, in 5e73a6048,\nbut without implementation, was not a great idea. It mostly defeats the\nidea of getting the compiler warnings - all the places already handle\nPG_COMPRESSION_LZ4/PG_COMPRESSION_ZSTD by throwing a pg_fatal. So you'd\nhave to grep for the options, inspect all the places or something like\nthat anyway. The warnings would only work for entirely new methods.\n\nHowever, I now also realize the compressor API in 0002 replaces all of\nthis with calls to a generic API callback, so trying to improve this was\npretty silly from me.\n\n\nPlease, fix the couple remaining details in v20, add the docs for the\ncallbacks, and I'll try to polish it and get it committed.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 20 Jan 2023 00:34:34 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Friday, January 20th, 2023 at 12:34 AM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> On 1/19/23 18:55, Tomas Vondra wrote:\n> \n> > Hi,\n> > \n> > On 1/19/23 17:42, gkokolatos@pm.me wrote:\n> > \n> > > ...\n> > > \n> > > Agreed. It was initially submitted as one patch. Then it was requested to be\n> > > split up in two parts, one to expand the use of the existing API and one to\n> > > replace with the new interface. Unfortunately the expansion of usage of the\n> > > existing API requires some tweaking, but that is not a very good reason for\n> > > the current patch set. I should have done a better job there.\n> > > \n> > > Please find v22 attach which combines back 0001 and 0002. It is missing the\n> > > documentation that was discussed above as I wanted to give a quick feedback.\n> > > Let me know if you think that the combined version is the one to move forward\n> > > with.\n> > \n> > Thanks, I'll take a look.\n> \n> \n> After taking a look and thinking about it a bit more, I think we should\n> keep the two parts separate. I think Michael (or whoever proposed) the\n> split was right, it makes the patches easier to grok.\n> \n\nExcellent. I will attempt a better split this time round.\n\n> \n> While reading the thread, I also noticed this:\n> \n> > By the way, I think that this 0002 should drop all the default clauses\n> > in the switches for the compression method so as we'd catch any\n> > missing code paths with compiler warnings if a new compression method\n> > is added in the future.\n> \n> \n> Now I realize why there were \"not yet implemented\" errors for lz4/zstd\n> in all the switches, and why after removing them you had to add a\n> default branch.\n> \n> We DON'T want a default branch, because the idea is that after adding a\n> new compression algorithm, we get warnings about switches not handling\n> it correctly.\n> \n> So I guess we should walk back this change too :-( It's probably easier\n> to go back to v20 from January 16, and redo the couple remaining things\n> I commented on.\n> \n\nSure.\n\n> \n> FWIW I think this is a hint that adding LZ4/ZSTD options, in 5e73a6048,\n> but without implementation, was not a great idea. It mostly defeats the\n> idea of getting the compiler warnings - all the places already handle\n> PG_COMPRESSION_LZ4/PG_COMPRESSION_ZSTD by throwing a pg_fatal. So you'd\n> have to grep for the options, inspect all the places or something like\n> that anyway. The warnings would only work for entirely new methods.\n> \n> However, I now also realize the compressor API in 0002 replaces all of\n> this with calls to a generic API callback, so trying to improve this was\n> pretty silly from me.\n\nI can try to do a better job at splitting things up.\n\n> \n> Please, fix the couple remaining details in v20, add the docs for the\n> callbacks, and I'll try to polish it and get it committed.\n\nExcellent. Allow me an attempt to polish and expect a new version soon.\n\nCheers,\n//Georgios\n\n> \n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 20 Jan 2023 15:03:30 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Friday, January 20th, 2023 at 12:34 AM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n>\n>\n> On 1/19/23 18:55, Tomas Vondra wrote:\n>\n> > Hi,\n> >\n> > On 1/19/23 17:42, gkokolatos@pm.me wrote:\n> >\n> > > ...\n> > >\n> > > Agreed. It was initially submitted as one patch. Then it was requested to be\n> > > split up in two parts, one to expand the use of the existing API and one to\n> > > replace with the new interface. Unfortunately the expansion of usage of the\n> > > existing API requires some tweaking, but that is not a very good reason for\n> > > the current patch set. I should have done a better job there.\n> > >\n> > > Please find v22 attach which combines back 0001 and 0002. It is missing the\n> > > documentation that was discussed above as I wanted to give a quick feedback.\n> > > Let me know if you think that the combined version is the one to move forward\n> > > with.\n> >\n> > Thanks, I'll take a look.\n>\n>\n> After taking a look and thinking about it a bit more, I think we should\n> keep the two parts separate. I think Michael (or whoever proposed) the\n> split was right, it makes the patches easier to grok.\n\nPlease find attached v23 which reintroduces the split.\n\n0001 is reworked to have a reduced footprint than before. Also in an attempt\nto facilitate the readability, 0002 splits the API's and the uncompressed\nimplementation in separate files.\n\n>\n> While reading the thread, I also noticed this:\n>\n> > By the way, I think that this 0002 should drop all the default clauses\n> > in the switches for the compression method so as we'd catch any\n> > missing code paths with compiler warnings if a new compression method\n> > is added in the future.\n>\n>\n> Now I realize why there were \"not yet implemented\" errors for lz4/zstd\n> in all the switches, and why after removing them you had to add a\n> default branch.\n>\n> We DON'T want a default branch, because the idea is that after adding a\n> new compression algorithm, we get warnings about switches not handling\n> it correctly.\n>\n> So I guess we should walk back this change too :-( It's probably easier\n> to go back to v20 from January 16, and redo the couple remaining things\n> I commented on.\n\nNo problem.\n\n> FWIW I think this is a hint that adding LZ4/ZSTD options, in 5e73a6048,\n> but without implementation, was not a great idea. It mostly defeats the\n> idea of getting the compiler warnings - all the places already handle\n> PG_COMPRESSION_LZ4/PG_COMPRESSION_ZSTD by throwing a pg_fatal. So you'd\n> have to grep for the options, inspect all the places or something like\n> that anyway. The warnings would only work for entirely new methods.\n>\n> However, I now also realize the compressor API in 0002 replaces all of\n> this with calls to a generic API callback, so trying to improve this was\n> pretty silly from me.\n>\n>\n> Please, fix the couple remaining details in v20, add the docs for the\n> callbacks, and I'll try to polish it and get it committed.\n\nThank you very much. Please find an attempt to comply with the requested\nchanges in the attached.\n\n>\n>\n> regards\n>\n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Mon, 23 Jan 2023 17:31:55 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\n> Please find attached v23 which reintroduces the split.\n> \n> 0001 is reworked to have a reduced footprint than before. Also in an attempt\n> to facilitate the readability, 0002 splits the API's and the uncompressed\n> implementation in separate files.\n\nThanks for updating the patch. Could you address the review comments I\nsent here ?\nhttps://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\n\nThanks,\n-- \nJustin\n\n\n","msg_date":"Mon, 23 Jan 2023 12:00:14 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n> \n> \n> On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\n> \n> > Please find attached v23 which reintroduces the split.\n> > \n> > 0001 is reworked to have a reduced footprint than before. Also in an attempt\n> > to facilitate the readability, 0002 splits the API's and the uncompressed\n> > implementation in separate files.\n> \n> \n> Thanks for updating the patch. Could you address the review comments I\n> sent here ?\n> https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\n\nPlease find v24 attached.\n\nCheers,\n//Georgios\n\n> \n> Thanks,\n> --\n> Justin","msg_date":"Tue, 24 Jan 2023 15:56:20 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Jan 24, 2023 at 03:56:20PM +0000, gkokolatos@pm.me wrote:\n> On Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\n> > \n> > > Please find attached v23 which reintroduces the split.\n> > > \n> > > 0001 is reworked to have a reduced footprint than before. Also in an attempt\n> > > to facilitate the readability, 0002 splits the API's and the uncompressed\n> > > implementation in separate files.\n> > \n> > Thanks for updating the patch. Could you address the review comments I\n> > sent here ?\n> > https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\n> \n> Please find v24 attached.\n\nThanks for updating the patch.\n\nIn 001, RestoreArchive() does:\n\n> -#ifndef HAVE_LIBZ\n> - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\n> - AH->PrintTocDataPtr != NULL)\n> + supports_compression = false;\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_NONE ||\n> + AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> + supports_compression = true;\n> +\n> + if (AH->PrintTocDataPtr != NULL)\n> {\n> for (te = AH->toc->next; te != AH->toc; te = te->next)\n> {\n> if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\n> - pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n> + {\n> +#ifndef HAVE_LIBZ\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> + supports_compression = false;\n> +#endif\n> + if (supports_compression == false)\n> + pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n> + }\n> }\n> }\n> -#endif\n\nThis first checks if the algorithm is implemented, and then checks if\nthe algorithm is supported by the current build - that confused me for a\nbit. It seems unnecessary to check for unimplemented algorithms before\nlooping. That also requires referencing both GZIP and LZ4 in two\nplaces.\n\nI think it could be written to avoid the need to change for added\ncompression algorithms:\n\n+ if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\n+ {\n+ /* Check if the compression algorithm is supported */\n+ pg_compress_specification spec;\n+ parse_compress_specification(AH->compression_spec.algorithm, NULL, &spec);\n+ if (spec->parse_error != NULL)\n+ pg_fatal(spec->parse_error);\n+ }\n\nOr maybe add a new function to compression.c to indicate whether a given\nalgorithm is supported.\n\nThat would also indicate *which* compression library isn't supported.\n\nOther than that, I think 001 is ready.\n\n002/003 use these names, which I think are too similar - initially I\ndidn't even realize there were two separate functions (each with a\nsecond stub function to handle the case of unsupported compression):\n\n+extern void InitCompressorGzip(CompressorState *cs, const pg_compress_specification compression_spec); \n+extern void InitCompressGzip(CompressFileHandle *CFH, const pg_compress_specification compression_spec); \n\n+extern void InitCompressorLZ4(CompressorState *cs, const pg_compress_specification compression_spec); \n+extern void InitCompressLZ4(CompressFileHandle *CFH, const pg_compress_specification compression_spec); \n\ntypo:\ns/not build with/not built with/\n\nShould AllocateCompressor() set cs->compression_spec, rather than doing\nit in each compressor ?\n\nThanks for considering.\n\n-- \nJustin\n\n\n","msg_date":"Tue, 24 Jan 2023 19:42:03 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Wednesday, January 25th, 2023 at 2:42 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n> \n> \n> On Tue, Jan 24, 2023 at 03:56:20PM +0000, gkokolatos@pm.me wrote:\n> \n> > On Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby pryzby@telsasoft.com wrote:\n> > \n> > > On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\n> > > \n> > > > Please find attached v23 which reintroduces the split.\n> > > > \n> > > > 0001 is reworked to have a reduced footprint than before. Also in an attempt\n> > > > to facilitate the readability, 0002 splits the API's and the uncompressed\n> > > > implementation in separate files.\n> > > \n> > > Thanks for updating the patch. Could you address the review comments I\n> > > sent here ?\n> > > https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\n> > \n> > Please find v24 attached.\n> \n> \n> Thanks for updating the patch.\n> \n> In 001, RestoreArchive() does:\n> \n> > -#ifndef HAVE_LIBZ\n> > - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\n> > - AH->PrintTocDataPtr != NULL)\n> > + supports_compression = false;\n> > + if (AH->compression_spec.algorithm == PG_COMPRESSION_NONE ||\n> > + AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> > + supports_compression = true;\n> > +\n> > + if (AH->PrintTocDataPtr != NULL)\n> > {\n> > for (te = AH->toc->next; te != AH->toc; te = te->next)\n> > {\n> > if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\n> > - pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n> > + {\n> > +#ifndef HAVE_LIBZ\n> > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> > + supports_compression = false;\n> > +#endif\n> > + if (supports_compression == false)\n> > + pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n> > + }\n> > }\n> > }\n> > -#endif\n> \n> \n> This first checks if the algorithm is implemented, and then checks if\n> the algorithm is supported by the current build - that confused me for a\n> bit. It seems unnecessary to check for unimplemented algorithms before\n> looping. That also requires referencing both GZIP and LZ4 in two\n> places.\n\nI am not certain that it is unnecessary, at least not in the way that is\ndescribed. The idea is that new compression methods can be added, without\nchanging the archive's version number. It is very possible that it is\nrequested to restore an archive compressed with a method not implemented\nin the current binary. The first check takes care of that and sets\nsupports_compression only for the supported versions. It is possible to\nenter the loop with supports_compression already set to false, for example\nbecause the archive was compressed with ZSTD, triggering the fatal error.\n\nOf course, one can throw the error before entering the loop, yet I think\nthat it does not help the readability of the code. IMHO it is easier to\nfollow if the error is thrown once during that check.\n\n> \n> I think it could be written to avoid the need to change for added\n> compression algorithms:\n> \n> + if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\n> \n> + {\n> + /* Check if the compression algorithm is supported */\n> + pg_compress_specification spec;\n> + parse_compress_specification(AH->compression_spec.algorithm, NULL, &spec);\n> \n> + if (spec->parse_error != NULL)\n> \n> + pg_fatal(spec->parse_error);\n> \n> + }\n\nI am not certain how that would work in the example with ZSTD above.\nIf I am not wrong, parse_compress_specification() will not throw an error\nif the codebase supports ZSTD, yet this specific pg_dump binary will not\nsupport it because ZSTD is not implemented. parse_compress_specification()\nis not aware of that and should not be aware of it, should it?\n\n> \n> Or maybe add a new function to compression.c to indicate whether a given\n> algorithm is supported.\n\nI am not certain how this would help, as compression.c is supposed to be\nused by multiple binaries while this is a pg_dump specific detail.\n\n> That would also indicate which compression library isn't supported.\n\nIf anything, I can suggest to throw an error much earlier, i.e. in ReadHead(),\nand remove altogether this check. On the other hand, I like the belts\nand suspenders approach because there are no more checks after this point.\n\n> Other than that, I think 001 is ready.\n\nThank you.\n\n> 002/003 use these names, which I think are too similar - initially I\n> didn't even realize there were two separate functions (each with a\n> second stub function to handle the case of unsupported compression):\n> \n> +extern void InitCompressorGzip(CompressorState *cs, const pg_compress_specification compression_spec);\n> +extern void InitCompressGzip(CompressFileHandle *CFH, const pg_compress_specification compression_spec);\n> \n> +extern void InitCompressorLZ4(CompressorState *cs, const pg_compress_specification compression_spec);\n> +extern void InitCompressLZ4(CompressFileHandle *CFH, const pg_compress_specification compression_spec);\n\nFair enough. Names are now updated.\n\n> \n> typo:\n> s/not build with/not built with/\n\nThank you.\n\n> \n> Should AllocateCompressor() set cs->compression_spec, rather than doing\n> it in each compressor ?\n\nI think that compression_spec should be owned by each compressor. With that\nin mind, it makes more sense to set it within each compressor. This is not\na hill I am willing to die on though.\n\nPlease find v25 attached.\n\n> \n> Thanks for considering.\n> \n> --\n> Justin","msg_date":"Wed, 25 Jan 2023 15:37:12 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\nOn 1/25/23 16:37, gkokolatos@pm.me wrote:\n> \n> \n> \n> \n> \n> ------- Original Message -------\n> On Wednesday, January 25th, 2023 at 2:42 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> \n> \n>>\n>>\n>> On Tue, Jan 24, 2023 at 03:56:20PM +0000, gkokolatos@pm.me wrote:\n>>\n>>> On Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby pryzby@telsasoft.com wrote:\n>>>\n>>>> On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\n>>>>\n>>>>> Please find attached v23 which reintroduces the split.\n>>>>>\n>>>>> 0001 is reworked to have a reduced footprint than before. Also in an attempt\n>>>>> to facilitate the readability, 0002 splits the API's and the uncompressed\n>>>>> implementation in separate files.\n>>>>\n>>>> Thanks for updating the patch. Could you address the review comments I\n>>>> sent here ?\n>>>> https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\n>>>\n>>> Please find v24 attached.\n>>\n>>\n>> Thanks for updating the patch.\n>>\n>> In 001, RestoreArchive() does:\n>>\n>>> -#ifndef HAVE_LIBZ\n>>> - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\n>>> - AH->PrintTocDataPtr != NULL)\n>>> + supports_compression = false;\n>>> + if (AH->compression_spec.algorithm == PG_COMPRESSION_NONE ||\n>>> + AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n>>> + supports_compression = true;\n>>> +\n>>> + if (AH->PrintTocDataPtr != NULL)\n>>> {\n>>> for (te = AH->toc->next; te != AH->toc; te = te->next)\n>>> {\n>>> if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\n>>> - pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n>>> + {\n>>> +#ifndef HAVE_LIBZ\n>>> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n>>> + supports_compression = false;\n>>> +#endif\n>>> + if (supports_compression == false)\n>>> + pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n>>> + }\n>>> }\n>>> }\n>>> -#endif\n>>\n>>\n>> This first checks if the algorithm is implemented, and then checks if\n>> the algorithm is supported by the current build - that confused me for a\n>> bit. It seems unnecessary to check for unimplemented algorithms before\n>> looping. That also requires referencing both GZIP and LZ4 in two\n>> places.\n> \n> I am not certain that it is unnecessary, at least not in the way that is\n> described. The idea is that new compression methods can be added, without\n> changing the archive's version number. It is very possible that it is\n> requested to restore an archive compressed with a method not implemented\n> in the current binary. The first check takes care of that and sets\n> supports_compression only for the supported versions. It is possible to\n> enter the loop with supports_compression already set to false, for example\n> because the archive was compressed with ZSTD, triggering the fatal error.\n> \n> Of course, one can throw the error before entering the loop, yet I think\n> that it does not help the readability of the code. IMHO it is easier to\n> follow if the error is thrown once during that check.\n> \n\nActually, I don't understand why 0001 moves the check into the loop. I\nmean, why not check HAVE_LIBZ before the loop?\n\n>>\n>> I think it could be written to avoid the need to change for added\n>> compression algorithms:\n>>\n>> + if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\n>>\n>> + {\n>> + /* Check if the compression algorithm is supported */\n>> + pg_compress_specification spec;\n>> + parse_compress_specification(AH->compression_spec.algorithm, NULL, &spec);\n>>\n>> + if (spec->parse_error != NULL)\n>>\n>> + pg_fatal(spec->parse_error);\n>>\n>> + }\n> \n> I am not certain how that would work in the example with ZSTD above.\n> If I am not wrong, parse_compress_specification() will not throw an error\n> if the codebase supports ZSTD, yet this specific pg_dump binary will not\n> support it because ZSTD is not implemented. parse_compress_specification()\n> is not aware of that and should not be aware of it, should it?\n> \n\nNot sure. What happens in a similar situation now? That is, when trying\nto deal with an archive gzip-compressed in a build without libz?\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 25 Jan 2023 18:28:49 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Jan 25, 2023 at 03:37:12PM +0000, gkokolatos@pm.me wrote:\n> Of course, one can throw the error before entering the loop, yet I think\n> that it does not help the readability of the code. IMHO it is easier to\n> follow if the error is thrown once during that check.\n\n> If anything, I can suggest to throw an error much earlier, i.e. in ReadHead(),\n> and remove altogether this check. On the other hand, I like the belts\n> and suspenders approach because there are no more checks after this point.\n\nWhile looking at this, I realized that commit 5e73a6048 introduced a\nregression:\n\n@@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\n\n- if (AH->compression != 0)\n- pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n+ if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n+ pg_fatal(\"archive is compressed, but this installation does not support compression\");\n\nBefore, it was possible to restore non-data chunks of a dump file, even\nif the current build didn't support its compression. But that's now\nimpossible - and it makes the code we're discussing in RestoreArchive()\nunreachable.\n\nI don't think we can currently test for that, since it requires creating a dump\nusing a build --with compression and then trying to restore using a build\n--without compression. The coverage report disagrees with me, though...\nhttps://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#3901\n\n> > I think it could be written to avoid the need to change for added\n> > compression algorithms:\n...\n> \n> I am not certain how that would work in the example with ZSTD above.\n> If I am not wrong, parse_compress_specification() will not throw an error\n> if the codebase supports ZSTD, yet this specific pg_dump binary will not\n> support it because ZSTD is not implemented. parse_compress_specification()\n> is not aware of that and should not be aware of it, should it?\n\nYou're right.\n\nI think the 001 patch should try to remove hardcoded references to\nLIBZ/GZIP, such that the later patches don't need to update those same\nplaces for LZ4. For example in ReadHead() and RestoreArchive(), and\nmaybe other places dealing with file extensions. Maybe that could be\ndone by adding a function specific to pg_dump indicating whether or not\nan algorithm is implemented and supported.\n\n-- \nJustin\n\n\n","msg_date":"Wed, 25 Jan 2023 12:00:20 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Wednesday, January 25th, 2023 at 6:28 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> \n> On 1/25/23 16:37, gkokolatos@pm.me wrote:\n> \n> > ------- Original Message -------\n> > On Wednesday, January 25th, 2023 at 2:42 AM, Justin Pryzby pryzby@telsasoft.com wrote:\n> > \n> > > On Tue, Jan 24, 2023 at 03:56:20PM +0000, gkokolatos@pm.me wrote:\n> > > \n> > > > On Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby pryzby@telsasoft.com wrote:\n> > > > \n> > > > > On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\n> > > > > \n> > > > > > Please find attached v23 which reintroduces the split.\n> > > > > > \n> > > > > > 0001 is reworked to have a reduced footprint than before. Also in an attempt\n> > > > > > to facilitate the readability, 0002 splits the API's and the uncompressed\n> > > > > > implementation in separate files.\n> > > > > \n> > > > > Thanks for updating the patch. Could you address the review comments I\n> > > > > sent here ?\n> > > > > https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\n> > > > \n> > > > Please find v24 attached.\n> > > \n> > > Thanks for updating the patch.\n> > > \n> > > In 001, RestoreArchive() does:\n> > > \n> > > > -#ifndef HAVE_LIBZ\n> > > > - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\n> > > > - AH->PrintTocDataPtr != NULL)\n> > > > + supports_compression = false;\n> > > > + if (AH->compression_spec.algorithm == PG_COMPRESSION_NONE ||\n> > > > + AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> > > > + supports_compression = true;\n> > > > +\n> > > > + if (AH->PrintTocDataPtr != NULL)\n> > > > {\n> > > > for (te = AH->toc->next; te != AH->toc; te = te->next)\n> > > > {\n> > > > if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\n> > > > - pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n> > > > + {\n> > > > +#ifndef HAVE_LIBZ\n> > > > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> > > > + supports_compression = false;\n> > > > +#endif\n> > > > + if (supports_compression == false)\n> > > > + pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n> > > > + }\n> > > > }\n> > > > }\n> > > > -#endif\n> > > \n> > > This first checks if the algorithm is implemented, and then checks if\n> > > the algorithm is supported by the current build - that confused me for a\n> > > bit. It seems unnecessary to check for unimplemented algorithms before\n> > > looping. That also requires referencing both GZIP and LZ4 in two\n> > > places.\n> > \n> > I am not certain that it is unnecessary, at least not in the way that is\n> > described. The idea is that new compression methods can be added, without\n> > changing the archive's version number. It is very possible that it is\n> > requested to restore an archive compressed with a method not implemented\n> > in the current binary. The first check takes care of that and sets\n> > supports_compression only for the supported versions. It is possible to\n> > enter the loop with supports_compression already set to false, for example\n> > because the archive was compressed with ZSTD, triggering the fatal error.\n> > \n> > Of course, one can throw the error before entering the loop, yet I think\n> > that it does not help the readability of the code. IMHO it is easier to\n> > follow if the error is thrown once during that check.\n> \n> \n> Actually, I don't understand why 0001 moves the check into the loop. I\n> mean, why not check HAVE_LIBZ before the loop?\n\nThe intention is to be able to restore archives that don't contain\ndata. In that case compression becomes irrelevant as only the data in\nan archive is compressed.\n\n> \n> > > I think it could be written to avoid the need to change for added\n> > > compression algorithms:\n> > > \n> > > + if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\n> > > \n> > > + {\n> > > + /* Check if the compression algorithm is supported */\n> > > + pg_compress_specification spec;\n> > > + parse_compress_specification(AH->compression_spec.algorithm, NULL, &spec);\n> > > \n> > > + if (spec->parse_error != NULL)\n> > > \n> > > + pg_fatal(spec->parse_error);\n> > > \n> > > + }\n> > \n> > I am not certain how that would work in the example with ZSTD above.\n> > If I am not wrong, parse_compress_specification() will not throw an error\n> > if the codebase supports ZSTD, yet this specific pg_dump binary will not\n> > support it because ZSTD is not implemented. parse_compress_specification()\n> > is not aware of that and should not be aware of it, should it?\n> \n> \n> Not sure. What happens in a similar situation now? That is, when trying\n> to deal with an archive gzip-compressed in a build without libz?\n\n\nIn case that there are no data chunks, the archive will be restored.\n\nCheers,\n//Georgios\n\n\n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 25 Jan 2023 19:50:49 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Wednesday, January 25th, 2023 at 7:00 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n> \n> \n> On Wed, Jan 25, 2023 at 03:37:12PM +0000, gkokolatos@pm.me wrote:\n> \n\n> While looking at this, I realized that commit 5e73a6048 introduced a\n> regression:\n> \n> @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\n> \n> - if (AH->compression != 0)\n> \n> - pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> \n> + pg_fatal(\"archive is compressed, but this installation does not support compression\");\n> \n> Before, it was possible to restore non-data chunks of a dump file, even\n> if the current build didn't support its compression. But that's now\n> impossible - and it makes the code we're discussing in RestoreArchive()\n> unreachable.\n\nNice catch!\n\nCheers,\n//Georgios\n\n> --\n> Justin\n\n\n","msg_date":"Wed, 25 Jan 2023 19:57:18 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Jan 25, 2023 at 07:57:18PM +0000, gkokolatos@pm.me wrote:\n> Nice catch!\n\nLet me see..\n--\nMichael","msg_date":"Thu, 26 Jan 2023 13:46:54 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Jan 25, 2023 at 12:00:20PM -0600, Justin Pryzby wrote:\n> While looking at this, I realized that commit 5e73a6048 introduced a\n> regression:\n> \n> @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\n> \n> - if (AH->compression != 0)\n> - pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> + pg_fatal(\"archive is compressed, but this installation does not support compression\");\n> \n> Before, it was possible to restore non-data chunks of a dump file, even\n> if the current build didn't support its compression. But that's now\n> impossible - and it makes the code we're discussing in RestoreArchive()\n> unreachable.\n\nRight. The impacts the possibility of looking at the header data,\nwhich is useful with pg_restore -l for example. On a dump that's been\ncompressed, pg_restore <= 15 would always print the TOC entries with\nor without compression support. On HEAD, this code prevents the\nheader lookup. All *nix or BSD platforms should have support for\nzlib, I hope.. Still that could be an issue on Windows, and this\nwould prevent folks to check the contents of the dumps after saving it\non a WIN32 host, so let's undo that.\n\nSo, I have been testing the attached with four sets of binaries from\n15/HEAD and with[out] zlib support, and this brings HEAD back to the\npre-15 state (header information able to show up, still failure when\nattempting to restore the dump's data without zlib).\n\n> I don't think we can currently test for that, since it requires creating a dump\n> using a build --with compression and then trying to restore using a build\n> --without compression.\n\nRight, the location of the data is in the header, and I don't see how\nyou would be able to do that without two sets of binaries at hand, but\nour tests run under the assumption that you have only one. Well,\nthat's not entirely true as well, as you could create a TAP test like\npg_upgrade that relies on a environment variable pointing to a second\nset of binaries. That's not worth the complication involved, IMO.\n\n> The coverage report disagrees with me, though...\n> https://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#3901\n\nIsn't that one of the tests like compression_gzip_plain?\n\nThoughts?\n--\nMichael","msg_date":"Thu, 26 Jan 2023 14:49:27 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Jan 26, 2023 at 02:49:27PM +0900, Michael Paquier wrote:\n> On Wed, Jan 25, 2023 at 12:00:20PM -0600, Justin Pryzby wrote:\n> > While looking at this, I realized that commit 5e73a6048 introduced a\n> > regression:\n> > \n> > @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\n> > \n> > - if (AH->compression != 0)\n> > - pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n> > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> > + pg_fatal(\"archive is compressed, but this installation does not support compression\");\n> > \n> > Before, it was possible to restore non-data chunks of a dump file, even\n> > if the current build didn't support its compression. But that's now\n> > impossible - and it makes the code we're discussing in RestoreArchive()\n> > unreachable.\n> \n> Right. The impacts the possibility of looking at the header data,\n> which is useful with pg_restore -l for example.\n\nIt's not just header data - it's schema and (I think) everything other\nthan table data.\n\n> > The coverage report disagrees with me, though...\n> > https://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#3901\n> \n> Isn't that one of the tests like compression_gzip_plain?\n\nI'm not sure what you mean. Plain dump is restored with psql and not\nwith pg_restore.\n\nMy line number was wrong:\nhttps://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#390\n\nWhat test would hit that code without rebuilding ?\n\n394 : #ifndef HAVE_LIBZ\n395 : if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\n\n> Thoughts?\n> #ifndef HAVE_LIBZ\n> \tif (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> -\t\tpg_fatal(\"archive is compressed, but this installation does not support compression\");\n> +\t\tpg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n\nYour patch is fine for now, but these errors should eventually specify\n*which* compression algorithm is unavailable. I think that should be a\npart of the 001 patch, ideally in a way that minimizes the number of\nplaces which need to be updated when adding an algorithm.\n\n-- \nJustin\n\n\n","msg_date":"Thu, 26 Jan 2023 00:28:46 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Thursday, January 26th, 2023 at 7:28 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n> \n> \n> On Thu, Jan 26, 2023 at 02:49:27PM +0900, Michael Paquier wrote:\n> \n> > On Wed, Jan 25, 2023 at 12:00:20PM -0600, Justin Pryzby wrote:\n> > \n> > > While looking at this, I realized that commit 5e73a6048 introduced a\n> > > regression:\n> > > \n> > > @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\n> > > \n> > > - if (AH->compression != 0)\n> > > - pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n> > > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> > > + pg_fatal(\"archive is compressed, but this installation does not support compression\");\n> > > \n> > > Before, it was possible to restore non-data chunks of a dump file, even\n> > > if the current build didn't support its compression. But that's now\n> > > impossible - and it makes the code we're discussing in RestoreArchive()\n> > > unreachable.\n> > \n> > Right. The impacts the possibility of looking at the header data,\n> > which is useful with pg_restore -l for example.\n> \n> \n> It's not just header data - it's schema and (I think) everything other\n> than table data.\n> \n> > > The coverage report disagrees with me, though...\n> > > https://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#3901\n> > \n> > Isn't that one of the tests like compression_gzip_plain?\n> \n> \n> I'm not sure what you mean. Plain dump is restored with psql and not\n> with pg_restore.\n> \n> My line number was wrong:\n> https://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#390\n> \n> What test would hit that code without rebuilding ?\n> \n> 394 : #ifndef HAVE_LIBZ\n> 395 : if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\n> \n> > Thoughts?\n> > #ifndef HAVE_LIBZ\n> > if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> > - pg_fatal(\"archive is compressed, but this installation does not support compression\");\n> > + pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n> \n> \n> Your patch is fine for now, but these errors should eventually specify\n> which compression algorithm is unavailable. I think that should be a\n> part of the 001 patch, ideally in a way that minimizes the number of\n> places which need to be updated when adding an algorithm.\n\nI gave this a little bit of thought. I think that ReadHead should not\nemit a warning, or at least not this warning as it is slightly misleading.\nIt implies that it will automatically turn off data restoration, which is\nfalse. Further ahead, the code will fail with a conflicting error message\nstating that the compression is not available.\n\nInstead, it would be cleaner both for the user and the maintainer to\nmove the check in RestoreArchive and make it the sole responsible for\nthis logic.\n\nPlease find v26 attached. 0001 does the above and 0002 addresses Justin's\ncomplaints regarding the code footprint.\n\n//Cheers,\nGeorgios \n\n\n> \n> --\n> Justin","msg_date":"Thu, 26 Jan 2023 11:24:47 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Jan 26, 2023 at 11:24:47AM +0000, gkokolatos@pm.me wrote:\n> I gave this a little bit of thought. I think that ReadHead should not\n> emit a warning, or at least not this warning as it is slightly misleading.\n> It implies that it will automatically turn off data restoration, which is\n> false. Further ahead, the code will fail with a conflicting error message\n> stating that the compression is not available.\n> \n> Instead, it would be cleaner both for the user and the maintainer to\n> move the check in RestoreArchive and make it the sole responsible for\n> this logic.\n\n- pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n+ pg_fatal(\"cannot restore data from compressed archive (compression not supported in this installation)\");\nHmm. I don't mind changing this part as you suggest.\n\n-#ifndef HAVE_LIBZ\n- if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n- pg_fatal(\"archive is compressed, but this installation does not support compression\");\n-#endif\nHowever I think that we'd better keep the warning, as it can offer a\nhint when using pg_restore -l not built with compression support if\nlooking at a dump that has been compressed.\n--\nMichael","msg_date":"Thu, 26 Jan 2023 20:53:28 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Thursday, January 26th, 2023 at 12:53 PM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Thu, Jan 26, 2023 at 11:24:47AM +0000, gkokolatos@pm.me wrote:\n> \n> > I gave this a little bit of thought. I think that ReadHead should not\n> > emit a warning, or at least not this warning as it is slightly misleading.\n> > It implies that it will automatically turn off data restoration, which is\n> > false. Further ahead, the code will fail with a conflicting error message\n> > stating that the compression is not available.\n> > \n> > Instead, it would be cleaner both for the user and the maintainer to\n> > move the check in RestoreArchive and make it the sole responsible for\n> > this logic.\n> \n> \n> - pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n> + pg_fatal(\"cannot restore data from compressed archive (compression not supported in this installation)\");\n> Hmm. I don't mind changing this part as you suggest.\n> \n> -#ifndef HAVE_LIBZ\n> - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> \n> - pg_fatal(\"archive is compressed, but this installation does not support compression\");\n> -#endif\n> However I think that we'd better keep the warning, as it can offer a\n> hint when using pg_restore -l not built with compression support if\n> looking at a dump that has been compressed.\n\nFair enough. Please find v27 attached.\n\nCheers,\n//Georgios\n\n\n> --\n> Michael","msg_date":"Thu, 26 Jan 2023 18:03:59 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Jan 25, 2023 at 07:57:18PM +0000, gkokolatos@pm.me wrote:\n> On Wednesday, January 25th, 2023 at 7:00 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > While looking at this, I realized that commit 5e73a6048 introduced a\n> > regression:\n> > \n> > @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\n> > \n> > - if (AH->compression != 0)\n> > \n> > - pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n> > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> > \n> > + pg_fatal(\"archive is compressed, but this installation does not support compression\");\n> > \n> > Before, it was possible to restore non-data chunks of a dump file, even\n> > if the current build didn't support its compression. But that's now\n> > impossible - and it makes the code we're discussing in RestoreArchive()\n> > unreachable.\n\nOn Thu, Jan 26, 2023 at 08:53:28PM +0900, Michael Paquier wrote:\n> On Thu, Jan 26, 2023 at 11:24:47AM +0000, gkokolatos@pm.me wrote:\n> > I gave this a little bit of thought. I think that ReadHead should not\n> > emit a warning, or at least not this warning as it is slightly misleading.\n> > It implies that it will automatically turn off data restoration, which is\n> > false. Further ahead, the code will fail with a conflicting error message\n> > stating that the compression is not available.\n> > \n> > Instead, it would be cleaner both for the user and the maintainer to\n> > move the check in RestoreArchive and make it the sole responsible for\n> > this logic.\n> \n> - pg_fatal(\"cannot restore from compressed archive (compression not supported in this installation)\");\n> + pg_fatal(\"cannot restore data from compressed archive (compression not supported in this installation)\");\n> Hmm. I don't mind changing this part as you suggest.\n> \n> -#ifndef HAVE_LIBZ\n> - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> - pg_fatal(\"archive is compressed, but this installation does not support compression\");\n> -#endif\n> However I think that we'd better keep the warning, as it can offer a\n> hint when using pg_restore -l not built with compression support if\n> looking at a dump that has been compressed.\n\nYeah. But the original log_warning text was better, and should be\nrestored:\n\n- if (AH->compression != 0)\n- pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n\nThat commit also added this to pg-dump.c:\n\n+ case PG_COMPRESSION_ZSTD:\n+ pg_fatal(\"compression with %s is not yet supported\", \"ZSTD\");\n+ break;\n+ case PG_COMPRESSION_LZ4:\n+ pg_fatal(\"compression with %s is not yet supported\", \"LZ4\");\n+ break;\n\nIn 002, that could be simplified by re-using the supports_compression()\nfunction. (And maybe the same in WriteDataToArchive()?)\n\n-- \nJustin\n\n\n","msg_date":"Thu, 26 Jan 2023 12:22:45 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Jan 26, 2023 at 12:22:45PM -0600, Justin Pryzby wrote:\n> Yeah. But the original log_warning text was better, and should be\n> restored:\n> \n> - if (AH->compression != 0)\n> - pg_log_warning(\"archive is compressed, but this installation does not support compression -- no data will be available\");\n\nYeah, this one's on me. So I have gone with the simplest solution and\napplied a fix to restore the original behavior, with the same warning\nshowing up.\n--\nMichael","msg_date":"Fri, 27 Jan 2023 10:23:55 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Jan 16, 2023 at 11:54:46AM +0900, Michael Paquier wrote:\n> On Sun, Jan 15, 2023 at 07:56:25PM -0600, Justin Pryzby wrote:\n> > On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\n> >> The functions changed by 0001 are cfopen[_write](),\n> >> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\n> >> to change these interfaces which basically exist to handle inputs?\n> > \n> > I changed to pass pg_compress_specification as a pointer, since that's\n> > the usual convention for structs, as followed by the existing uses of\n> > pg_compress_specification.\n> \n> Okay, but what do we gain here? It seems to me that this introduces\n> the risk that a careless change in one of the internal routines if\n> they change slight;ly compress_spec, hence impacting any of their\n> callers? Or is that fixing an actual bug (except if I am missing your\n> point, that does not seem to be the case)? \n\nTo circle back to this: I was not saying there's any bug. The proposed\nchange was only to follow normal and existing normal conventions for\npassing structs. It could also be a pointer to const. It's fine with\nme if you say that it's intentional how it's written already.\n\n> >> Is there some benefit in changing compression_spec within the\n> >> internals of these routines before going back one layer down to their\n> >> callers? Changing the compression_spec on-the-fly in these internal\n> >> paths could be risky, actually, no?\n> > \n> > I think what you're saying is that if the spec is passed as a pointer,\n> > then the called functions shouldn't set spec->algorithm=something.\n> \n> Yes. HEAD makes sure of that, 0001 would not prevent that. So I am a\n> bit confused in seeing how this is a benefit.\n\n\n","msg_date":"Thu, 26 Jan 2023 21:56:32 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Jan 26, 2023 at 12:22:45PM -0600, Justin Pryzby wrote:\n> That commit also added this to pg-dump.c:\n> \n> + case PG_COMPRESSION_ZSTD:\n> + pg_fatal(\"compression with %s is not yet supported\", \"ZSTD\");\n> + break;\n> + case PG_COMPRESSION_LZ4:\n> + pg_fatal(\"compression with %s is not yet supported\", \"LZ4\");\n> + break;\n> \n> In 002, that could be simplified by re-using the supports_compression()\n> function. (And maybe the same in WriteDataToArchive()?)\n\nThe first patch aims to minimize references to \".gz\" and \"GZIP\" and\nZLIB. pg_backup_directory.c comments still refers to \".gz\". I think\nthe patch should ideally change to refer to \"the compressed file\nextension\" (similar to compress_io.c), avoiding the need to update it\nlater.\n\nI think the file extension stuff could be generalized, so it doesn't\nneed to be updated in multiple places (pg_backup_directory.c and\ncompress_io.c). Maybe it's useful to add a function to return the\nextension of a given compression method. It could go in compression.c,\nand be useful in basebackup.\n\nFor the 2nd patch:\n\nI might be in the minority, but I still think some references to \"gzip\"\nshould say \"zlib\":\n\n+} GzipCompressorState;\n+\n+/* Private routines that support gzip compressed data I/O */\n+static void\n+DeflateCompressorGzip(ArchiveHandle *AH, CompressorState *cs, bool flush)\n\nIn my mind, three things here are misleading, because it doesn't use\ngzip headers:\n\n| GzipCompressorState, DeflateCompressorGzip, \"gzip compressed\".\n\nThis comment is about exactly that:\n\n * underlying stream. The second API is a wrapper around fopen/gzopen and\n * friends, providing an interface similar to those, but abstracts away\n * the possible compression. Both APIs use libz for the compression, but\n * the second API uses gzip headers, so the resulting files can be easily\n * manipulated with the gzip utility.\n\nAIUI, Michael says that it's fine that the user-facing command-line\noptions use \"-Z gzip\" (even though the \"custom\" format doesn't use gzip\nheaders). I'm okay with that, as long as that's discussed/understood.\n\n-- \nJustin\n\n\n","msg_date":"Fri, 27 Jan 2023 11:23:20 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Friday, January 27th, 2023 at 6:23 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n> \n> \n> On Thu, Jan 26, 2023 at 12:22:45PM -0600, Justin Pryzby wrote:\n> \n> > That commit also added this to pg-dump.c:\n> > \n> > + case PG_COMPRESSION_ZSTD:\n> > + pg_fatal(\"compression with %s is not yet supported\", \"ZSTD\");\n> > + break;\n> > + case PG_COMPRESSION_LZ4:\n> > + pg_fatal(\"compression with %s is not yet supported\", \"LZ4\");\n> > + break;\n> > \n> > In 002, that could be simplified by re-using the supports_compression()\n> > function. (And maybe the same in WriteDataToArchive()?)\n> \n> \n> The first patch aims to minimize references to \".gz\" and \"GZIP\" and\n> ZLIB. pg_backup_directory.c comments still refers to \".gz\". I think\n> the patch should ideally change to refer to \"the compressed file\n> extension\" (similar to compress_io.c), avoiding the need to update it\n> later.\n> \n> I think the file extension stuff could be generalized, so it doesn't\n> need to be updated in multiple places (pg_backup_directory.c and\n> compress_io.c). Maybe it's useful to add a function to return the\n> extension of a given compression method. It could go in compression.c,\n> and be useful in basebackup.\n> \n> For the 2nd patch:\n> \n> I might be in the minority, but I still think some references to \"gzip\"\n> should say \"zlib\":\n> \n> +} GzipCompressorState;\n> +\n> +/* Private routines that support gzip compressed data I/O */\n> +static void\n> +DeflateCompressorGzip(ArchiveHandle *AH, CompressorState *cs, bool flush)\n> \n> In my mind, three things here are misleading, because it doesn't use\n> gzip headers:\n> \n> | GzipCompressorState, DeflateCompressorGzip, \"gzip compressed\".\n> \n> This comment is about exactly that:\n> \n> * underlying stream. The second API is a wrapper around fopen/gzopen and\n> * friends, providing an interface similar to those, but abstracts away\n> * the possible compression. Both APIs use libz for the compression, but\n> * the second API uses gzip headers, so the resulting files can be easily\n> * manipulated with the gzip utility.\n> \n> AIUI, Michael says that it's fine that the user-facing command-line\n> options use \"-Z gzip\" (even though the \"custom\" format doesn't use gzip\n> headers). I'm okay with that, as long as that's discussed/understood.\n> \n\nThank you for the input Justin. I am currently waiting for input from a\nthird person to get some conclusion. I thought that it should be stated\nbefore my inactiveness is considered as indifference, which is not.\n\nCheers,\n//Georgios\n\n> --\n> Justin\n\n\n","msg_date":"Tue, 31 Jan 2023 09:00:56 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Jan 31, 2023 at 09:00:56AM +0000, gkokolatos@pm.me wrote:\n> > In my mind, three things here are misleading, because it doesn't use\n> > gzip headers:\n> > \n> > | GzipCompressorState, DeflateCompressorGzip, \"gzip compressed\".\n> > \n> > This comment is about exactly that:\n> > \n> > * underlying stream. The second API is a wrapper around fopen/gzopen and\n> > * friends, providing an interface similar to those, but abstracts away\n> > * the possible compression. Both APIs use libz for the compression, but\n> > * the second API uses gzip headers, so the resulting files can be easily\n> > * manipulated with the gzip utility.\n> > \n> > AIUI, Michael says that it's fine that the user-facing command-line\n> > options use \"-Z gzip\" (even though the \"custom\" format doesn't use gzip\n> > headers). I'm okay with that, as long as that's discussed/understood.\n> \n> Thank you for the input Justin. I am currently waiting for input from a\n> third person to get some conclusion. I thought that it should be stated\n> before my inactiveness is considered as indifference, which is not.\n\nI'm not sure what there is to lose by making the names more accurate -\nespecially since they're private/internal-only.\n\nTomas marked himself as a committer, so maybe could comment.\n\nIt'd be nice to also come to some conclusion about whether -Fc -Z gzip\nis confusing (due to not actually using gzip).\n\nBTW, do you intend to merge this for v16 ? I verified in earlier patch\nversions that tests all pass with lz4 as the default compression method.\nAnd checked that gzip output is compatible with before, and that old\ndumps restore correctly, and there's no memory leaks or other errors.\n\n-- \nJustin\n\n\n","msg_date":"Tue, 7 Feb 2023 20:37:18 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Jan 27, 2023 2:04 AM gkokolatos@pm.me <gkokolatos@pm.me> wrote:\r\n> \r\n> ------- Original Message -------\r\n> On Thursday, January 26th, 2023 at 12:53 PM, Michael Paquier\r\n> <michael@paquier.xyz> wrote:\r\n> \r\n> \r\n> >\r\n> >\r\n> > On Thu, Jan 26, 2023 at 11:24:47AM +0000, gkokolatos@pm.me wrote:\r\n> >\r\n> > > I gave this a little bit of thought. I think that ReadHead should not\r\n> > > emit a warning, or at least not this warning as it is slightly misleading.\r\n> > > It implies that it will automatically turn off data restoration, which is\r\n> > > false. Further ahead, the code will fail with a conflicting error message\r\n> > > stating that the compression is not available.\r\n> > >\r\n> > > Instead, it would be cleaner both for the user and the maintainer to\r\n> > > move the check in RestoreArchive and make it the sole responsible for\r\n> > > this logic.\r\n> >\r\n> >\r\n> > - pg_fatal(\"cannot restore from compressed archive (compression not\r\n> supported in this installation)\");\r\n> > + pg_fatal(\"cannot restore data from compressed archive (compression not\r\n> supported in this installation)\");\r\n> > Hmm. I don't mind changing this part as you suggest.\r\n> >\r\n> > -#ifndef HAVE_LIBZ\r\n> > - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\r\n> >\r\n> > - pg_fatal(\"archive is compressed, but this installation does not support\r\n> compression\");\r\n> > -#endif\r\n> > However I think that we'd better keep the warning, as it can offer a\r\n> > hint when using pg_restore -l not built with compression support if\r\n> > looking at a dump that has been compressed.\r\n> \r\n> Fair enough. Please find v27 attached.\r\n> \r\n\r\nHi,\r\n\r\nI am interested in this feature and tried the patch. While reading the comments,\r\nI noticed some minor things that could possibly be improved (in v27-0003 patch).\r\n\r\n1.\r\n+\t/*\r\n+\t * Open a file for writing.\r\n+\t *\r\n+\t * 'mode' can be one of ''w', 'wb', 'a', and 'ab'. Requrires an already\r\n+\t * initialized CompressFileHandle.\r\n+\t */\r\n+\tint\t\t\t(*open_write_func) (const char *path, const char *mode,\r\n+\t\t\t\t\t\t\t\t\tCompressFileHandle *CFH);\r\n\r\nThere is a redundant single quote in front of 'w'.\r\n\r\n2.\r\n/*\r\n * Callback function for WriteDataToArchive. Writes one block of (compressed)\r\n * data to the archive.\r\n */\r\n/*\r\n * Callback function for ReadDataFromArchive. To keep things simple, we\r\n * always read one compressed block at a time.\r\n */\r\n\r\nShould the function names in the comments be updated?\r\n\r\nWriteDataToArchive\r\n->\r\nwriteData\r\n\r\nReadDataFromArchive\r\n->\r\nreadData\r\n\r\n3.\r\n+\tAssert(strcmp(mode, \"r\") == 0 || strcmp(mode, \"rb\") == 0);\r\n\r\nCould we use PG_BINARY_R instead of \"r\" and \"rb\" here?\r\n\r\nRegards,\r\nShi Yu\r\n\r\n","msg_date":"Wed, 15 Feb 2023 13:51:21 +0000","msg_from":"\"shiy.fnst@fujitsu.com\" <shiy.fnst@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Wednesday, February 15th, 2023 at 2:51 PM, shiy.fnst@fujitsu.com <shiy.fnst@fujitsu.com> wrote:\n\n\n> \n> Hi,\n> \n> I am interested in this feature and tried the patch. While reading the comments,\n> I noticed some minor things that could possibly be improved (in v27-0003 patch).\n\nThank you very much for the interest. Please find a rebased v28 attached. Due to\nthe rebase, 0001 of v27 is no longer relevant and has been removed. Your comments\nare applied on v28-0002.\n\n> \n> 1.\n> + /*\n> + * Open a file for writing.\n> + *\n> + * 'mode' can be one of ''w', 'wb', 'a', and 'ab'. Requrires an already\n> + * initialized CompressFileHandle.\n> + */\n> + int (*open_write_func) (const char *path, const char *mode,\n> + CompressFileHandle CFH);\n> \n> There is a redundant single quote in front of 'w'.\n\nFixed.\n\n> \n> 2.\n> /\n> * Callback function for WriteDataToArchive. Writes one block of (compressed)\n> * data to the archive.\n> /\n> /\n> * Callback function for ReadDataFromArchive. To keep things simple, we\n> * always read one compressed block at a time.\n> */\n> \n> Should the function names in the comments be updated?\n\nAgreed. Fixed.\n\n> \n> 3.\n> + Assert(strcmp(mode, \"r\") == 0 || strcmp(mode, \"rb\") == 0);\n> \n> Could we use PG_BINARY_R instead of \"r\" and \"rb\" here?\n\nWe could and we should. Using PG_BINARY_R has the added benefit\nof needing only one strcmp() call. Fixed.\n\nCheers,\n//Georgios\n\n> \n> Regards,\n> Shi Yu","msg_date":"Thu, 16 Feb 2023 11:38:00 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"RE: Add LZ4 compression in pg_dump"},{"msg_contents":"Hi Georgios,\n\nI spent some time looking at the patch again, and IMO it's RFC. But I\nneed some help with the commit messages - I updated 0001 and 0002 but I\nwasn't quite sure what some of the stuff meant to say and/or it seemed\nmaybe coming from an earlier patch version and obsolete.\n\nCould you go over them and check if I got it right? Also feel free to\nupdate the list of reviewers (I compiled that from substantial reviews\non the thread).\n\nThe 0003 commit message seems somewhat confusing - I added some XXX\nlines asking about unclear stuff.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Sun, 19 Feb 2023 18:10:13 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Some little updates since I last checked:\n\n+ * This file also includes the implementation when compression is none for\n+ * both API's.\n\n=> this comment is obsolete.\n\ns/deffer/infer/ ?\nor determine ?\nThis typo occurs multiple times.\n\ncurrently this includes only \".gz\"\n=> remove this phase from the 002 patch (or at least update it in 003).\n\ndeferred by iteratively\n=> inferred?\n\ns/Requrires/Requires/\ntwice.\n\ns/occured/occurred/\n\ns/disc/disk/ ?\nProbably unimportant, but \"disc\" isn't used anywhere else.\n\n\"compress file handle\"\n=> maybe these should say \"compressed\"\n\nsupports_compression():\nSince this is an exported function, it should probably be called\npgdump_supports_compresion.\n\n\n","msg_date":"Sun, 19 Feb 2023 13:43:45 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Sunday, February 19th, 2023 at 6:10 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n>\n>\n> Hi Georgios,\n>\n> I spent some time looking at the patch again, and IMO it's RFC. But I\n> need some help with the commit messages - I updated 0001 and 0002 but I\n> wasn't quite sure what some of the stuff meant to say and/or it seemed\n> maybe coming from an earlier patch version and obsolete.\n\nThank you very much Tomas! Indeed I have not being paying any attention\nat the commit messages.\n\n> Could you go over them and check if I got it right? Also feel free to\n> update the list of reviewers (I compiled that from substantial reviews\n> on the thread).\n\nDone. Rachel has been correctly identified as author in the relevant parts\nup to commit 98fe74218d. After that, she had offered review comments and I\nhave taken the liberty to add her as a reviewer through out.\n\nAlso I think that Shi Yu should be credited as a reviewer of 0003.\n\n>\n> The 0003 commit message seems somewhat confusing - I added some XXX\n> lines asking about unclear stuff.\n\nPlease find in the attached v30 an updated message, as well as an amended\nreviewer list. Also v30 addresses the final comments raised by Justin.\n\nCheers,\n//Georgios\n\n> regards\n>\n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Mon, 20 Feb 2023 13:09:22 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Thanks for v30 with the updated commit messages. I've pushed 0001 after\nfixing a comment typo and removing (I think) an unnecessary change in an\nerror message.\n\nI'll give the buildfarm a bit of time before pushing 0002 and 0003.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 23 Feb 2023 16:26:35 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 2/23/23 16:26, Tomas Vondra wrote:\n> Thanks for v30 with the updated commit messages. I've pushed 0001 after\n> fixing a comment typo and removing (I think) an unnecessary change in an\n> error message.\n> \n> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\n> \n\nI've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\nand marked the CF entry as committed. Thanks for the patch!\n\nI wonder how difficult would it be to add the zstd compression, so that\nwe don't have the annoying \"unsupported\" cases.\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 23 Feb 2023 21:24:46 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Feb 23, 2023 at 09:24:46PM +0100, Tomas Vondra wrote:\n> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\n> and marked the CF entry as committed. Thanks for the patch!\n\nA big thanks from me to everyone involved.\n\n> I wonder how difficult would it be to add the zstd compression, so that\n> we don't have the annoying \"unsupported\" cases.\n\nI'll send a patch soon. I first submitted patches for that 2 years ago\n(before PGDG was ready to add zstd).\nhttps://commitfest.postgresql.org/31/2888/\n\n-- \nJustin\n\n\n","msg_date":"Thu, 23 Feb 2023 19:51:16 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Feb 23, 2023 at 07:51:16PM -0600, Justin Pryzby wrote:\n> On Thu, Feb 23, 2023 at 09:24:46PM +0100, Tomas Vondra wrote:\n>> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\n>> and marked the CF entry as committed. Thanks for the patch!\n> \n> A big thanks from me to everyone involved.\n\nWow, nice! The APIs are clear to follow.\n\n> I'll send a patch soon. I first submitted patches for that 2 years ago\n> (before PGDG was ready to add zstd).\n> https://commitfest.postgresql.org/31/2888/\n\nThanks. It should be straight-forward to see that in 16, I guess.\n--\nMichael","msg_date":"Fri, 24 Feb 2023 13:35:14 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Friday, February 24th, 2023 at 5:35 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Thu, Feb 23, 2023 at 07:51:16PM -0600, Justin Pryzby wrote:\n> \n> > On Thu, Feb 23, 2023 at 09:24:46PM +0100, Tomas Vondra wrote:\n> > \n> > > I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\n> > > and marked the CF entry as committed. Thanks for the patch!\n> > \n> > A big thanks from me to everyone involved.\n> \n> \n> Wow, nice! The APIs are clear to follow.\n\nI am out of words, thank you all so very much. I learned a lot. \n\n> \n> > I'll send a patch soon. I first submitted patches for that 2 years ago\n> > (before PGDG was ready to add zstd).\n> > https://commitfest.postgresql.org/31/2888/\n> \n> \n> Thanks. It should be straight-forward to see that in 16, I guess.\n> --\n> Michael\n\n\n","msg_date":"Fri, 24 Feb 2023 09:21:32 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"I have some fixes (attached) and questions while polishing the patch for\nzstd compression. The fixes are small and could be integrated with the\npatch for zstd, but could be applied independently.\n\n- I'm unclear about get_error_func(). That's called in three places\n from pg_backup_directory.c, after failures from write_func(), to\n supply an compression-specific error message to pg_fatal(). But it's\n not being used outside of directory format, nor for errors for other\n function pointers, or even for all errors in write_func(). Is there\n some reason why each compression method's write_func() shouldn't call\n pg_fatal() directly, with its compression-specific message ?\n\n- I still think supports_compression() should be renamed, or made into a\n static function in the necessary file. The main reason is that it's\n more clear what it indicates - whether compression is \"implemented by\n pgdump\" and not whether compression is \"supported by this postgres\n build\". It also seems possible that we'd want to add a function\n called something like supports_compression(), indicating whether the\n algorithm is supported by the current build. It'd be better if pgdump\n didn't subjugate that name.\n\n- Finally, the \"Nothing to do in the default case\" comment comes from\n Michael's commit 5e73a6048:\n\n+ /*\n+ * Custom and directory formats are compressed by default with gzip when\n+ * available, not the others.\n+ */\n+ if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\n+ !user_compression_defined)\n {\n #ifdef HAVE_LIBZ\n- if (archiveFormat == archCustom || archiveFormat == archDirectory)\n- compressLevel = Z_DEFAULT_COMPRESSION;\n- else\n+ parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\n+ &compression_spec);\n+#else\n+ /* Nothing to do in the default case */\n #endif\n- compressLevel = 0;\n }\n\n\nAs the comment says: for -Fc and -Fd, the compression is set to zlib, if\nenabled, and when not otherwise specified by the user.\n\nBefore 5e73a6048, this set compressLevel=0 for -Fp and -Ft, *and* when\nzlib was unavailable.\n\nBut I'm not sure why there's now an empty \"#else\". I also don't know\nwhat \"the default case\" refers to.\n\nMaybe the best thing here is to move the preprocessor #if, since it's no\nlonger in the middle of a runtime conditional:\n\n #ifdef HAVE_LIBZ\n+ if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\n+ !user_compression_defined)\n+ parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\n+ &compression_spec);\n #endif\n\n...but that elicits a warning about \"variable set but not used\"...\n\n-- \nJustin","msg_date":"Fri, 24 Feb 2023 23:02:14 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n> I have some fixes (attached) and questions while polishing the patch for\n> zstd compression. The fixes are small and could be integrated with the\n> patch for zstd, but could be applied independently.\n\nOne more - WriteDataToArchiveGzip() says:\n\n+ if (cs->compression_spec.level == 0)\n+ pg_fatal(\"requested to compress the archive yet no level was specified\");\n\nThat was added at e9960732a. \n\nBut if you specify gzip:0, the compression level is already enforced by\nvalidate_compress_specification(), before hitting gzip.c:\n\n| pg_dump: error: invalid compression specification: compression algorithm \"gzip\" expects a compression level between 1 and 9 (default at -1)\n\n5e73a6048 intended that to work as before, and you *can* specify -Z0:\n\n The change is backward-compatible, hence specifying only an integer\n leads to no compression for a level of 0 and gzip compression when the\n level is greater than 0.\n\n $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\n /dev/stdin: ASCII text\n\nRight now, I think that pg_fatal in gzip.c is dead code - that was first\nadded in the patch version sent on 21 Dec 2022.\n\n-- \nJustin","msg_date":"Sat, 25 Feb 2023 08:05:53 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 2/25/23 06:02, Justin Pryzby wrote:\n> I have some fixes (attached) and questions while polishing the patch for\n> zstd compression. The fixes are small and could be integrated with the\n> patch for zstd, but could be applied independently.\n> \n> - I'm unclear about get_error_func(). That's called in three places\n> from pg_backup_directory.c, after failures from write_func(), to\n> supply an compression-specific error message to pg_fatal(). But it's\n> not being used outside of directory format, nor for errors for other\n> function pointers, or even for all errors in write_func(). Is there\n> some reason why each compression method's write_func() shouldn't call\n> pg_fatal() directly, with its compression-specific message ?\n> \n\nI think there are a couple more places that might/should call\nget_error_func(). For example ahwrite() in pg_backup_archiver.c now\nsimply does\n\n if (bytes_written != size * nmemb)\n WRITE_ERROR_EXIT;\n\nbut perhaps it should call get_error_func() too. There are probably\nother places that call write_func() and should use get_error_func().\n\n> - I still think supports_compression() should be renamed, or made into a\n> static function in the necessary file. The main reason is that it's\n> more clear what it indicates - whether compression is \"implemented by\n> pgdump\" and not whether compression is \"supported by this postgres\n> build\". It also seems possible that we'd want to add a function\n> called something like supports_compression(), indicating whether the\n> algorithm is supported by the current build. It'd be better if pgdump\n> didn't subjugate that name.\n> \n\nIf we choose to rename this to have pgdump_ prefix, fine with me. But I\ndon't think there's a realistic chance of conflict, as it's restricted\nto pgdump header etc. And it's not part of an API, so I guess we could\nrename that in the future if needed.\n\n> - Finally, the \"Nothing to do in the default case\" comment comes from\n> Michael's commit 5e73a6048:\n> \n> + /*\n> + * Custom and directory formats are compressed by default with gzip when\n> + * available, not the others.\n> + */\n> + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\n> + !user_compression_defined)\n> {\n> #ifdef HAVE_LIBZ\n> - if (archiveFormat == archCustom || archiveFormat == archDirectory)\n> - compressLevel = Z_DEFAULT_COMPRESSION;\n> - else\n> + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\n> + &compression_spec);\n> +#else\n> + /* Nothing to do in the default case */\n> #endif\n> - compressLevel = 0;\n> }\n> \n> \n> As the comment says: for -Fc and -Fd, the compression is set to zlib, if\n> enabled, and when not otherwise specified by the user.\n> \n> Before 5e73a6048, this set compressLevel=0 for -Fp and -Ft, *and* when\n> zlib was unavailable.\n> \n> But I'm not sure why there's now an empty \"#else\". I also don't know\n> what \"the default case\" refers to.\n> \n> Maybe the best thing here is to move the preprocessor #if, since it's no\n> longer in the middle of a runtime conditional:\n> \n> #ifdef HAVE_LIBZ\n> + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\n> + !user_compression_defined)\n> + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\n> + &compression_spec);\n> #endif\n> \n> ...but that elicits a warning about \"variable set but not used\"...\n> \n\nNot sure, I need to think about this a bit.\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Sun, 26 Feb 2023 15:59:41 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, Feb 25, 2023 at 08:05:53AM -0600, Justin Pryzby wrote:\n> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n> > I have some fixes (attached) and questions while polishing the patch for\n> > zstd compression. The fixes are small and could be integrated with the\n> > patch for zstd, but could be applied independently.\n> \n> One more - WriteDataToArchiveGzip() says:\n\nOne more again.\n\nThe LZ4 path is using non-streaming mode, which compresses each block\nwithout persistent state, giving poor compression for -Fc compared with\n-Fp. If the data is highly compressible, the difference can be orders\nof magnitude.\n\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fp |wc -c\n12351763\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\n21890708\n\nThat's not true for gzip:\n\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fc |wc -c\n2118869\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fp |wc -c\n2115832\n\nThe function ought to at least use streaming mode, so each block/row\nisn't compressioned in isolation. 003 is a simple patch to use\nstreaming mode, which improves the -Fc case:\n\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\n15178283\n\nHowever, that still flushes the compression buffer, writing a block\nheader, for every row. With a single-column table, pg_dump -Fc -Z lz4\nstill outputs ~10% *more* data than with no compression at all. And\nthat's for compressible data.\n\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z lz4 |wc -c\n12890296\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z none |wc -c\n11890296\n\nI think this should use the LZ4F API with frames, which are buffered to\navoid outputting a header for every single row. The LZ4F format isn't\ncompatible with the LZ4 format, so (unlike changing to the streaming\nAPI) that's not something we can change in a bugfix release. I consider\nthis an Opened Item.\n\nWith the LZ4F API in 004, -Fp and -Fc are essentially the same size\n(like gzip). (Oh, and the output is three times smaller, too.)\n\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fp |wc -c\n4155448\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fc |wc -c\n4156548\n\n-- \nJustin","msg_date":"Sun, 26 Feb 2023 22:49:10 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Sunday, February 26th, 2023 at 3:59 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> \n> \n> On 2/25/23 06:02, Justin Pryzby wrote:\n> \n> > I have some fixes (attached) and questions while polishing the patch for\n> > zstd compression. The fixes are small and could be integrated with the\n> > patch for zstd, but could be applied independently.\n> > \n> > - I'm unclear about get_error_func(). That's called in three places\n> > from pg_backup_directory.c, after failures from write_func(), to\n> > supply an compression-specific error message to pg_fatal(). But it's\n> > not being used outside of directory format, nor for errors for other\n> > function pointers, or even for all errors in write_func(). Is there\n> > some reason why each compression method's write_func() shouldn't call\n> > pg_fatal() directly, with its compression-specific message ?\n> \n> \n> I think there are a couple more places that might/should call\n> get_error_func(). For example ahwrite() in pg_backup_archiver.c now\n> simply does\n> \n> if (bytes_written != size * nmemb)\n> WRITE_ERROR_EXIT;\n> \n> but perhaps it should call get_error_func() too. There are probably\n> other places that call write_func() and should use get_error_func().\n\nAgreed, calling get_error_func() would be preferable to a fatal error. It\nshould be the caller of the api who decides how to proceed.\n\n> \n> > - I still think supports_compression() should be renamed, or made into a\n> > static function in the necessary file. The main reason is that it's\n> > more clear what it indicates - whether compression is \"implemented by\n> > pgdump\" and not whether compression is \"supported by this postgres\n> > build\". It also seems possible that we'd want to add a function\n> > called something like supports_compression(), indicating whether the\n> > algorithm is supported by the current build. It'd be better if pgdump\n> > didn't subjugate that name.\n> \n> \n> If we choose to rename this to have pgdump_ prefix, fine with me. But I\n> don't think there's a realistic chance of conflict, as it's restricted\n> to pgdump header etc. And it's not part of an API, so I guess we could\n> rename that in the future if needed.\n\nAgreed, it is very unrealistic that one will include that header file anywhere\nbut within pg_dump. Also. I think that adding a prefix, \"pgdump\", \"pg_dump\",\nor similar does not add value and subtracts readability. \n\n> \n> > - Finally, the \"Nothing to do in the default case\" comment comes from\n> > Michael's commit 5e73a6048:\n> > \n> > + /*\n> > + * Custom and directory formats are compressed by default with gzip when\n> > + * available, not the others.\n> > + /\n> > + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\n> > + !user_compression_defined)\n> > {\n> > #ifdef HAVE_LIBZ\n> > - if (archiveFormat == archCustom || archiveFormat == archDirectory)\n> > - compressLevel = Z_DEFAULT_COMPRESSION;\n> > - else\n> > + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\n> > + &compression_spec);\n> > +#else\n> > + / Nothing to do in the default case */\n> > #endif\n> > - compressLevel = 0;\n> > }\n> > \n> > As the comment says: for -Fc and -Fd, the compression is set to zlib, if\n> > enabled, and when not otherwise specified by the user.\n> > \n> > Before 5e73a6048, this set compressLevel=0 for -Fp and -Ft, and when\n> > zlib was unavailable.\n> > \n> > But I'm not sure why there's now an empty \"#else\". I also don't know\n> > what \"the default case\" refers to.\n> > \n> > Maybe the best thing here is to move the preprocessor #if, since it's no\n> > longer in the middle of a runtime conditional:\n> > \n> > #ifdef HAVE_LIBZ\n> > + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\n> > + !user_compression_defined)\n> > + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\n> > + &compression_spec);\n> > #endif\n> > \n> > ...but that elicits a warning about \"variable set but not used\"...\n> \n> \n> Not sure, I need to think about this a bit.\n\nNot having warnings is preferable, isn't it? I can understand the confusion\non the message though. Maybe a phrasing like:\n/* Nothing to do for the default case when LIBZ is not available */\nis easier to understand.\n\nCheers,\n//Georgios\n\n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company\n\n\n","msg_date":"Mon, 27 Feb 2023 14:33:04 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Saturday, February 25th, 2023 at 3:05 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n> \n> \n> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n> \n> > I have some fixes (attached) and questions while polishing the patch for\n> > zstd compression. The fixes are small and could be integrated with the\n> > patch for zstd, but could be applied independently.\n\n\nPlease find some comments on the rest of the fixes patch that Tomas has not\ncommented on.\n\n can be compressed with the <application>gzip</application> or\n- <application>lz4</application>tool.\n+ <application>lz4</application> tools.\n\n+1\n\n The compression method can be set to <literal>gzip</literal> or\n- <literal>lz4</literal> or <literal>none</literal> for no compression.\n+ <literal>lz4</literal>, or <literal>none</literal> for no compression.\n\nI am not a native English speaker. Yet I think that if one adds commas\nin one of the options, then one should add commas to all the options.\nNamely, the aboveis missing a comma between gzip and lz4. However I\nthink that not having any commas still works grammatically and\nsyntactically.\n\n- /*\n- * A level of zero simply copies the input one block at the time. This\n- * is probably not what the user wanted when calling this interface.\n- */\n- if (cs->compression_spec.level == 0)\n- pg_fatal(\"requested to compress the archive yet no level was specified\");\n\n\nI disagree with change. WriteDataToArchiveGzip() is far away from\nwhat ever the code in pg_dump.c is doing. Any non valid values for\nlevel will emit an error in when the proper gzip/zlib code is\ncalled. A zero value however, will not emit such error. Having the\nextra check there is a future proof guarantee in a very low cost.\nFurthermore, it quickly informs the reader of the code about that\nspecific value helping with readability and comprehension.\n\nIf any change is required, something for which I vote strongly\nagainst, I would at least recommend to replace it with an\nassertion.\n\n- * Initialize a compress file stream. Deffer the compression algorithm\n+ * Initialize a compress file stream. Infer the compression algorithm\n\n:+1:\n\n- # Skip command-level tests for gzip if there is no support for it.\n+ # Skip command-level tests for gzip/lz4 if they're not supported.\n\nWe will be back at that again soon. Maybe change to:\n\nSkip command-level test for unsupported compression methods\n\nTo include everything.\n\n\n- ($pgdump_runs{$run}->{compile_option} eq 'gzip' && !$supports_gzip) ||\n- ($pgdump_runs{$run}->{compile_option} eq 'lz4' && !$supports_lz4))\n+ (($pgdump_runs{$run}->{compile_option} eq 'gzip' && !$supports_gzip) ||\n+ ($pgdump_runs{$run}->{compile_option} eq 'lz4' && !$supports_lz4)))\n\nGood catch, :+1:\n\nCheers,\n//Georgios\n\n> --\n> Justin\n\n\n","msg_date":"Mon, 27 Feb 2023 14:56:08 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Feb 23, 2023 at 09:24:46PM +0100, Tomas Vondra wrote:\n> On 2/23/23 16:26, Tomas Vondra wrote:\n> > Thanks for v30 with the updated commit messages. I've pushed 0001 after\n> > fixing a comment typo and removing (I think) an unnecessary change in an\n> > error message.\n> > \n> > I'll give the buildfarm a bit of time before pushing 0002 and 0003.\n> > \n> \n> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\n> and marked the CF entry as committed. Thanks for the patch!\n\nI found that e9960732a broke writing of empty gzip-compressed data,\nspecifically LOs. pg_dump succeeds, but then the restore fails:\n\npostgres=# SELECT lo_create(1234);\nlo_create | 1234\n\n$ time ./src/bin/pg_dump/pg_dump -h /tmp -d postgres -Fc |./src/bin/pg_dump/pg_restore -f /dev/null -v \npg_restore: implied data-only restore\npg_restore: executing BLOB 1234\npg_restore: processing BLOBS\npg_restore: restoring large object with OID 1234\npg_restore: error: could not uncompress data: (null)\n\nThe inline patch below fixes it, but you won't be able to apply it\ndirectly, as it's on top of other patches which rename the functions\nback to \"Zlib\" and rearranges the functions to their original order, to\nallow running:\n\ngit diff --diff-algorithm=minimal -w e9960732a~:./src/bin/pg_dump/compress_io.c ./src/bin/pg_dump/compress_gzip.c\n\nThe current function order avoids 3 lines of declarations, but it's\nobviously pretty useful to be able to run that diff command. I already\nargued for not calling the functions \"Gzip\" on the grounds that the name\nwas inaccurate.\n\nI'd want to create an empty large object in src/test/sql/largeobject.sql\nto exercise this tested during pgupgrade. But unfortunately that\ndoesn't use -Fc, so this isn't hit. Empty input is an important enough\ntest case to justify a tap test, if there's no better way.\n\ndiff --git a/src/bin/pg_dump/compress_gzip.c b/src/bin/pg_dump/compress_gzip.c\nindex f3f5e87c9a8..68f3111b2fe 100644\n--- a/src/bin/pg_dump/compress_gzip.c\n+++ b/src/bin/pg_dump/compress_gzip.c\n@@ -55,6 +55,32 @@ InitCompressorZlib(CompressorState *cs,\n \tgzipcs = (ZlibCompressorState *) pg_malloc0(sizeof(ZlibCompressorState));\n \n \tcs->private_data = gzipcs;\n+\n+\tif (cs->writeF)\n+\t{\n+\t\tz_streamp\tzp;\n+\t\tzp = gzipcs->zp = (z_streamp) pg_malloc0(sizeof(z_stream));\n+\t\tzp->zalloc = Z_NULL;\n+\t\tzp->zfree = Z_NULL;\n+\t\tzp->opaque = Z_NULL;\n+\n+\t\t/*\n+\t\t * outsize is the buffer size we tell zlib it can output to. We\n+\t\t * actually allocate one extra byte because some routines want to append a\n+\t\t * trailing zero byte to the zlib output.\n+\t\t */\n+\n+\t\tgzipcs->outbuf = pg_malloc(ZLIB_OUT_SIZE + 1);\n+\t\tgzipcs->outsize = ZLIB_OUT_SIZE;\n+\n+\t\tif (deflateInit(gzipcs->zp, cs->compression_spec.level) != Z_OK)\n+\t\t\tpg_fatal(\"could not initialize compression library: %s\",\n+\t\t\t\t\tzp->msg);\n+\n+\t\t/* Just be paranoid - maybe End is called after Start, with no Write */\n+\t\tzp->next_out = gzipcs->outbuf;\n+\t\tzp->avail_out = gzipcs->outsize;\n+\t}\n }\n \n static void\n@@ -63,7 +89,7 @@ EndCompressorZlib(ArchiveHandle *AH, CompressorState *cs)\n \tZlibCompressorState *gzipcs = (ZlibCompressorState *) cs->private_data;\n \tz_streamp\tzp;\n \n-\tif (gzipcs->zp)\n+\tif (cs->writeF != NULL)\n \t{\n \t\tzp = gzipcs->zp;\n \t\tzp->next_in = NULL;\n@@ -131,29 +157,6 @@ WriteDataToArchiveZlib(ArchiveHandle *AH, CompressorState *cs,\n \t\t\t\t\t const void *data, size_t dLen)\n {\n \tZlibCompressorState *gzipcs = (ZlibCompressorState *) cs->private_data;\n-\tz_streamp\tzp;\n-\n-\tif (!gzipcs->zp)\n-\t{\n-\t\tzp = gzipcs->zp = (z_streamp) pg_malloc(sizeof(z_stream));\n-\t\tzp->zalloc = Z_NULL;\n-\t\tzp->zfree = Z_NULL;\n-\t\tzp->opaque = Z_NULL;\n-\n-\t\t/*\n-\t\t * outsize is the buffer size we tell zlib it can output to. We\n-\t\t * actually allocate one extra byte because some routines want to\n-\t\t * append a trailing zero byte to the zlib output.\n-\t\t */\n-\t\tgzipcs->outbuf = pg_malloc(ZLIB_OUT_SIZE + 1);\n-\t\tgzipcs->outsize = ZLIB_OUT_SIZE;\n-\n-\t\tif (deflateInit(zp, cs->compression_spec.level) != Z_OK)\n-\t\t\tpg_fatal(\"could not initialize compression library: %s\", zp->msg);\n-\n-\t\tzp->next_out = gzipcs->outbuf;\n-\t\tzp->avail_out = gzipcs->outsize;\n-\t}\n \n \tgzipcs->zp->next_in = (void *) unconstify(void *, data);\n \tgzipcs->zp->avail_in = dLen;","msg_date":"Tue, 28 Feb 2023 17:58:34 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Feb 28, 2023 at 05:58:34PM -0600, Justin Pryzby wrote:\n> I found that e9960732a broke writing of empty gzip-compressed data,\n> specifically LOs. pg_dump succeeds, but then the restore fails:\n\nThe number of issues you have been reporting here begins to worries\nme.. How many of them have you found? Is it right to assume that all\nof them have basically 03d02f5 as oldest origin point?\n--\nMichael","msg_date":"Wed, 1 Mar 2023 16:24:59 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Wednesday, March 1st, 2023 at 12:58 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n\n\n\n> I found that e9960732a broke writing of empty gzip-compressed data,\n> specifically LOs. pg_dump succeeds, but then the restore fails:\n> \n> postgres=# SELECT lo_create(1234);\n> lo_create | 1234\n> \n> $ time ./src/bin/pg_dump/pg_dump -h /tmp -d postgres -Fc |./src/bin/pg_dump/pg_restore -f /dev/null -v\n> pg_restore: implied data-only restore\n> pg_restore: executing BLOB 1234\n> pg_restore: processing BLOBS\n> pg_restore: restoring large object with OID 1234\n> pg_restore: error: could not uncompress data: (null)\n> \n\nThank you for looking. This was an untested case.\n\n> The inline patch below fixes it, but you won't be able to apply it\n> directly, as it's on top of other patches which rename the functions\n> back to \"Zlib\" and rearranges the functions to their original order, to\n> allow running:\n> \n> git diff --diff-algorithm=minimal -w e9960732a~:./src/bin/pg_dump/compress_io.c ./src/bin/pg_dump/compress_gzip.c\n> \n\nPlease find a patch attached that can be applied directly.\n\n> The current function order avoids 3 lines of declarations, but it's\n> obviously pretty useful to be able to run that diff command. I already\n> argued for not calling the functions \"Gzip\" on the grounds that the name\n> was inaccurate.\n\nI have no idea why we are back on the naming issue. I stand by the name\nbecause in my humble opinion helps the code reader. There is a certain\nuniformity when the compression_spec.algorithm and the compressor\nfunctions match as the following code sample shows.\n\n if (compression_spec.algorithm == PG_COMPRESSION_NONE) \n InitCompressorNone(cs, compression_spec);\n else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\n InitCompressorGzip(cs, compression_spec);\n else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\n InitCompressorLZ4(cs, compression_spec); \n \nWhen the reader wants to see what happens when the PG_COMPRESSION_XXX\nis set, has to simply search for the XXX part. I think that this is\njustification enough for the use of the names.\n\n> \n> I'd want to create an empty large object in src/test/sql/largeobject.sql\n> to exercise this tested during pgupgrade. But unfortunately that\n> doesn't use -Fc, so this isn't hit. Empty input is an important enough\n> test case to justify a tap test, if there's no better way.\n\nPlease find in the attached a test case that exercises this codepath.\n\nCheers,\n//Georgios","msg_date":"Wed, 01 Mar 2023 13:39:14 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 3/1/23 08:24, Michael Paquier wrote:\n> On Tue, Feb 28, 2023 at 05:58:34PM -0600, Justin Pryzby wrote:\n>> I found that e9960732a broke writing of empty gzip-compressed data,\n>> specifically LOs. pg_dump succeeds, but then the restore fails:\n> \n> The number of issues you have been reporting here begins to worries\n> me.. How many of them have you found? Is it right to assume that all\n> of them have basically 03d02f5 as oldest origin point?\n\nAFAICS a lot of the issues are more a discussion about wording in a\ncouple places, whether it's nicer to do A or B, name the functions\ndifferently or what.\n\nI'm aware of three genuine issues that I intend to fix shortly:\n\n1) incorrect \"if\" condition in a TAP test\n\n2) failure when compressing empty LO (which we had no test for)\n\n3) change in handling \"compression level = 0\" (which I believe should be\nmade to behave like before)\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 1 Mar 2023 15:33:00 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 3/1/23 14:39, gkokolatos@pm.me wrote:\n> \n> \n> \n> \n> \n> ------- Original Message -------\n> On Wednesday, March 1st, 2023 at 12:58 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> \n> \n> \n>> I found that e9960732a broke writing of empty gzip-compressed data,\n>> specifically LOs. pg_dump succeeds, but then the restore fails:\n>>\n>> postgres=# SELECT lo_create(1234);\n>> lo_create | 1234\n>>\n>> $ time ./src/bin/pg_dump/pg_dump -h /tmp -d postgres -Fc |./src/bin/pg_dump/pg_restore -f /dev/null -v\n>> pg_restore: implied data-only restore\n>> pg_restore: executing BLOB 1234\n>> pg_restore: processing BLOBS\n>> pg_restore: restoring large object with OID 1234\n>> pg_restore: error: could not uncompress data: (null)\n>>\n> \n> Thank you for looking. This was an untested case.\n> \n\nYeah :-(\n\n>> The inline patch below fixes it, but you won't be able to apply it\n>> directly, as it's on top of other patches which rename the functions\n>> back to \"Zlib\" and rearranges the functions to their original order, to\n>> allow running:\n>>\n>> git diff --diff-algorithm=minimal -w e9960732a~:./src/bin/pg_dump/compress_io.c ./src/bin/pg_dump/compress_gzip.c\n>>\n> \n> Please find a patch attached that can be applied directly.\n> \n>> The current function order avoids 3 lines of declarations, but it's\n>> obviously pretty useful to be able to run that diff command. I already\n>> argued for not calling the functions \"Gzip\" on the grounds that the name\n>> was inaccurate.\n> \n> I have no idea why we are back on the naming issue. I stand by the name\n> because in my humble opinion helps the code reader. There is a certain\n> uniformity when the compression_spec.algorithm and the compressor\n> functions match as the following code sample shows.\n> \n> if (compression_spec.algorithm == PG_COMPRESSION_NONE) \n> InitCompressorNone(cs, compression_spec);\n> else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> InitCompressorGzip(cs, compression_spec);\n> else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\n> InitCompressorLZ4(cs, compression_spec); \n> \n> When the reader wants to see what happens when the PG_COMPRESSION_XXX\n> is set, has to simply search for the XXX part. I think that this is\n> justification enough for the use of the names.\n> \n\nI don't recall the previous discussion about the naming, but I'm not\nsure why would it be inaccurate. We call it 'gzip' pretty much\neverywhere, and I agree with Georgios there's it helps to make this\nconsistent with the PG_COMPRESSION_ stuff.\n\nThe one thing that concerned me while reviewing it earlier was that it\nmight make the backpatcheing harder. But that's mostly irrelevant due to\nall the other changes I think.\n\n>>\n>> I'd want to create an empty large object in src/test/sql/largeobject.sql\n>> to exercise this tested during pgupgrade. But unfortunately that\n>> doesn't use -Fc, so this isn't hit. Empty input is an important enough\n>> test case to justify a tap test, if there's no better way.\n> \n> Please find in the attached a test case that exercises this codepath.\n> \n\nThanks. That seems correct to me, but I find it somewhat confusing,\nbecause we now have\n\n DeflateCompressorInit vs. InitCompressorGzip\n\n DeflateCompressorEnd vs. EndCompressorGzip\n\n DeflateCompressorData - The name doesn't really say what it does (would\n be better to have a verb in there, I think).\n\nI wonder if we can make this somehow clearer?\n\nAlso, InitCompressorGzip says this:\n\n /*\n * If the caller has defined a write function, prepare the necessary\n * state. Avoid initializing during the first write call, because End\n * may be called without ever writing any data.\n */\n if (cs->writeF)\n DeflateCompressorInit(cs);\n\nDoes it actually make sense to not have writeF defined in some cases?\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 1 Mar 2023 16:52:49 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 2/27/23 15:56, gkokolatos@pm.me wrote:\n> \n> \n> \n> \n> \n> ------- Original Message -------\n> On Saturday, February 25th, 2023 at 3:05 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> \n> \n>>\n>>\n>> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n>>\n>>> I have some fixes (attached) and questions while polishing the patch for\n>>> zstd compression. The fixes are small and could be integrated with the\n>>> patch for zstd, but could be applied independently.\n> \n> \n> Please find some comments on the rest of the fixes patch that Tomas has not\n> commented on.\n> \n> can be compressed with the <application>gzip</application> or\n> - <application>lz4</application>tool.\n> + <application>lz4</application> tools.\n> \n> +1\n> \n> The compression method can be set to <literal>gzip</literal> or\n> - <literal>lz4</literal> or <literal>none</literal> for no compression.\n> + <literal>lz4</literal>, or <literal>none</literal> for no compression.\n> \n> I am not a native English speaker. Yet I think that if one adds commas\n> in one of the options, then one should add commas to all the options.\n> Namely, the aboveis missing a comma between gzip and lz4. However I\n> think that not having any commas still works grammatically and\n> syntactically.\n> \n\nI pushed a fix with most of these wording changes. As for this comma, I\nbelieve the correct style is\n\n a, b, or c\n\nAt least that's what the other places in the pg_dump.sgml file do.\n\n> - ($pgdump_runs{$run}->{compile_option} eq 'gzip' && !$supports_gzip) ||\n> - ($pgdump_runs{$run}->{compile_option} eq 'lz4' && !$supports_lz4))\n> + (($pgdump_runs{$run}->{compile_option} eq 'gzip' && !$supports_gzip) ||\n> + ($pgdump_runs{$run}->{compile_option} eq 'lz4' && !$supports_lz4)))\n> \n\nPushed a fix for this too.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 1 Mar 2023 17:14:08 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 2/25/23 15:05, Justin Pryzby wrote:\n> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n>> I have some fixes (attached) and questions while polishing the patch for\n>> zstd compression. The fixes are small and could be integrated with the\n>> patch for zstd, but could be applied independently.\n> \n> One more - WriteDataToArchiveGzip() says:\n> \n> + if (cs->compression_spec.level == 0)\n> + pg_fatal(\"requested to compress the archive yet no level was specified\");\n> \n> That was added at e9960732a. \n> \n> But if you specify gzip:0, the compression level is already enforced by\n> validate_compress_specification(), before hitting gzip.c:\n> \n> | pg_dump: error: invalid compression specification: compression algorithm \"gzip\" expects a compression level between 1 and 9 (default at -1)\n> \n> 5e73a6048 intended that to work as before, and you *can* specify -Z0:\n> \n> The change is backward-compatible, hence specifying only an integer\n> leads to no compression for a level of 0 and gzip compression when the\n> level is greater than 0.\n> \n> $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\n> /dev/stdin: ASCII text\n> \n\nFWIW I agree we should make this backwards-compatible - accept \"0\" and\ntreat it as no compression.\n\nGeorgios, can you prepare a patch doing that?\n\n\nregards\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 1 Mar 2023 17:20:05 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 2/27/23 05:49, Justin Pryzby wrote:\n> On Sat, Feb 25, 2023 at 08:05:53AM -0600, Justin Pryzby wrote:\n>> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n>>> I have some fixes (attached) and questions while polishing the patch for\n>>> zstd compression. The fixes are small and could be integrated with the\n>>> patch for zstd, but could be applied independently.\n>>\n>> One more - WriteDataToArchiveGzip() says:\n> \n> One more again.\n> \n> The LZ4 path is using non-streaming mode, which compresses each block\n> without persistent state, giving poor compression for -Fc compared with\n> -Fp. If the data is highly compressible, the difference can be orders\n> of magnitude.\n> \n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fp |wc -c\n> 12351763\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\n> 21890708\n> \n> That's not true for gzip:\n> \n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fc |wc -c\n> 2118869\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fp |wc -c\n> 2115832\n> \n> The function ought to at least use streaming mode, so each block/row\n> isn't compressioned in isolation. 003 is a simple patch to use\n> streaming mode, which improves the -Fc case:\n> \n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\n> 15178283\n> \n> However, that still flushes the compression buffer, writing a block\n> header, for every row. With a single-column table, pg_dump -Fc -Z lz4\n> still outputs ~10% *more* data than with no compression at all. And\n> that's for compressible data.\n> \n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z lz4 |wc -c\n> 12890296\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z none |wc -c\n> 11890296\n> \n> I think this should use the LZ4F API with frames, which are buffered to\n> avoid outputting a header for every single row. The LZ4F format isn't\n> compatible with the LZ4 format, so (unlike changing to the streaming\n> API) that's not something we can change in a bugfix release. I consider\n> this an Opened Item.\n> \n> With the LZ4F API in 004, -Fp and -Fc are essentially the same size\n> (like gzip). (Oh, and the output is three times smaller, too.)\n> \n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fp |wc -c\n> 4155448\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fc |wc -c\n> 4156548\n> \n\nThanks. Those are definitely interesting improvements/optimizations!\n\nI suggest we track them as a separate patch series - please add them to\nthe CF app (I guess you'll have to add them to 2023-07 at this point,\nbut we can get them in, I think).\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 1 Mar 2023 17:39:54 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Mar 01, 2023 at 01:39:14PM +0000, gkokolatos@pm.me wrote:\n> On Wednesday, March 1st, 2023 at 12:58 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\n> \n> > The current function order avoids 3 lines of declarations, but it's\n> > obviously pretty useful to be able to run that diff command. I already\n> > argued for not calling the functions \"Gzip\" on the grounds that the name\n> > was inaccurate.\n> \n> I have no idea why we are back on the naming issue. I stand by the name\n> because in my humble opinion helps the code reader. There is a certain\n> uniformity when the compression_spec.algorithm and the compressor\n> functions match as the following code sample shows.\n\nI mentioned that it's because this allows usefully running \"diff\"\nagainst the previous commits.\n\n> if (compression_spec.algorithm == PG_COMPRESSION_NONE) \n> InitCompressorNone(cs, compression_spec);\n> else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> InitCompressorGzip(cs, compression_spec);\n> else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\n> InitCompressorLZ4(cs, compression_spec); \n> \n> When the reader wants to see what happens when the PG_COMPRESSION_XXX\n> is set, has to simply search for the XXX part. I think that this is\n> justification enough for the use of the names.\n\nYou're right about that.\n\nBut (with the exception of InitCompressorGzip), I'm referring to the\nnaming of internal functions, static to gzip.c, so renaming can't be\nsaid to cause a loss of clarity.\n\n> > I'd want to create an empty large object in src/test/sql/largeobject.sql\n> > to exercise this tested during pgupgrade. But unfortunately that\n> > doesn't use -Fc, so this isn't hit. Empty input is an important enough\n> > test case to justify a tap test, if there's no better way.\n> \n> Please find in the attached a test case that exercises this codepath.\n\nThanks for writing it.\n\nThis patch could be an opportunity to improve the \"diff\" output, without\nrenaming anything.\n\nThe old order of functions was:\n-InitCompressorZlib\n-EndCompressorZlib\n-DeflateCompressorZlib\n-WriteDataToArchiveZlib\n-ReadDataFromArchiveZlib\n\nIf you put DeflateCompressorEnd immediately after DeflateCompressorInit,\ndiff works nicely. You'll have to add at least one declaration, which\nseems very worth it.\n\n-- \nJustin\n\n\n","msg_date":"Wed, 1 Mar 2023 11:08:17 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Wednesday, March 1st, 2023 at 5:20 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> \n> \n> On 2/25/23 15:05, Justin Pryzby wrote:\n> \n> > On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n> > \n> > > I have some fixes (attached) and questions while polishing the patch for\n> > > zstd compression. The fixes are small and could be integrated with the\n> > > patch for zstd, but could be applied independently.\n> > \n> > One more - WriteDataToArchiveGzip() says:\n> > \n> > + if (cs->compression_spec.level == 0)\n> > + pg_fatal(\"requested to compress the archive yet no level was specified\");\n> > \n> > That was added at e9960732a.\n> > \n> > But if you specify gzip:0, the compression level is already enforced by\n> > validate_compress_specification(), before hitting gzip.c:\n> > \n> > | pg_dump: error: invalid compression specification: compression algorithm \"gzip\" expects a compression level between 1 and 9 (default at -1)\n> > \n> > 5e73a6048 intended that to work as before, and you can specify -Z0:\n> > \n> > The change is backward-compatible, hence specifying only an integer\n> > leads to no compression for a level of 0 and gzip compression when the\n> > level is greater than 0.\n> > \n> > $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\n> > /dev/stdin: ASCII text\n> \n> \n> FWIW I agree we should make this backwards-compatible - accept \"0\" and\n> treat it as no compression.\n> \n> Georgios, can you prepare a patch doing that?\n\nPlease find a patch attached. However I am a bit at a loss, the backwards\ncompatible behaviour has not changed. Passing -Z0/--compress=0 does produce\na non compressed output. So I am not really certain as to what broke and\nneeds fixing.\n\nWhat commit 5e73a6048 did fail to do, is test the backwards compatible\nbehaviour. The attached amends it.\n\nCheers,\n//Georgios\n\n> \n> \n> regards\n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Thu, 02 Mar 2023 16:34:23 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Mar 01, 2023 at 05:20:05PM +0100, Tomas Vondra wrote:\n> On 2/25/23 15:05, Justin Pryzby wrote:\n> > On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n> >> I have some fixes (attached) and questions while polishing the patch for\n> >> zstd compression. The fixes are small and could be integrated with the\n> >> patch for zstd, but could be applied independently.\n> > \n> > One more - WriteDataToArchiveGzip() says:\n> > \n> > + if (cs->compression_spec.level == 0)\n> > + pg_fatal(\"requested to compress the archive yet no level was specified\");\n> > \n> > That was added at e9960732a. \n> > \n> > But if you specify gzip:0, the compression level is already enforced by\n> > validate_compress_specification(), before hitting gzip.c:\n> > \n> > | pg_dump: error: invalid compression specification: compression algorithm \"gzip\" expects a compression level between 1 and 9 (default at -1)\n> > \n> > 5e73a6048 intended that to work as before, and you *can* specify -Z0:\n> > \n> > The change is backward-compatible, hence specifying only an integer\n> > leads to no compression for a level of 0 and gzip compression when the\n> > level is greater than 0.\n> > \n> > $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\n> > /dev/stdin: ASCII text\n> \n> FWIW I agree we should make this backwards-compatible - accept \"0\" and\n> treat it as no compression.\n> \n> Georgios, can you prepare a patch doing that?\n\nI think maybe Tomas misunderstood. What I was trying to say is that -Z\n0 *is* accepted to mean no compression. This part wasn't quoted, but I\nsaid:\n\n> Right now, I think that pg_fatal in gzip.c is dead code - that was first\n> added in the patch version sent on 21 Dec 2022.\n\nIf you run the diff command that I've been talking about, you'll see\nthat InitCompressorZlib was almost unchanged - e9960732 is essentially a\nrefactoring. I don't think it's desirable to add a pg_fatal() in a\nfunction that's otherwise nearly-unchanged. The fact that it's\nnearly-unchanged is a good thing: it simplifies reading of what changed.\nIf someone wants to add a pg_fatal() in that code path, it'd be better\ndone in its own commit, with a separate message explaining the change.\n\nIf you insist on changing anything here, you might add an assertion (as\nyou said earlier) along with a comment like\n/* -Z 0 uses the \"None\" compressor rather than zlib with no compression */\n\n-- \nJustin\n\n\n","msg_date":"Thu, 2 Mar 2023 11:18:24 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 3/2/23 18:18, Justin Pryzby wrote:\n> On Wed, Mar 01, 2023 at 05:20:05PM +0100, Tomas Vondra wrote:\n>> On 2/25/23 15:05, Justin Pryzby wrote:\n>>> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n>>>> I have some fixes (attached) and questions while polishing the patch for\n>>>> zstd compression. The fixes are small and could be integrated with the\n>>>> patch for zstd, but could be applied independently.\n>>>\n>>> One more - WriteDataToArchiveGzip() says:\n>>>\n>>> + if (cs->compression_spec.level == 0)\n>>> + pg_fatal(\"requested to compress the archive yet no level was specified\");\n>>>\n>>> That was added at e9960732a. \n>>>\n>>> But if you specify gzip:0, the compression level is already enforced by\n>>> validate_compress_specification(), before hitting gzip.c:\n>>>\n>>> | pg_dump: error: invalid compression specification: compression algorithm \"gzip\" expects a compression level between 1 and 9 (default at -1)\n>>>\n>>> 5e73a6048 intended that to work as before, and you *can* specify -Z0:\n>>>\n>>> The change is backward-compatible, hence specifying only an integer\n>>> leads to no compression for a level of 0 and gzip compression when the\n>>> level is greater than 0.\n>>>\n>>> $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\n>>> /dev/stdin: ASCII text\n>>\n>> FWIW I agree we should make this backwards-compatible - accept \"0\" and\n>> treat it as no compression.\n>>\n>> Georgios, can you prepare a patch doing that?\n> \n> I think maybe Tomas misunderstood. What I was trying to say is that -Z\n> 0 *is* accepted to mean no compression. This part wasn't quoted, but I\n> said:\n> \n\nAh, I see. Well, I also tried but with \"-Z gzip:0\" (and not -Z 0), and\nthat does fail:\n\n error: invalid compression specification: compression algorithm \"gzip\"\n expects a compression level between 1 and 9 (default at -1)\n\nIt's a bit weird these two cases behave differently, when both translate\nto the same default compression method (gzip).\n\n>> Right now, I think that pg_fatal in gzip.c is dead code - that was first\n>> added in the patch version sent on 21 Dec 2022.\n> \n> If you run the diff command that I've been talking about, you'll see\n> that InitCompressorZlib was almost unchanged - e9960732 is essentially a\n> refactoring. I don't think it's desirable to add a pg_fatal() in a\n> function that's otherwise nearly-unchanged. The fact that it's\n> nearly-unchanged is a good thing: it simplifies reading of what changed.\n> If someone wants to add a pg_fatal() in that code path, it'd be better\n> done in its own commit, with a separate message explaining the change.\n> \n> If you insist on changing anything here, you might add an assertion (as\n> you said earlier) along with a comment like\n> /* -Z 0 uses the \"None\" compressor rather than zlib with no compression */\n> \n\nYeah, a comment would be helpful.\n\nAlso, after thinking about it a bit more maybe having the unreachable\npg_fatal() is not a good thing, as it will just confuse people (I'd\ncertainly assume having such check means there's a way in which case it\nmight trigger.). Maybe an assert would be better?\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 2 Mar 2023 19:27:00 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Mar 01, 2023 at 04:52:49PM +0100, Tomas Vondra wrote:\n> Thanks. That seems correct to me, but I find it somewhat confusing,\n> because we now have\n> \n> DeflateCompressorInit vs. InitCompressorGzip\n> \n> DeflateCompressorEnd vs. EndCompressorGzip\n> \n> DeflateCompressorData - The name doesn't really say what it does (would\n> be better to have a verb in there, I think).\n> \n> I wonder if we can make this somehow clearer?\n\nTo move things along, I updated Georgios' patch:\n\nRename DeflateCompressorData() to DeflateCompressorCommon();\nRearrange functions to their original order allowing a cleaner diff to the prior code;\nChange pg_fatal() to an assertion+comment;\nUpdate the commit message and fix a few typos;\n\n> Also, InitCompressorGzip says this:\n> \n> /*\n> * If the caller has defined a write function, prepare the necessary\n> * state. Avoid initializing during the first write call, because End\n> * may be called without ever writing any data.\n> */\n> if (cs->writeF)\n> DeflateCompressorInit(cs);\n>\n> Does it actually make sense to not have writeF defined in some cases?\n\nInitCompressor is being called for either reading or writing, either of\nwhich could be null:\n\nsrc/bin/pg_dump/pg_backup_custom.c: ctx->cs = AllocateCompressor(AH->compression_spec,\nsrc/bin/pg_dump/pg_backup_custom.c- NULL,\nsrc/bin/pg_dump/pg_backup_custom.c- _CustomWriteFunc);\n--\nsrc/bin/pg_dump/pg_backup_custom.c: cs = AllocateCompressor(AH->compression_spec,\nsrc/bin/pg_dump/pg_backup_custom.c- _CustomReadFunc, NULL);\n\nIt's confusing that the comment says \"Avoid initializing...\". What it\nreally means is \"Initialize eagerly...\". But that makes more sense in\nthe context of the commit message for this bugfix than in a comment. So\nI changed that too.\n\n+ /* If deflation was initialized, finalize it */ \n+ if (cs->private_data) \n+ DeflateCompressorEnd(AH, cs); \n\nMaybe it'd be more clear if this used \"if (cs->writeF)\", like in the\ninit function ?\n\n-- \nJustin","msg_date":"Tue, 7 Mar 2023 19:45:20 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Mar 01, 2023 at 05:39:54PM +0100, Tomas Vondra wrote:\n> On 2/27/23 05:49, Justin Pryzby wrote:\n> > On Sat, Feb 25, 2023 at 08:05:53AM -0600, Justin Pryzby wrote:\n> >> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n> >>> I have some fixes (attached) and questions while polishing the patch for\n> >>> zstd compression. The fixes are small and could be integrated with the\n> >>> patch for zstd, but could be applied independently.\n> >>\n> >> One more - WriteDataToArchiveGzip() says:\n> > \n> > One more again.\n> > \n> > The LZ4 path is using non-streaming mode, which compresses each block\n> > without persistent state, giving poor compression for -Fc compared with\n> > -Fp. If the data is highly compressible, the difference can be orders\n> > of magnitude.\n> > \n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fp |wc -c\n> > 12351763\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\n> > 21890708\n> > \n> > That's not true for gzip:\n> > \n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fc |wc -c\n> > 2118869\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fp |wc -c\n> > 2115832\n> > \n> > The function ought to at least use streaming mode, so each block/row\n> > isn't compressioned in isolation. 003 is a simple patch to use\n> > streaming mode, which improves the -Fc case:\n> > \n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\n> > 15178283\n> > \n> > However, that still flushes the compression buffer, writing a block\n> > header, for every row. With a single-column table, pg_dump -Fc -Z lz4\n> > still outputs ~10% *more* data than with no compression at all. And\n> > that's for compressible data.\n> > \n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z lz4 |wc -c\n> > 12890296\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z none |wc -c\n> > 11890296\n> > \n> > I think this should use the LZ4F API with frames, which are buffered to\n> > avoid outputting a header for every single row. The LZ4F format isn't\n> > compatible with the LZ4 format, so (unlike changing to the streaming\n> > API) that's not something we can change in a bugfix release. I consider\n> > this an Opened Item.\n> > \n> > With the LZ4F API in 004, -Fp and -Fc are essentially the same size\n> > (like gzip). (Oh, and the output is three times smaller, too.)\n> > \n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fp |wc -c\n> > 4155448\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fc |wc -c\n> > 4156548\n> \n> Thanks. Those are definitely interesting improvements/optimizations!\n> \n> I suggest we track them as a separate patch series - please add them to\n> the CF app (I guess you'll have to add them to 2023-07 at this point,\n> but we can get them in, I think).\n\nThanks for looking. I'm not sure if I'm the best person to write/submit\nthe patch to implement that for LZ4. Georgios, would you want to take\non this change ?\n\nI think that needs to be changed for v16, since 1) LZ4F works so much\nbetter like this, and 2) we can't change it later without breaking\ncompatibility of the dumpfiles by changing the header with another name\nother than \"lz4\". Also, I imagine we'd want to continue supporting the\nability to *restore* a dumpfile using the old(current) format, which\nwould be untestable code unless we also preserved the ability to write\nit somehow (like -Z lz4-old).\n\nOne issue is that LZ4F_createCompressionContext() and\nLZ4F_compressBegin() ought to be called in InitCompressorLZ4(). It\nseems like it might *need* to be called there to avoid exactly the kind\nof issue that I reported with empty LOs with gzip. But\nInitCompressorLZ4() isn't currently passed the ArchiveHandle, so can't\nwrite the header. And LZ4CompressorState has a simple char *buf, and\nnot an more elaborate data structure like zlib. You could work around\nthat by keeping storing the \"len\" of the existing buffer, and flushing\nit in EndCompressorLZ4(), but that adds needless complexity to the Write\nand End functions. Maybe the Init function should be passed the AH.\n\n-- \nJustin\n\n\n","msg_date":"Thu, 9 Mar 2023 10:15:55 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 3/9/23 17:15, Justin Pryzby wrote:\n> On Wed, Mar 01, 2023 at 05:39:54PM +0100, Tomas Vondra wrote:\n>> On 2/27/23 05:49, Justin Pryzby wrote:\n>>> On Sat, Feb 25, 2023 at 08:05:53AM -0600, Justin Pryzby wrote:\n>>>> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\n>>>>> I have some fixes (attached) and questions while polishing the patch for\n>>>>> zstd compression. The fixes are small and could be integrated with the\n>>>>> patch for zstd, but could be applied independently.\n>>>>\n>>>> One more - WriteDataToArchiveGzip() says:\n>>>\n>>> One more again.\n>>>\n>>> The LZ4 path is using non-streaming mode, which compresses each block\n>>> without persistent state, giving poor compression for -Fc compared with\n>>> -Fp. If the data is highly compressible, the difference can be orders\n>>> of magnitude.\n>>>\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fp |wc -c\n>>> 12351763\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\n>>> 21890708\n>>>\n>>> That's not true for gzip:\n>>>\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fc |wc -c\n>>> 2118869\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fp |wc -c\n>>> 2115832\n>>>\n>>> The function ought to at least use streaming mode, so each block/row\n>>> isn't compressioned in isolation. 003 is a simple patch to use\n>>> streaming mode, which improves the -Fc case:\n>>>\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\n>>> 15178283\n>>>\n>>> However, that still flushes the compression buffer, writing a block\n>>> header, for every row. With a single-column table, pg_dump -Fc -Z lz4\n>>> still outputs ~10% *more* data than with no compression at all. And\n>>> that's for compressible data.\n>>>\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z lz4 |wc -c\n>>> 12890296\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z none |wc -c\n>>> 11890296\n>>>\n>>> I think this should use the LZ4F API with frames, which are buffered to\n>>> avoid outputting a header for every single row. The LZ4F format isn't\n>>> compatible with the LZ4 format, so (unlike changing to the streaming\n>>> API) that's not something we can change in a bugfix release. I consider\n>>> this an Opened Item.\n>>>\n>>> With the LZ4F API in 004, -Fp and -Fc are essentially the same size\n>>> (like gzip). (Oh, and the output is three times smaller, too.)\n>>>\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fp |wc -c\n>>> 4155448\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fc |wc -c\n>>> 4156548\n>>\n>> Thanks. Those are definitely interesting improvements/optimizations!\n>>\n>> I suggest we track them as a separate patch series - please add them to\n>> the CF app (I guess you'll have to add them to 2023-07 at this point,\n>> but we can get them in, I think).\n> \n> Thanks for looking. I'm not sure if I'm the best person to write/submit\n> the patch to implement that for LZ4. Georgios, would you want to take\n> on this change ?\n> \n> I think that needs to be changed for v16, since 1) LZ4F works so much\n> better like this, and 2) we can't change it later without breaking\n> compatibility of the dumpfiles by changing the header with another name\n> other than \"lz4\". Also, I imagine we'd want to continue supporting the\n> ability to *restore* a dumpfile using the old(current) format, which\n> would be untestable code unless we also preserved the ability to write\n> it somehow (like -Z lz4-old).\n> \n\nI'm a bit confused about the lz4 vs. lz4f stuff, TBH. If we switch to\nlz4f, doesn't that mean it (e.g. restore) won't work on systems that\nonly have older lz4 version? What would/should happen if we take backup\ncompressed with lz4f, an then try restoring it on an older system where\nlz4 does not support lz4f?\n\nMaybe if lz4f format is incompatible with regular lz4, we should treat\nit as a separate compression method 'lz4f'?\n\nI'm mostly afk until the end of the week, but I tried searching for lz4f\ninfo - the results are not particularly enlightening, unfortunately.\n\nAFAICS this only applies to lz4f stuff. Or would the streaming mode be a\nbreaking change too?\n\n> One issue is that LZ4F_createCompressionContext() and\n> LZ4F_compressBegin() ought to be called in InitCompressorLZ4(). It\n> seems like it might *need* to be called there to avoid exactly the kind\n> of issue that I reported with empty LOs with gzip. But\n> InitCompressorLZ4() isn't currently passed the ArchiveHandle, so can't\n> write the header. And LZ4CompressorState has a simple char *buf, and\n> not an more elaborate data structure like zlib. You could work around\n> that by keeping storing the \"len\" of the existing buffer, and flushing\n> it in EndCompressorLZ4(), but that adds needless complexity to the Write\n> and End functions. Maybe the Init function should be passed the AH.\n> \n\nNot sure, but looking at GzipCompressorState I see the only extra thing\nit has (compared to LZ4CompressorState) is \"z_streamp\". I can't\nexperiment with this until the end of this week, so perhaps that's not\nworkable, but wouldn't it be better to add a similar field into\nLZ4CompressorState? Passing AH to the init function seems like a\nviolation of abstraction.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 9 Mar 2023 18:58:20 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Mar 09, 2023 at 06:58:20PM +0100, Tomas Vondra wrote:\n> I'm a bit confused about the lz4 vs. lz4f stuff, TBH. If we switch to\n> lz4f, doesn't that mean it (e.g. restore) won't work on systems that\n> only have older lz4 version? What would/should happen if we take backup\n> compressed with lz4f, an then try restoring it on an older system where\n> lz4 does not support lz4f?\n\nYou seem to be thinking about LZ4F as a weird, new innovation I'm\nexperimenting with, but compress_lz4.c already uses LZ4F for its \"file\"\nAPI. LZ4F is also what's written by the lz4 CLI tool, and I found that\nLZ4F has been included in the library for ~8 years:\n\nhttps://github.com/lz4/lz4/releases?page=2\nr126 Dec 24, 2014\nNew : lz4frame API is now integrated into liblz4\n\n> Maybe if lz4f format is incompatible with regular lz4, we should treat\n> it as a separate compression method 'lz4f'?\n> \n> I'm mostly afk until the end of the week, but I tried searching for lz4f\n> info - the results are not particularly enlightening, unfortunately.\n> \n> AFAICS this only applies to lz4f stuff. Or would the streaming mode be a\n> breaking change too?\n\nStreaming mode outputs the same format as the existing code, but gives\nbetter compression. We could (theoretically) change it in a bugfix\nrelease, and old output would still be restorable (I think new output\nwould even be restorable with the old versions of pg_restore).\n\nBut that's not true for LZ4F. The benefit there is that it avoids\noutputing a separate block for each row. That's essential for narrow\ntables, for which the block header currently being written has an\noverhead several times larger than the data.\n\n-- \nJustin\n\n\n","msg_date":"Fri, 10 Mar 2023 07:05:49 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Mar 10, 2023 at 07:05:49AM -0600, Justin Pryzby wrote:\n> On Thu, Mar 09, 2023 at 06:58:20PM +0100, Tomas Vondra wrote:\n>> I'm a bit confused about the lz4 vs. lz4f stuff, TBH. If we switch to\n>> lz4f, doesn't that mean it (e.g. restore) won't work on systems that\n>> only have older lz4 version? What would/should happen if we take backup\n>> compressed with lz4f, an then try restoring it on an older system where\n>> lz4 does not support lz4f?\n> \n> You seem to be thinking about LZ4F as a weird, new innovation I'm\n> experimenting with, but compress_lz4.c already uses LZ4F for its \"file\"\n> API.\n\nNote: we already use lz4 frames in pg_receivewal (for WAL) and\npg_basebackup (bbstreamer).\n--\nMichael","msg_date":"Sat, 11 Mar 2023 08:00:55 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Hello,\n23.02.2023 23:24, Tomas Vondra wrote:\n> On 2/23/23 16:26, Tomas Vondra wrote:\n>> Thanks for v30 with the updated commit messages. I've pushed 0001 after\n>> fixing a comment typo and removing (I think) an unnecessary change in an\n>> error message.\n>>\n>> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\n>>\n> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\n> and marked the CF entry as committed. Thanks for the patch!\n>\n> I wonder how difficult would it be to add the zstd compression, so that\n> we don't have the annoying \"unsupported\" cases.\n\nWith the patch 0003 committed, a single warning -Wtype-limits appeared in the\nmaster branch:\n$ CPPFLAGS=\"-Og -Wtype-limits\" ./configure --with-lz4 -q && make -s -j8\ncompress_lz4.c: In function ‘LZ4File_gets’:\ncompress_lz4.c:492:19: warning: comparison of unsigned expression in ‘< 0’ is \nalways false [-Wtype-limits]\n 492 | if (dsize < 0)\n |\n(I wonder, is it accidental that there no other places that triggers\nthe warning, or some buildfarm animals had this check enabled before?)\n\nIt is not a false positive as can be proved by the 002_pg_dump.pl modified as\nfollows:\n- program => $ENV{'LZ4'},\n+ program => 'mv',\n args => [\n- '-z', '-f', '--rm',\n\"$tempdir/compression_lz4_dir/blobs.toc\",\n\"$tempdir/compression_lz4_dir/blobs.toc.lz4\",\n ],\n },\nA diagnostic logging added shows:\nLZ4File_gets() after LZ4File_read_internal; dsize: 18446744073709551615\n\nand pg_restore fails with:\nerror: invalid line in large object TOC file \n\".../src/bin/pg_dump/tmp_check/tmp_test_22ri/compression_lz4_dir/blobs.toc\": \"????\"\n\nBest regards,\nAlexander\n\n\n","msg_date":"Sat, 11 Mar 2023 09:00:00 +0300","msg_from":"Alexander Lakhin <exclusion@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Saturday, March 11th, 2023 at 7:00 AM, Alexander Lakhin <exclusion@gmail.com> wrote:\n\n> Hello,\n> 23.02.2023 23:24, Tomas Vondra wrote:\n> \n> > On 2/23/23 16:26, Tomas Vondra wrote:\n> > \n> > > Thanks for v30 with the updated commit messages. I've pushed 0001 after\n> > > fixing a comment typo and removing (I think) an unnecessary change in an\n> > > error message.\n> > > \n> > > I'll give the buildfarm a bit of time before pushing 0002 and 0003.\n> > \n> > I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\n> > and marked the CF entry as committed. Thanks for the patch!\n> > \n> > I wonder how difficult would it be to add the zstd compression, so that\n> > we don't have the annoying \"unsupported\" cases.\n> \n> \n> With the patch 0003 committed, a single warning -Wtype-limits appeared in the\n> master branch:\n> $ CPPFLAGS=\"-Og -Wtype-limits\" ./configure --with-lz4 -q && make -s -j8\n> compress_lz4.c: In function ‘LZ4File_gets’:\n> compress_lz4.c:492:19: warning: comparison of unsigned expression in ‘< 0’ is\n> always false [-Wtype-limits]\n> 492 | if (dsize < 0)\n> |\n\nThank you Alexander. Please find attached an attempt to address it.\n\n> (I wonder, is it accidental that there no other places that triggers\n> the warning, or some buildfarm animals had this check enabled before?)\n\nI can not answer about the buildfarms. Do you think that adding an explicit\ncheck for this warning in meson would help? I am a bit uncertain as I think\nthat type-limits are included in extra.\n\n@@ -1748,6 +1748,7 @@ common_warning_flags = [\n '-Wshadow=compatible-local',\n # This was included in -Wall/-Wformat in older GCC versions\n '-Wformat-security',\n+ '-Wtype-limits',\n ]\n\n> \n> It is not a false positive as can be proved by the 002_pg_dump.pl modified as\n> follows:\n> - program => $ENV{'LZ4'},\n> \n> + program => 'mv',\n> \n> args => [\n> \n> - '-z', '-f', '--rm',\n> \"$tempdir/compression_lz4_dir/blobs.toc\",\n> \"$tempdir/compression_lz4_dir/blobs.toc.lz4\",\n> ],\n> },\n\nCorrect, it is not a false positive. The existing testing framework provides\nlimited support for exercising error branches. Especially so when those are\ndependent on generated output. \n\n> A diagnostic logging added shows:\n> LZ4File_gets() after LZ4File_read_internal; dsize: 18446744073709551615\n> \n> and pg_restore fails with:\n> error: invalid line in large object TOC file\n> \".../src/bin/pg_dump/tmp_check/tmp_test_22ri/compression_lz4_dir/blobs.toc\": \"????\"\n\nIt is a good thing that the restore fails with bad input. Yet it should\nhave failed earlier. The attached makes certain it does fail earlier. \n\nCheers,\n//Georgios\n\n> \n> Best regards,\n> Alexander","msg_date":"Sat, 11 Mar 2023 10:50:06 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Hi Georgios,\n\n11.03.2023 13:50, gkokolatos@pm.me wrote:\n> I can not answer about the buildfarms. Do you think that adding an explicit\n> check for this warning in meson would help? I am a bit uncertain as I think\n> that type-limits are included in extra.\n>\n> @@ -1748,6 +1748,7 @@ common_warning_flags = [\n> '-Wshadow=compatible-local',\n> # This was included in -Wall/-Wformat in older GCC versions\n> '-Wformat-security',\n> + '-Wtype-limits',\n> ]\nI'm not sure that I can promote additional checks (or determine where\nto put them), but if some patch introduces a warning of a type that wasn't\npresent before, I think it's worth to eliminate the warning (if it is\nsensible) to keep the source code check baseline at the same level\nor even lift it up gradually.\nI've also found that the same commit introduced a single instance of\nthe analyzer-possible-null-argument warning:\nCPPFLAGS=\"-Og -fanalyzer -Wno-analyzer-malloc-leak -Wno-analyzer-file-leak \n-Wno-analyzer-null-dereference -Wno-analyzer-shift-count-overflow \n-Wno-analyzer-free-of-non-heap -Wno-analyzer-null-argument \n-Wno-analyzer-double-free -Wanalyzer-possible-null-argument\" ./configure \n--with-lz4 -q && make -s -j8\ncompress_io.c: In function ‘hasSuffix’:\ncompress_io.c:158:47: warning: use of possibly-NULL ‘filename’ where non-null \nexpected [CWE-690] [-Wanalyzer-possible-null-argument]\n 158 | int filenamelen = strlen(filename);\n | ^~~~~~~~~~~~~~~~\n ‘InitDiscoverCompressFileHandle’: events 1-3\n...\n\n(I use gcc-11.3.)\nAs I can see, many existing uses of strdup() are followed by a check for\nnull result, so maybe it's a common practice and a similar check should\nbe added in InitDiscoverCompressFileHandle().\n(There also a couple of other warnings introduced with the lz4 compression\npatches, but those ones are not unique, so I maybe they aren't worth fixing.)\n\n>> It is a good thing that the restore fails with bad input. Yet it should\n>> have failed earlier. The attached makes certain it does fail earlier.\n>>\nThanks! Your patch definitely fixes the issue.\n\nBest regards,\nAlexander\n\n\n","msg_date":"Sat, 11 Mar 2023 21:00:00 +0300","msg_from":"Alexander Lakhin <exclusion@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 11.03.23 07:00, Alexander Lakhin wrote:\n> Hello,\n> 23.02.2023 23:24, Tomas Vondra wrote:\n>> On 2/23/23 16:26, Tomas Vondra wrote:\n>>> Thanks for v30 with the updated commit messages. I've pushed 0001 after\n>>> fixing a comment typo and removing (I think) an unnecessary change in an\n>>> error message.\n>>>\n>>> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\n>>>\n>> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\n>> and marked the CF entry as committed. Thanks for the patch!\n>>\n>> I wonder how difficult would it be to add the zstd compression, so that\n>> we don't have the annoying \"unsupported\" cases.\n> \n> With the patch 0003 committed, a single warning -Wtype-limits appeared \n> in the\n> master branch:\n> $ CPPFLAGS=\"-Og -Wtype-limits\" ./configure --with-lz4 -q && make -s -j8\n> compress_lz4.c: In function ‘LZ4File_gets’:\n> compress_lz4.c:492:19: warning: comparison of unsigned expression in ‘< \n> 0’ is always false [-Wtype-limits]\n> 492 | if (dsize < 0)\n> |\n> (I wonder, is it accidental that there no other places that triggers\n> the warning, or some buildfarm animals had this check enabled before?)\n\nI think there is an underlying problem in this code that it dances back \nand forth between size_t and int in an unprincipled way.\n\nIn the code that triggers the warning, dsize is size_t. dsize is the \nreturn from LZ4File_read_internal(), which is declared to return int. \nThe variable that LZ4File_read_internal() returns in the success case is \nsize_t, but in case of an error it returns -1. (So the code that is \nwarning is meaning to catch this error case, but it won't ever work.) \nFurther below LZ4File_read_internal() calls LZ4File_read_overflow(), \nwhich is declared to return int, but in some cases it returns \nfs->overflowlen, which is size_t.\n\nThis should be cleaned up.\n\nAFAICT, the upstream API in lz4.h uses int for size values, but \nlz4frame.h uses size_t, so I don't know what the correct approach is.\n\n\n","msg_date":"Sun, 12 Mar 2023 11:07:33 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 3/12/23 11:07, Peter Eisentraut wrote:\n> On 11.03.23 07:00, Alexander Lakhin wrote:\n>> Hello,\n>> 23.02.2023 23:24, Tomas Vondra wrote:\n>>> On 2/23/23 16:26, Tomas Vondra wrote:\n>>>> Thanks for v30 with the updated commit messages. I've pushed 0001 after\n>>>> fixing a comment typo and removing (I think) an unnecessary change\n>>>> in an\n>>>> error message.\n>>>>\n>>>> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\n>>>>\n>>> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\n>>> and marked the CF entry as committed. Thanks for the patch!\n>>>\n>>> I wonder how difficult would it be to add the zstd compression, so that\n>>> we don't have the annoying \"unsupported\" cases.\n>>\n>> With the patch 0003 committed, a single warning -Wtype-limits appeared\n>> in the\n>> master branch:\n>> $ CPPFLAGS=\"-Og -Wtype-limits\" ./configure --with-lz4 -q && make -s -j8\n>> compress_lz4.c: In function ‘LZ4File_gets’:\n>> compress_lz4.c:492:19: warning: comparison of unsigned expression in\n>> ‘< 0’ is always false [-Wtype-limits]\n>> 492 | if (dsize < 0)\n>> |\n>> (I wonder, is it accidental that there no other places that triggers\n>> the warning, or some buildfarm animals had this check enabled before?)\n> \n> I think there is an underlying problem in this code that it dances back\n> and forth between size_t and int in an unprincipled way.\n> \n> In the code that triggers the warning, dsize is size_t. dsize is the\n> return from LZ4File_read_internal(), which is declared to return int.\n> The variable that LZ4File_read_internal() returns in the success case is\n> size_t, but in case of an error it returns -1. (So the code that is\n> warning is meaning to catch this error case, but it won't ever work.)\n> Further below LZ4File_read_internal() calls LZ4File_read_overflow(),\n> which is declared to return int, but in some cases it returns\n> fs->overflowlen, which is size_t.\n> \n\nI agree. I just got home so I looked at this only very briefly, but I\nthink it's clearly wrong to assign the LZ4File_read_internal() result to\na size_t variable (and it seems to me LZ4File_gets does the same mistake\nwith LZ4File_read_internal() result).\n\nI'll get this fixed early next week, I'm too tired to do that now\nwithout likely causing further issues.\n\n> This should be cleaned up.\n> \n> AFAICT, the upstream API in lz4.h uses int for size values, but\n> lz4frame.h uses size_t, so I don't know what the correct approach is.\n\nYeah, that's a good point. I think Justin is right we should be using\nthe LZ4F stuff, so ultimately we'll probably switch to size_t. But IMO\nit's definitely better to correct the current code first, and only then\nswitch to LZ4F (from one correct state to another).\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Sun, 12 Mar 2023 19:34:02 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 3/11/23 11:50, gkokolatos@pm.me wrote:\n> ------- Original Message -------\n> On Saturday, March 11th, 2023 at 7:00 AM, Alexander Lakhin <exclusion@gmail.com> wrote:\n> \n>> Hello,\n>> 23.02.2023 23:24, Tomas Vondra wrote:\n>>\n>>> On 2/23/23 16:26, Tomas Vondra wrote:\n>>>\n>>>> Thanks for v30 with the updated commit messages. I've pushed 0001 after\n>>>> fixing a comment typo and removing (I think) an unnecessary change in an\n>>>> error message.\n>>>>\n>>>> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\n>>>\n>>> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\n>>> and marked the CF entry as committed. Thanks for the patch!\n>>>\n>>> I wonder how difficult would it be to add the zstd compression, so that\n>>> we don't have the annoying \"unsupported\" cases.\n>>\n>>\n>> With the patch 0003 committed, a single warning -Wtype-limits appeared in the\n>> master branch:\n>> $ CPPFLAGS=\"-Og -Wtype-limits\" ./configure --with-lz4 -q && make -s -j8\n>> compress_lz4.c: In function ‘LZ4File_gets’:\n>> compress_lz4.c:492:19: warning: comparison of unsigned expression in ‘< 0’ is\n>> always false [-Wtype-limits]\n>> 492 | if (dsize < 0)\n>> |\n> \n> Thank you Alexander. Please find attached an attempt to address it.\n> \n>> (I wonder, is it accidental that there no other places that triggers\n>> the warning, or some buildfarm animals had this check enabled before?)\n> \n> I can not answer about the buildfarms. Do you think that adding an explicit\n> check for this warning in meson would help? I am a bit uncertain as I think\n> that type-limits are included in extra.\n> \n> @@ -1748,6 +1748,7 @@ common_warning_flags = [\n> '-Wshadow=compatible-local',\n> # This was included in -Wall/-Wformat in older GCC versions\n> '-Wformat-security',\n> + '-Wtype-limits',\n> ]\n> \n>>\n>> It is not a false positive as can be proved by the 002_pg_dump.pl modified as\n>> follows:\n>> - program => $ENV{'LZ4'},\n>>\n>> + program => 'mv',\n>>\n>> args => [\n>>\n>> - '-z', '-f', '--rm',\n>> \"$tempdir/compression_lz4_dir/blobs.toc\",\n>> \"$tempdir/compression_lz4_dir/blobs.toc.lz4\",\n>> ],\n>> },\n> \n> Correct, it is not a false positive. The existing testing framework provides\n> limited support for exercising error branches. Especially so when those are\n> dependent on generated output. \n> \n>> A diagnostic logging added shows:\n>> LZ4File_gets() after LZ4File_read_internal; dsize: 18446744073709551615\n>>\n>> and pg_restore fails with:\n>> error: invalid line in large object TOC file\n>> \".../src/bin/pg_dump/tmp_check/tmp_test_22ri/compression_lz4_dir/blobs.toc\": \"????\"\n> \n> It is a good thing that the restore fails with bad input. Yet it should\n> have failed earlier. The attached makes certain it does fail earlier. \n> \n\nThanks for the patch.\n\nI did look if there are other places that might have the same issue, and\nI think there are - see attached 0002. For example LZ4File_write is\ndeclared to return size_t, but then it also does\n\n if (LZ4F_isError(status))\n {\n fs->errcode = status;\n return -1;\n }\n\nThat won't work :-(\n\nAnd these issues may not be restricted to lz4 code - Gzip_write is\ndeclared to return size_t, but it does\n\n return gzwrite(gzfp, ptr, size);\n\nand gzwrite returns int. Although, maybe that's correct, because\ngzwrite() is \"0 on error\" so maybe this is fine ...\n\nHowever, Gzip_read assigns gzread() to size_t, and that does not seem\ngreat. It probably will still trigger the following pg_fatal() because\nit'd be very lucky to match the expected 'size', but it's confusing.\n\n\nI wonder whether CompressorState should use int or size_t for the\nread_func/write_func callbacks. I guess no option is perfect, i.e. no\ndata type will work for all compression libraries we might use (lz4 uses\nint while lz4f uses size_t, to there's that).\n\nIt's a bit weird the \"open\" functions return int and the read/write\nsize_t. Maybe we should stick to int, which is what the old functions\n(cfwrite etc.) did.\n\n\nBut I think the actual problem here is that the API does not clearly\ndefine how errors are communicated. I mean, it's nice to return the\nvalue returned by the library function without \"mangling\" it by\nconversion to size_t, but what if the libraries communicate errors in\ndifferent way? Some may return \"0\" while others may return \"-1\".\n\nI think the right approach is to handle all library errors and not just\nlet them through. So Gzip_write() needs to check the return value, and\neither call pg_fatal() or translate it to an error defined by the API.\n\nFor example we might say \"returns 0 on error\" and then translate all\nlibrary-specific errors to that.\n\n\nWhile looking at the code I realized a couple function comments don't\nsay what's returned in case of error, etc. So 0004 adds those.\n\n0003 is a couple minor assorted comments/questions:\n\n- Should we move ZLIB_OUT_SIZE/ZLIB_IN_SIZE to compress_gzip.c?\n\n- Why are LZ4 buffer sizes different (ZLIB has both 4kB)?\n\n- I wonder if we actually need LZ4F_HEADER_SIZE_MAX? Is it even possible\nfor LZ4F_compressBound to return value this small (especially for 16kB\ninput buffer)?\n\n\n\nregards\n\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 13 Mar 2023 21:21:11 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Hi Justin,\n\nThanks for the patch.\n\nOn 3/8/23 02:45, Justin Pryzby wrote:\n> On Wed, Mar 01, 2023 at 04:52:49PM +0100, Tomas Vondra wrote:\n>> Thanks. That seems correct to me, but I find it somewhat confusing,\n>> because we now have\n>>\n>> DeflateCompressorInit vs. InitCompressorGzip\n>>\n>> DeflateCompressorEnd vs. EndCompressorGzip\n>>\n>> DeflateCompressorData - The name doesn't really say what it does (would\n>> be better to have a verb in there, I think).\n>>\n>> I wonder if we can make this somehow clearer?\n> \n> To move things along, I updated Georgios' patch:\n> \n> Rename DeflateCompressorData() to DeflateCompressorCommon();\n\nHmmm, I don't find \"common\" any clearer than \"data\" :-( There needs to\nat least be a comment explaining what \"common\" does.\n\n> Rearrange functions to their original order allowing a cleaner diff to the prior code;\n\nOK. I wasn't very enthusiastic about this initially, but after thinking\nabout it a bit I think it's meaningful to make diffs clearer. But I\ndon't see much difference with/without the patch. The\n\ngit diff --diff-algorithm=minimal -w\ne9960732a~:src/bin/pg_dump/compress_io.c src/bin/pg_dump/compress_gzip.c\n\nProduces ~25k diff with/without the patch. What am I doing wrong?\n\n> Change pg_fatal() to an assertion+comment;\n\nYeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\ncould add such protections against \"impossible\" stuff to a zillion other\nplaces and the confusion likely outweighs the benefits.\n\n> Update the commit message and fix a few typos;\n> \n\nThanks. I don't want to annoy you too much, but could you split the\npatch into the \"empty-data\" fix and all the other changes (rearranging\nfunctions etc.)? I'd rather not mix those in the same commit.\n\n\n>> Also, InitCompressorGzip says this:\n>>\n>> /*\n>> * If the caller has defined a write function, prepare the necessary\n>> * state. Avoid initializing during the first write call, because End\n>> * may be called without ever writing any data.\n>> */\n>> if (cs->writeF)\n>> DeflateCompressorInit(cs);\n>>\n>> Does it actually make sense to not have writeF defined in some cases?\n> \n> InitCompressor is being called for either reading or writing, either of\n> which could be null:\n> \n> src/bin/pg_dump/pg_backup_custom.c: ctx->cs = AllocateCompressor(AH->compression_spec,\n> src/bin/pg_dump/pg_backup_custom.c- NULL,\n> src/bin/pg_dump/pg_backup_custom.c- _CustomWriteFunc);\n> --\n> src/bin/pg_dump/pg_backup_custom.c: cs = AllocateCompressor(AH->compression_spec,\n> src/bin/pg_dump/pg_backup_custom.c- _CustomReadFunc, NULL);\n> \n> It's confusing that the comment says \"Avoid initializing...\". What it\n> really means is \"Initialize eagerly...\". But that makes more sense in\n> the context of the commit message for this bugfix than in a comment. So\n> I changed that too.\n> \n> + /* If deflation was initialized, finalize it */ \n> + if (cs->private_data) \n> + DeflateCompressorEnd(AH, cs); \n> \n> Maybe it'd be more clear if this used \"if (cs->writeF)\", like in the\n> init function ?\n> \n\nYeah, if the two checks are equivalent, it'd be better to stick to the\nsame check everywhere.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Mon, 13 Mar 2023 22:47:12 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n------- Original Message -------\nOn Monday, March 13th, 2023 at 10:47 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n\n> \n> > Change pg_fatal() to an assertion+comment;\n> \n> \n> Yeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\n> could add such protections against \"impossible\" stuff to a zillion other\n> places and the confusion likely outweighs the benefits.\n> \n\nA minor note to add is to not ignore the lessons learned from a7885c9bb.\n\nFor example, as the testing framework stands, one can not test that the\ncontents of the custom format are indeed compressed. One can infer it by\nexamining the header of the produced dump and searching for the\ncompression flag. The code responsible for writing the header and the\ncode responsible for actually dealing with data, is not the same. Also,\nthe compression library itself will happily read and write uncompressed\ndata.\n\nA pg_fatal, assertion, or similar, is the only guard rail against this\nkind of error. Without it, the tests will continue passing even after\ne.g. a wrong initialization of the API. It was such a case that lead to\na7885c9bb in the first place. I do think that we wish it to be an\n\"impossible\" case. Also it will be an untested case with some history\nwithout such a guard rail.\n\nOf course I will not object to removing it, if you think that is more\nconfusing than useful.\n\nCheers,\n//Georgios\n\n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company\n\n\n","msg_date":"Tue, 14 Mar 2023 11:07:25 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Monday, March 13th, 2023 at 9:21 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n>\n>\n>\n>\n> On 3/11/23 11:50, gkokolatos@pm.me wrote:\n>\n> > ------- Original Message -------\n> > On Saturday, March 11th, 2023 at 7:00 AM, Alexander Lakhin exclusion@gmail.com wrote:\n> >\n> > > Hello,\n> > > 23.02.2023 23:24, Tomas Vondra wrote:\n\n>\n>\n> Thanks for the patch.\n>\n> I did look if there are other places that might have the same issue, and\n> I think there are - see attached 0002. For example LZ4File_write is\n> declared to return size_t, but then it also does\n>\n> if (LZ4F_isError(status))\n> {\n> fs->errcode = status;\n>\n> return -1;\n> }\n>\n> That won't work :-(\n\nYou are right. It is confusing.\n\n>\n> And these issues may not be restricted to lz4 code - Gzip_write is\n> declared to return size_t, but it does\n>\n> return gzwrite(gzfp, ptr, size);\n>\n> and gzwrite returns int. Although, maybe that's correct, because\n> gzwrite() is \"0 on error\" so maybe this is fine ...\n>\n> However, Gzip_read assigns gzread() to size_t, and that does not seem\n> great. It probably will still trigger the following pg_fatal() because\n> it'd be very lucky to match the expected 'size', but it's confusing.\n\nAgreed.\n\n>\n>\n> I wonder whether CompressorState should use int or size_t for the\n> read_func/write_func callbacks. I guess no option is perfect, i.e. no\n> data type will work for all compression libraries we might use (lz4 uses\n> int while lz4f uses size_t, to there's that).\n>\n> It's a bit weird the \"open\" functions return int and the read/write\n> size_t. Maybe we should stick to int, which is what the old functions\n> (cfwrite etc.) did.\n>\nYou are right. These functions are modeled by the open/fread/\nfwrite etc, and they have kept the return types of these ones. Their\ncallers do check the return value of read_func and write_func against\nthe requested size of bytes to be transferred.\n\n>\n> But I think the actual problem here is that the API does not clearly\n> define how errors are communicated. I mean, it's nice to return the\n> value returned by the library function without \"mangling\" it by\n> conversion to size_t, but what if the libraries communicate errors in\n> different way? Some may return \"0\" while others may return \"-1\".\n\nAgreed.\n\n>\n> I think the right approach is to handle all library errors and not just\n> let them through. So Gzip_write() needs to check the return value, and\n> either call pg_fatal() or translate it to an error defined by the API.\n\nIt makes sense. It will change some of the behaviour of the callers,\nmostly on what constitutes an error, and what error message is emitted.\nThis is a reasonable change though.\n\n>\n> For example we might say \"returns 0 on error\" and then translate all\n> library-specific errors to that.\n\nOk.\n\n> While looking at the code I realized a couple function comments don't\n> say what's returned in case of error, etc. So 0004 adds those.\n>\n> 0003 is a couple minor assorted comments/questions:\n>\n> - Should we move ZLIB_OUT_SIZE/ZLIB_IN_SIZE to compress_gzip.c?\n\nIt would make things clearer.\n\n> - Why are LZ4 buffer sizes different (ZLIB has both 4kB)?\n\nClearly some comments are needed, if the difference makes sense.\n\n> - I wonder if we actually need LZ4F_HEADER_SIZE_MAX? Is it even possible\n> for LZ4F_compressBound to return value this small (especially for 16kB\n> input buffer)?\n>\n\nI would recommend to keep it. Earlier versions of LZ4F_HEADER_SIZE_MAX\ndo not have it. Later versions do advise to use it.\n\nWould you mind me trying to come with a patch to address your points?\n\nCheers,\n//Georgios\n\n>\n>\n> regards\n>\n>\n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company\n\n\n","msg_date":"Tue, 14 Mar 2023 15:18:33 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 3/14/23 16:18, gkokolatos@pm.me wrote:\n> ...> Would you mind me trying to come with a patch to address your points?\n> \n\nThat'd be great, thanks. Please keep it split into smaller patches - two\nmight work, with one patch for \"cosmetic\" changes and the other tweaking\nthe API error-handling stuff.\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Tue, 14 Mar 2023 16:32:56 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 3/14/23 12:07, gkokolatos@pm.me wrote:\n> \n> \n> ------- Original Message -------\n> On Monday, March 13th, 2023 at 10:47 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n> \n> \n> \n>>\n>>> Change pg_fatal() to an assertion+comment;\n>>\n>>\n>> Yeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\n>> could add such protections against \"impossible\" stuff to a zillion other\n>> places and the confusion likely outweighs the benefits.\n>>\n> \n> A minor note to add is to not ignore the lessons learned from a7885c9bb.\n> \n> For example, as the testing framework stands, one can not test that the\n> contents of the custom format are indeed compressed. One can infer it by\n> examining the header of the produced dump and searching for the\n> compression flag. The code responsible for writing the header and the\n> code responsible for actually dealing with data, is not the same. Also,\n> the compression library itself will happily read and write uncompressed\n> data.\n> \n> A pg_fatal, assertion, or similar, is the only guard rail against this\n> kind of error. Without it, the tests will continue passing even after\n> e.g. a wrong initialization of the API. It was such a case that lead to\n> a7885c9bb in the first place. I do think that we wish it to be an\n> \"impossible\" case. Also it will be an untested case with some history\n> without such a guard rail.\n> \n\nSo is the pg_fatal() a dead code or not? My understanding was it's not\nreally reachable, and the main purpose is to remind people this is not\npossible. Or am I mistaken/confused?\n\nIf it's reachable, can we test it? AFAICS we don't, per the coverage\nreports.\n\nIf it's just a protection against incorrect API initialization, then an\nassert is the right solution, I think. With proper comment. But can't we\nactually verify that *during* the initialization?\n\nAlso, how come WriteDataToArchiveLZ4() doesn't need this protection too?\nOr is that due to gzip being the default compression method?\n\n> Of course I will not object to removing it, if you think that is more\n> confusing than useful.\n> \n\nNot sure, I have a feeling I don't quite understand in what situation\nthis actually helps.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Tue, 14 Mar 2023 16:52:20 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Mar 13, 2023 at 10:47:12PM +0100, Tomas Vondra wrote:\n> > Rearrange functions to their original order allowing a cleaner diff to the prior code;\n> \n> OK. I wasn't very enthusiastic about this initially, but after thinking\n> about it a bit I think it's meaningful to make diffs clearer. But I\n> don't see much difference with/without the patch. The\n> \n> git diff --diff-algorithm=minimal -w e9960732a~:src/bin/pg_dump/compress_io.c src/bin/pg_dump/compress_gzip.c\n> \n> Produces ~25k diff with/without the patch. What am I doing wrong?\n\nDo you mean 25 kB of diff ? I agree that the statistics of the diff\noutput don't change a lot:\n\n 1 file changed, 201 insertions(+), 570 deletions(-)\n 1 file changed, 198 insertions(+), 548 deletions(-)\n\nBut try reading the diff while looking for the cause of a bug. It's the\ndifference between reading 50, two-line changes, and reading a hunk that\nreplaces 100 lines with a different 100 lines, with empty/unrelated\nlines randomly thrown in as context.\n\nWhen the diff is readable, the pg_fatal() also stands out.\n\n> > Change pg_fatal() to an assertion+comment;\n> \n> Yeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\n> could add such protections against \"impossible\" stuff to a zillion other\n> places and the confusion likely outweighs the benefits.\n> \n> > Update the commit message and fix a few typos;\n> \n> Thanks. I don't want to annoy you too much, but could you split the\n> patch into the \"empty-data\" fix and all the other changes (rearranging\n> functions etc.)? I'd rather not mix those in the same commit.\n\nI don't know if that makes sense? The \"empty-data\" fix creates a new\nfunction called DeflateCompressorInit(). My proposal was to add the new\nfunction in the same place in the file as it used to be.\n\nThe patch also moves the pg_fatal() that's being removed. I don't think\nit's going to look any cleaner to read a history involving the\npg_fatal() first being added, then moved, then removed. Anyway, I'll\nwait while the community continues discussion about the pg_fatal().\n\n-- \nJustin\n\n\n","msg_date":"Wed, 15 Mar 2023 19:20:59 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Tuesday, March 14th, 2023 at 4:32 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> \n> \n> On 3/14/23 16:18, gkokolatos@pm.me wrote:\n> \n> > ...> Would you mind me trying to come with a patch to address your points?\n> \n> \n> That'd be great, thanks. Please keep it split into smaller patches - two\n> might work, with one patch for \"cosmetic\" changes and the other tweaking\n> the API error-handling stuff.\n\nPlease find attached a set for it. I will admit that the splitting in the\nseries might not be ideal and what you requested. It is split on what\nseemed as a logical units. Please advice how a better split can look like.\n\n0001 is unifying types and return values on the API\n0002 is addressing the constant definitions\n0003 is your previous 0004 adding comments\n\nAs far as the error handling is concerned, you had said upthread:\n\n> I think the right approach is to handle all library errors and not just\n> let them through. So Gzip_write() needs to check the return value, and\n> either call pg_fatal() or translate it to an error defined by the API.\n\nWhile working on it, I thought it would be clearer and more consistent\nfor the pg_fatal() to be called by the caller of the individual functions.\nEach individual function can keep track of the specifics of the error\ninternally. Then the caller upon detecting that there was an error by\nchecking the return value, can call pg_fatal() with a uniform error\nmessage and then add the specifics by calling the get_error_func().\n\nThoughts?\n\nCheers,\n//Georgios\n\n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Thu, 16 Mar 2023 17:04:16 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 3/16/23 18:04, gkokolatos@pm.me wrote:\n> \n> ------- Original Message -------\n> On Tuesday, March 14th, 2023 at 4:32 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n>>\n>> On 3/14/23 16:18, gkokolatos@pm.me wrote:\n>>\n>>> ...> Would you mind me trying to come with a patch to address your points?\n>>\n>>\n>> That'd be great, thanks. Please keep it split into smaller patches - two\n>> might work, with one patch for \"cosmetic\" changes and the other tweaking\n>> the API error-handling stuff.\n> \n> Please find attached a set for it. I will admit that the splitting in the\n> series might not be ideal and what you requested. It is split on what\n> seemed as a logical units. Please advice how a better split can look like.\n> \n> 0001 is unifying types and return values on the API\n> 0002 is addressing the constant definitions\n> 0003 is your previous 0004 adding comments\n> \n\nThanks. I think the split seems reasonable - the goal was to not mix\ndifferent changes, and from that POV it works.\n\nI'm not sure I understand the Gzip_read/Gzip_write changes in 0001. I\nmean, gzread/gzwrite returns int, so how does renaming the size_t\nvariable solve the issue of negative values for errors? I mean, this\n\n- size_t ret;\n+ size_t gzret;\n\n- ret = gzread(gzfp, ptr, size);\n+ gzret = gzread(gzfp, ptr, size);\n\nmeans we still lost the information gzread() returned a negative value,\nno? We'll still probably trigger an error, but it's a bit weird.\n\n\nISTM all this kinda assumes we're processing chunks of memory small\nenough that we'll never actually overflow int - I did check what the\ncode in 15 does, and it seems use int and size_t quite arbitrarily.\n\nFor example cfread() seems quite sane:\n\n int\n cfread(void *ptr, int size, cfp *fp)\n {\n int ret;\n ...\n ret = gzread(fp->compressedfp, ptr, size);\n ...\n return ret;\n }\n\nbut then _PrintFileData() happily stashes it into a size_t, ignoring the\nsignedness. Surely, if\n\n static void\n _PrintFileData(ArchiveHandle *AH, char *filename)\n {\n size_t cnt;\n ...\n while ((cnt = cfread(buf, buflen, cfp)))\n {\n ahwrite(buf, 1, cnt, AH);\n }\n ...\n }\n\nUnless I'm missing something, if gzread() ever returns -1 or some other\nnegative error value, we'll cast it to size_t, while condition will\nevaluate to \"true\" and we'll happily chew on some random chunk of data.\n\nSo the confusion is (at least partially) a preexisting issue ...\n\nFor gzwrite() it seems to be fine, because that only returns 0 on error.\nOTOH it's defined to take 'int size' but then we happily pass size_t\nvalues to it.\n\nAs I wrote earlier, this apparently assumes we never need to deal with\nbuffers larger than int, and I don't think we have the ambition to relax\nthat (I'm not sure it's even needed / possible).\n\n\nI see the read/write functions are now defined as int, but we only ever\nreturn 0/1 from them, and then interpret that as bool. Why not to define\nit like that? I don't think we need to adhere to the custom that\neverything returns \"int\". This is an internal API. Or if we want to\nstick to int, I'd define meaningful \"nice\" constants for 0/1.\n\n\n\n0002 seems fine to me. I see you've ditched the idea of having two\nseparate buffers, and replaced them with DEFAULT_IO_BUFFER_SIZE. Fine\nwith me, although I wonder if this might have negative impact on\nperformance or something (but I doubt that).\n\n0003 seems fine too.\n\n\n> As far as the error handling is concerned, you had said upthread:\n> \n>> I think the right approach is to handle all library errors and not just\n>> let them through. So Gzip_write() needs to check the return value, and\n>> either call pg_fatal() or translate it to an error defined by the API.\n> \n> While working on it, I thought it would be clearer and more consistent\n> for the pg_fatal() to be called by the caller of the individual functions.\n> Each individual function can keep track of the specifics of the error\n> internally. Then the caller upon detecting that there was an error by\n> checking the return value, can call pg_fatal() with a uniform error\n> message and then add the specifics by calling the get_error_func().\n> \n\nI agree it's cleaner the way you did it.\n\nI was thinking that with each compression function handling error\ninternally, the callers would not need to do that. But I haven't\nrealized there's logic to detect ENOSPC and so on, and we'd need to\nduplicate that in every compression func.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 16 Mar 2023 22:20:10 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 3/16/23 01:20, Justin Pryzby wrote:\n> On Mon, Mar 13, 2023 at 10:47:12PM +0100, Tomas Vondra wrote:\n>>> Rearrange functions to their original order allowing a cleaner diff to the prior code;\n>>\n>> OK. I wasn't very enthusiastic about this initially, but after thinking\n>> about it a bit I think it's meaningful to make diffs clearer. But I\n>> don't see much difference with/without the patch. The\n>>\n>> git diff --diff-algorithm=minimal -w e9960732a~:src/bin/pg_dump/compress_io.c src/bin/pg_dump/compress_gzip.c\n>>\n>> Produces ~25k diff with/without the patch. What am I doing wrong?\n> \n> Do you mean 25 kB of diff ?\n\nYes, if you redirect the git-diff to a file, it's a 25kB file.\n\n> I agree that the statistics of the diff output don't change a lot:\n> \n> 1 file changed, 201 insertions(+), 570 deletions(-)\n> 1 file changed, 198 insertions(+), 548 deletions(-)\n> \n> But try reading the diff while looking for the cause of a bug. It's the\n> difference between reading 50, two-line changes, and reading a hunk that\n> replaces 100 lines with a different 100 lines, with empty/unrelated\n> lines randomly thrown in as context.\n> \n> When the diff is readable, the pg_fatal() also stands out.\n> \n\nI don't know, maybe I'm doing something wrong or maybe I just am bad at\nlooking at diffs, but if I apply the patch you submitted on 8/3 and do\nthe git-diff above (output attached), it seems pretty incomprehensible\nto me :-( I don't see 50 two-line changes (I certainly wouldn't be able\nto identify the root cause of the bug based on that).\n\n>>> Change pg_fatal() to an assertion+comment;\n>>\n>> Yeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\n>> could add such protections against \"impossible\" stuff to a zillion other\n>> places and the confusion likely outweighs the benefits.\n>>\n>>> Update the commit message and fix a few typos;\n>>\n>> Thanks. I don't want to annoy you too much, but could you split the\n>> patch into the \"empty-data\" fix and all the other changes (rearranging\n>> functions etc.)? I'd rather not mix those in the same commit.\n> \n> I don't know if that makes sense? The \"empty-data\" fix creates a new\n> function called DeflateCompressorInit(). My proposal was to add the new\n> function in the same place in the file as it used to be.\n> \n\nGot it. In that case I agree it's fine to do that in a single commit.\n\n> The patch also moves the pg_fatal() that's being removed. I don't think\n> it's going to look any cleaner to read a history involving the\n> pg_fatal() first being added, then moved, then removed. Anyway, I'll\n> wait while the community continues discussion about the pg_fatal().\n> \n\nI think the agreement was to replace the pg_fatal with and assert, and I\nsee your patch already does that.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Thu, 16 Mar 2023 23:30:50 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Mar 16, 2023 at 11:30:50PM +0100, Tomas Vondra wrote:\n> On 3/16/23 01:20, Justin Pryzby wrote:\n> > But try reading the diff while looking for the cause of a bug. It's the\n> > difference between reading 50, two-line changes, and reading a hunk that\n> > replaces 100 lines with a different 100 lines, with empty/unrelated\n> > lines randomly thrown in as context.\n>\n> I don't know, maybe I'm doing something wrong or maybe I just am bad at\n> looking at diffs, but if I apply the patch you submitted on 8/3 and do\n> the git-diff above (output attached), it seems pretty incomprehensible\n> to me :-( I don't see 50 two-line changes (I certainly wouldn't be able\n> to identify the root cause of the bug based on that).\n\nIt's true that most of the diff is still incomprehensible...\n\nBut look at the part relevant to the \"empty-data\" bug:\n\n[... incomprehensible changes elided ...]\n> static void\n> -InitCompressorZlib(CompressorState *cs, int level)\n> +DeflateCompressorInit(CompressorState *cs)\n> {\n> +\tGzipCompressorState *gzipcs;\n> \tz_streamp\tzp;\n> \n> -\tzp = cs->zp = (z_streamp) pg_malloc(sizeof(z_stream));\n> +\tgzipcs = (GzipCompressorState *) pg_malloc0(sizeof(GzipCompressorState));\n> +\tzp = gzipcs->zp = (z_streamp) pg_malloc(sizeof(z_stream));\n> \tzp->zalloc = Z_NULL;\n> \tzp->zfree = Z_NULL;\n> \tzp->opaque = Z_NULL;\n> \n> \t/*\n> -\t * zlibOutSize is the buffer size we tell zlib it can output to. We\n> -\t * actually allocate one extra byte because some routines want to append a\n> -\t * trailing zero byte to the zlib output.\n> +\t * outsize is the buffer size we tell zlib it can output to. We actually\n> +\t * allocate one extra byte because some routines want to append a trailing\n> +\t * zero byte to the zlib output.\n> \t */\n> -\tcs->zlibOut = (char *) pg_malloc(ZLIB_OUT_SIZE + 1);\n> -\tcs->zlibOutSize = ZLIB_OUT_SIZE;\n> +\tgzipcs->outbuf = pg_malloc(ZLIB_OUT_SIZE + 1);\n> +\tgzipcs->outsize = ZLIB_OUT_SIZE;\n> \n> -\tif (deflateInit(zp, level) != Z_OK)\n> -\t\tpg_fatal(\"could not initialize compression library: %s\",\n> -\t\t\t\t zp->msg);\n> +\t/* -Z 0 uses the \"None\" compressor -- not zlib with no compression */\n> +\tAssert(cs->compression_spec.level != 0);\n> +\n> +\tif (deflateInit(zp, cs->compression_spec.level) != Z_OK)\n> +\t\tpg_fatal(\"could not initialize compression library: %s\", zp->msg);\n> \n> \t/* Just be paranoid - maybe End is called after Start, with no Write */\n> -\tzp->next_out = (void *) cs->zlibOut;\n> -\tzp->avail_out = cs->zlibOutSize;\n> +\tzp->next_out = gzipcs->outbuf;\n> +\tzp->avail_out = gzipcs->outsize;\n> +\n> +\t/* Keep track of gzipcs */\n> +\tcs->private_data = gzipcs;\n> }\n> \n> static void\n> -EndCompressorZlib(ArchiveHandle *AH, CompressorState *cs)\n> +DeflateCompressorEnd(ArchiveHandle *AH, CompressorState *cs)\n> {\n> -\tz_streamp\tzp = cs->zp;\n> +\tGzipCompressorState *gzipcs = (GzipCompressorState *) cs->private_data;\n> +\tz_streamp\tzp;\n> \n> +\tzp = gzipcs->zp;\n> \tzp->next_in = NULL;\n> \tzp->avail_in = 0;\n> \n> \t/* Flush any remaining data from zlib buffer */\n> -\tDeflateCompressorZlib(AH, cs, true);\n> +\tDeflateCompressorCommon(AH, cs, true);\n> \n> \tif (deflateEnd(zp) != Z_OK)\n> \t\tpg_fatal(\"could not close compression stream: %s\", zp->msg);\n> \n> -\tfree(cs->zlibOut);\n> -\tfree(cs->zp);\n> +\tpg_free(gzipcs->outbuf);\n> +\tpg_free(gzipcs->zp);\n> +\tpg_free(gzipcs);\n> +\tcs->private_data = NULL;\n> }\n> \n> static void\n> -DeflateCompressorZlib(ArchiveHandle *AH, CompressorState *cs, bool flush)\n> +DeflateCompressorCommon(ArchiveHandle *AH, CompressorState *cs, bool flush)\n> {\n> -\tz_streamp\tzp = cs->zp;\n> -\tchar\t *out = cs->zlibOut;\n> +\tGzipCompressorState *gzipcs = (GzipCompressorState *) cs->private_data;\n> +\tz_streamp\tzp = gzipcs->zp;\n> +\tvoid\t *out = gzipcs->outbuf;\n> \tint\t\t\tres = Z_OK;\n> \n> -\twhile (cs->zp->avail_in != 0 || flush)\n> +\twhile (gzipcs->zp->avail_in != 0 || flush)\n> \t{\n> \t\tres = deflate(zp, flush ? Z_FINISH : Z_NO_FLUSH);\n> \t\tif (res == Z_STREAM_ERROR)\n> \t\t\tpg_fatal(\"could not compress data: %s\", zp->msg);\n> -\t\tif ((flush && (zp->avail_out < cs->zlibOutSize))\n> +\t\tif ((flush && (zp->avail_out < gzipcs->outsize))\n> \t\t\t|| (zp->avail_out == 0)\n> \t\t\t|| (zp->avail_in != 0)\n> \t\t\t)\n> @@ -289,18 +122,18 @@ DeflateCompressorZlib(ArchiveHandle *AH, CompressorState *cs, bool flush)\n> \t\t\t * chunk is the EOF marker in the custom format. This should never\n> \t\t\t * happen but ...\n> \t\t\t */\n> -\t\t\tif (zp->avail_out < cs->zlibOutSize)\n> +\t\t\tif (zp->avail_out < gzipcs->outsize)\n> \t\t\t{\n> \t\t\t\t/*\n> \t\t\t\t * Any write function should do its own error checking but to\n> \t\t\t\t * make sure we do a check here as well ...\n> \t\t\t\t */\n> -\t\t\t\tsize_t\t\tlen = cs->zlibOutSize - zp->avail_out;\n> +\t\t\t\tsize_t\t\tlen = gzipcs->outsize - zp->avail_out;\n> \n> -\t\t\t\tcs->writeF(AH, out, len);\n> +\t\t\t\tcs->writeF(AH, (char *) out, len);\n> \t\t\t}\n> -\t\t\tzp->next_out = (void *) out;\n> -\t\t\tzp->avail_out = cs->zlibOutSize;\n> +\t\t\tzp->next_out = out;\n> +\t\t\tzp->avail_out = gzipcs->outsize;\n> \t\t}\n> \n> \t\tif (res == Z_STREAM_END)\n> @@ -309,16 +142,26 @@ DeflateCompressorZlib(ArchiveHandle *AH, CompressorState *cs, bool flush)\n> }\n> \n> static void\n> -WriteDataToArchiveZlib(ArchiveHandle *AH, CompressorState *cs,\n> -\t\t\t\t\t const char *data, size_t dLen)\n> +EndCompressorGzip(ArchiveHandle *AH, CompressorState *cs)\n> {\n> -\tcs->zp->next_in = (void *) unconstify(char *, data);\n> -\tcs->zp->avail_in = dLen;\n> -\tDeflateCompressorZlib(AH, cs, false);\n> +\t/* If deflation was initialized, finalize it */\n> +\tif (cs->private_data)\n> +\t\tDeflateCompressorEnd(AH, cs);\n> }\n> \n> static void\n> -ReadDataFromArchiveZlib(ArchiveHandle *AH, ReadFunc readF)\n> +WriteDataToArchiveGzip(ArchiveHandle *AH, CompressorState *cs,\n> +\t\t\t\t\t const void *data, size_t dLen)\n> +{\n> +\tGzipCompressorState *gzipcs = (GzipCompressorState *) cs->private_data;\n> +\n> +\tgzipcs->zp->next_in = (void *) unconstify(void *, data);\n> +\tgzipcs->zp->avail_in = dLen;\n> +\tDeflateCompressorCommon(AH, cs, false);\n> +}\n> +\n> +static void\n> +ReadDataFromArchiveGzip(ArchiveHandle *AH, CompressorState *cs)\n> {\n> \tz_streamp\tzp;\n> \tchar\t *out;\n> @@ -342,7 +185,7 @@ ReadDataFromArchiveZlib(ArchiveHandle *AH, ReadFunc readF)\n> \t\t\t\t zp->msg);\n> \n> \t/* no minimal chunk size for zlib */\n> -\twhile ((cnt = readF(AH, &buf, &buflen)))\n> +\twhile ((cnt = cs->readF(AH, &buf, &buflen)))\n> \t{\n> \t\tzp->next_in = (void *) buf;\n> \t\tzp->avail_in = cnt;\n> @@ -382,389 +225,196 @@ ReadDataFromArchiveZlib(ArchiveHandle *AH, ReadFunc readF)\n> \tfree(out);\n> \tfree(zp);\n> }\n[... more incomprehensible changes elided ...]\n\n\n","msg_date":"Thu, 16 Mar 2023 17:58:42 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 3/16/23 23:58, Justin Pryzby wrote:\n> On Thu, Mar 16, 2023 at 11:30:50PM +0100, Tomas Vondra wrote:\n>> On 3/16/23 01:20, Justin Pryzby wrote:\n>>> But try reading the diff while looking for the cause of a bug. It's the\n>>> difference between reading 50, two-line changes, and reading a hunk that\n>>> replaces 100 lines with a different 100 lines, with empty/unrelated\n>>> lines randomly thrown in as context.\n>>\n>> I don't know, maybe I'm doing something wrong or maybe I just am bad at\n>> looking at diffs, but if I apply the patch you submitted on 8/3 and do\n>> the git-diff above (output attached), it seems pretty incomprehensible\n>> to me :-( I don't see 50 two-line changes (I certainly wouldn't be able\n>> to identify the root cause of the bug based on that).\n> \n> It's true that most of the diff is still incomprehensible...\n> \n> But look at the part relevant to the \"empty-data\" bug:\n> \n\nWell, yeah. If you know where to look, and if you squint just the right\nway, then you can see any bug. I don't think I'd be able to spot the bug\nin the diff unless I knew in advance what the bug is.\n\nThat being said, I don't object to moving the function etc. Unless there\nare alternative ideas how to fix the empty-data issue, I'll get this\ncommitted after playing with it a bit more.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 17 Mar 2023 03:02:22 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Thursday, March 16th, 2023 at 10:20 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> \n> \n> On 3/16/23 18:04, gkokolatos@pm.me wrote:\n> \n> > ------- Original Message -------\n> > On Tuesday, March 14th, 2023 at 4:32 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\n> > \n> > > On 3/14/23 16:18, gkokolatos@pm.me wrote:\n> > > \n> > > > ...> Would you mind me trying to come with a patch to address your points?\n> > > \n> > > That'd be great, thanks. Please keep it split into smaller patches - two\n> > > might work, with one patch for \"cosmetic\" changes and the other tweaking\n> > > the API error-handling stuff.\n> > \n> > Please find attached a set for it. I will admit that the splitting in the\n> > series might not be ideal and what you requested. It is split on what\n> > seemed as a logical units. Please advice how a better split can look like.\n> > \n> > 0001 is unifying types and return values on the API\n> > 0002 is addressing the constant definitions\n> > 0003 is your previous 0004 adding comments\n> \n> \n> Thanks. I think the split seems reasonable - the goal was to not mix\n> different changes, and from that POV it works.\n> \n> I'm not sure I understand the Gzip_read/Gzip_write changes in 0001. I\n> mean, gzread/gzwrite returns int, so how does renaming the size_t\n> variable solve the issue of negative values for errors? I mean, this\n> \n> - size_t ret;\n> + size_t gzret;\n> \n> - ret = gzread(gzfp, ptr, size);\n> + gzret = gzread(gzfp, ptr, size);\n> \n> means we still lost the information gzread() returned a negative value,\n> no? We'll still probably trigger an error, but it's a bit weird.\n\nYou are obviously correct. My bad, I miss-read the return type of gzread().\n\nPlease find an amended version attached.\n\n> Unless I'm missing something, if gzread() ever returns -1 or some other\n> negative error value, we'll cast it to size_t, while condition will\n> evaluate to \"true\" and we'll happily chew on some random chunk of data.\n> \n> So the confusion is (at least partially) a preexisting issue ...\n> \n> For gzwrite() it seems to be fine, because that only returns 0 on error.\n> OTOH it's defined to take 'int size' but then we happily pass size_t\n> values to it.\n> \n> As I wrote earlier, this apparently assumes we never need to deal with\n> buffers larger than int, and I don't think we have the ambition to relax\n> that (I'm not sure it's even needed / possible).\n\nAgreed.\n\n\n> I see the read/write functions are now defined as int, but we only ever\n> return 0/1 from them, and then interpret that as bool. Why not to define\n> it like that? I don't think we need to adhere to the custom that\n> everything returns \"int\". This is an internal API. Or if we want to\n> stick to int, I'd define meaningful \"nice\" constants for 0/1.\n\nThe return types are now booleans and the callers have been made aware.\n\n\n> 0002 seems fine to me. I see you've ditched the idea of having two\n> separate buffers, and replaced them with DEFAULT_IO_BUFFER_SIZE. Fine\n> with me, although I wonder if this might have negative impact on\n> performance or something (but I doubt that).\n> \n\nI doubt that too. Thank you.\n\n> 0003 seems fine too.\n\nThank you.\n\n\n> > As far as the error handling is concerned, you had said upthread:\n> > \n> > > I think the right approach is to handle all library errors and not just\n> > > let them through. So Gzip_write() needs to check the return value, and\n> > > either call pg_fatal() or translate it to an error defined by the API.\n> > \n> > While working on it, I thought it would be clearer and more consistent\n> > for the pg_fatal() to be called by the caller of the individual functions.\n> > Each individual function can keep track of the specifics of the error\n> > internally. Then the caller upon detecting that there was an error by\n> > checking the return value, can call pg_fatal() with a uniform error\n> > message and then add the specifics by calling the get_error_func().\n> \n> \n> I agree it's cleaner the way you did it.\n> \n> I was thinking that with each compression function handling error\n> internally, the callers would not need to do that. But I haven't\n> realized there's logic to detect ENOSPC and so on, and we'd need to\n> duplicate that in every compression func.\n> \n\nIf you agree, I can prepare a patch to improve on the error handling\naspect of the API as a separate thread, since here we are trying to\nfocus on correctness.\n\nCheers,\n//Georgios\n\n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Fri, 17 Mar 2023 15:43:58 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 3/17/23 16:43, gkokolatos@pm.me wrote:\n>>\n>> ...\n>>\n>> I agree it's cleaner the way you did it.\n>>\n>> I was thinking that with each compression function handling error\n>> internally, the callers would not need to do that. But I haven't\n>> realized there's logic to detect ENOSPC and so on, and we'd need to\n>> duplicate that in every compression func.\n>>\n> \n> If you agree, I can prepare a patch to improve on the error handling\n> aspect of the API as a separate thread, since here we are trying to\n> focus on correctness.\n> \n\nYes, that makes sense. There are far too many patches in this thread\nalready ...\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Fri, 17 Mar 2023 17:56:58 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Hi,\n\nI was preparing to get the 3 cleanup patches pushed, so I\nupdated/reworded the commit messages a bit (attached, please check).\n\nBut I noticed the commit message for 0001 said:\n\n In passing save the appropriate errno in LZ4File_open_write in\n case that the caller is not using the API's get_error_func.\n\nI think that's far too low-level for a commit message, it'd be much more\nappropriate for a comment at the function.\n\nHowever, do we even need this behavior? I was looking for code calling\nthis function without using get_error_func(), but no luck. And if there\nis such caller, shouldn't we fix it to use get_error_func()?\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 20 Mar 2023 21:48:58 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, Mar 17, 2023 at 03:43:58PM +0000, gkokolatos@pm.me wrote:\n> From a174cdff4ec8aad59f5bcc7e8d52218a14fe56fc Mon Sep 17 00:00:00 2001\n> From: Georgios Kokolatos <gkokolatos@pm.me>\n> Date: Fri, 17 Mar 2023 14:45:58 +0000\n> Subject: [PATCH v3 1/3] Improve type handling in pg_dump's compress file API\n\n> -int\n> +bool\n> EndCompressFileHandle(CompressFileHandle *CFH)\n> {\n> -\tint\t\t\tret = 0;\n> +\tbool\t\tret = 0;\n\nShould say \"= false\" ?\n\n> \t/*\n> \t * Write 'size' bytes of data into the file from 'ptr'.\n> +\t *\n> +\t * Returns true on success and false on error.\n> +\t */\n> + bool (*write_func) (const void *ptr, size_t size,\n\n> - * Get a pointer to a string that describes an error that occurred during a\n> - * compress file handle operation.\n> + * Get a pointer to a string that describes an error that occurred during\n> + * a compress file handle operation.\n> */\n> const char *(*get_error_func) (CompressFileHandle *CFH);\n\nThis should mention that the error accessible in error_func() applies (only) to\nwrite_func() ?\n\nAs long as this touches pg_backup_directory.c you could update the\nheader comment to refer to \"compressed extensions\", not just .gz.\n\nI noticed that EndCompressorLZ4() tests \"if (LZ4cs)\", but that should\nalways be true.\n\nI was able to convert the zstd patch to this new API with no issue.\n\n-- \nJustin\n\n\n","msg_date":"Mon, 20 Mar 2023 17:40:32 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Hi,\n\nI looked at this again, and I realized I misunderstood the bit about\nerrno in LZ4File_open_write a bit. I now see it simply just brings the\nfunction in line with Gzip_open_write(), so that the callers can just do\npg_fatal(\"%m\"). I still think the special \"errno\" handling in this one\nplace feels a bit random, and handling it by get_error_func() would be\nnicer, but we can leave that for a separate patch - no need to block\nthese changes because of that.\n\nSo pushed all three parts, after updating the commit messages a bit.\n\nThis leaves the empty-data issue (which we have a fix for) and the\nswitch to LZ4F. And then the zstd part.\n\n\nOn 3/20/23 23:40, Justin Pryzby wrote:\n> On Fri, Mar 17, 2023 at 03:43:58PM +0000, gkokolatos@pm.me wrote:\n>> From a174cdff4ec8aad59f5bcc7e8d52218a14fe56fc Mon Sep 17 00:00:00 2001\n>> From: Georgios Kokolatos <gkokolatos@pm.me>\n>> Date: Fri, 17 Mar 2023 14:45:58 +0000\n>> Subject: [PATCH v3 1/3] Improve type handling in pg_dump's compress file API\n> \n>> -int\n>> +bool\n>> EndCompressFileHandle(CompressFileHandle *CFH)\n>> {\n>> -\tint\t\t\tret = 0;\n>> +\tbool\t\tret = 0;\n> \n> Should say \"= false\" ?\n> \n\nRight, fixed.\n\n>> \t/*\n>> \t * Write 'size' bytes of data into the file from 'ptr'.\n>> +\t *\n>> +\t * Returns true on success and false on error.\n>> +\t */\n>> + bool (*write_func) (const void *ptr, size_t size,\n> \n>> - * Get a pointer to a string that describes an error that occurred during a\n>> - * compress file handle operation.\n>> + * Get a pointer to a string that describes an error that occurred during\n>> + * a compress file handle operation.\n>> */\n>> const char *(*get_error_func) (CompressFileHandle *CFH);\n> \n> This should mention that the error accessible in error_func() applies (only) to\n> write_func() ?\n> \n> As long as this touches pg_backup_directory.c you could update the\n> header comment to refer to \"compressed extensions\", not just .gz.\n> \n> I noticed that EndCompressorLZ4() tests \"if (LZ4cs)\", but that should\n> always be true.\n> \n\nI haven't done these two things. We can/should do that, but it didn't\nfit into the three patches.\n\n> I was able to convert the zstd patch to this new API with no issue.\n> \n\nGood to hear.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Thu, 23 Mar 2023 18:10:27 +0100","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n\n> \n> So pushed all three parts, after updating the commit messages a bit.\n\nThank you very much.\n\n> \n> This leaves the empty-data issue (which we have a fix for) and the\n> switch to LZ4F. And then the zstd part.\n\nPlease expect promptly a patch for the switch to frames.\n\nCheers,\n//Georgios\n\n\n\n","msg_date":"Fri, 24 Mar 2023 09:30:28 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Thursday, March 16th, 2023 at 11:30 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> \n> \n> On 3/16/23 01:20, Justin Pryzby wrote:\n> \n> > On Mon, Mar 13, 2023 at 10:47:12PM +0100, Tomas Vondra wrote:\n> > \n> > > \n> > > Thanks. I don't want to annoy you too much, but could you split the\n> > > patch into the \"empty-data\" fix and all the other changes (rearranging\n> > > functions etc.)? I'd rather not mix those in the same commit.\n> > \n> > I don't know if that makes sense? The \"empty-data\" fix creates a new\n> > function called DeflateCompressorInit(). My proposal was to add the new\n> > function in the same place in the file as it used to be.\n> \n> \n> Got it. In that case I agree it's fine to do that in a single commit.\n\nFor what is worth, I think that this patch should get a +1 and get in. It\nsolves the empty writes problem and includes a test to a previous untested\ncase.\n\nCheers,\n//Georgios\n\n> \n> > The patch also moves the pg_fatal() that's being removed. I don't think\n> > it's going to look any cleaner to read a history involving the\n> > pg_fatal() first being added, then moved, then removed. Anyway, I'll\n> > wait while the community continues discussion about the pg_fatal().\n> \n> \n> I think the agreement was to replace the pg_fatal with and assert, and I\n> see your patch already does that.\n> \n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company\n\n\n","msg_date":"Mon, 27 Mar 2023 22:34:52 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me <gkokolatos@pm.me> wrote:\n\n> \n> ------- Original Message -------\n> On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\n> \n> > This leaves the empty-data issue (which we have a fix for) and the\n> > switch to LZ4F. And then the zstd part.\n> \n> Please expect promptly a patch for the switch to frames.\n\nPlease find the expected patch attached. Note that the bulk of the\npatch is code unification, variable renaming to something more\nappropriate, and comment addition. These are changes that are not\nstrictly necessary to switch to LZ4F. I do believe that are\nessential for code hygiene after the switch and they do belong\non the same commit. \n\nCheers,\n//Georgios\n\n> \n> Cheers,\n> //Georgios","msg_date":"Tue, 28 Mar 2023 16:07:30 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 3/28/23 18:07, gkokolatos@pm.me wrote:\n> \n> \n> \n> \n> \n> ------- Original Message -------\n> On Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me <gkokolatos@pm.me> wrote:\n> \n>>\n>> ------- Original Message -------\n>> On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\n>>\n>>> This leaves the empty-data issue (which we have a fix for) and the\n>>> switch to LZ4F. And then the zstd part.\n>>\n>> Please expect promptly a patch for the switch to frames.\n> \n> Please find the expected patch attached. Note that the bulk of the\n> patch is code unification, variable renaming to something more\n> appropriate, and comment addition. These are changes that are not\n> strictly necessary to switch to LZ4F. I do believe that are\n> essential for code hygiene after the switch and they do belong\n> on the same commit. \n> \n\nThanks!\n\nI agree the renames & cleanup are appropriate - it'd be silly to stick\nto misleading naming etc. Would it make sense to split the patch into\ntwo, to separate the renames and the switch to lz4f?\n\nThat'd make it the changes necessary for lz4f switch clearer.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Tue, 28 Mar 2023 18:40:03 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Mar 28, 2023 at 06:40:03PM +0200, Tomas Vondra wrote:\n> On 3/28/23 18:07, gkokolatos@pm.me wrote:\n> > ------- Original Message -------\n> > On Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me <gkokolatos@pm.me> wrote:\n> > \n> >> ------- Original Message -------\n> >> On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\n> >>\n> >>> This leaves the empty-data issue (which we have a fix for) and the\n> >>> switch to LZ4F. And then the zstd part.\n> >>\n> >> Please expect promptly a patch for the switch to frames.\n> > \n> > Please find the expected patch attached. Note that the bulk of the\n> > patch is code unification, variable renaming to something more\n> > appropriate, and comment addition. These are changes that are not\n> > strictly necessary to switch to LZ4F. I do believe that are\n> > essential for code hygiene after the switch and they do belong\n> > on the same commit. \n> \n> Thanks!\n> \n> I agree the renames & cleanup are appropriate - it'd be silly to stick\n> to misleading naming etc. Would it make sense to split the patch into\n> two, to separate the renames and the switch to lz4f?\n> That'd make it the changes necessary for lz4f switch clearer.\n\nI don't think so. Did you mean separate commits only for review ?\n\nThe patch is pretty readable - the File API has just some renames, and\nthe compressor API is what's being replaced, which isn't going to be any\nmore clear.\n\n@Georgeos: did you consider using a C union in LZ4State, to separate the\nparts used by the different APIs ?\n\n-- \nJustin\n\n\n","msg_date":"Tue, 28 Mar 2023 13:47:21 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 3/28/23 18:07, gkokolatos@pm.me wrote:\n> \n> ------- Original Message -------\n> On Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me <gkokolatos@pm.me> wrote:\n> \n>>\n>> ------- Original Message -------\n>> On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\n>>\n>>> This leaves the empty-data issue (which we have a fix for) and the\n>>> switch to LZ4F. And then the zstd part.\n>>\n>> Please expect promptly a patch for the switch to frames.\n> \n> Please find the expected patch attached. Note that the bulk of the\n> patch is code unification, variable renaming to something more\n> appropriate, and comment addition. These are changes that are not\n> strictly necessary to switch to LZ4F. I do believe that are\n> essential for code hygiene after the switch and they do belong\n> on the same commit. \n> \n\n\nI think the patch is fine, but I'm wondering if the renames shouldn't go\na bit further. It removes references to LZ4File struct, but there's a\nbunch of functions with LZ4File_ prefix. Why not to simply use LZ4_\nprefix? We don't have GzipFile either.\n\nSure, it might be a bit confusing because lz4.h uses LZ4_ prefix, but\nthen we probably should not define LZ4_compressor_init ...\n\nAlso, maybe the comments shouldn't use \"File API\" when compress_io.c\ncalls that \"Compressed stream API\".\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 29 Mar 2023 00:02:00 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 3/28/23 00:34, gkokolatos@pm.me wrote:\n> \n> ...\n>\n>> Got it. In that case I agree it's fine to do that in a single commit.\n> \n> For what is worth, I think that this patch should get a +1 and get in. It\n> solves the empty writes problem and includes a test to a previous untested\n> case.\n> \n\nPushed, after updating / rewording the commit message a little bit.\n\nThanks!\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 29 Mar 2023 02:40:15 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Wednesday, March 29th, 2023 at 12:02 AM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> On 3/28/23 18:07, gkokolatos@pm.me wrote:\n> \n> > ------- Original Message -------\n> > On Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me gkokolatos@pm.me wrote:\n> > \n> > > ------- Original Message -------\n> > > On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\n> > > \n> > > > This leaves the empty-data issue (which we have a fix for) and the\n> > > > switch to LZ4F. And then the zstd part.\n> > > \n> > > Please expect promptly a patch for the switch to frames.\n> > \n> > Please find the expected patch attached. Note that the bulk of the\n> > patch is code unification, variable renaming to something more\n> > appropriate, and comment addition. These are changes that are not\n> > strictly necessary to switch to LZ4F. I do believe that are\n> > essential for code hygiene after the switch and they do belong\n> > on the same commit.\n> \n> \n> I think the patch is fine, but I'm wondering if the renames shouldn't go\n> a bit further. It removes references to LZ4File struct, but there's a\n> bunch of functions with LZ4File_ prefix. Why not to simply use LZ4_\n> prefix? We don't have GzipFile either.\n> \n> Sure, it might be a bit confusing because lz4.h uses LZ4_ prefix, but\n> then we probably should not define LZ4_compressor_init ...\n\nThis is a good point. The initial thought was that since lz4.h is now\nremoved, such ambiguity will not be present. In v2 of the patch the\nfunction is renamed to `LZ4State_compression_init` since this name\ndescribes better its purpose. It initializes the LZ4State for\ncompression.\n\nAs for the LZ4File_ prefix, I have no objections. Please find the\nprefix changed to LZ4Stream_. For the record, the word 'File' is not\nunique to the lz4 implementation. The common data structure used by\nthe API in compress_io.h:\n\n typedef struct CompressFileHandle CompressFileHandle; \n\nThe public functions for this API are named:\n\n InitCompressFileHandle\n InitDiscoverCompressFileHandle\n EndCompressFileHandle\n\nAnd within InitCompressFileHandle the pattern is:\n\n if (compression_spec.algorithm == PG_COMPRESSION_NONE)\n InitCompressFileHandleNone(CFH, compression_spec);\n else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\n InitCompressFileHandleGzip(CFH, compression_spec);\n else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\n InitCompressFileHandleLZ4(CFH, compression_spec);\n\nIt was felt that a prefix was required due to the inclusion 'lz4.h'\nheader where naming functions as 'LZ4_' would be wrong. The prefix\n'LZ4File_' seemed to be in line with the naming of the rest of\nthe relevant functions and structures. Other compressions, gzip and\nnone, did not face the same issue.\n\nTo conclude, I think that having a prefix is slightly preferred\nover not having one. Since the prefix `LZ4File_` is not desired,\nI propose `LZ4Stream_` in v2.\n\nI will not object to dismissing the argument and drop `File` from\nthe prefix, if so requested.\n\n> \n> Also, maybe the comments shouldn't use \"File API\" when compress_io.c\n> calls that \"Compressed stream API\".\n\nDone.\n\nCheers,\n//Georgios\n\n> \n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Fri, 31 Mar 2023 09:19:47 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 3/31/23 11:19, gkokolatos@pm.me wrote:\n> \n>> ...\n>>\n>>\n>> I think the patch is fine, but I'm wondering if the renames shouldn't go\n>> a bit further. It removes references to LZ4File struct, but there's a\n>> bunch of functions with LZ4File_ prefix. Why not to simply use LZ4_\n>> prefix? We don't have GzipFile either.\n>>\n>> Sure, it might be a bit confusing because lz4.h uses LZ4_ prefix, but\n>> then we probably should not define LZ4_compressor_init ...\n> \n> This is a good point. The initial thought was that since lz4.h is now\n> removed, such ambiguity will not be present. In v2 of the patch the\n> function is renamed to `LZ4State_compression_init` since this name\n> describes better its purpose. It initializes the LZ4State for\n> compression.\n> \n> As for the LZ4File_ prefix, I have no objections. Please find the\n> prefix changed to LZ4Stream_. For the record, the word 'File' is not\n> unique to the lz4 implementation. The common data structure used by\n> the API in compress_io.h:\n> \n> typedef struct CompressFileHandle CompressFileHandle; \n> \n> The public functions for this API are named:\n> \n> InitCompressFileHandle\n> InitDiscoverCompressFileHandle\n> EndCompressFileHandle\n> \n> And within InitCompressFileHandle the pattern is:\n> \n> if (compression_spec.algorithm == PG_COMPRESSION_NONE)\n> InitCompressFileHandleNone(CFH, compression_spec);\n> else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\n> InitCompressFileHandleGzip(CFH, compression_spec);\n> else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\n> InitCompressFileHandleLZ4(CFH, compression_spec);\n> \n> It was felt that a prefix was required due to the inclusion 'lz4.h'\n> header where naming functions as 'LZ4_' would be wrong. The prefix\n> 'LZ4File_' seemed to be in line with the naming of the rest of\n> the relevant functions and structures. Other compressions, gzip and\n> none, did not face the same issue.\n> \n> To conclude, I think that having a prefix is slightly preferred\n> over not having one. Since the prefix `LZ4File_` is not desired,\n> I propose `LZ4Stream_` in v2.\n> \n> I will not object to dismissing the argument and drop `File` from\n> the prefix, if so requested.\n> \n\nThanks.\n\nI think the LZ4Stream prefix is reasonable, so let's roll with that. I\ncleaned up the patch a little bit (mostly comment tweaks, etc.), updated\nthe commit message and pushed it.\n\nThe main tweak I did is renaming all the LZ4State variables from \"fs\" to\nstate. The old name referred to the now abandoned \"file state\", but\nafter the rename to LZ4State that seems confusing. Some of the places\nalready used \"state\"and it's easier to know \"state\" is always LZ4State,\nso let's keep it consistent.\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Sat, 1 Apr 2023 01:34:16 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, Feb 27, 2023 at 02:33:04PM +0000, gkokolatos@pm.me wrote:\n> > > - Finally, the \"Nothing to do in the default case\" comment comes from\n> > > Michael's commit 5e73a6048:\n> > > \n> > > + /*\n> > > + * Custom and directory formats are compressed by default with gzip when\n> > > + * available, not the others.\n> > > + /\n> > > + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\n> > > + !user_compression_defined)\n> > > {\n> > > #ifdef HAVE_LIBZ\n> > > - if (archiveFormat == archCustom || archiveFormat == archDirectory)\n> > > - compressLevel = Z_DEFAULT_COMPRESSION;\n> > > - else\n> > > + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\n> > > + &compression_spec);\n> > > +#else\n> > > + / Nothing to do in the default case */\n> > > #endif\n> > > - compressLevel = 0;\n> > > }\n> > > \n> > > As the comment says: for -Fc and -Fd, the compression is set to zlib, if\n> > > enabled, and when not otherwise specified by the user.\n> > > \n> > > Before 5e73a6048, this set compressLevel=0 for -Fp and -Ft, and when\n> > > zlib was unavailable.\n> > > \n> > > But I'm not sure why there's now an empty \"#else\". I also don't know\n> > > what \"the default case\" refers to.\n> > > \n> > > Maybe the best thing here is to move the preprocessor #if, since it's no\n> > > longer in the middle of a runtime conditional:\n> > > \n> > > #ifdef HAVE_LIBZ\n> > > + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\n> > > + !user_compression_defined)\n> > > + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\n> > > + &compression_spec);\n> > > #endif\n> > > \n> > > ...but that elicits a warning about \"variable set but not used\"...\n> > \n> > \n> > Not sure, I need to think about this a bit.\n\n> /* Nothing to do for the default case when LIBZ is not available */\n> is easier to understand.\n\nMaybe I would write it as: \"if zlib is unavailable, default to no\ncompression\". But I think that's best done in the leading comment, and\nnot inside an empty preprocessor #else.\n\nI was hoping Michael would comment on this.\nThe placement and phrasing of the comment makes no sense to me.\n\n-- \nJustin\n\n\n","msg_date":"Tue, 11 Apr 2023 19:41:11 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Apr 11, 2023 at 07:41:11PM -0500, Justin Pryzby wrote:\n> Maybe I would write it as: \"if zlib is unavailable, default to no\n> compression\". But I think that's best done in the leading comment, and\n> not inside an empty preprocessor #else.\n> \n> I was hoping Michael would comment on this.\n\n(Sorry for the late reply, somewhat missed that.)\n\n> The placement and phrasing of the comment makes no sense to me.\n\nYes, this comment gives no value as it stands. I would be tempted to\nfollow the suggestion to group the whole code block in a single ifdef,\nincluding the check, and remove this comment. Like the attached\nperhaps?\n--\nMichael","msg_date":"Wed, 12 Apr 2023 10:07:08 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Apr 12, 2023 at 10:07:08AM +0900, Michael Paquier wrote:\n> On Tue, Apr 11, 2023 at 07:41:11PM -0500, Justin Pryzby wrote:\n> > Maybe I would write it as: \"if zlib is unavailable, default to no\n> > compression\". But I think that's best done in the leading comment, and\n> > not inside an empty preprocessor #else.\n> > \n> > I was hoping Michael would comment on this.\n> \n> (Sorry for the late reply, somewhat missed that.)\n> \n> > The placement and phrasing of the comment makes no sense to me.\n> \n> Yes, this comment gives no value as it stands. I would be tempted to\n> follow the suggestion to group the whole code block in a single ifdef,\n> including the check, and remove this comment. Like the attached\n> perhaps?\n\n+1\n\n\n","msg_date":"Tue, 11 Apr 2023 20:19:59 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, Apr 11, 2023 at 08:19:59PM -0500, Justin Pryzby wrote:\n> On Wed, Apr 12, 2023 at 10:07:08AM +0900, Michael Paquier wrote:\n>> Yes, this comment gives no value as it stands. I would be tempted to\n>> follow the suggestion to group the whole code block in a single ifdef,\n>> including the check, and remove this comment. Like the attached\n>> perhaps?\n> \n> +1\n\nLet me try this one again, as the previous patch would cause a warning\nunder --without:-zlib as user_compression_defined would be unused. We\ncould do something like the attached instead. It means doing twice\nparse_compress_specification() for the non-zlib path, still we are\nalready doing so for the zlib path.\n\nIf there are other ideas, feel free.\n--\nMichael","msg_date":"Thu, 13 Apr 2023 07:23:48 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Apr 13, 2023 at 07:23:48AM +0900, Michael Paquier wrote:\n> On Tue, Apr 11, 2023 at 08:19:59PM -0500, Justin Pryzby wrote:\n> > On Wed, Apr 12, 2023 at 10:07:08AM +0900, Michael Paquier wrote:\n> >> Yes, this comment gives no value as it stands. I would be tempted to\n> >> follow the suggestion to group the whole code block in a single ifdef,\n> >> including the check, and remove this comment. Like the attached\n> >> perhaps?\n> > \n> > +1\n> \n> Let me try this one again, as the previous patch would cause a warning\n> under --without:-zlib as user_compression_defined would be unused. We\n> could do something like the attached instead. It means doing twice\n> parse_compress_specification() for the non-zlib path, still we are\n> already doing so for the zlib path.\n> \n> If there are other ideas, feel free.\n\nI don't think you need to call parse_compress_specification(NONE).\nAs you wrote it, if zlib is unavailable, there's no parse(NONE) call,\neven for directory and custom formats. And there's no parse(NONE) call\nfor plan format when zlib is available.\n\nThe old way had preprocessor #if around both the \"if\" and \"else\" - is\nthat what you meant ?\n\nIf you don't insist on calling parse(NONE), the only change is to remove\nthe empty #else, which was my original patch.\n\n\"if no compression specification has been specified\" is redundant with\n\"by default\", and causes \"not the others\" to dangle.\n\nIf I were to rewrite the comment, it'd say:\n\n+ * When gzip is available, custom and directory formats are compressed by\n+ * default\n\n\n","msg_date":"Wed, 12 Apr 2023 17:52:40 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Apr 12, 2023 at 05:52:40PM -0500, Justin Pryzby wrote:\n> I don't think you need to call parse_compress_specification(NONE).\n> As you wrote it, if zlib is unavailable, there's no parse(NONE) call,\n> even for directory and custom formats. And there's no parse(NONE) call\n> for plan format when zlib is available.\n\nYeah, that's not necessary, but I was wondering if it made the code a\nbit cleaner, or else the non-zlib path would rely on the default\ncompression method string.\n\n> The old way had preprocessor #if around both the \"if\" and \"else\" - is\n> that what you meant?\n> \n> If you don't insist on calling parse(NONE), the only change is to remove\n> the empty #else, which was my original patch.\n\nRemoving the empty else has as problem to create an empty if block,\nwhich could be itself a cause of warnings?\n\n> If I were to rewrite the comment, it'd say:\n> \n> + * When gzip is available, custom and directory formats are compressed by\n> + * default\n\nOkay.\n--\nMichael","msg_date":"Thu, 13 Apr 2023 09:37:06 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Thu, Apr 13, 2023 at 09:37:06AM +0900, Michael Paquier wrote:\n> > If you don't insist on calling parse(NONE), the only change is to remove\n> > the empty #else, which was my original patch.\n> \n> Removing the empty else has as problem to create an empty if block,\n> which could be itself a cause of warnings?\n\nI doubt it - in the !HAVE_LIBZ case, it's currently an \"if\" statement\nwith nothing but a comment, which isn't a problem.\n\nI think the only issue with an empty \"if\" is when you have no braces,\nlike:\n\n\tif (...)\n#if ...\n\t\tsomething;\n#endif\n\n\t// problem here //\n\n-- \nJustin\n\n\n","msg_date":"Wed, 12 Apr 2023 19:53:53 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Apr 12, 2023 at 07:53:53PM -0500, Justin Pryzby wrote:\n> I doubt it - in the !HAVE_LIBZ case, it's currently an \"if\" statement\n> with nothing but a comment, which isn't a problem.\n> \n> I think the only issue with an empty \"if\" is when you have no braces,\n> like:\n> \n> \tif (...)\n> #if ...\n> \t\tsomething;\n> #endif\n> \n> \t// problem here //\n\n(My apologies for the late reply.)\n\nStill it could be easily messed up, and that's not a style that\nreally exists in the tree, either, because there are always #else\nblocks set up in such cases. Another part that makes me a bit\nuncomfortable is that we would still call twice\nparse_compress_specification(), something that should not happen but\nwe are doing so on HEAD because the default compression_algorithm_str\nis \"none\" and we want to enforce \"gzip\" for custom and directory\nformats when building with zlib.\n\nWhat about just moving this block a bit up, just before the\ncompression spec parsing, then? If we set compression_algorithm_str,\nthe specification is compiled with the expected default, once instead\nof twice.\n--\nMichael","msg_date":"Tue, 25 Apr 2023 15:02:18 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\n\n\n\n------- Original Message -------\nOn Tuesday, April 25th, 2023 at 8:02 AM, Michael Paquier <michael@paquier.xyz> wrote:\n\n\n> \n> \n> On Wed, Apr 12, 2023 at 07:53:53PM -0500, Justin Pryzby wrote:\n> \n> > I doubt it - in the !HAVE_LIBZ case, it's currently an \"if\" statement\n> > with nothing but a comment, which isn't a problem.\n> > \n> > I think the only issue with an empty \"if\" is when you have no braces,\n> > like:\n> > \n> > if (...)\n> > #if ...\n> > something;\n> > #endif\n> > \n> > // problem here //\n> \n> \n> (My apologies for the late reply.)\n> \n> Still it could be easily messed up, and that's not a style that\n> really exists in the tree, either, because there are always #else\n> blocks set up in such cases. Another part that makes me a bit\n> uncomfortable is that we would still call twice\n> parse_compress_specification(), something that should not happen but\n> we are doing so on HEAD because the default compression_algorithm_str\n> is \"none\" and we want to enforce \"gzip\" for custom and directory\n> formats when building with zlib.\n> \n> What about just moving this block a bit up, just before the\n> compression spec parsing, then? If we set compression_algorithm_str,\n> the specification is compiled with the expected default, once instead\n> of twice.\n\nFor what is worth, I think this would be the best approach. +1\n\nCheers,\n//Georgios\n\n> --\n> Michael\n\n\n","msg_date":"Wed, 26 Apr 2023 08:50:46 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Wed, Apr 26, 2023 at 08:50:46AM +0000, gkokolatos@pm.me wrote:\n> For what is worth, I think this would be the best approach. +1\n\nThanks. I have gone with that, then!\n--\nMichael","msg_date":"Thu, 27 Apr 2023 13:38:48 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"23.03.2023 20:10, Tomas Vondra wrote:\n> So pushed all three parts, after updating the commit messages a bit.\n>\n> This leaves the empty-data issue (which we have a fix for) and the\n> switch to LZ4F. And then the zstd part.\n>\n\nI'm sorry that I haven't noticed/checked that before, but when trying to\nperform check-world with Valgrind I've discovered another issue presumably\nrelated to LZ4File_gets().\nWhen running under Valgrind:\nPROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\nI get:\n...\n[07:07:11.683](0.000s) ok 1939 - compression_lz4_dir: glob check for \n.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\n# Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql \n.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\n\n==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\n==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\n==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\n==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\n==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\n==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\n==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\n==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\n==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\n==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\n==00:00:00:00.579 2811926==\n...\n\nIt looks like the line variable returned by gets_func() here is not\nnull-terminated:\n while ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\n {\n...\n if (sscanf(line, \"%u %\" CppAsString2(MAXPGPATH) \"s\\n\", &oid, lofname) != 2)\n...\nAnd Valgrind doesn't like it.\n\nBest regards,\nAlexander\n\n\n","msg_date":"Fri, 5 May 2023 09:00:01 +0300","msg_from":"Alexander Lakhin <exclusion@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Friday, May 5th, 2023 at 8:00 AM, Alexander Lakhin <exclusion@gmail.com> wrote:\n\n\n> \n> \n> 23.03.2023 20:10, Tomas Vondra wrote:\n> \n> > So pushed all three parts, after updating the commit messages a bit.\n> > \n> > This leaves the empty-data issue (which we have a fix for) and the\n> > switch to LZ4F. And then the zstd part.\n> \n> \n> I'm sorry that I haven't noticed/checked that before, but when trying to\n> perform check-world with Valgrind I've discovered another issue presumably\n> related to LZ4File_gets().\n> When running under Valgrind:\n> PROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\n> I get:\n> ...\n> 07:07:11.683 ok 1939 - compression_lz4_dir: glob check for\n> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\n> # Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql\n> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\n> \n> ==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\n> ==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\n> ==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\n> ==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\n> ==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\n> ==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\n> ==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\n> ==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\n> ==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\n> ==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\n> ==00:00:00:00.579 2811926==\n> ...\n> \n> It looks like the line variable returned by gets_func() here is not\n> null-terminated:\n> while ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\n> \n> {\n> ...\n> if (sscanf(line, \"%u %\" CppAsString2(MAXPGPATH) \"s\\n\", &oid, lofname) != 2)\n> ...\n> And Valgrind doesn't like it.\n> \n\nValgrind is correct to not like it. LZ4Stream_gets() got modeled after\ngets() when it should have been modeled after fgets().\n\nPlease find a patch attached to address it.\n\nCheers,\n//Georgios\n\n> Best regards,\n> Alexander","msg_date":"Fri, 05 May 2023 10:02:12 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 2023-05-05 Fr 06:02, gkokolatos@pm.me wrote:\n>\n>\n>\n>\n> ------- Original Message -------\n> On Friday, May 5th, 2023 at 8:00 AM, Alexander Lakhin<exclusion@gmail.com> wrote:\n>\n>\n>>\n>> 23.03.2023 20:10, Tomas Vondra wrote:\n>>\n>>> So pushed all three parts, after updating the commit messages a bit.\n>>>\n>>> This leaves the empty-data issue (which we have a fix for) and the\n>>> switch to LZ4F. And then the zstd part.\n>>\n>> I'm sorry that I haven't noticed/checked that before, but when trying to\n>> perform check-world with Valgrind I've discovered another issue presumably\n>> related to LZ4File_gets().\n>> When running under Valgrind:\n>> PROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\n>> I get:\n>> ...\n>> 07:07:11.683 ok 1939 - compression_lz4_dir: glob check for\n>> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\n>> # Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql\n>> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\n>>\n>> ==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\n>> ==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\n>> ==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\n>> ==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\n>> ==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\n>> ==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\n>> ==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\n>> ==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\n>> ==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\n>> ==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\n>> ==00:00:00:00.579 2811926==\n>> ...\n>>\n>> It looks like the line variable returned by gets_func() here is not\n>> null-terminated:\n>> while ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\n>>\n>> {\n>> ...\n>> if (sscanf(line, \"%u %\" CppAsString2(MAXPGPATH) \"s\\n\", &oid, lofname) != 2)\n>> ...\n>> And Valgrind doesn't like it.\n>>\n> Valgrind is correct to not like it. LZ4Stream_gets() got modeled after\n> gets() when it should have been modeled after fgets().\n>\n> Please find a patch attached to address it.\n>\n>\n\nIsn't using memset here a bit wasteful? Why not just put a null at the \nend after calling LZ4Stream_read_internal(), which tells you how many \nbytes it has written?\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB:https://www.enterprisedb.com\n\n\n\n\n\n\n\n\nOn 2023-05-05 Fr 06:02,\n gkokolatos@pm.me wrote:\n\n\n\n\n\n\n\n------- Original Message -------\nOn Friday, May 5th, 2023 at 8:00 AM, Alexander Lakhin <exclusion@gmail.com> wrote:\n\n\n\n\n\n\n23.03.2023 20:10, Tomas Vondra wrote:\n\n\n\nSo pushed all three parts, after updating the commit messages a bit.\n\nThis leaves the empty-data issue (which we have a fix for) and the\nswitch to LZ4F. And then the zstd part.\n\n\n\n\nI'm sorry that I haven't noticed/checked that before, but when trying to\nperform check-world with Valgrind I've discovered another issue presumably\nrelated to LZ4File_gets().\nWhen running under Valgrind:\nPROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\nI get:\n...\n07:07:11.683 ok 1939 - compression_lz4_dir: glob check for\n.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\n# Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql\n.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\n\n==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\n==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\n==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\n==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\n==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\n==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\n==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\n==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\n==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\n==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\n==00:00:00:00.579 2811926==\n...\n\nIt looks like the line variable returned by gets_func() here is not\nnull-terminated:\nwhile ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\n\n{\n...\nif (sscanf(line, \"%u %\" CppAsString2(MAXPGPATH) \"s\\n\", &oid, lofname) != 2)\n...\nAnd Valgrind doesn't like it.\n\n\n\n\nValgrind is correct to not like it. LZ4Stream_gets() got modeled after\ngets() when it should have been modeled after fgets().\n\nPlease find a patch attached to address it.\n\n\n\n\n\n\nIsn't using memset here a bit wasteful? Why not just put a null\n at the end after calling LZ4Stream_read_internal(), which tells\n you how many bytes it has written?\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com","msg_date":"Fri, 5 May 2023 09:23:37 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Friday, May 5th, 2023 at 3:23 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\n\n> On 2023-05-05 Fr 06:02, gkokolatos@pm.me wrote:\n>\n>> ------- Original Message -------\n>> On Friday, May 5th, 2023 at 8:00 AM, Alexander Lakhin\n>> [<exclusion@gmail.com>](mailto:exclusion@gmail.com)\n>> wrote:\n>>\n>>> 23.03.2023 20:10, Tomas Vondra wrote:\n>>>\n>>>> So pushed all three parts, after updating the commit messages a bit.\n>>>>\n>>>> This leaves the empty-data issue (which we have a fix for) and the\n>>>> switch to LZ4F. And then the zstd part.\n>>>\n>>> I'm sorry that I haven't noticed/checked that before, but when trying to\n>>> perform check-world with Valgrind I've discovered another issue presumably\n>>> related to LZ4File_gets().\n>>> When running under Valgrind:\n>>> PROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\n>>> I get:\n>>> ...\n>>> 07:07:11.683 ok 1939 - compression_lz4_dir: glob check for\n>>> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\n>>> # Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql\n>>> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\n>>>\n>>> ==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\n>>> ==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\n>>> ==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\n>>> ==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\n>>> ==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\n>>> ==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\n>>> ==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\n>>> ==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\n>>> ==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\n>>> ==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\n>>> ==00:00:00:00.579 2811926==\n>>> ...\n>>>\n>>> It looks like the line variable returned by gets_func() here is not\n>>> null-terminated:\n>>> while ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\n>>>\n>>> {\n>>> ...\n>>> if (sscanf(line, \"%u %\" CppAsString2(MAXPGPATH) \"s\\n\", &oid, lofname) != 2)\n>>> ...\n>>> And Valgrind doesn't like it.\n>>\n>> Valgrind is correct to not like it. LZ4Stream_gets() got modeled after\n>> gets() when it should have been modeled after fgets().\n>>\n>> Please find a patch attached to address it.\n>\n> Isn't using memset here a bit wasteful? Why not just put a null at the end after calling LZ4Stream_read_internal(), which tells you how many bytes it has written?\n\nGood point. I thought about it before submitting the patch. I concluded that given the complexity and operations involved in LZ4Stream_read_internal() and the rest of the pg_dump/pg_restore code, the memset() call will be negligible. However from the readability point of view, the function is a bit cleaner with the memset().\n\nI will not object to any suggestion though, as this is a very trivial point. Please find attached a v2 of the patch following the suggested approach.\n\nCheers,\n\n//Georgios\n\n> cheers\n>\n> andrew\n>\n> --\n> Andrew Dunstan\n> EDB:\n> https://www.enterprisedb.com","msg_date":"Fri, 05 May 2023 14:13:28 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:\n> Good point. I thought about it before submitting the patch. I\n> concluded that given the complexity and operations involved in\n> LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore\n> code, the memset() call will be negligible. However from the\n> readability point of view, the function is a bit cleaner with the\n> memset(). \n> \n> I will not object to any suggestion though, as this is a very\n> trivial point. Please find attached a v2 of the patch following the\n> suggested approach.\n\nPlease note that an open item has been added for this stuff.\n--\nMichael","msg_date":"Sat, 6 May 2023 11:51:19 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sat, May 6, 2023 at 04:51, Michael Paquier <[michael@paquier.xyz](mailto:On Sat, May 6, 2023 at 04:51, Michael Paquier <<a href=)> wrote:\n\n> On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:\n>> Good point. I thought about it before submitting the patch. I\n>> concluded that given the complexity and operations involved in\n>> LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore\n>> code, the memset() call will be negligible. However from the\n>> readability point of view, the function is a bit cleaner with the\n>> memset().\n>>\n>> I will not object to any suggestion though, as this is a very\n>> trivial point. Please find attached a v2 of the patch following the\n>> suggested approach.\n>\n> Please note that an open item has been added for this stuff.\n\nThank you but I am not certain I know what that means. Can you please explain?\n\nCheers,\n//Georgios\n\n> --\n> Michael\n On Sat, May 6, 2023 at 04:51, Michael Paquier <michael@paquier.xyz> wrote: On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:> Good point. I thought about it before submitting the patch. I> concluded that given the complexity and operations involved in> LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore> code, the memset() call will be negligible. However from the> readability point of view, the function is a bit cleaner with the> memset(). > > I will not object to any suggestion though, as this is a very> trivial point. Please find attached a v2 of the patch following the> suggested approach.Please note that an open item has been added for this stuff.Thank you but I am not certain I know what that means. Can you please explain?Cheers,//Georgios--Michael","msg_date":"Sun, 07 May 2023 15:01:52 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sun, May 07, 2023 at 03:01:52PM +0000, gkokolatos@pm.me wrote:\n> Thank you but I am not certain I know what that means. Can you please explain?\n\nIt means that this thread has been added to the following list:\nhttps://wiki.postgresql.org/wiki/PostgreSQL_16_Open_Items#Open_Issues\n\npg_dump/compress_lz4.c is new as of PostgreSQL 16, and this patch is\nfixing a deficiency. That's just a way outside of the commit fest to\ntrack any problems and make sure these are fixed before the release\nhappens.\n--\nMichael","msg_date":"Mon, 8 May 2023 07:26:23 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:\n> Good point. I thought about it before submitting the patch. I\n> concluded that given the complexity and operations involved in\n> LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore\n> code, the memset() call will be negligible. However from the\n> readability point of view, the function is a bit cleaner with the\n> memset(). \n> \n> I will not object to any suggestion though, as this is a very\n> trivial point. Please find attached a v2 of the patch following the\n> suggested approach.\n\nHmm. I was looking at this patch, and what you are trying to do\nsounds rather right to keep a parallel with the gzip and zstd code\npaths.\n\nLooking at the code of gzread.c, gzgets() enforces a null-termination\non the string read. Still, isn't that something we'd better enforce\nin read_none() as well? compress_io.h lists this as a requirement of\nthe callback, and Zstd_gets() does so already. read_none() does not\nenforce that, unfortunately.\n\n+ /* No work needs to be done for a zero-sized output buffer */\n+ if (size <= 0)\n+ return 0;\n\nIndeed. This should be OK.\n\n- ret = LZ4Stream_read_internal(state, ptr, size, true);\n+ Assert(size > 1);\n\nThe addition of this assertion is a bit surprising, and this is\ninconsistent with Zstd_gets where a length of 1 is authorized. We\nshould be more consistent across all the callbacks, IMO, not less, so\nas we apply the same API contract across all the compression methods.\n\nWhile testing this patch, I have triggered an error pointing out that\nthe decompression path of LZ4 is broken for table data. I can\nreproduce that with a dump of the regression database, as of:\nmake installcheck\npg_dump --format=d --file=dump_lz4 --compress=lz4 regression\ncreatedb regress_lz4\npg_restore --format=d -d regress_lz4 dump_lz4\npg_restore: error: COPY failed for table \"clstr_tst\": ERROR: extra data after last expected column\nCONTEXT: COPY clstr_tst, line 15: \"32\t6\tseis\txyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzy...\"\npg_restore: warning: errors ignored on restore: 1\n\nThis does not show up with gzip or zstd, and the patch does not\ninfluence the result. In short it shows up with and without the\npatch, on HEAD. That does not look really stable :/\n--\nMichael","msg_date":"Mon, 8 May 2023 09:31:06 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"Michael Paquier <michael@paquier.xyz> writes:\n> While testing this patch, I have triggered an error pointing out that\n> the decompression path of LZ4 is broken for table data. I can\n> reproduce that with a dump of the regression database, as of:\n> make installcheck\n> pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\n> createdb regress_lz4\n> pg_restore --format=d -d regress_lz4 dump_lz4\n> pg_restore: error: COPY failed for table \"clstr_tst\": ERROR: extra data after last expected column\n> CONTEXT: COPY clstr_tst, line 15: \"32\t6\tseis\txyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzy...\"\n> pg_restore: warning: errors ignored on restore: 1\n\nUgh. Reproduced here ... so we need an open item for this.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 07 May 2023 21:09:25 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Sun, May 07, 2023 at 09:09:25PM -0400, Tom Lane wrote:\n> Ugh. Reproduced here ... so we need an open item for this.\n\nYep. Already added.\n--\nMichael","msg_date":"Mon, 8 May 2023 10:13:55 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"I wrote:\n> Michael Paquier <michael@paquier.xyz> writes:\n>> While testing this patch, I have triggered an error pointing out that\n>> the decompression path of LZ4 is broken for table data. I can\n>> reproduce that with a dump of the regression database, as of:\n>> make installcheck\n>> pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\n\n> Ugh. Reproduced here ... so we need an open item for this.\n\nBTW, it seems to work with --format=c.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 07 May 2023 21:16:51 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 5/7/23 17:01, gkokolatos@pm.me wrote:\n> \n> \n> \n> On Sat, May 6, 2023 at 04:51, Michael Paquier <michael@paquier.xyz\n> <mailto:On Sat, May 6, 2023 at 04:51, Michael Paquier <<a href=>> wrote:\n>> On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:\n>> > Good point. I thought about it before submitting the patch. I\n>> > concluded that given the complexity and operations involved in\n>> > LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore\n>> > code, the memset() call will be negligible. However from the\n>> > readability point of view, the function is a bit cleaner with the\n>> > memset().\n>> >\n>> > I will not object to any suggestion though, as this is a very\n>> > trivial point. Please find attached a v2 of the patch following the\n>> > suggested approach.\n>>\n>> Please note that an open item has been added for this stuff.\n> Thank you but I am not certain I know what that means. Can you please\n> explain?\n> \n\nIt means it was added to the list of items we need to fix before PG16\ngets out:\n\nhttps://wiki.postgresql.org/wiki/PostgreSQL_16_Open_Items\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Mon, 8 May 2023 03:34:15 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Monday, May 8th, 2023 at 3:16 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n\n\n> \n> \n> I wrote:\n> \n> > Michael Paquier michael@paquier.xyz writes:\n> > \n> > > While testing this patch, I have triggered an error pointing out that\n> > > the decompression path of LZ4 is broken for table data. I can\n> > > reproduce that with a dump of the regression database, as of:\n> > > make installcheck\n> > > pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\n> \n> > Ugh. Reproduced here ... so we need an open item for this.\n> \n> \n> BTW, it seems to work with --format=c.\n> \n\nThank you for the extra tests. It seems that exists a gap in the test\ncoverage. Please find a patch attached that is addressing the issue\nand attempt to provide tests for it.\n\nCheers,\n//Georgios\n\n> regards, tom lane","msg_date":"Mon, 08 May 2023 16:19:43 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 5/8/23 03:16, Tom Lane wrote:\n> I wrote:\n>> Michael Paquier <michael@paquier.xyz> writes:\n>>> While testing this patch, I have triggered an error pointing out that\n>>> the decompression path of LZ4 is broken for table data. I can\n>>> reproduce that with a dump of the regression database, as of:\n>>> make installcheck\n>>> pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\n> \n>> Ugh. Reproduced here ... so we need an open item for this.\n> \n> BTW, it seems to work with --format=c.\n> \n\nThe LZ4Stream_write() forgot to move the pointer to the next chunk, so\nit was happily decompressing the initial chunk over and over. A bit\nembarrassing oversight :-(\n\nThe custom format calls WriteDataToArchiveLZ4(), which was correct.\n\nThe attached patch fixes this for me.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company","msg_date":"Mon, 8 May 2023 20:00:39 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 5/8/23 18:19, gkokolatos@pm.me wrote:\n> \n> \n> \n> \n> \n> ------- Original Message -------\n> On Monday, May 8th, 2023 at 3:16 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> \n> \n>>\n>>\n>> I wrote:\n>>\n>>> Michael Paquier michael@paquier.xyz writes:\n>>>\n>>>> While testing this patch, I have triggered an error pointing out that\n>>>> the decompression path of LZ4 is broken for table data. I can\n>>>> reproduce that with a dump of the regression database, as of:\n>>>> make installcheck\n>>>> pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\n>>\n>>> Ugh. Reproduced here ... so we need an open item for this.\n>>\n>>\n>> BTW, it seems to work with --format=c.\n>>\n> \n> Thank you for the extra tests. It seems that exists a gap in the test\n> coverage. Please find a patch attached that is addressing the issue\n> and attempt to provide tests for it.\n> \n\nSeems I'm getting messages with a delay - this is mostly the same fix I\nended up with, not realizing you already posted a fix.\n\nI don't think we need the local \"in\" variable - the pointer parameter is\nlocal in the function, so we can modify it directly (with a cast).\nWriteDataToArchiveLZ4 does it that way too.\n\nThe tests are definitely a good idea. I wonder if we should add a\ncomment to DEFAULT_IO_BUFFER_SIZE mentioning that if we choose to\nincrease the value in the future, we needs to tweak the tests too to use\nmore data in order to exercise the buffering etc. Maybe it's obvious?\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Mon, 8 May 2023 20:20:42 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Monday, May 8th, 2023 at 8:20 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> \n> \n> On 5/8/23 18:19, gkokolatos@pm.me wrote:\n> \n> > ------- Original Message -------\n> > On Monday, May 8th, 2023 at 3:16 AM, Tom Lane tgl@sss.pgh.pa.us wrote:\n> > \n> > > I wrote:\n> > > \n> > > > Michael Paquier michael@paquier.xyz writes:\n> > > > \n> > > > > While testing this patch, I have triggered an error pointing out that\n> > > > > the decompression path of LZ4 is broken for table data. I can\n> > > > > reproduce that with a dump of the regression database, as of:\n> > > > > make installcheck\n> > > > > pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\n> > > \n> > > > Ugh. Reproduced here ... so we need an open item for this.\n> > > \n> > > BTW, it seems to work with --format=c.\n> > \n> > Thank you for the extra tests. It seems that exists a gap in the test\n> > coverage. Please find a patch attached that is addressing the issue\n> > and attempt to provide tests for it.\n> \n> \n> Seems I'm getting messages with a delay - this is mostly the same fix I\n> ended up with, not realizing you already posted a fix.\n\nThank you very much for looking.\n\n> I don't think we need the local \"in\" variable - the pointer parameter is\n> local in the function, so we can modify it directly (with a cast).\n> WriteDataToArchiveLZ4 does it that way too.\n\nSure, patch updated.\n \n> The tests are definitely a good idea.\n\nThank you.\n\n> I wonder if we should add a\n> comment to DEFAULT_IO_BUFFER_SIZE mentioning that if we choose to\n> increase the value in the future, we needs to tweak the tests too to use\n> more data in order to exercise the buffering etc. Maybe it's obvious?\n> \n\nYou are right. Added a comment both in the header and in the test.\n\nI hope v2 gets closer to closing the open item for this.\n\nCheers,\n//Georgios\n\n\n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Mon, 08 May 2023 19:53:28 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Mon, May 08, 2023 at 08:00:39PM +0200, Tomas Vondra wrote:\n> The LZ4Stream_write() forgot to move the pointer to the next chunk, so\n> it was happily decompressing the initial chunk over and over. A bit\n> embarrassing oversight :-(\n> \n> The custom format calls WriteDataToArchiveLZ4(), which was correct.\n> \n> The attached patch fixes this for me.\n\nOuch. So this was corrupting the dumps and the compression when\ntrying to write more than two chunks at once, not the decompression\nsteps. That addresses the issue here as well, thanks!\n--\nMichael","msg_date":"Tue, 9 May 2023 07:10:55 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 5/9/23 00:10, Michael Paquier wrote:\n> On Mon, May 08, 2023 at 08:00:39PM +0200, Tomas Vondra wrote:\n>> The LZ4Stream_write() forgot to move the pointer to the next chunk, so\n>> it was happily decompressing the initial chunk over and over. A bit\n>> embarrassing oversight :-(\n>>\n>> The custom format calls WriteDataToArchiveLZ4(), which was correct.\n>>\n>> The attached patch fixes this for me.\n> \n> Ouch. So this was corrupting the dumps and the compression when\n> trying to write more than two chunks at once, not the decompression\n> steps. That addresses the issue here as well, thanks!\n\nYeah. Thanks for the report, should have been found during review.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Tue, 9 May 2023 14:54:31 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"------- Original Message -------\nOn Tuesday, May 9th, 2023 at 2:54 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\n\n\n> \n> \n> On 5/9/23 00:10, Michael Paquier wrote:\n> \n> > On Mon, May 08, 2023 at 08:00:39PM +0200, Tomas Vondra wrote:\n> > \n> > > The LZ4Stream_write() forgot to move the pointer to the next chunk, so\n> > > it was happily decompressing the initial chunk over and over. A bit\n> > > embarrassing oversight :-(\n> > > \n> > > The custom format calls WriteDataToArchiveLZ4(), which was correct.\n> > > \n> > > The attached patch fixes this for me.\n> > \n> > Ouch. So this was corrupting the dumps and the compression when\n> > trying to write more than two chunks at once, not the decompression\n> > steps. That addresses the issue here as well, thanks!\n> \n> \n> Yeah. Thanks for the report, should have been found during review.\n\nThank you both for looking. A small consolation is that now there are\ntests for this case.\n\nMoving on to the other open item for this, please find attached v2\nof the patch as requested.\n\nCheers,\n//Georgios\n\n> \n> \n> regards\n> \n> --\n> Tomas Vondra\n> EnterpriseDB: http://www.enterprisedb.com\n> The Enterprise PostgreSQL Company","msg_date":"Tue, 09 May 2023 14:12:44 +0000","msg_from":"gkokolatos@pm.me","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, May 09, 2023 at 02:12:44PM +0000, gkokolatos@pm.me wrote:\n> Thank you both for looking. A small consolation is that now there are\n> tests for this case.\n\n+1, noticing that was pure luck ;)\n\nWorth noting that the patch posted in [1] has these tests, not the\nversion posted in [2].\n\n+ create_sql => 'INSERT INTO dump_test.test_compression_method (col1) '\n+ . 'SELECT string_agg(a::text, \\'\\') FROM generate_series(1,4096) a;',\n\nYep, good and cheap idea to check for longer chunks. That should be\nenough to loop twice.\n\n[1]: https://www.postgresql.org/message-id/SYTRcNgtAbzyn3y3IInh1x-UfNTKMNpnFvI3mr6SyqyVf3PkaDsMy_cpKKgsl3_HdLy2MFAH4zwjxDmFfiLO8rWtSiJWBtqT06OMjeNo4GA=@pm.me\n[2]: https://www.postgresql.org/message-id/f735df01-0bb4-2fbc-1297-73a520cfc534@enterprisedb.com\n\n> Moving on to the other open item for this, please find attached v2\n> of the patch as requested.\n\nDid you notice the comments of [3] about the second patch that aims to\nadd the null termination in the line from the LZ4 fgets() callback?\n\n[3]: https://www.postgresql.org/message-id/ZFhCyn4Gm2eu60rB@paquier.xyz\n--\nMichael","msg_date":"Wed, 10 May 2023 11:22:10 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On Tue, May 09, 2023 at 02:54:31PM +0200, Tomas Vondra wrote:\n> Yeah. Thanks for the report, should have been found during review.\n\nTomas, are you planning to do something by the end of this week for\nbeta1? Or do you need some help of any kind?\n--\nMichael","msg_date":"Wed, 17 May 2023 15:18:58 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"On 5/17/23 08:18, Michael Paquier wrote:\n> On Tue, May 09, 2023 at 02:54:31PM +0200, Tomas Vondra wrote:\n>> Yeah. Thanks for the report, should have been found during review.\n> \n> Tomas, are you planning to do something by the end of this week for\n> beta1? Or do you need some help of any kind?\n\nI'll take care of it.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 17 May 2023 10:59:02 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"},{"msg_contents":"\n\nOn 5/17/23 10:59, Tomas Vondra wrote:\n> On 5/17/23 08:18, Michael Paquier wrote:\n>> On Tue, May 09, 2023 at 02:54:31PM +0200, Tomas Vondra wrote:\n>>> Yeah. Thanks for the report, should have been found during review.\n>>\n>> Tomas, are you planning to do something by the end of this week for\n>> beta1? Or do you need some help of any kind?\n> \n> I'll take care of it.\n> \n\nFWIW I've pushed fixes for both open issues associated with the pg_dump\ncompression. I'll keep an eye on the buildfarm, but hopefully that'll do\nit for beta1.\n\n\nregards\n\n-- \nTomas Vondra\nEnterpriseDB: http://www.enterprisedb.com\nThe Enterprise PostgreSQL Company\n\n\n","msg_date":"Wed, 17 May 2023 17:02:22 +0200","msg_from":"Tomas Vondra <tomas.vondra@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Add LZ4 compression in pg_dump"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nplease find attached a patchset which adds lz4 compression in pg_dump.\\n\\nThe first commit does the heavy lifting required for additional compression methods.\\nIt expands testing coverage for the already supported gzip compression. Commit\\nbf9aa490db introduced cfp in compress_io.{c,h} with the intent of unifying\\ncompression related code and allow for the introduction of additional archive\\nformats. However, pg_backup_archiver.c was not using that API. This commit\\nteaches pg_backup_archiver.c about cfp and is using it through out.\\n\\nFurthermore, compression was chosen based on the value of the level passed\\nas an argument during the invocation of pg_dump or some hardcoded defaults. This\\ndoes not scale for more than one compression methods. Now the method used for\\ncompression can be explicitly requested during command invocation, or set during\\nhardcoded defaults. Then it is stored in the relevant structs and passed in the\\nrelevant functions, along side compression level which has lost it's special\\nmeaning. The method for compression is not yet stored in the actual archive.\\nThis is done in the next commit which does introduce a new method.\\n\\nThe previously named CompressionAlgorithm enum is changed for\\nCompressionMethod so that it matches better similar variables found through out\\nthe code base.\\n\\nIn a fashion similar to the binary for pg_basebackup, the method for compression\\nis passed using the already existing -Z/--compress parameter of pg_dump. The\\nlegacy format and behaviour is maintained. Additionally, the user can explicitly\\npass a requested method and optionaly the level to be used after a semicolon,e.g. --compress=gzip:6\\n\\nThe second commit adds LZ4 compression in pg_dump and pg_restore.\\n\\nWithin compress_io.{c,h} there are two distinct APIs exposed, the streaming API\\nand a file API. The first one, is aimed at inlined use cases and thus simple\\nlz4.h calls can be used directly. The second one is generating output, or is\\nparsing input, which can be read/generated via the lz4 utility.\\n\\nIn the later case, the API is using an opaque wrapper around a file stream,\\nwhich aquired via fopen() or gzopen() respectively. It would then provide\\nwrappers around fread(), fwrite(), fgets(), fgetc(), feof(), and fclose(); or\\ntheir gz equivallents. However the LZ4F api does not provide this functionality.\\nSo this has been implemented localy.\\n\\nIn order to maintain the API compatibility a new structure LZ4File is\\nintroduced. It is responsible for keeping state and any yet unused generated\\ncontent. The later is required when the generated decompressed output, exceeds\\nthe caller's buffer capacity.\\n\\nCustom compressed archives need to now store the compression method in their\\nheader. This requires a bump in the version number. The level of compression is\\nstill stored in the dump, though admittedly is of no apparent use.\\n\\nThe series is authored by me. Rachel Heaton helped out with the expansion\\nof the testing coverage, testing in different platforms and providing debug information\\non those, as well as native speaker wording.\\n\\nCheers,\\n//Georgios\",\n \"msg_date\": \"Fri, 25 Feb 2022 12:05:31 +0000\",\n \"msg_from\": \"Georgios <gkokolatos@protonmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 12:05:31PM +0000, Georgios wrote:\\n> The first commit does the heavy lifting required for additional compression methods.\\n> It expands testing coverage for the already supported gzip compression. Commit\\n> bf9aa490db introduced cfp in compress_io.{c,h} with the intent of unifying\\n> compression related code and allow for the introduction of additional archive\\n> formats. However, pg_backup_archiver.c was not using that API. This commit\\n> teaches pg_backup_archiver.c about cfp and is using it through out.\\n\\nThanks for the patch. I have a few high-level comments.\\n\\n+ # Do not use --no-sync to give test coverage for data sync.\\n+ compression_gzip_directory_format => {\\n+ test_key => 'compression',\\n\\nThe tests for GZIP had better be split into their own commit, as\\nthat's a coverage improvement for the existing code.\\n\\nI was assuming that this was going to be much larger :)\\n\\n+/* Routines that support LZ4 compressed data I/O */\\n+#ifdef HAVE_LIBLZ4\\n+static void InitCompressorLZ4(CompressorState *cs);\\n+static void ReadDataFromArchiveLZ4(ArchiveHandle *AH, ReadFunc\\nreadF);\\n+static void WriteDataToArchiveLZ4(ArchiveHandle *AH, CompressorState *cs,\\n+ const char *data, size_t dLen);\\n+static void EndCompressorLZ4(ArchiveHandle *AH, CompressorState *cs);\\n+#endif\\n\\nHmm. This is the same set of APIs as ZLIB and NONE to init, read,\\nwrite and end, but for the LZ4 compressor (NONE has no init/end).\\nWouldn't it be better to refactor the existing pg_dump code to have a\\ncentral structure holding all the function definitions in a common\\nstructure so as all those function signatures are set in stone in the\\nshape of a catalog of callbacks, making the addition of more\\ncompression formats easier? I would imagine that we'd split the code\\nof each compression method into their own file with their own context\\ndata. This would lead to a removal of compress_io.c, with its entry\\npoints ReadDataFromArchive(), WriteDataToArchive() & co replaced by\\npointers to each per-compression callback.\\n\\n> Furthermore, compression was chosen based on the value of the level passed\\n> as an argument during the invocation of pg_dump or some hardcoded defaults. This\\n> does not scale for more than one compression methods. Now the method used for\\n> compression can be explicitly requested during command invocation, or set during\\n> hardcoded defaults. Then it is stored in the relevant structs and passed in the\\n> relevant functions, along side compression level which has lost it's special\\n> meaning. The method for compression is not yet stored in the actual archive.\\n> This is done in the next commit which does introduce a new method.\\n\\nThat's one thing Robert was arguing about with pg_basebackup, so that\\nwould be consistent, and the option set is backward-compatible as far\\nas I get it by reading the code.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 1 Mar 2022 17:04:35 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"The patch is failing on cfbot/freebsd.\\nhttp://cfbot.cputube.org/georgios-kokolatos.html\\n\\nAlso, I wondered if you'd looked at the \\\"high compression\\\" interfaces in\\nlz4hc.h ? Should pg_dump use that ?\\n\\nOn Fri, Feb 25, 2022 at 08:03:40AM -0600, Justin Pryzby wrote:\\n> Thanks for working on this. Your 0001 looks similar to what I did for zstd 1-2\\n> years ago.\\n> https://commitfest.postgresql.org/32/2888/\\n> \\n> I rebased and attached the latest patches I had in case they're useful to you.\\n> I'd like to see zstd included in pg_dump eventually, but it was too much work\\n> to shepherd the patches. Now that seems reasonable for pg16.\\n> \\n> With the other compression patches I've worked on, we've used an extra patch\\n> with changes the default to the new compression algorithm, to force cfbot to\\n> exercize the new code.\\n> \\n> Do you know the process with commitfests and cfbot ?\\n> There's also this, which allows running the tests on cirrus before mailing the\\n> patch to the hackers list.\\n> ./src/tools/ci/README\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Mar 2022 10:10:39 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"It seems development on this has stalled. If there's no further work\\nhappening I guess I'll mark the patch returned with feedback. Feel\\nfree to resubmit it to the next CF when there's progress.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 01:20:47 -0400\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 01:20:47AM -0400, Greg Stark wrote:\\n> It seems development on this has stalled. If there's no further work\\n> happening I guess I'll mark the patch returned with feedback. Feel\\n> free to resubmit it to the next CF when there's progress.\\n\\nSince it's a reasonably large patch (and one that I had myself started before)\\nand it's only been 20some days since (minor) review comments, and since the\\nfocus right now is on committing features, and not reviewing new patches, and\\nthis patch is new one month ago, and its 0002 not intended for pg15, therefor\\nI'm moving it to the next CF, where I hope to work with its authors to progress\\nit.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 08:22:53 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 6:22 AM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> On Fri, Mar 25, 2022 at 01:20:47AM -0400, Greg Stark wrote:\\n> > It seems development on this has stalled. If there's no further work\\n> > happening I guess I'll mark the patch returned with feedback. Feel\\n> > free to resubmit it to the next CF when there's progress.\\n>\\n> Since it's a reasonably large patch (and one that I had myself started before)\\n> and it's only been 20some days since (minor) review comments, and since the\\n> focus right now is on committing features, and not reviewing new patches, and\\n> this patch is new one month ago, and its 0002 not intended for pg15, therefor\\n> I'm moving it to the next CF, where I hope to work with its authors to progress\\n> it.\\n>\\nHi Folks,\\n\\nHere is an updated patchset from Georgios, with minor assistance from myself.\\nThe comments above should be addressed, but please let us know if\\nthere are other things to go over. A functional change in this\\npatchset is when `--compress=none` is passed to pg_dump, it will not\\ncompress for directory type (previously, it would use gzip if\\npresent). The previous default behavior is retained.\\n\\n- Rachel\",\n \"msg_date\": \"Fri, 25 Mar 2022 16:13:08 -0700\",\n \"msg_from\": \"Rachel Heaton <rachelmheaton@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n------- Original Message -------\\n\\nOn Saturday, March 26th, 2022 at 12:13 AM, Rachel Heaton <rachelmheaton@gmail.com> wrote:\\n\\n> On Fri, Mar 25, 2022 at 6:22 AM Justin Pryzby pryzby@telsasoft.com wrote:\\n>\\n> > On Fri, Mar 25, 2022 at 01:20:47AM -0400, Greg Stark wrote:\\n> >\\n> > > It seems development on this has stalled. If there's no further work\\n> > > happening I guess I'll mark the patch returned with feedback. Feel\\n> > > free to resubmit it to the next CF when there's progress.\\n\\nWe had some progress yet we didn't want to distract the list with too many\\nemails. Of course, it seemed stalled to the outside observer, yet I simply\\nwanted to set the record straight and say that we are actively working on it.\\n\\n> >\\n> > Since it's a reasonably large patch (and one that I had myself started before)\\n> > and it's only been 20some days since (minor) review comments, and since the\\n> > focus right now is on committing features, and not reviewing new patches, and\\n> > this patch is new one month ago, and its 0002 not intended for pg15, therefor\\n> > I'm moving it to the next CF, where I hope to work with its authors to progress\\n> > it.\\n\\nThank you. It is much appreciated. We will sent updates when the next commitfest\\nstarts in July as to not distract from the 15 work. Then, we can take it from there.\\n\\n>\\n> Hi Folks,\\n>\\n> Here is an updated patchset from Georgios, with minor assistance from myself.\\n> The comments above should be addressed, but please let us know if\\n\\nA small amendment to the above statement. This patchset does not include the\\nrefactoring of compress_io suggested by Mr Paquier in the same thread, as it is\\nmissing documentation. An updated version will be sent to include those changes\\non the next commitfest.\\n\\n> there are other things to go over. A functional change in this\\n> patchset is when `--compress=none` is passed to pg_dump, it will not\\n> compress for directory type (previously, it would use gzip if\\n> present). The previous default behavior is retained.\\n>\\n> - Rachel\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Mar 2022 23:43:17 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 25, 2022 at 11:43:17PM +0000, gkokolatos@pm.me wrote:\\n> On Saturday, March 26th, 2022 at 12:13 AM, Rachel Heaton <rachelmheaton@gmail.com> wrote:\\n>> Here is an updated patchset from Georgios, with minor assistance from myself.\\n>> The comments above should be addressed, but please let us know if\\n> \\n> A small amendment to the above statement. This patchset does not include the\\n> refactoring of compress_io suggested by Mr Paquier in the same thread, as it is\\n> missing documentation. An updated version will be sent to include those changes\\n> on the next commitfest.\\n\\nThe refactoring using callbacks would make the code much cleaner IMO\\nin the long term, with zstd waiting in the queue. Now, I see some\\npieces of the patch set that could be merged now without waiting for\\nthe development cycle of 16 to begin, as of 0001 to add more tests and\\n0002.\\n\\nI have a question about 0002, actually. What has led you to the\\nconclusion that this code is dead and could be removed?\\n--\\nMichael\",\n \"msg_date\": \"Sat, 26 Mar 2022 14:57:50 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 26, 2022 at 02:57:50PM +0900, Michael Paquier wrote:\\n> I have a question about 0002, actually. What has led you to the\\n> conclusion that this code is dead and could be removed?\\n\\nSee 0001 and the manpage.\\n\\n+ 'pg_dump: compression is not supported by tar archive format');\\n\\nWhen I submitted a patch to support zstd, I spent awhile trying to make\\ncompression work with tar, but it's a significant effort and better done\\nseparately.\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Mar 2022 01:14:41 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\\n\\nI ran into that on an ubuntu LTS, so I don't think it's so old that it\\nshouldn't be handled more gracefully. LZ4 should either have an explicit\\nversion check, or else shouldn't depend on that feature (or should define a\\nsafe fallback version if the library header doesn't define it).\\n\\nhttps://packages.ubuntu.com/liblz4-1\\n\\n0003: typo: of legacy => or legacy\\n\\nThere are a large number of ifdefs being added here - it'd be nice to minimize\\nthat. basebackup was organized to use separate files, which is one way.\\n\\n$ git grep -c 'ifdef .*LZ4' src/bin/pg_dump/compress_io.c\\nsrc/bin/pg_dump/compress_io.c:19\\n\\nIn last year's CF entry, I had made a union within CompressorState. LZ4\\ndoesn't need z_streamp (and ztsd will need ZSTD_outBuffer, ZSTD_inBuffer,\\nZSTD_CStream).\\n\\n0002: I wonder if you're able to re-use any of the basebackup parsing stuff\\nfrom commit ffd53659c. You're passing both the compression method *and* level.\\nI think there should be a structure which includes both. In the future, that\\ncan also handle additional options. I hope to re-use these same things for\\nwal_compression=method:level.\\n\\nYou renamed this:\\n\\n|- COMPR_ALG_LIBZ\\n|-} CompressionAlgorithm;\\n|+ COMPRESSION_GZIP,\\n|+} CompressionMethod;\\n\\n..But I don't think that's an improvement. If you were to change it, it should\\nsay something like PGDUMP_COMPRESS_ZLIB, since there are other compression\\nstructs and typedefs. zlib is not idential to gzip, which uses a different\\nheader, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\\n\\nThe cf* changes in pg_backup_archiver could be split out into a separate\\ncommit. It's strictly a code simplification - not just preparation for more\\ncompression algorithms. The commit message should \\\"See also:\\nbf9aa490db24b2334b3595ee33653bf2fe39208c\\\".\\n\\nThe changes in 0002 for cfopen_write seem insufficient:\\n|+ if (compressionMethod == COMPRESSION_NONE)\\n|+ fp = cfopen(path, mode, compressionMethod, 0);\\n| else\\n| {\\n| #ifdef HAVE_LIBZ\\n| char *fname;\\n| \\n| fname = psprintf(\\\"%s.gz\\\", path);\\n|- fp = cfopen(fname, mode, compression);\\n|+ fp = cfopen(fname, mode, compressionMethod, compressionLevel);\\n| free_keep_errno(fname);\\n| #else\\n\\nThe only difference between the LIBZ and uncompressed case is the file\\nextension, and it'll be the only difference with LZ4 too. So I suggest to\\nfirst handle the file extension, and the rest of the code path is not\\nconditional on the compression method. I don't think cfopen_write even needs\\nHAVE_LIBZ - can't you handle that in cfopen_internal() ?\\n\\nThis patch rejects -Z0, which ought to be accepted:\\n./src/bin/pg_dump/pg_dump -h /tmp regression -Fc -Z0 |wc\\npg_dump: error: can only specify -Z/--compress [LEVEL] when method is set\\n\\nYour 0003 patch shouldn't reference LZ4:\\n+#ifndef HAVE_LIBLZ4\\n+ if (*compressionMethod == COMPRESSION_LZ4)\\n+ supports_compression = false;\\n+#endif\\n\\nThe 0004 patch renames zlibOutSize to outsize - I think the patch series should\\nbe constructed such as to minimize the size of the method-specific patches. I\\nsay this anticipating also adding support for zstd. The preliminary patches\\nshould have all the boring stuff. It would help for reviewing to keep the \\npatches split up, or to enumerate all the boring things that are being renamed\\n(like change OutputContext to cfp, rename zlibOutSize, ...).\\n\\n0004: The include should use <lz4.h> and not \\\"lz4.h\\\"\\n\\nfreebsd/cfbot is failing.\\n\\nI suggested off-list to add an 0099 patch to change LZ4 to the default, to\\nexercise it more on CI.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Mar 2022 11:21:56 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\\n> You're passing both the compression method *and* level. I think there should\\n> be a structure which includes both. In the future, that can also handle\\n> additional options.\\n\\nI'm not sure if there's anything worth saving, but I did that last year with\\n0003-Support-multiple-compression-algs-levels-opts.patch\\nI sent a rebased copy off-list.\\nhttps://www.postgresql.org/message-id/flat/20210104025321.GA9712@telsasoft.com#ca1b9f9d3552c87fa874731cad9d8391\\n\\n|\\tfatal(\\\"not built with LZ4 support\\\");\\n|\\tfatal(\\\"not built with lz4 support\\\");\\n\\nPlease use consistent capitalization of \\\"lz4\\\" - then the compiler can optimize\\naway duplicate strings.\\n\\n> 0004: The include should use <lz4.h> and not \\\"lz4.h\\\"\\n\\nAlso, use USE_LZ4 rather than HAVE_LIBLZ4, per 75eae0908.\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Mar 2022 13:33:36 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"> On 26 Mar 2022, at 17:21, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n> I suggested off-list to add an 0099 patch to change LZ4 to the default, to\\n> exercise it more on CI.\\n\\nNo need to change the defaults in autoconf for that. The CFBot uses the cirrus\\nfile in the tree so changing what the job includes can be easily done (assuming\\nthe CFBot hasn't changed this recently which I think it hasn't). I used that\\ntrick in the NSS patchset to add a completely new job for --with-ssl=nss beside\\nthe --with-ssl=openssl job.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Mar 2022 00:37:27 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 27, 2022 at 12:37:27AM +0100, Daniel Gustafsson wrote:\\n> > On 26 Mar 2022, at 17:21, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> \\n> > I suggested off-list to add an 0099 patch to change LZ4 to the default, to\\n> > exercise it more on CI.\\n> \\n> No need to change the defaults in autoconf for that. The CFBot uses the cirrus\\n> file in the tree so changing what the job includes can be easily done (assuming\\n> the CFBot hasn't changed this recently which I think it hasn't). I used that\\n> trick in the NSS patchset to add a completely new job for --with-ssl=nss beside\\n> the --with-ssl=openssl job.\\n\\nI think you misunderstood - I'm suggesting not only to use with-lz4 (which was\\nalways true since 93d973494), but to change pg_dump -Fc and -Fd to use LZ4 by\\ndefault (the same as I suggested for toast_compression, wal_compression, and\\nagain in last year's patch to add zstd compression to pg_dump, for which\\npostgres was not ready).\\n\\n@@ -781,6 +807,11 @@ main(int argc, char **argv)\\n compress.alg = COMPR_ALG_LIBZ;\\n compress.level = Z_DEFAULT_COMPRESSION;\\n #endif\\n+\\n+#ifdef USE_ZSTD\\n+ compress.alg = COMPR_ALG_ZSTD; // Set default for testing purposes\\n+ compress.level = ZSTD_CLEVEL_DEFAULT;\\n+#endif\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Mar 2022 18:51:34 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 26, 2022 at 12:22 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\\n> from commit ffd53659c. You're passing both the compression method *and* level.\\n> I think there should be a structure which includes both. In the future, that\\n> can also handle additional options. I hope to re-use these same things for\\n> wal_compression=method:level.\\n\\nYeah, we should really try to use that infrastructure instead of\\ninventing a bunch of different ways to do it. It might require some\\nrenaming here and there, and I'm not sure whether we really want to\\ntry to rush all this into the current release, but I think we should\\nfind a way to get it done.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Mar 2022 10:13:00 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 27, 2022 at 10:13:00AM -0400, Robert Haas wrote:\\n> On Sat, Mar 26, 2022 at 12:22 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\\n> > from commit ffd53659c. You're passing both the compression method *and* level.\\n> > I think there should be a structure which includes both. In the future, that\\n> > can also handle additional options. I hope to re-use these same things for\\n> > wal_compression=method:level.\\n> \\n> Yeah, we should really try to use that infrastructure instead of\\n> inventing a bunch of different ways to do it. It might require some\\n> renaming here and there, and I'm not sure whether we really want to\\n> try to rush all this into the current release, but I think we should\\n> find a way to get it done.\\n\\nIt seems like something a whole lot like parse_compress_options() should be in\\ncommon/. Nobody wants to write it again, and I couldn't convince myself to\\ncopy it when I looked at using it for wal_compression.\\n\\nMaybe it should take an argument which specifies the default algorithm to use\\nfor input of a numeric \\\"level\\\". And reject such input if not specified, since\\nwal_compression has never taken a \\\"level\\\", so it's not useful or desirable to\\nhave that default to some new algorithm.\\n\\nI could write this down if you want, although I'm not sure how/if you intend\\nother people to use bc_algorithm and bc_algorithm. I don't think it's\\nimportant to do for v15, but it seems like it could be done after featue\\nfreeze. pg_dump+lz4 is targetting v16, although there's a cleanup patch that\\ncould also go in before branching.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Mar 2022 11:06:35 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"> On 27 Mar 2022, at 00:51, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> \\n> On Sun, Mar 27, 2022 at 12:37:27AM +0100, Daniel Gustafsson wrote:\\n>>> On 26 Mar 2022, at 17:21, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>> \\n>>> I suggested off-list to add an 0099 patch to change LZ4 to the default, to\\n>>> exercise it more on CI.\\n>> \\n>> No need to change the defaults in autoconf for that. The CFBot uses the cirrus\\n>> file in the tree so changing what the job includes can be easily done (assuming\\n>> the CFBot hasn't changed this recently which I think it hasn't). I used that\\n>> trick in the NSS patchset to add a completely new job for --with-ssl=nss beside\\n>> the --with-ssl=openssl job.\\n> \\n> I think you misunderstood - I'm suggesting not only to use with-lz4 (which was\\n> always true since 93d973494), but to change pg_dump -Fc and -Fd to use LZ4 by\\n> default (the same as I suggested for toast_compression, wal_compression, and\\n> again in last year's patch to add zstd compression to pg_dump, for which\\n> postgres was not ready).\\n\\nRight, I clearly misunderstood, thanks for the clarification.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Mar 2022 21:55:52 +0200\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sun, Mar 27, 2022 at 12:06 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> Maybe it should take an argument which specifies the default algorithm to use\\n> for input of a numeric \\\"level\\\". And reject such input if not specified, since\\n> wal_compression has never taken a \\\"level\\\", so it's not useful or desirable to\\n> have that default to some new algorithm.\\n\\nThat sounds odd to me. Wouldn't it be rather confusing if a bare\\ninteger meant gzip for one case and lz4 for another?\\n\\n> I could write this down if you want, although I'm not sure how/if you intend\\n> other people to use bc_algorithm and bc_algorithm. I don't think it's\\n> important to do for v15, but it seems like it could be done after featue\\n> freeze. pg_dump+lz4 is targetting v16, although there's a cleanup patch that\\n> could also go in before branching.\\n\\nWell, I think the first thing we should do is get rid of enum\\nWalCompressionMethod and use enum WalCompression instead. They've got\\nthe same elements and very similar names, but the WalCompressionMethod\\nones just have names like COMPRESSION_NONE, which is too generic,\\nwhereas WalCompressionMethod uses WAL_COMPRESSION_NONE, which is\\nbetter. Then I think we should also rename the COMPR_ALG_* constants\\nin pg_dump.h to names like DUMP_COMPRESSION_*. Once we do that we've\\ngot rid of all the unprefixed things that purport to be a list of\\ncompression algorithms.\\n\\nThen, if people are willing to adopt the syntax that the\\nbackup_compression.c/h stuff supports as a project standard (+1 from\\nme) we can go the other way and rename that stuff to be more generic,\\ntaking backup out of the name.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Mar 2022 08:36:15 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 28, 2022 at 08:36:15AM -0400, Robert Haas wrote:\\n> Well, I think the first thing we should do is get rid of enum\\n> WalCompressionMethod and use enum WalCompression instead. They've got\\n> the same elements and very similar names, but the WalCompressionMethod\\n> ones just have names like COMPRESSION_NONE, which is too generic,\\n> whereas WalCompressionMethod uses WAL_COMPRESSION_NONE, which is\\n> better. Then I think we should also rename the COMPR_ALG_* constants\\n> in pg_dump.h to names like DUMP_COMPRESSION_*. Once we do that we've\\n> got rid of all the unprefixed things that purport to be a list of\\n> compression algorithms.\\n\\nYes, having a centralized enum for the compression method would make\\nsense, along with the routines to parse and get the compression method\\nnames. At least that would be one step towards more unity in\\nsrc/common/.\\n\\n> Then, if people are willing to adopt the syntax that the\\n> backup_compression.c/h stuff supports as a project standard (+1 from\\n> me) we can go the other way and rename that stuff to be more generic,\\n> taking backup out of the name.\\n\\nI am not sure about the specification part which is only used by base\\nbackups that has no client-server requirements, so option values would\\nstill require their own grammar.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 29 Mar 2022 14:03:12 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, Mar 26, 2022 at 01:14:41AM -0500, Justin Pryzby wrote:\\n> See 0001 and the manpage.\\n> \\n> + 'pg_dump: compression is not supported by tar archive format');\\n> \\n> When I submitted a patch to support zstd, I spent awhile trying to make\\n> compression work with tar, but it's a significant effort and better done\\n> separately.\\n\\nWow. This stuff is old enough to vote (c3e18804), dead since its\\nintroduction. There is indeed an argument for removing that, it is\\nnot good to keep around that that has never been stressed and/or\\nused. Upon review, the cleanup done looks correct, as we have never\\nbeen able to generate .dat.gz files in for a dump in the tar format.\\n\\n+ command_fails_like(\\n+ [ 'pg_dump', '--compress', '1', '--format', 'tar' ],\\nThis addition depending on HAVE_LIBZ is a good thing as a reminder of\\nany work that could be done in 0002. Now that's waiting for 20 years\\nso I would not hold my breath on this support. I think that this\\ncould be just applied first, with 0002 on top of it, as a first\\nimprovement.\\n\\n+ compress_cmd => [\\n+ $ENV{'GZIP_PROGRAM'},\\nPatch 0001 is missing and update of pg_dump's Makefile to pass down\\nthis environment variable to the test scripts, no?\\n\\n+ compress_cmd => [\\n+ $ENV{'GZIP_PROGRAM'},\\n+ '-f',\\n[...]\\n+ $ENV{'GZIP_PROGRAM'},\\n+ '-k', '-d',\\n-f and -d are available everywhere I looked at, but is -k/--keep a\\nportable choice with a gzip command? I don't see this option in\\nOpenBSD, for one. So this test is going to cause problems on those\\nbuildfarm machines, at least. Couldn't this part be replaced by a\\nsimple --test to check that what has been compressed is in correct\\nshape? We know that this works, based on our recent experiences with\\nthe other tests.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 29 Mar 2022 16:27:32 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\n\\nOn Tuesday, March 29th, 2022 at 9:27 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Sat, Mar 26, 2022 at 01:14:41AM -0500, Justin Pryzby wrote:\\n> > See 0001 and the manpage.\\n> > + 'pg_dump: compression is not supported by tar archive format');\\n> > When I submitted a patch to support zstd, I spent awhile trying to make\\n> > compression work with tar, but it's a significant effort and better done\\n> > separately.\\n>\\n> Wow. This stuff is old enough to vote (c3e18804), dead since its\\n> introduction. There is indeed an argument for removing that, it is\\n> not good to keep around that that has never been stressed and/or\\n> used. Upon review, the cleanup done looks correct, as we have never\\n> been able to generate .dat.gz files in for a dump in the tar format.\\n\\nCorrect. My driving force behind it was to ease up the cleanup/refactoring\\nwork that follows, by eliminating the callers of the GZ*() macros.\\n\\n> + command_fails_like(\\n>\\n> + [ 'pg_dump', '--compress', '1', '--format', 'tar' ],\\n> This addition depending on HAVE_LIBZ is a good thing as a reminder of\\n> any work that could be done in 0002. Now that's waiting for 20 years\\n> so I would not hold my breath on this support. I think that this\\n> could be just applied first, with 0002 on top of it, as a first\\n> improvement.\\n\\nExcellent, thank you.\\n\\n> + compress_cmd => [\\n> + $ENV{'GZIP_PROGRAM'},\\n> Patch 0001 is missing and update of pg_dump's Makefile to pass down\\n> this environment variable to the test scripts, no?\\n\\nAgreed. It was not properly moved forward. Fixed.\\n\\n> + compress_cmd => [\\n> + $ENV{'GZIP_PROGRAM'},\\n> + '-f',\\n> [...]\\n> + $ENV{'GZIP_PROGRAM'},\\n> + '-k', '-d',\\n> -f and -d are available everywhere I looked at, but is -k/--keep a\\n> portable choice with a gzip command? I don't see this option in\\n> OpenBSD, for one. So this test is going to cause problems on those\\n> buildfarm machines, at least. Couldn't this part be replaced by a\\n> simple --test to check that what has been compressed is in correct\\n> shape? We know that this works, based on our recent experiences with\\n> the other tests.\\n\\nI would argue that the simple '--test' will not do in this case, as the\\nTAP tests do need a file named <test>.sql to compare the contents with.\\nThis file is generated either directly by pg_dump itself, or by running\\npg_restore on pg_dump's output. In the case of compression pg_dump will\\ngenerate a <test>.sql.<compression program suffix> which can not be\\nused in the comparison tests. So the intention of this block, is not to\\nsimply test for validity, but to also decompress pg_dump's output for it\\nto be able to be used.\\n\\nI updated the patch to simply remove the '-k' flag.\\n\\nPlease find v3 attached. (only 0001 and 0002 are relevant, 0003 and\\n0004 are only for reference and are currently under active modification).\\n\\nCheers,\\n//Georgios\",\n \"msg_date\": \"Tue, 29 Mar 2022 09:46:27 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 29, 2022 at 1:03 AM Michael Paquier <michael@paquier.xyz> wrote:\\n> > Then, if people are willing to adopt the syntax that the\\n> > backup_compression.c/h stuff supports as a project standard (+1 from\\n> > me) we can go the other way and rename that stuff to be more generic,\\n> > taking backup out of the name.\\n>\\n> I am not sure about the specification part which is only used by base\\n> backups that has no client-server requirements, so option values would\\n> still require their own grammar.\\n\\nI don't know what you mean by this. I think the specification stuff\\ncould be reused in a lot of places. If you can ask for a base backup\\nwith zstd:level=3,long=1,fancystuff=yes or whatever we end up with,\\nwhy not enable exactly the same for every other place that uses\\ncompression? I don't know what \\\"client-server requirements\\\" is or what\\nthat has to do with this.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 09:14:03 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 29, 2022 at 09:46:27AM +0000, gkokolatos@pm.me wrote:\\n> On Tuesday, March 29th, 2022 at 9:27 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n>> On Sat, Mar 26, 2022 at 01:14:41AM -0500, Justin Pryzby wrote:\\n>> Wow. This stuff is old enough to vote (c3e18804), dead since its\\n>> introduction. There is indeed an argument for removing that, it is\\n>> not good to keep around that that has never been stressed and/or\\n>> used. Upon review, the cleanup done looks correct, as we have never\\n>> been able to generate .dat.gz files in for a dump in the tar format.\\n> \\n> Correct. My driving force behind it was to ease up the cleanup/refactoring\\n> work that follows, by eliminating the callers of the GZ*() macros.\\n\\nMakes sense to me.\\n\\n>> + command_fails_like(\\n>>\\n>> + [ 'pg_dump', '--compress', '1', '--format', 'tar' ],\\n>> This addition depending on HAVE_LIBZ is a good thing as a reminder of\\n>> any work that could be done in 0002. Now that's waiting for 20 years\\n>> so I would not hold my breath on this support. I think that this\\n>> could be just applied first, with 0002 on top of it, as a first\\n>> improvement.\\n> \\n> Excellent, thank you.\\n\\nI have applied the test for --compress and --format=tar, separating it\\nfrom the rest.\\n\\nWhile moving on with 0002, I have noticed the following in\\n_StartBlob():\\n if (AH->compression != 0)\\n sfx = \\\".gz\\\";\\n else\\n sfx = \\\"\\\";\\n\\t\\nShouldn't this bit also be simplified, adding a fatal() like the other\\ncode paths, for safety?\\n\\n>> + compress_cmd => [\\n>> + $ENV{'GZIP_PROGRAM'},\\n>> + '-f',\\n>> [...]\\n>> + $ENV{'GZIP_PROGRAM'},\\n>> + '-k', '-d',\\n>> -f and -d are available everywhere I looked at, but is -k/--keep a\\n>> portable choice with a gzip command? I don't see this option in\\n>> OpenBSD, for one. So this test is going to cause problems on those\\n>> buildfarm machines, at least. Couldn't this part be replaced by a\\n>> simple --test to check that what has been compressed is in correct\\n>> shape? We know that this works, based on our recent experiences with\\n>> the other tests.\\n> \\n> I would argue that the simple '--test' will not do in this case, as the\\n> TAP tests do need a file named <test>.sql to compare the contents with.\\n> This file is generated either directly by pg_dump itself, or by running\\n> pg_restore on pg_dump's output. In the case of compression pg_dump will\\n> generate a <test>.sql.<compression program suffix> which can not be\\n> used in the comparison tests. So the intention of this block, is not to\\n> simply test for validity, but to also decompress pg_dump's output for it\\n> to be able to be used.\\n\\nAhh, I see, thanks. I would add a comment about that in the area of\\ncompression_gzip_plain_format.\\n\\n+ my $supports_compression = check_pg_config(\\\"#define HAVE_LIBZ 1\\\");\\n\\nThis part could be moved within the if block a couple of lines down.\\n\\n+ my $compress_program = $ENV{GZIP_PROGRAM};\\n\\nIt seems to me that it is enough to rely on {compress_cmd}, hence\\nthere should be no need for $compress_program, no?\\n\\nIt seems to me that we should have a description for compress_cmd at\\nthe top of 002_pg_dump.pl (close to \\\"Definition of the pg_dump runs to\\nmake\\\"). There is an order dependency with restore_cmd.\\n\\n> I updated the patch to simply remove the '-k' flag.\\n\\nOkay.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 30 Mar 2022 14:54:34 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 29, 2022 at 09:14:03AM -0400, Robert Haas wrote:\\n> I don't know what you mean by this. I think the specification stuff\\n> could be reused in a lot of places. If you can ask for a base backup\\n> with zstd:level=3,long=1,fancystuff=yes or whatever we end up with,\\n> why not enable exactly the same for every other place that uses\\n> compression? I don't know what \\\"client-server requirements\\\" is or what\\n> that has to do with this.\\n\\nOh. I think that I got confused here. I saw the backup component in\\nthe file name and this has been associated with the client/server\\nchoice that can be done in the options of pg_basebackup. But\\nparse_bc_specification() does not include any knowledge about that:\\npg_basebackup does this job in parse_compress_options(). I agree that\\nit looks possible to reuse that stuff in more places than just base\\nbackups.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 30 Mar 2022 15:36:02 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\n\\nOn Wednesday, March 30th, 2022 at 7:54 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Tue, Mar 29, 2022 at 09:46:27AM +0000, gkokolatos@pm.me wrote:\\n> > On Tuesday, March 29th, 2022 at 9:27 AM, Michael Paquier michael@paquier.xyz wrote:\\n> > > On Sat, Mar 26, 2022 at 01:14:41AM -0500, Justin Pryzby wrote:\\n> > > + command_fails_like(\\n> > > + [ 'pg_dump', '--compress', '1', '--format', 'tar' ],\\n> > > This addition depending on HAVE_LIBZ is a good thing as a reminder of\\n> > > any work that could be done in 0002. Now that's waiting for 20 years\\n> > > so I would not hold my breath on this support. I think that this\\n> > > could be just applied first, with 0002 on top of it, as a first\\n> > > improvement.\\n> >\\n> > Excellent, thank you.\\n>\\n> I have applied the test for --compress and --format=tar, separating it\\n> from the rest.\\n\\nThank you.\\n\\n> While moving on with 0002, I have noticed the following in\\n>\\n> _StartBlob():\\n> if (AH->compression != 0)\\n> sfx = \\\".gz\\\";\\n> else\\n> sfx = \\\"\\\";\\n>\\n> Shouldn't this bit also be simplified, adding a fatal() like the other\\n> code paths, for safety?\\n\\nAgreed. Fixed.\\n\\n> > > + compress_cmd => [\\n> > > + $ENV{'GZIP_PROGRAM'},\\n> > > + '-f',\\n> > > [...]\\n> > > + $ENV{'GZIP_PROGRAM'},\\n> > > + '-k', '-d',\\n> > > -f and -d are available everywhere I looked at, but is -k/--keep a\\n> > > portable choice with a gzip command? I don't see this option in\\n> > > OpenBSD, for one. So this test is going to cause problems on those\\n> > > buildfarm machines, at least. Couldn't this part be replaced by a\\n> > > simple --test to check that what has been compressed is in correct\\n> > > shape? We know that this works, based on our recent experiences with\\n> > > the other tests.\\n> >\\n> > I would argue that the simple '--test' will not do in this case, as the\\n> > TAP tests do need a file named <test>.sql to compare the contents with.\\n> > This file is generated either directly by pg_dump itself, or by running\\n> > pg_restore on pg_dump's output. In the case of compression pg_dump will\\n> > generate a <test>.sql.<compression program suffix> which can not be\\n> > used in the comparison tests. So the intention of this block, is not to\\n> > simply test for validity, but to also decompress pg_dump's output for it\\n> > to be able to be used.\\n>\\n> Ahh, I see, thanks. I would add a comment about that in the area of\\n> compression_gzip_plain_format.\\n\\nAgreed. Comment added.\\n\\n> + my $supports_compression = check_pg_config(\\\"#define HAVE_LIBZ 1\\\");\\n>\\n> This part could be moved within the if block a couple of lines down.\\n\\nI moved it instead out of the for loop above to not have to call it on\\neach iteration.\\n\\n> + my $compress_program = $ENV{GZIP_PROGRAM};\\n>\\n> It seems to me that it is enough to rely on {compress_cmd}, hence\\n> there should be no need for $compress_program, no?\\n\\nMaybe not. We don't want to the tests to fail if the utility is not\\ninstalled. That becomes even more evident as more methods are added.\\nHowever I realized that the presence of the environmental variable does\\nnot guarrantee that the utility is actually installed. In the attached,\\nthe existance of the utility is based on the return value of system_log().\\n\\n> It seems to me that we should have a description for compress_cmd at\\n> the top of 002_pg_dump.pl (close to \\\"Definition of the pg_dump runs to\\n> make\\\"). There is an order dependency with restore_cmd.\\n\\nAgreed. Comment added.\\n\\nCheers,\\n//Georgios\",\n \"msg_date\": \"Wed, 30 Mar 2022 15:32:55 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2022 at 03:32:55PM +0000, gkokolatos@pm.me wrote:\\n> On Wednesday, March 30th, 2022 at 7:54 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n>> While moving on with 0002, I have noticed the following in\\n>>\\n>> _StartBlob():\\n>> if (AH->compression != 0)\\n>> sfx = \\\".gz\\\";\\n>> else\\n>> sfx = \\\"\\\";\\n>>\\n>> Shouldn't this bit also be simplified, adding a fatal() like the other\\n>> code paths, for safety?\\n> \\n> Agreed. Fixed.\\n\\nOkay. 0002 looks fine as-is, and I don't mind the extra fatal()\\ncalls. These could be asserts but that's not a big deal one way or\\nthe other. And the cleanup is now applied.\\n\\n>> + my $compress_program = $ENV{GZIP_PROGRAM};\\n>>\\n>> It seems to me that it is enough to rely on {compress_cmd}, hence\\n>> there should be no need for $compress_program, no?\\n> \\n> Maybe not. We don't want to the tests to fail if the utility is not\\n> installed. That becomes even more evident as more methods are added.\\n> However I realized that the presence of the environmental variable does\\n> not guarrantee that the utility is actually installed. In the attached,\\n> the existance of the utility is based on the return value of system_log().\\n\\nHmm. [.. thinks ..] The thing that's itching me here is that you\\nalign the concept of compression with gzip, but that's not going to be\\ntrue once more compression options are added to pg_dump, and that\\nwould make $supports_compression and $compress_program_exists\\nincorrect. Perhaps the right answer would be to rename all that with\\na suffix like \\\"_gzip\\\" to make a difference? Or would there be enough\\ncontrol with a value of \\\"compression_gzip\\\" instead of \\\"compression\\\" in\\ntest_key?\\n\\n+my $compress_program_exists = (system_log(\\\"$ENV{GZIP_PROGRAM}\\\", '-h',\\n+ '>', '/dev/null') == 0);\\nDo we need this command execution at all? In all the other tests, we\\nrely on a simple \\\"if (!defined $gzip || $gzip eq '');\\\", so we could do\\nthe same here.\\n\\nA last thing is that we should perhaps make a clear difference between\\nthe check that looks at if the code has been built with zlib and the\\ncheck for the presence of GZIP_PROGRAM, as it can be useful in some\\nenvironments to be able to run pg_dump built with zlib, even if the\\nGZIP_PROGRAM command does not exist (I don't think this can be the\\ncase, but other tests are flexible). As of now, the patch relies on\\npg_dump enforcing uncompression if building under --without-zlib even\\nif --compress/-Z is used, but that also means that those compression\\ntests overlap with the existing tests in this case. Wouldn't it be\\nmore consistent to check after $supports_compression when executing\\nthe dump command for test_key = \\\"compression[_gzip]\\\"? This would mean\\nkeeping GZIP_PROGRAM as sole check when executing the compression\\ncommand.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 31 Mar 2022 11:34:52 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thursday, March 31st, 2022 at 4:34 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n> On Wed, Mar 30, 2022 at 03:32:55PM +0000, gkokolatos@pm.me wrote:\\n> > On Wednesday, March 30th, 2022 at 7:54 AM, Michael Paquier michael@paquier.xyz wrote:\\n>\\n> Okay. 0002 looks fine as-is, and I don't mind the extra fatal()\\n> calls. These could be asserts but that's not a big deal one way or\\n> the other. And the cleanup is now applied.\\n\\nThank you very much.\\n\\n> > > + my $compress_program = $ENV{GZIP_PROGRAM};\\n> > > It seems to me that it is enough to rely on {compress_cmd}, hence\\n> > > there should be no need for $compress_program, no?\\n> >\\n> > Maybe not. We don't want to the tests to fail if the utility is not\\n> > installed. That becomes even more evident as more methods are added.\\n> > However I realized that the presence of the environmental variable does\\n> > not guarrantee that the utility is actually installed. In the attached,\\n> > the existance of the utility is based on the return value of system_log().\\n>\\n> Hmm. [.. thinks ..] The thing that's itching me here is that you\\n> align the concept of compression with gzip, but that's not going to be\\n> true once more compression options are added to pg_dump, and that\\n> would make $supports_compression and $compress_program_exists\\n> incorrect. Perhaps the right answer would be to rename all that with\\n> a suffix like \\\"_gzip\\\" to make a difference? Or would there be enough\\n> control with a value of \\\"compression_gzip\\\" instead of \\\"compression\\\" in\\n> test_key?\\n\\nI understand the itch. Indeed when LZ4 is added as compression method, this\\nblock changes slightly. I went with the minimum amount changed. Please find\\nin 0001 of the attached this variable renamed as $gzip_program_exist. I thought\\nthat as prefix it will match better the already used $ENV{GZIP_PROGRAM}.\\n\\n> +my $compress_program_exists = (system_log(\\\"$ENV{GZIP_PROGRAM}\\\", '-h',\\n> + '>', '/dev/null') == 0);\\n>\\n> Do we need this command execution at all? In all the other tests, we\\n> rely on a simple \\\"if (!defined $gzip || $gzip eq '');\\\", so we could do\\n> the same here.\\n\\nYou are very correct that we are using the simple version, and that is what\\nit was included in the previous versions of the current patch. However, I\\ndid notice that the variable is hard-coded in Makefile.global.in and it does\\nnot go through configure. By now, gzip is considered an essential package\\nin most installations, and this hard-code makes sense. Though I did remove\\nthe utility from my system, (apt remove gzip) and tried the test with the\\nsimple \\\"if (!defined $gzip || $gzip eq '');\\\", which predictably failed. For\\nthis, I went with the system call, it is not too expensive and is rather\\nreliable.\\n\\nIt is true that the rest of the TAP tests that use this, e.g. in pg_basebackup,\\nalso failed. There is an argument to go simple and I will be happy to revert\\nto the previous version.\\n\\n> A last thing is that we should perhaps make a clear difference between\\n> the check that looks at if the code has been built with zlib and the\\n> check for the presence of GZIP_PROGRAM, as it can be useful in some\\n> environments to be able to run pg_dump built with zlib, even if the\\n> GZIP_PROGRAM command does not exist (I don't think this can be the\\n> case, but other tests are flexible).\\n\\nYou are very correct. We do that already in the current patch. Note that we skip\\nthe test only when we specifically have to execute a compression command. Not\\nall compression tests define such command, exactly so that we can test those\\ncases as well. The point of using an external utility program is in order to\\nextend the coverage in previously untested yet supported scenarios, e.g. manual\\ncompression of the *.toc files.\\n\\nAlso in the case where it will actually skip the compression command because the\\ngzip program is not present, it will execute the pg_dump command first.\\n\\n> As of now, the patch relies on\\n> pg_dump enforcing uncompression if building under --without-zlib even\\n> if --compress/-Z is used, but that also means that those compression\\n> tests overlap with the existing tests in this case. Wouldn't it be\\n> more consistent to check after $supports_compression when executing\\n> the dump command for test_key = \\\"compression[_gzip]\\\"? This would mean\\n> keeping GZIP_PROGRAM as sole check when executing the compression\\n> command.\\n\\nI can see the overlap case. Yet, I understand the test_key as serving different\\npurpose, as it is a key of %tests and %full_runs. I do not expect the database\\ncontent of the generated dump to change based on which compression method is used.\\n\\nIn the next round, I can see one explitcly requesting --compress=none to override\\ndefaults. There is a benefit to group the tests for this scenario under the same\\ntest_key, i.e. compression.\\n\\nAlso there will be cases where if the program exists, yet the codebase is compiled\\nwithout support for the method. Then compress_cmd or the restore_cmd that follows\\nwill fail. For example, in the plain output, if we try to uncompress the generated\\nthe test will fail with 'gzip: <filename> not in gzip format'. In the directory\\nformat the compress_cmd will compress the *.toc files, but the restore_cmd will\\nfail because it does not build with support for them.\\n\\nIn the attached version, I propose that the compression_cmd is converted into\\na hash. It contains two keys, the program and the arguments. Maybe it is easier\\nto read than before or than simply grabbing the first element of the array.\\n\\nCheers,\\n//Georgios\",\n \"msg_date\": \"Fri, 01 Apr 2022 15:06:40 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 01, 2022 at 03:06:40PM +0000, gkokolatos@pm.me wrote:\\n> I understand the itch. Indeed when LZ4 is added as compression method, this\\n> block changes slightly. I went with the minimum amount changed. Please find\\n> in 0001 of the attached this variable renamed as $gzip_program_exist. I thought\\n> that as prefix it will match better the already used $ENV{GZIP_PROGRAM}.\\n\\nHmm. I have spent some time on that, and upon review I really think\\nthat we should skip the tests marked as dedicated to the gzip\\ncompression entirely if the build is not compiled with this option,\\nrather than letting the code run a dump for nothing in some cases,\\nrelying on the default to uncompress the contents in others. In the\\nlatter case, it happens that we have already some checks like\\ndefaults_custom_format, but you already mentioned that.\\n\\nWe should also skip the later parts of the tests if the compression\\nprogram does not exist as we rely on it, but only if the command does\\nnot exist. This will count for LZ4.\\n\\n> I can see the overlap case. Yet, I understand the test_key as serving different\\n> purpose, as it is a key of %tests and %full_runs. I do not expect the database\\n> content of the generated dump to change based on which compression method is used.\\n\\nContrary to the current LZ4 tests in pg_dump, what we have here is a\\ncheck for a command-level run and not a data-level check. So what's\\nintroduced is a new concept, and we need a new way to control if the\\ntests should be entirely skipped or not, particularly if we finish by\\nnot using test_key to make the difference. Perhaps the best way to\\naddress that is to have a new keyword in the $runs structure. The\\nattached defines a new compile_option, that can be completed later for\\nnew compression methods introduced in the tests. So the idea is to\\nmark all the tests related to compression with the same test_key, and\\nthe tests can be skipped depending on what compile_option requires.\\n\\n> In the attached version, I propose that the compression_cmd is converted into\\n> a hash. It contains two keys, the program and the arguments. Maybe it is easier\\n> to read than before or than simply grabbing the first element of the array.\\n\\nSplitting the program and its arguments makes sense.\\n\\nAt the end I am finishing with the attached. I also saw an overlap\\nwith the addition of --jobs for the directory format vs not using the\\noption, so I have removed the case where --jobs was not used in the\\ndirectory format.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 5 Apr 2022 10:34:49 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n------- Original Message -------\\n\\nOn Tuesday, April 5th, 2022 at 3:34 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n> On Fri, Apr 01, 2022 at 03:06:40PM +0000, gkokolatos@pm.me wrote:\\n\\n> Splitting the program and its arguments makes sense.\\n\\nGreat.\\n\\n> At the end I am finishing with the attached. I also saw an overlap\\n> with the addition of --jobs for the directory format vs not using the\\n> option, so I have removed the case where --jobs was not used in the\\n> directory format.\\n\\nThank you. I agree with the attached and I will carry it forward to the\\nrest of the patchset.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\\n\\n\\n\",\n \"msg_date\": \"Tue, 05 Apr 2022 07:13:35 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 05, 2022 at 07:13:35AM +0000, gkokolatos@pm.me wrote:\\n> Thank you. I agree with the attached and I will carry it forward to the\\n> rest of the patchset.\\n\\nNo need to carry it forward anymore, I think ;)\\n--\\nMichael\",\n \"msg_date\": \"Tue, 5 Apr 2022 19:55:05 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n------- Original Message -------\\n\\nOn Tuesday, April 5th, 2022 at 12:55 PM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On Tue, Apr 05, 2022 at 07:13:35AM +0000, gkokolatos@pm.me wrote:\\n> No need to carry it forward anymore, I think ;)\\n\\nThank you for committing!\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\\n\\n\\n\",\n \"msg_date\": \"Tue, 05 Apr 2022 11:12:28 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nWill you be able to send a rebased patch for the next CF ?\\n\\nIf you update for the review comments I sent in March, I'll plan to do another\\nround of review.\\n\\nOn Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\\n> LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\\n> \\n> I ran into that on an ubuntu LTS, so I don't think it's so old that it\\n> shouldn't be handled more gracefully. LZ4 should either have an explicit\\n> version check, or else shouldn't depend on that feature (or should define a\\n> safe fallback version if the library header doesn't define it).\\n> \\n> https://packages.ubuntu.com/liblz4-1\\n> \\n> 0003: typo: of legacy => or legacy\\n> \\n> There are a large number of ifdefs being added here - it'd be nice to minimize\\n> that. basebackup was organized to use separate files, which is one way.\\n> \\n> $ git grep -c 'ifdef .*LZ4' src/bin/pg_dump/compress_io.c\\n> src/bin/pg_dump/compress_io.c:19\\n> \\n> In last year's CF entry, I had made a union within CompressorState. LZ4\\n> doesn't need z_streamp (and ztsd will need ZSTD_outBuffer, ZSTD_inBuffer,\\n> ZSTD_CStream).\\n> \\n> 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\\n> from commit ffd53659c. You're passing both the compression method *and* level.\\n> I think there should be a structure which includes both. In the future, that\\n> can also handle additional options. I hope to re-use these same things for\\n> wal_compression=method:level.\\n> \\n> You renamed this:\\n> \\n> |- COMPR_ALG_LIBZ\\n> |-} CompressionAlgorithm;\\n> |+ COMPRESSION_GZIP,\\n> |+} CompressionMethod;\\n> \\n> ..But I don't think that's an improvement. If you were to change it, it should\\n> say something like PGDUMP_COMPRESS_ZLIB, since there are other compression\\n> structs and typedefs. zlib is not idential to gzip, which uses a different\\n> header, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\\n> \\n> The cf* changes in pg_backup_archiver could be split out into a separate\\n> commit. It's strictly a code simplification - not just preparation for more\\n> compression algorithms. The commit message should \\\"See also:\\n> bf9aa490db24b2334b3595ee33653bf2fe39208c\\\".\\n> \\n> The changes in 0002 for cfopen_write seem insufficient:\\n> |+ if (compressionMethod == COMPRESSION_NONE)\\n> |+ fp = cfopen(path, mode, compressionMethod, 0);\\n> | else\\n> | {\\n> | #ifdef HAVE_LIBZ\\n> | char *fname;\\n> | \\n> | fname = psprintf(\\\"%s.gz\\\", path);\\n> |- fp = cfopen(fname, mode, compression);\\n> |+ fp = cfopen(fname, mode, compressionMethod, compressionLevel);\\n> | free_keep_errno(fname);\\n> | #else\\n> \\n> The only difference between the LIBZ and uncompressed case is the file\\n> extension, and it'll be the only difference with LZ4 too. So I suggest to\\n> first handle the file extension, and the rest of the code path is not\\n> conditional on the compression method. I don't think cfopen_write even needs\\n> HAVE_LIBZ - can't you handle that in cfopen_internal() ?\\n> \\n> This patch rejects -Z0, which ought to be accepted:\\n> ./src/bin/pg_dump/pg_dump -h /tmp regression -Fc -Z0 |wc\\n> pg_dump: error: can only specify -Z/--compress [LEVEL] when method is set\\n> \\n> Your 0003 patch shouldn't reference LZ4:\\n> +#ifndef HAVE_LIBLZ4\\n> + if (*compressionMethod == COMPRESSION_LZ4)\\n> + supports_compression = false;\\n> +#endif\\n> \\n> The 0004 patch renames zlibOutSize to outsize - I think the patch series should\\n> be constructed such as to minimize the size of the method-specific patches. I\\n> say this anticipating also adding support for zstd. The preliminary patches\\n> should have all the boring stuff. It would help for reviewing to keep the \\n> patches split up, or to enumerate all the boring things that are being renamed\\n> (like change OutputContext to cfp, rename zlibOutSize, ...).\\n> \\n> 0004: The include should use <lz4.h> and not \\\"lz4.h\\\"\\n> \\n> freebsd/cfbot is failing.\\n> \\n> I suggested off-list to add an 0099 patch to change LZ4 to the default, to\\n> exercise it more on CI.\\n\\nOn Sat, Mar 26, 2022 at 01:33:36PM -0500, Justin Pryzby wrote:\\n> On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\\n> > You're passing both the compression method *and* level. I think there should\\n> > be a structure which includes both. In the future, that can also handle\\n> > additional options.\\n> \\n> I'm not sure if there's anything worth saving, but I did that last year with\\n> 0003-Support-multiple-compression-algs-levels-opts.patch\\n> I sent a rebased copy off-list.\\n> https://www.postgresql.org/message-id/flat/20210104025321.GA9712@telsasoft.com#ca1b9f9d3552c87fa874731cad9d8391\\n> \\n> |\\tfatal(\\\"not built with LZ4 support\\\");\\n> |\\tfatal(\\\"not built with lz4 support\\\");\\n> \\n> Please use consistent capitalization of \\\"lz4\\\" - then the compiler can optimize\\n> away duplicate strings.\\n> \\n> > 0004: The include should use <lz4.h> and not \\\"lz4.h\\\"\\n> \\n> Also, use USE_LZ4 rather than HAVE_LIBLZ4, per 75eae0908.\\n\\n\\n\",\n \"msg_date\": \"Sun, 26 Jun 2022 10:55:32 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Sunday, June 26th, 2022 at 5:55 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n>\\n>\\n> Hi,\\n>\\n> Will you be able to send a rebased patch for the next CF ?\\n\\nThank you for taking an interest in the PR. The plan is indeed to sent\\na new version.\\n\\n> If you update for the review comments I sent in March, I'll plan to do another\\n> round of review.\\n\\nThank you.\\n\\n>\\n> On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\\n>\\n> > LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\\n> >\\n> > I ran into that on an ubuntu LTS, so I don't think it's so old that it\\n> > shouldn't be handled more gracefully. LZ4 should either have an explicit\\n> > version check, or else shouldn't depend on that feature (or should define a\\n> > safe fallback version if the library header doesn't define it).\\n> >\\n> > https://packages.ubuntu.com/liblz4-1\\n> >\\n> > 0003: typo: of legacy => or legacy\\n> >\\n> > There are a large number of ifdefs being added here - it'd be nice to minimize\\n> > that. basebackup was organized to use separate files, which is one way.\\n> >\\n> > $ git grep -c 'ifdef .*LZ4' src/bin/pg_dump/compress_io.c\\n> > src/bin/pg_dump/compress_io.c:19\\n> >\\n> > In last year's CF entry, I had made a union within CompressorState. LZ4\\n> > doesn't need z_streamp (and ztsd will need ZSTD_outBuffer, ZSTD_inBuffer,\\n> > ZSTD_CStream).\\n> >\\n> > 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\\n> > from commit ffd53659c. You're passing both the compression method and level.\\n> > I think there should be a structure which includes both. In the future, that\\n> > can also handle additional options. I hope to re-use these same things for\\n> > wal_compression=method:level.\\n> >\\n> > You renamed this:\\n> >\\n> > |- COMPR_ALG_LIBZ\\n> > |-} CompressionAlgorithm;\\n> > |+ COMPRESSION_GZIP,\\n> > |+} CompressionMethod;\\n> >\\n> > ..But I don't think that's an improvement. If you were to change it, it should\\n> > say something like PGDUMP_COMPRESS_ZLIB, since there are other compression\\n> > structs and typedefs. zlib is not idential to gzip, which uses a different\\n> > header, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\\n> >\\n> > The cf* changes in pg_backup_archiver could be split out into a separate\\n> > commit. It's strictly a code simplification - not just preparation for more\\n> > compression algorithms. The commit message should \\\"See also:\\n> > bf9aa490db24b2334b3595ee33653bf2fe39208c\\\".\\n> >\\n> > The changes in 0002 for cfopen_write seem insufficient:\\n> > |+ if (compressionMethod == COMPRESSION_NONE)\\n> > |+ fp = cfopen(path, mode, compressionMethod, 0);\\n> > | else\\n> > | {\\n> > | #ifdef HAVE_LIBZ\\n> > | char *fname;\\n> > |\\n> > | fname = psprintf(\\\"%s.gz\\\", path);\\n> > |- fp = cfopen(fname, mode, compression);\\n> > |+ fp = cfopen(fname, mode, compressionMethod, compressionLevel);\\n> > | free_keep_errno(fname);\\n> > | #else\\n> >\\n> > The only difference between the LIBZ and uncompressed case is the file\\n> > extension, and it'll be the only difference with LZ4 too. So I suggest to\\n> > first handle the file extension, and the rest of the code path is not\\n> > conditional on the compression method. I don't think cfopen_write even needs\\n> > HAVE_LIBZ - can't you handle that in cfopen_internal() ?\\n> >\\n> > This patch rejects -Z0, which ought to be accepted:\\n> > ./src/bin/pg_dump/pg_dump -h /tmp regression -Fc -Z0 |wc\\n> > pg_dump: error: can only specify -Z/--compress [LEVEL] when method is set\\n> >\\n> > Your 0003 patch shouldn't reference LZ4:\\n> > +#ifndef HAVE_LIBLZ4\\n> > + if (*compressionMethod == COMPRESSION_LZ4)\\n> > + supports_compression = false;\\n> > +#endif\\n> >\\n> > The 0004 patch renames zlibOutSize to outsize - I think the patch series should\\n> > be constructed such as to minimize the size of the method-specific patches. I\\n> > say this anticipating also adding support for zstd. The preliminary patches\\n> > should have all the boring stuff. It would help for reviewing to keep the\\n> > patches split up, or to enumerate all the boring things that are being renamed\\n> > (like change OutputContext to cfp, rename zlibOutSize, ...).\\n> >\\n> > 0004: The include should use <lz4.h> and not \\\"lz4.h\\\"\\n> >\\n> > freebsd/cfbot is failing.\\n> >\\n> > I suggested off-list to add an 0099 patch to change LZ4 to the default, to\\n> > exercise it more on CI.\\n>\\n>\\n> On Sat, Mar 26, 2022 at 01:33:36PM -0500, Justin Pryzby wrote:\\n>\\n> > On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\\n> >\\n> > > You're passing both the compression method and level. I think there should\\n> > > be a structure which includes both. In the future, that can also handle\\n> > > additional options.\\n> >\\n> > I'm not sure if there's anything worth saving, but I did that last year with\\n> > 0003-Support-multiple-compression-algs-levels-opts.patch\\n> > I sent a rebased copy off-list.\\n> > https://www.postgresql.org/message-id/flat/20210104025321.GA9712@telsasoft.com#ca1b9f9d3552c87fa874731cad9d8391\\n> >\\n> > | fatal(\\\"not built with LZ4 support\\\");\\n> > | fatal(\\\"not built with lz4 support\\\");\\n> >\\n> > Please use consistent capitalization of \\\"lz4\\\" - then the compiler can optimize\\n> > away duplicate strings.\\n> >\\n> > > 0004: The include should use <lz4.h> and not \\\"lz4.h\\\"\\n> >\\n> > Also, use USE_LZ4 rather than HAVE_LIBLZ4, per 75eae0908.\\n>\\n>\\n\\n\\n\",\n \"msg_date\": \"Tue, 28 Jun 2022 08:14:25 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Sunday, June 26th, 2022 at 5:55 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n>\\n> Hi,\\n>\\n> Will you be able to send a rebased patch for the next CF ?\\n\\nPlease find a rebased and heavily refactored patchset. Since parts of this\\npatchset were already committed, I restarted numbering. I am not certain if\\nthis is the preferred way. This makes alignment with previous comments a bit\\nharder\\n\\n> If you update for the review comments I sent in March, I'll plan to do another\\n> round of review.\\n\\nI have updated for \\\"some\\\" of the comments. This is not an unwillingness to\\nincorporate those specific comments. Simply this patchset had started to divert\\nheavily already based on comments from Mr. Paquier who had already requested for\\nthe APIs to be refactored to use function pointers. This is happening in 0002 of\\nthe patchset. 0001 of the patchset is using the new compression.h under common.\\n\\nThis patchset should be considered a late draft, as commentary, documentation,\\nand some finer details are not yet finalized; because I am expecting the proposed\\nrefactor to receive a wealth of comments. It would be helpful to understand if\\nthe proposed direction is something worth to be worked upon, before moving to the\\nfiner details.\\n\\nFor what is worth, I am the sole author of the current patchset.\\n\\nCheers,\\n//Georgios\\n\\n\\n> On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\\n>\\n> > LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\\n> >\\n> > I ran into that on an ubuntu LTS, so I don't think it's so old that it\\n> > shouldn't be handled more gracefully. LZ4 should either have an explicit\\n> > version check, or else shouldn't depend on that feature (or should define a\\n> > safe fallback version if the library header doesn't define it).\\n> >\\n> > https://packages.ubuntu.com/liblz4-1\\n> >\\n> > 0003: typo: of legacy => or legacy\\n> >\\n> > There are a large number of ifdefs being added here - it'd be nice to minimize\\n> > that. basebackup was organized to use separate files, which is one way.\\n> >\\n> > $ git grep -c 'ifdef .*LZ4' src/bin/pg_dump/compress_io.c\\n> > src/bin/pg_dump/compress_io.c:19\\n> >\\n> > In last year's CF entry, I had made a union within CompressorState. LZ4\\n> > doesn't need z_streamp (and ztsd will need ZSTD_outBuffer, ZSTD_inBuffer,\\n> > ZSTD_CStream).\\n> >\\n> > 0002: I wonder if you're able to re-use any of the basebackup parsing stuff\\n> > from commit ffd53659c. You're passing both the compression method and level.\\n> > I think there should be a structure which includes both. In the future, that\\n> > can also handle additional options. I hope to re-use these same things for\\n> > wal_compression=method:level.\\n> >\\n> > You renamed this:\\n> >\\n> > |- COMPR_ALG_LIBZ\\n> > |-} CompressionAlgorithm;\\n> > |+ COMPRESSION_GZIP,\\n> > |+} CompressionMethod;\\n> >\\n> > ..But I don't think that's an improvement. If you were to change it, it should\\n> > say something like PGDUMP_COMPRESS_ZLIB, since there are other compression\\n> > structs and typedefs. zlib is not idential to gzip, which uses a different\\n> > header, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\\n> >\\n> > The cf* changes in pg_backup_archiver could be split out into a separate\\n> > commit. It's strictly a code simplification - not just preparation for more\\n> > compression algorithms. The commit message should \\\"See also:\\n> > bf9aa490db24b2334b3595ee33653bf2fe39208c\\\".\\n> >\\n> > The changes in 0002 for cfopen_write seem insufficient:\\n> > |+ if (compressionMethod == COMPRESSION_NONE)\\n> > |+ fp = cfopen(path, mode, compressionMethod, 0);\\n> > | else\\n> > | {\\n> > | #ifdef HAVE_LIBZ\\n> > | char *fname;\\n> > |\\n> > | fname = psprintf(\\\"%s.gz\\\", path);\\n> > |- fp = cfopen(fname, mode, compression);\\n> > |+ fp = cfopen(fname, mode, compressionMethod, compressionLevel);\\n> > | free_keep_errno(fname);\\n> > | #else\\n> >\\n> > The only difference between the LIBZ and uncompressed case is the file\\n> > extension, and it'll be the only difference with LZ4 too. So I suggest to\\n> > first handle the file extension, and the rest of the code path is not\\n> > conditional on the compression method. I don't think cfopen_write even needs\\n> > HAVE_LIBZ - can't you handle that in cfopen_internal() ?\\n> >\\n> > This patch rejects -Z0, which ought to be accepted:\\n> > ./src/bin/pg_dump/pg_dump -h /tmp regression -Fc -Z0 |wc\\n> > pg_dump: error: can only specify -Z/--compress [LEVEL] when method is set\\n> >\\n> > Your 0003 patch shouldn't reference LZ4:\\n> > +#ifndef HAVE_LIBLZ4\\n> > + if (*compressionMethod == COMPRESSION_LZ4)\\n> > + supports_compression = false;\\n> > +#endif\\n> >\\n> > The 0004 patch renames zlibOutSize to outsize - I think the patch series should\\n> > be constructed such as to minimize the size of the method-specific patches. I\\n> > say this anticipating also adding support for zstd. The preliminary patches\\n> > should have all the boring stuff. It would help for reviewing to keep the\\n> > patches split up, or to enumerate all the boring things that are being renamed\\n> > (like change OutputContext to cfp, rename zlibOutSize, ...).\\n> >\\n> > 0004: The include should use <lz4.h> and not \\\"lz4.h\\\"\\n> >\\n> > freebsd/cfbot is failing.\\n> >\\n> > I suggested off-list to add an 0099 patch to change LZ4 to the default, to\\n> > exercise it more on CI.\\n>\\n>\\n> On Sat, Mar 26, 2022 at 01:33:36PM -0500, Justin Pryzby wrote:\\n>\\n> > On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\\n> >\\n> > > You're passing both the compression method and level. I think there should\\n> > > be a structure which includes both. In the future, that can also handle\\n> > > additional options.\\n> >\\n> > I'm not sure if there's anything worth saving, but I did that last year with\\n> > 0003-Support-multiple-compression-algs-levels-opts.patch\\n> > I sent a rebased copy off-list.\\n> > https://www.postgresql.org/message-id/flat/20210104025321.GA9712@telsasoft.com#ca1b9f9d3552c87fa874731cad9d8391\\n> >\\n> > | fatal(\\\"not built with LZ4 support\\\");\\n> > | fatal(\\\"not built with lz4 support\\\");\\n> >\\n> > Please use consistent capitalization of \\\"lz4\\\" - then the compiler can optimize\\n> > away duplicate strings.\\n> >\\n> > > 0004: The include should use <lz4.h> and not \\\"lz4.h\\\"\\n> >\\n> > Also, use USE_LZ4 rather than HAVE_LIBLZ4, per 75eae0908.\\n>\\n>\",\n \"msg_date\": \"Tue, 05 Jul 2022 13:22:47 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"This is a review of 0001.\\n\\nOn Tue, Jul 05, 2022 at 01:22:47PM +0000, gkokolatos@pm.me wrote:\\n> Simply this patchset had started to divert\\n> heavily already based on comments from Mr. Paquier who had already requested for\\n> the APIs to be refactored to use function pointers. This is happening in 0002 of\\n> the patchset.\\n\\nI said something about reducing ifdefs, but I'm having trouble finding what\\nMichael said about this ?\\n\\n> > On Sat, Mar 26, 2022 at 11:21:56AM -0500, Justin Pryzby wrote:\\n> >\\n> > > LZ4F_HEADER_SIZE_MAX isn't defined in old LZ4.\\n> > >\\n> > > I ran into that on an ubuntu LTS, so I don't think it's so old that it\\n> > > shouldn't be handled more gracefully. LZ4 should either have an explicit\\n> > > version check, or else shouldn't depend on that feature (or should define a\\n> > > safe fallback version if the library header doesn't define it).\\n> > > https://packages.ubuntu.com/liblz4-1\\n\\nThe constant still seems to be used without defining a fallback or a minimum version.\\n\\n> > > 0003: typo: of legacy => or legacy\\n\\nThis is still there\\n\\n> > > You renamed this:\\n> > >\\n> > > |- COMPR_ALG_LIBZ\\n> > > |-} CompressionAlgorithm;\\n> > > |+ COMPRESSION_GZIP,\\n> > > |+} CompressionMethod;\\n> > >\\n> > > ..But I don't think that's an improvement. If you were to change it, it should\\n> > > say something like PGDUMP_COMPRESS_ZLIB, since there are other compression\\n> > > structs and typedefs. zlib is not idential to gzip, which uses a different\\n> > > header, so in WriteDataToArchive(), LIBZ is correct, and GZIP is incorrect.\\n\\nThis comment still applies - zlib's gz* functions are \\\"gzip\\\" but the others are\\n\\\"zlib\\\". https://zlib.net/manual.html\\n\\nThat affects both the 0001 and 0002 patches.\\n\\nActually, I think that \\\"gzip\\\" should not be the name of the user-facing option,\\nsince (except for \\\"plain\\\" format) it isn't using gzip.\\n\\n+Robert, since this suggests amending parse_compress_algorithm(). Maybe \\\"zlib\\\"\\nshould be parsed the same way as \\\"gzip\\\" - I don't think we ever expose both to\\na user, but in some cases (basebackup and pg_dump -Fp -Z1) the output is \\\"gzip\\\"\\nand in some cases NO it's zlib (pg_dump -Fc -Z1).\\n\\n> > > The cf* changes in pg_backup_archiver could be split out into a separate\\n> > > commit. It's strictly a code simplification - not just preparation for more\\n> > > compression algorithms. The commit message should \\\"See also:\\n> > > bf9aa490db24b2334b3595ee33653bf2fe39208c\\\".\\n\\nI still think this could be an early, 0000 patch.\\n\\n> > > freebsd/cfbot is failing.\\n\\nThis is still failing for bsd, windows and compiler warnings.\\nWindows also has compiler warnings.\\nhttp://cfbot.cputube.org/georgios-kokolatos.html\\n\\nPlease see: src/tools/ci/README, which you can use to run check-world on 4 OS\\nby pushing a branch to github.\\n\\n> > > I suggested off-list to add an 0099 patch to change LZ4 to the default, to\\n> > > exercise it more on CI.\\n\\nWhat about this ? I think the patch needs to pass CI on all 4 OS with\\ndefault=zlib and default=lz4.\\n\\n> > On Sat, Mar 26, 2022 at 01:33:36PM -0500, Justin Pryzby wrote:\\n\\n> @@ -254,7 +251,12 @@ CreateArchive(const char *FileSpec, const ArchiveFormat fmt,\\n> Archive *\\n> OpenArchive(const char *FileSpec, const ArchiveFormat fmt)\\n> {\\n> -\\tArchiveHandle *AH = _allocAH(FileSpec, fmt, 0, true, archModeRead, setupRestoreWorker);\\n> +\\tArchiveHandle *AH;\\n> +\\tpg_compress_specification compress_spec;\\n\\nShould this be initialized to {0} ?\\n\\n> @@ -969,6 +969,8 @@ NewRestoreOptions(void)\\n> \\topts->format = archUnknown;\\n> \\topts->cparams.promptPassword = TRI_DEFAULT;\\n> \\topts->dumpSections = DUMP_UNSECTIONED;\\n> +\\topts->compress_spec.algorithm = PG_COMPRESSION_NONE;\\n> +\\topts->compress_spec.level = INT_MIN;\\n\\nWhy INT_MIN ?\\n\\n> @@ -1115,23 +1117,28 @@ PrintTOCSummary(Archive *AHX)\\n> \\tArchiveHandle *AH = (ArchiveHandle *) AHX;\\n> \\tRestoreOptions *ropt = AH->public.ropt;\\n> \\tTocEntry *te;\\n> +\\tpg_compress_specification out_compress_spec;\\n\\nShould have {0} ?\\nI suggest to write it like my 2020 patch for this, which says:\\nno_compression = {0};\\n\\n> \\t/* Open stdout with no compression for AH output handle */\\n> -\\tAH->gzOut = 0;\\n> -\\tAH->OF = stdout;\\n> +\\tout_compress_spec.algorithm = PG_COMPRESSION_NONE;\\n> +\\tAH->OF = cfdopen(dup(fileno(stdout)), PG_BINARY_A, out_compress_spec);\\n\\nIdeally this should check the success of dup().\\n\\n> @@ -3776,21 +3746,25 @@ ReadHead(ArchiveHandle *AH)\\n> +\\tif (AH->compress_spec.level != INT_MIN)\\n\\nWhy is it testing the level and not the algorithm ?\\n\\n> --- a/src/bin/pg_dump/pg_backup_custom.c\\n> +++ b/src/bin/pg_dump/pg_backup_custom.c\\n> @@ -298,7 +298,7 @@ _StartData(ArchiveHandle *AH, TocEntry *te)\\n> \\t_WriteByte(AH, BLK_DATA);\\t/* Block type */\\n> \\tWriteInt(AH, te->dumpId);\\t/* For sanity check */\\n> \\n> -\\tctx->cs = AllocateCompressor(AH->compression, _CustomWriteFunc);\\n> +\\tctx->cs = AllocateCompressor(AH->compress_spec, _CustomWriteFunc);\\n\\nIs it necessary to rename the data structure ?\\nIf not, this file can remain unchanged.\\n\\n> --- a/src/bin/pg_dump/pg_backup_directory.c\\n> +++ b/src/bin/pg_dump/pg_backup_directory.c\\n> @@ -573,6 +574,7 @@ _CloseArchive(ArchiveHandle *AH)\\n> \\tif (AH->mode == archModeWrite)\\n> \\t{\\n> \\t\\tcfp\\t\\t *tocFH;\\n> +\\t\\tpg_compress_specification compress_spec;\\n\\nShould use {0} ?\\n\\n> @@ -639,12 +642,14 @@ static void\\n> _StartBlobs(ArchiveHandle *AH, TocEntry *te)\\n> {\\n> \\tlclContext *ctx = (lclContext *) AH->formatData;\\n> +\\tpg_compress_specification compress_spec;\\n\\nSame\\n\\n> +\\t/*\\n> +\\t * Custom and directory formats are compressed by default (zlib), others\\n> +\\t * not\\n> +\\t */\\n> +\\tif (user_compression_defined == false)\\n\\nShould be: !user_compression_defined\\n\\nYour 0001+0002 patches (without 0003) fail to compile:\\n\\npg_backup_directory.c: In function ‘_ReadByte’:\\npg_backup_directory.c:519:12: error: ‘CompressFileHandle’ {aka ‘struct CompressFileHandle’} has no member named ‘_IO_getc’\\n 519 | return CFH->getc(CFH);\\n | ^~\\npg_backup_directory.c:520:1: warning: control reaches end of non-void function [-Wreturn-type]\\n 520 | }\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Jul 2022 10:13:28 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 05, 2022 at 01:22:47PM +0000, gkokolatos@pm.me wrote:\\n> I have updated for \\\"some\\\" of the comments. This is not an unwillingness to\\n> incorporate those specific comments. Simply this patchset had started to divert\\n> heavily already based on comments from Mr. Paquier who had already requested for\\n> the APIs to be refactored to use function pointers. This is happening in 0002 of\\n> the patchset. 0001 of the patchset is using the new compression.h under common.\\n> \\n> This patchset should be considered a late draft, as commentary, documentation,\\n> and some finer details are not yet finalized; because I am expecting the proposed\\n> refactor to receive a wealth of comments. It would be helpful to understand if\\n> the proposed direction is something worth to be worked upon, before moving to the\\n> finer details.\\n\\nI have read through the patch set, and I like a lot the separation you\\nare doing here with CompressFileHandle where a compression method has\\nto specify a full set of callbacks depending on the actions that need\\nto be taken. One advantage, as you patch shows, is that you reduce\\nthe dependency of each code path depending on the compression method,\\nwith #ifdefs and such located mostly into their own file structure, so\\nas adding a new compression method becomes really easier. These\\ncallbacks are going to require much more documentation to describe\\nwhat anybody using them should expect from them, and perhaps they\\ncould be renamed in a more generic way as the currect names come from\\nPOSIX (say read_char(), read_string()?), even if this patch has just\\ninherited the names coming from pg_dump itself, but this can be tuned\\nover and over.\\n\\nThe split into three parts as of 0001 to plug into pg_dump the new\\ncompression option set, 0002 to introduce the callbacks and 0003 to\\nadd LZ4, building on the two first parts, makes sense to me. 0001 and\\n0002 could be done in a reversed order as they are mostly independent,\\nthis order is fine as-is.\\n\\nIn short, I am fine with the proposed approach.\\n\\n+#define K_VERS_1_15 MAKE_ARCHIVE_VERSION(1, 15, 0) /* add compressionMethod\\n+ * in header */\\nIndeed, the dump format needs a version bump for this information.\\n\\n+static bool\\n+parse_compression_option(const char *opt,\\n+ pg_compress_specification *compress_spec)\\nThis parsing logic in pg_dump.c looks a lot like what pg_receivewal.c\\ndoes with its parse_compress_options() where, for compatibility:\\n- If only a number is given:\\n-- Assume no compression if level is 0.\\n-- Assume gzip with given compression if level > 0.\\n- If a string is found, assume a full spec, with optionally a level.\\nSo some consolidation could be done between both.\\n\\nBy the way, I can see that GZCLOSE(), etc. are still defined in\\ncompress_io.h but they are not used.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 21 Jul 2022 14:04:18 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"This entry has been waiting on author input for a while (our current\\nthreshold is roughly two weeks), so I've marked it Returned with\\nFeedback.\\n\\nOnce you think the patchset is ready for review again, you (or any\\ninterested party) can resurrect the patch entry by visiting\\n\\n https://commitfest.postgresql.org/38/3571/\\n\\nand changing the status to \\\"Needs Review\\\", and then changing the\\nstatus again to \\\"Move to next CF\\\". (Don't forget the second step;\\nhopefully we will have streamlined this in the near future!)\\n\\nThanks,\\n--Jacob\\n\\n\\n\",\n \"msg_date\": \"Tue, 2 Aug 2022 11:42:03 -0700\",\n \"msg_from\": \"Jacob Champion <jchampion@timescale.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Thank you for your work during commitfest.\\r\\n\\r\\nThe patch is still in development. Given vacation status, expect the next patches to be ready for the November commitfest.\\r\\nFor now it has moved to the September one. Further action will be taken then as needed.\\r\\n\\r\\nEnjoy the rest of the summer!\",\n \"msg_date\": \"Fri, 05 Aug 2022 14:23:45 +0000\",\n \"msg_from\": \"Georgios Kokolatos <gkokolatos@protonmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Checking if you'll be able to submit new patches soon ?\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Nov 2022 08:28:03 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 2, 2022 at 14:28, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n> Checking if you'll be able to submit new patches soon ?\\n\\nThank you for checking up. Expect new versions within this commitfest cycle.\\n On Wed, Nov 2, 2022 at 14:28, Justin Pryzby <pryzby@telsasoft.com> wrote: Checking if you'll be able to submit new patches soon ?Thank you for checking up. Expect new versions within this commitfest cycle.\",\n \"msg_date\": \"Sun, 06 Nov 2022 14:53:12 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Aug 05, 2022 at 02:23:45PM +0000, Georgios Kokolatos wrote:\\n> Thank you for your work during commitfest.\\n> \\n> The patch is still in development. Given vacation status, expect the next patches to be ready for the November commitfest.\\n> For now it has moved to the September one. Further action will be taken then as needed.\\n\\nOn Sun, Nov 06, 2022 at 02:53:12PM +0000, gkokolatos@pm.me wrote:\\n> On Wed, Nov 2, 2022 at 14:28, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > Checking if you'll be able to submit new patches soon ?\\n> \\n> Thank you for checking up. Expect new versions within this commitfest cycle.\\n\\nHi,\\n\\nI think this patch record should be closed for now. You can re-open the\\nexisting patch record once a patch is ready to be reviewed.\\n\\nThe commitfest is a time for committing/reviewing patches that were\\npreviously submitted, but there's no new patch since July. Making a\\npatch available for review at the start of the commitfest seems like a\\nrequirement for current patch records (same as for new patch records).\\n\\nI wrote essentially the same patch as your early patches 2 years ago\\n(before postgres was ready to consider new compression algorithms), so\\nI'm happy to review a new patch when it's available, regardless of its\\nstatus in the cfapp.\\n\\nBTW, some of my own review comments from March weren't addressed.\\nPlease check. Also, in February, I asked if you knew how to use\\ncirrusci to run checks on cirrusci, but the patches still had\\ncompilation errors and warnings on various OS.\\n\\nhttps://cirrus-ci.com/github/postgresql-cfbot/postgresql/commitfest/39/3571\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sun, 20 Nov 2022 11:26:11 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sun, Nov 20, 2022 at 11:26:11AM -0600, Justin Pryzby wrote:\\n> I think this patch record should be closed for now. You can re-open the\\n> existing patch record once a patch is ready to be reviewed.\\n\\nIndeed. As of things are, this is just a dead entry in the CF which\\nwould be confusing. I have marked it as RwF.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 21 Nov 2022 08:13:16 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Monday, November 21st, 2022 at 12:13 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Sun, Nov 20, 2022 at 11:26:11AM -0600, Justin Pryzby wrote:\\n> \\n> > I think this patch record should be closed for now. You can re-open the\\n> > existing patch record once a patch is ready to be reviewed.\\n> \\n> \\n> Indeed. As of things are, this is just a dead entry in the CF which\\n> would be confusing. I have marked it as RwF.\\n\\nThank you for closing it.\\n\\nFor the record I am currently working on it simply unsure if I should submit\\nWIP patches and add noise to the list or wait until it is in a state that I\\nfeel that the comments have been addressed.\\n\\nA new version that I feel that is in a decent enough state for review should\\nbe ready within this week. I am happy to drop the patch if you think I should\\nnot work on it though.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\\n\\n\\n\",\n \"msg_date\": \"Tue, 22 Nov 2022 10:00:47 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 22, 2022 at 10:00:47AM +0000, gkokolatos@pm.me wrote:\\n> A new version that I feel that is in a decent enough state for review should\\n> be ready within this week. I am happy to drop the patch if you think I should\\n> not work on it though.\\n\\nIf you can post a new version of the patch, that's fine, of course.\\nI'll be happy to look over it more.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 22 Nov 2022 19:49:13 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 22, 2022 at 10:00:47AM +0000, gkokolatos@pm.me wrote:\\n> For the record I am currently working on it simply unsure if I should submit\\n> WIP patches and add noise to the list or wait until it is in a state that I\\n> feel that the comments have been addressed.\\n> \\n> A new version that I feel that is in a decent enough state for review should\\n> be ready within this week. I am happy to drop the patch if you think I should\\n> not work on it though.\\n\\nI hope you'll want to continue work on it. The patch record is like a\\nrequest for review, so it's closed if there's nothing ready to review.\\n\\nI think you should re-send patches (and update the CF app) as often as\\nthey're ready for more review. Your 001 commit (which is almost the\\nsame as what I wrote 2 years ago) still needs to account for some review\\ncomments, and the whole patch set ought to pass cirrusci tests. At that\\npoint, you'll be ready for another round of review, even if there's\\nknown TODO/FIXME items in later patches.\\n\\nBTW I saw that you updated your branch on github. You'll need to make\\nthe corresponding changes to ./meson.build that you made to ./Makefile.\\nhttps://wiki.postgresql.org/wiki/Meson_for_patch_authors\\nhttps://wiki.postgresql.org/wiki/Meson \\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Nov 2022 14:52:22 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Tuesday, November 22nd, 2022 at 11:49 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Tue, Nov 22, 2022 at 10:00:47AM +0000, gkokolatos@pm.me wrote:\\n> \\n> > A new version that I feel that is in a decent enough state for review should\\n> > be ready within this week. I am happy to drop the patch if you think I should\\n> > not work on it though.\\n> \\n> \\n> If you can post a new version of the patch, that's fine, of course.\\n> I'll be happy to look over it more.\\n\\nThank you Michael (and Justin). Allow me to present v8.\\n\\nThe focus of this version of this series is 0001 and 0002.\\n\\nAdmittedly 0001 could be presented in a separate thread though given its size and\\nproximity to the topic, I present it here.\\n\\nIn an earlier review you spotted the similarity between pg_dump's and pg_receivewal's\\nparsing of compression options. However there exists a substantial difference in the\\nbehaviour of the two programs; one treats the lack of support for the requested\\nalgorithm as a fatal error, whereas the other does not. The existing functions in\\ncommon/compression.c do not account for the later. 0002 proposes an implementation\\nfor this. It's usefulness is shown in 0003.\\n\\nPlease consider 0003-0005 as work in progress. They are differences from v7 yet they\\nmay contain unaddressed comments for now.\\n\\nA welcome feedback would be in splitting and/or reordering of 0003-0005. I think\\nthat they now split in coherent units and are presented in a logical order. Let me\\nknow if you disagree and where should the breakpoints be.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\",\n \"msg_date\": \"Mon, 28 Nov 2022 16:32:43 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Nov 28, 2022 at 04:32:43PM +0000, gkokolatos@pm.me wrote:\\n> The focus of this version of this series is 0001 and 0002.\\n> \\n> Admittedly 0001 could be presented in a separate thread though given its size and\\n> proximity to the topic, I present it here.\\n\\nI don't mind. This was a hole in meson.build, so nice catch! I have\\nnoticed a second defect with pg_verifybackup for all the commands, and\\napplied both at the same time.\\n\\n> In an earlier review you spotted the similarity between pg_dump's and pg_receivewal's\\n> parsing of compression options. However there exists a substantial difference in the\\n> behaviour of the two programs; one treats the lack of support for the requested\\n> algorithm as a fatal error, whereas the other does not. The existing functions in\\n> common/compression.c do not account for the later. 0002 proposes an implementation\\n> for this. It's usefulness is shown in 0003.\\n\\nIn what does it matter? The logic in compression.c provides an error\\nwhen looking at a spec or validating it, but the caller is free to\\nconsume it as it wants because this is shared between the frontend and\\nthe backend, and that includes consuming it as a warning rather than a\\nahrd failure. If we don't want to issue an error and force\\nnon-compression if attempting to use a compression method not\\nsupported in pg_dump, that's fine by me as a historical behavior, but\\nI don't see why these routines have any need to be split more as\\nproposed in 0002.\\n\\nSaying that, I do agree that it would be nice to remove the\\nduplication between the option parsing of pg_basebackup and\\npg_receivewal. Your patch is very close to that, actually, and it\\noccured to me that if we move the check on \\\"server-\\\" and \\\"client-\\\" in\\npg_basebackup to be just before the integer-only check then we can\\nconsolidate the whole thing.\\n\\nAttached is an alternative that does not sacrifice the pluggability of\\nthe existing routines while allowing 0003~ to still use them (I don't\\nreally want to move around the checks on the supported build options\\nnow in parse_compress_specification(), that was hard enough to settle\\non this location). On top of that, pg_basebackup is able to cope with\\nthe case of --compress=0 already, enforcing \\\"none\\\" (BaseBackup could\\nbe simplified a bit more before StartLogStreamer). This refactoring\\nshaves a little bit of code.\\n\\n> Please consider 0003-0005 as work in progress. They are differences from v7 yet they\\n> may contain unaddressed comments for now.\\n\\nOkay.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 29 Nov 2022 15:19:17 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 29, 2022 at 03:19:17PM +0900, Michael Paquier wrote:\\n> Attached is an alternative that does not sacrifice the pluggability of\\n> the existing routines while allowing 0003~ to still use them (I don't\\n> really want to move around the checks on the supported build options\\n> now in parse_compress_specification(), that was hard enough to settle\\n> on this location). On top of that, pg_basebackup is able to cope with\\n> the case of --compress=0 already, enforcing \\\"none\\\" (BaseBackup could\\n> be simplified a bit more before StartLogStreamer). This refactoring\\n> shaves a little bit of code.\\n\\nOne thing that I forgot to mention is that this refactoring would\\ntreat things like server-N, client-N as valid grammars (in this case N\\ntakes precedence over an optional detail string), implying that N = 0\\nis \\\"none\\\" and N > 0 is gzip, so that makes for an extra grammar flavor\\nwithout impacting the existing ones. I am not sure that it is worth\\ndocumenting, still worth mentioning.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 29 Nov 2022 16:08:27 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Tuesday, November 29th, 2022 at 7:19 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Mon, Nov 28, 2022 at 04:32:43PM +0000, gkokolatos@pm.me wrote:\\n> \\n> > The focus of this version of this series is 0001 and 0002.\\n> > \\n> > Admittedly 0001 could be presented in a separate thread though given its size and\\n> > proximity to the topic, I present it here.\\n> \\n> \\n> I don't mind. This was a hole in meson.build, so nice catch! I have\\n> noticed a second defect with pg_verifybackup for all the commands, and\\n> applied both at the same time.\\n\\nThank you.\\n\\n> \\n> > In an earlier review you spotted the similarity between pg_dump's and pg_receivewal's\\n> > parsing of compression options. However there exists a substantial difference in the\\n> > behaviour of the two programs; one treats the lack of support for the requested\\n> > algorithm as a fatal error, whereas the other does not. The existing functions in\\n> > common/compression.c do not account for the later. 0002 proposes an implementation\\n> > for this. It's usefulness is shown in 0003.\\n> \\n> \\n> In what does it matter? The logic in compression.c provides an error\\n> when looking at a spec or validating it, but the caller is free to\\n> consume it as it wants because this is shared between the frontend and\\n> the backend, and that includes consuming it as a warning rather than a\\n> ahrd failure. If we don't want to issue an error and force\\n> non-compression if attempting to use a compression method not\\n> supported in pg_dump, that's fine by me as a historical behavior, but\\n> I don't see why these routines have any need to be split more as\\n> proposed in 0002.\\n\\nI understand. The reason for the change in the routines was because it was\\nimpossible to distinguish a genuine parse error from a missing library in\\nparse_compress_specification(). If the zlib library is missing, then both\\n'--compress=gzip:garbage' and '--compress=gzip:7' would populate the\\nparse_error member of the struct and subsequent calls to\\nvalidate_compress_specification() would error out, although only one of\\nthe two options is truly an error. Historically the code would fail on\\ninvalid input regardless of whether the library was present or not.\\n\\n> Saying that, I do agree that it would be nice to remove the\\n> duplication between the option parsing of pg_basebackup and\\n> pg_receivewal. Your patch is very close to that, actually, and it\\n> occured to me that if we move the check on \\\"server-\\\" and \\\"client-\\\" in\\n> pg_basebackup to be just before the integer-only check then we can\\n> consolidate the whole thing.\\n\\nGreat. I did notice the possible benefit but chose to not tread too far\\noff the necessary in my patch.\\n \\n> Attached is an alternative that does not sacrifice the pluggability of\\n> the existing routines while allowing 0003~ to still use them (I don't\\n> really want to move around the checks on the supported build options\\n> now in parse_compress_specification(), that was hard enough to settle\\n> on this location).\\n\\nYeah, I thought that it would be a hard sell, hence an \\\"earlier\\\"\\nversion.\\n\\nThe attached version 10, contains verbatim your proposed v9 as 0001.\\nThen 0002 is switching a bit the parsing order in pg_dump and will not\\nfail as described above on missing libraries. Now, it will first parse\\nthe algorithm, discard it when unsupported, and only parse the rest of\\nthe option if the algorithm is supported. Granted it is a bit 'uglier'\\nwith the preprocessing blocks, yet it maintains most of the historic\\nbehaviour without altering the common compression interfaces. Now, as\\nshown in 001_basic.pl, invalid detail will fail only if the algorithm\\nis supported.\\n\\n> On top of that, pg_basebackup is able to cope with\\n> the case of --compress=0 already, enforcing \\\"none\\\" (BaseBackup could\\n> be simplified a bit more before StartLogStreamer). This refactoring\\n> shaves a little bit of code.\\n> \\n> > Please consider 0003-0005 as work in progress. They are differences from v7 yet they\\n> > may contain unaddressed comments for now.\\n> \\n> \\n> Okay.\\n\\nThank you. Please advice if is preferable to split 0002 in two parts.\\nI think not but I will happily do so if you think otherwise.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\",\n \"msg_date\": \"Tue, 29 Nov 2022 12:10:46 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Nov 29, 2022 at 12:10:46PM +0000, gkokolatos@pm.me wrote:\\n> Thank you. Please advice if is preferable to split 0002 in two parts.\\n> I think not but I will happily do so if you think otherwise.\\n\\nThis one makes me curious. What kind of split are you talking about?\\nIf it makes the code review and the git history cleaner and easier, I\\nam usually a lot in favor of such incremental changes. As far as I\\ncan see, there is the switch from the compression integer to\\ncompression specification as one thing. The second thing is the\\nrefactoring of cfclose() and these routines, paving the way for 0003.\\nHmm, it may be cleaner to move the switch to the compression spec in\\none patch, and move the logic around cfclose() to its own, paving the\\nway to 0003.\\n\\nBy the way, I think that this 0002 should drop all the default clauses\\nin the switches for the compression method so as we'd catch any\\nmissing code paths with compiler warnings if a new compression method\\nis added in the future.\\n\\nAnyway, I have applied 0001, adding you as a primary author because\\nyou did most of it with only tweaks from me for pg_basebackup. The\\ndocs of pg_basebackup have been amended to mention the slight change\\nin grammar, affecting the case where we do not have a detail string.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 30 Nov 2022 09:50:34 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Wednesday, November 30th, 2022 at 1:50 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Tue, Nov 29, 2022 at 12:10:46PM +0000, gkokolatos@pm.me wrote:\\n> \\n> > Thank you. Please advice if is preferable to split 0002 in two parts.\\n> > I think not but I will happily do so if you think otherwise.\\n> \\n> \\n> This one makes me curious. What kind of split are you talking about?\\n> If it makes the code review and the git history cleaner and easier, I\\n> am usually a lot in favor of such incremental changes. As far as I\\n> can see, there is the switch from the compression integer to\\n> compression specification as one thing. The second thing is the\\n> refactoring of cfclose() and these routines, paving the way for 0003.\\n> Hmm, it may be cleaner to move the switch to the compression spec in\\n> one patch, and move the logic around cfclose() to its own, paving the\\n> way to 0003.\\n\\nFair enough. The atteched v11 does that. 0001 introduces compression\\nspecification and is using it throughout. 0002 paves the way to the\\nnew interface by homogenizing the use of cfp. 0003 introduces the new\\nAPI and stores the compression algorithm in the custom format header\\ninstead of the compression level integer. Finally 0004 adds support for\\nLZ4.\\n\\nBesides the version bump in 0003 which can possibly be split out and\\nas an independent and earlier step, I think that the patchset consists\\nof coherent units.\\n\\n> By the way, I think that this 0002 should drop all the default clauses\\n> in the switches for the compression method so as we'd catch any\\n> missing code paths with compiler warnings if a new compression method\\n> is added in the future.\\n\\nSure. \\n\\n> Anyway, I have applied 0001, adding you as a primary author because\\n> you did most of it with only tweaks from me for pg_basebackup. The\\n> docs of pg_basebackup have been amended to mention the slight change\\n> in grammar, affecting the case where we do not have a detail string.\\n\\nVery kind of you, thank you.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\",\n \"msg_date\": \"Wed, 30 Nov 2022 17:11:44 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Nov 30, 2022 at 05:11:44PM +0000, gkokolatos@pm.me wrote:\\n> Fair enough. The atteched v11 does that. 0001 introduces compression\\n> specification and is using it throughout. 0002 paves the way to the\\n> new interface by homogenizing the use of cfp. 0003 introduces the new\\n> API and stores the compression algorithm in the custom format header\\n> instead of the compression level integer. Finally 0004 adds support for\\n> LZ4.\\n\\nI have been looking at 0001, and.. Hmm. I am really wondering\\nwhether it would not be better to just nuke this warning into orbit.\\nThis stuff enforces non-compression even if -Z has been used to a\\nnon-default value. This has been moved to its current location by\\ncae2bb1 as of this thread:\\nhttps://www.postgresql.org/message-id/20160526.185551.242041780.horiguchi.kyotaro%40lab.ntt.co.jp\\n\\nHowever, this is only active if -Z is used when not building with\\nzlib. At the end, it comes down to whether we want to prioritize the\\nportability of pg_dump commands specifying a -Z/--compress across\\nenvironments knowing that these may or may not be built with zlib,\\nvs the amount of simplification/uniformity we would get across the\\nbinaries in the tree once we switch everything to use the compression\\nspecifications. Now that pg_basebackup and pg_receivewal are managed\\nby compression specifications, and that we'd want more compression\\noptions for pg_dump, I would tend to do the latter and from now on\\ncomplain if attempting to do a pg_dump -Z under --without-zlib with a\\ncompression level > 0. zlib is also widely available, and we don't\\ndocument the fact that non-compression is enforced in this case,\\neither. (Two TAP tests with the custom format had to be tweaked.)\\n\\nAs per the patch, it is true that we do not need to bump the format of\\nthe dump archives, as we can still store only the compression level\\nand guess the method from it. I have added some notes about that in\\nReadHead and WriteHead to not forget.\\n\\nMost of the changes are really-straight forward, and it has resisted\\nmy tests, so I think that this is in a rather-commitable shape as-is.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 1 Dec 2022 11:05:54 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Thursday, December 1st, 2022 at 3:05 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Wed, Nov 30, 2022 at 05:11:44PM +0000, gkokolatos@pm.me wrote:\\n> \\n> > Fair enough. The atteched v11 does that. 0001 introduces compression\\n> > specification and is using it throughout. 0002 paves the way to the\\n> > new interface by homogenizing the use of cfp. 0003 introduces the new\\n> > API and stores the compression algorithm in the custom format header\\n> > instead of the compression level integer. Finally 0004 adds support for\\n> > LZ4.\\n> \\n> \\n> I have been looking at 0001, and.. Hmm. I am really wondering\\n> whether it would not be better to just nuke this warning into orbit.\\n> This stuff enforces non-compression even if -Z has been used to a\\n> non-default value. This has been moved to its current location by\\n> cae2bb1 as of this thread:\\n> https://www.postgresql.org/message-id/20160526.185551.242041780.horiguchi.kyotaro%40lab.ntt.co.jp\\n> \\n> However, this is only active if -Z is used when not building with\\n> zlib. At the end, it comes down to whether we want to prioritize the\\n> portability of pg_dump commands specifying a -Z/--compress across\\n> environments knowing that these may or may not be built with zlib,\\n> vs the amount of simplification/uniformity we would get across the\\n> binaries in the tree once we switch everything to use the compression\\n> specifications. Now that pg_basebackup and pg_receivewal are managed\\n> by compression specifications, and that we'd want more compression\\n> options for pg_dump, I would tend to do the latter and from now on\\n> complain if attempting to do a pg_dump -Z under --without-zlib with a\\n> compression level > 0. zlib is also widely available, and we don't\\n> document the fact that non-compression is enforced in this case,\\n> either. (Two TAP tests with the custom format had to be tweaked.)\\n\\nFair enough. Thank you for looking. However I have a small comment\\non your new patch.\\n\\n- /* Custom and directory formats are compressed by default, others not */\\n- if (compressLevel == -1)\\n- {\\n-#ifdef HAVE_LIBZ\\n- if (archiveFormat == archCustom || archiveFormat == archDirectory)\\n- compressLevel = Z_DEFAULT_COMPRESSION;\\n- else\\n-#endif\\n- compressLevel = 0;\\n- }\\n\\n\\nNuking the warning from orbit and changing the behaviour around disabling\\nthe requested compression when the libraries are not present, should not\\nmean that we need to change the behaviour of default values for different\\nformats. Please find v13 attached which reinstates it.\\n\\nWhich in itself it got me looking and wondering why the tests succeeded.\\nThe only existing test covering that path is `defaults_dir_format` in\\n`002_pg_dump.pl`. However as the test is currently written it does not\\ncheck whether the output was compressed. The restore command would succeed\\nin either case. A simple `gzip -t -r` against the directory will not\\nsuffice to test it, because there exist files which are never compressed\\nin this format (.toc). A little bit more involved test case would need\\nto be written, yet before I embark to this journey, I would like to know\\nif you would agree to reinstate the defaults for those formats.\\n\\n> \\n> As per the patch, it is true that we do not need to bump the format of\\n> the dump archives, as we can still store only the compression level\\n> and guess the method from it. I have added some notes about that in\\n> ReadHead and WriteHead to not forget.\\n\\nAgreed. A minor suggestion if you may.\\n\\n #ifndef HAVE_LIBZ\\n- if (AH->compression != 0)\\n+ if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n #endif\\n\\nIt would seem a more consistent to error out in this case. We do error\\nin all other cases where the compression is not available.\\n\\n> \\n> Most of the changes are really-straight forward, and it has resisted\\n> my tests, so I think that this is in a rather-commitable shape as-is.\\n\\nThank you.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\",\n \"msg_date\": \"Thu, 01 Dec 2022 14:58:35 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 01, 2022 at 02:58:35PM +0000, gkokolatos@pm.me wrote:\\n> Nuking the warning from orbit and changing the behaviour around disabling\\n> the requested compression when the libraries are not present, should not\\n> mean that we need to change the behaviour of default values for different\\n> formats. Please find v13 attached which reinstates it.\\n\\nGah, thanks! And this default behavior is documented as dependent on\\nthe compilation as well.\\n\\n> Which in itself it got me looking and wondering why the tests succeeded.\\n> The only existing test covering that path is `defaults_dir_format` in\\n> `002_pg_dump.pl`. However as the test is currently written it does not\\n> check whether the output was compressed. The restore command would succeed\\n> in either case. A simple `gzip -t -r` against the directory will not\\n> suffice to test it, because there exist files which are never compressed\\n> in this format (.toc). A little bit more involved test case would need\\n> to be written, yet before I embark to this journey, I would like to know\\n> if you would agree to reinstate the defaults for those formats.\\n\\nOn top of my mind, I briefly recall that -r is not that portable. And\\nthe toc format makes the files generated non-deterministic as these\\nuse OIDs..\\n\\n[.. thinks ..]\\n\\nWe are going to need a new thing here, as compress_cmd cannot be\\ndirectly used. What if we used only an array of glob()-able elements?\\nLet's say \\\"expected_contents\\\" that could include a \\\"dir_path/*.gz\\\"\\nconditional on $supports_gzip? glob() can only be calculated when the\\ntest is run as the file names cannot be known beforehand :/\\n\\n>> As per the patch, it is true that we do not need to bump the format of\\n>> the dump archives, as we can still store only the compression level\\n>> and guess the method from it. I have added some notes about that in\\n>> ReadHead and WriteHead to not forget.\\n> \\n> Agreed. A minor suggestion if you may.\\n> \\n> #ifndef HAVE_LIBZ\\n> - if (AH->compression != 0)\\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n> #endif\\n> \\n> It would seem a more consistent to error out in this case. We do error\\n> in all other cases where the compression is not available.\\n\\nMakes sense.\\n\\nI have gone through the patch again, and applied it. Thanks!\\n--\\nMichael\",\n \"msg_date\": \"Fri, 2 Dec 2022 10:56:26 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Friday, December 2nd, 2022 at 2:56 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n> On top of my mind, I briefly recall that -r is not that portable. And\\n> the toc format makes the files generated non-deterministic as these\\n> use OIDs..\\n> \\n> [.. thinks ..]\\n> \\n> We are going to need a new thing here, as compress_cmd cannot be\\n> directly used. What if we used only an array of glob()-able elements?\\n> Let's say \\\"expected_contents\\\" that could include a \\\"dir_path/*.gz\\\"\\n> conditional on $supports_gzip? glob() can only be calculated when the\\n> test is run as the file names cannot be known beforehand :/\\n\\nYou are very correct. However one can glob after the fact. Please find\\n0001 of the attached v14 which attempts to implement it.\\n\\n> I have gone through the patch again, and applied it. Thanks!\\n\\nThank you. Please find the rest of of the patchset series rebased on top\\nof it. I dare to say that 0002 is in a state worth of your consideration.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\",\n \"msg_date\": \"Fri, 02 Dec 2022 16:15:10 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Dec 02, 2022 at 04:15:10PM +0000, gkokolatos@pm.me wrote:\\n> You are very correct. However one can glob after the fact. Please find\\n> 0001 of the attached v14 which attempts to implement it.\\n\\n+ if ($pgdump_runs{$run}->{glob_pattern})\\n+ {\\n+ my $glob_pattern = $pgdump_runs{$run}->{glob_pattern};\\n+ my @glob_output = glob($glob_pattern);\\n+ is(scalar(@glob_output) > 0, 1, \\\"glob pattern matched\\\")\\n+ }\\n\\nWhile this is correct in checking that the contents are compressed\\nunder --with-zlib, this also removes the coverage where we make sure\\nthat this command is able to complete under --without-zlib without\\ncompressing any of the table data files. Hence my point from\\nupthread: this test had better not use compile_option, but change\\nglob_pattern depending on if the build uses zlib or not.\\n\\nIn order to check this behavior with defaults_custom_format, perhaps\\nwe could just remove the -Z6 from it or add an extra command for its\\ndefault behavior?\\n--\\nMichael\",\n \"msg_date\": \"Sat, 3 Dec 2022 11:45:30 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, Dec 03, 2022 at 11:45:30AM +0900, Michael Paquier wrote:\\n> While this is correct in checking that the contents are compressed\\n> under --with-zlib, this also removes the coverage where we make sure\\n> that this command is able to complete under --without-zlib without\\n> compressing any of the table data files. Hence my point from\\n> upthread: this test had better not use compile_option, but change\\n> glob_pattern depending on if the build uses zlib or not.\\n\\nIn short, I mean something like the attached. I have named the flag\\ncontent_patterns, and switched it to an array so as we can check that\\ntoc.dat is always uncompression and that the other data files are\\nalways uncompressed.\\n\\n> In order to check this behavior with defaults_custom_format, perhaps\\n> we could just remove the -Z6 from it or add an extra command for its\\n> default behavior?\\n\\nThis is slightly more complicated as there is just one file generated\\nfor the compression and non-compression cases, so I have let that as\\nit is now.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 5 Dec 2022 16:05:23 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Monday, December 5th, 2022 at 8:05 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Sat, Dec 03, 2022 at 11:45:30AM +0900, Michael Paquier wrote:\\n> \\n> > While this is correct in checking that the contents are compressed\\n> > under --with-zlib, this also removes the coverage where we make sure\\n> > that this command is able to complete under --without-zlib without\\n> > compressing any of the table data files. Hence my point from\\n> > upthread: this test had better not use compile_option, but change\\n> > glob_pattern depending on if the build uses zlib or not.\\n> \\n> In short, I mean something like the attached. I have named the flag\\n> content_patterns, and switched it to an array so as we can check that\\n> toc.dat is always uncompression and that the other data files are\\n> always uncompressed.\\n\\nI see. This approach is much better than my proposal, thanks. If you\\nallow me, I find 'content_patterns' to be slightly ambiguous. While is\\ntrue that it refers to the contents of a directory, it is not the\\ncontents of the dump that it is examining. I took the liberty of proposing\\nan alternative name in the attached v16.\\n\\nI also took the liberty of applying the test pattern when it the dump\\nis explicitly compressed.\\n\\n> > In order to check this behavior with defaults_custom_format, perhaps\\n> > we could just remove the -Z6 from it or add an extra command for its\\n> > default behavior?\\n> \\n> This is slightly more complicated as there is just one file generated\\n> for the compression and non-compression cases, so I have let that as\\n> it is now.\\n\\nI was thinking a bit more about this. I think that we can use the list\\nTOC option of pg_restore. This option will first print out the header\\ninfo which contains the compression. Perl utils already support to \\nparse the generated output of a command. Please find an attempt to do\\nso in the attached. The benefits of having some testing for this case\\nbecome a bit more obvious in 0004 of the patchset, when lz4 is\\nintroduced.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\",\n \"msg_date\": \"Mon, 05 Dec 2022 12:48:28 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 05, 2022 at 12:48:28PM +0000, gkokolatos@pm.me wrote:\\n> I also took the liberty of applying the test pattern when it the dump\\n> is explicitly compressed.\\n\\nSticking with glob_patterns is fine by me.\\n\\n> I was thinking a bit more about this. I think that we can use the list\\n> TOC option of pg_restore. This option will first print out the header\\n> info which contains the compression. Perl utils already support to \\n> parse the generated output of a command. Please find an attempt to do\\n> so in the attached. The benefits of having some testing for this case\\n> become a bit more obvious in 0004 of the patchset, when lz4 is\\n> introduced.\\n\\nThis is where the fun is. What you are doing here is more complete,\\nand we would make sure that the custom and data directory would always\\nsee their contents compressed by default. And it would have caught\\nthe bug you mentioned upthread for the custom format.\\n\\nI have kept things as you proposed at the end, added a few comments,\\ndocumented the new command_like and an extra command_like for\\ndefaults_dir_format. Glad to see this addressed, thanks!\\n--\\nMichael\",\n \"msg_date\": \"Tue, 6 Dec 2022 09:22:57 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Tuesday, December 6th, 2022 at 1:22 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Mon, Dec 05, 2022 at 12:48:28PM +0000, gkokolatos@pm.me wrote:\\n> \\n> This is where the fun is. What you are doing here is more complete,\\n> and we would make sure that the custom and data directory would always\\n> see their contents compressed by default. And it would have caught\\n> the bug you mentioned upthread for the custom format.\\n\\nThank you very much Michael.\\n\\n> I have kept things as you proposed at the end, added a few comments,\\n> documented the new command_like and an extra command_like for\\n> defaults_dir_format. Glad to see this addressed, thanks!\\n\\nPlease find attached v17, which builds on top of what is already\\ncommitted. I dare to think 0001 as ready to be reviewed. 0002 is\\nalso complete albeit with some documentation gaps.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\",\n \"msg_date\": \"Tue, 06 Dec 2022 15:52:01 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"001: still refers to \\\"gzip\\\", which is correct for -Fp and -Fd but not\\nfor -Fc, for which it's more correct to say \\\"zlib\\\". That affects the\\nname of the function, structures, comments, etc. I'm not sure if it's\\nan issue to re-use the basebackup compression routines here. Maybe we\\nshould accept \\\"-Zn\\\" for zlib output (-Fc), but reject \\\"gzip:9\\\", which\\nI'm sure some will find confusing, as it does not output. Maybe 001\\nshould be split into a patch to re-use the existing \\\"cfp\\\" interface\\n(which is a clear win), and 002 to re-use the basebackup interfaces for\\nuser input and constants, etc.\\n\\n001 still doesn't compile on freebsd, and 002 doesn't compile on\\nwindows. Have you checked test results from cirrusci on your private\\ngithub account ?\\n\\n002 says:\\n+ save_errno = errno; \\n+ errno = save_errno; \\n\\nI suppose that's intended to wrap the preceding library call.\\n\\n002 breaks \\\"pg_dump -Fc -Z2\\\" because (I think) AllocateCompressor()\\ndoesn't store the passed-in compression_spec.\\n\\n003 still uses <lz4.h> and not \\\"lz4.h\\\".\\n\\nEarlier this year I also suggested to include an 999 patch to change to\\nuse LZ4 as the default compression, to exercise the new code under CI.\\nI suggest to re-open the cf patch entry after that passes tests on all\\nplatforms and when it's ready for more review.\\n\\nBTW, some of these review comments are the same as what I sent earlier\\nthis year.\\n\\nhttps://www.postgresql.org/message-id/20220326162156.GI28503%40telsasoft.com\\nhttps://www.postgresql.org/message-id/20220705151328.GQ13040%40telsasoft.com\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sat, 17 Dec 2022 17:26:15 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, Dec 17, 2022 at 05:26:15PM -0600, Justin Pryzby wrote:\\n> 001: still refers to \\\"gzip\\\", which is correct for -Fp and -Fd but not\\n> for -Fc, for which it's more correct to say \\\"zlib\\\".\\n\\nOr should we begin by changing all these existing \\\"not built with zlib \\nsupport\\\" error strings to the more generic \\\"this build does not\\nsupport compression with %s\\\" to reduce the number of messages to\\ntranslate? That would bring consistency with the other tools dealing\\nwith compression.\\n\\n> That affects the\\n> name of the function, structures, comments, etc. I'm not sure if it's\\n> an issue to re-use the basebackup compression routines here. Maybe we\\n> should accept \\\"-Zn\\\" for zlib output (-Fc), but reject \\\"gzip:9\\\", which\\n> I'm sure some will find confusing, as it does not output. Maybe 001\\n> should be split into a patch to re-use the existing \\\"cfp\\\" interface\\n> (which is a clear win), and 002 to re-use the basebackup interfaces for\\n> user input and constants, etc.\\n> \\n> 001 still doesn't compile on freebsd, and 002 doesn't compile on\\n> windows. Have you checked test results from cirrusci on your private\\n> github account ?\\n\\nFYI, I have re-added an entry to the CF app to get some automated\\ncoverage:\\nhttps://commitfest.postgresql.org/41/3571/\\n\\nOn MinGW, a complain about the open() callback, which I guess ought to\\nbe avoided with a rename:\\n[00:16:37.254] compress_gzip.c:356:38: error: macro \\\"open\\\" passed 4 arguments, but takes just 3\\n[00:16:37.254] 356 | ret = CFH->open(fname, -1, mode, CFH);\\n[00:16:37.254] | ^\\n[00:16:37.254] In file included from ../../../src/include/c.h:1309,\\n[00:16:37.254] from ../../../src/include/postgres_fe.h:25,\\n[00:16:37.254] from compress_gzip.c:15:\\n\\nOn MSVC, some declaration conflicts, for a similar issue:\\n[00:12:31.966] ../src/bin/pg_dump/compress_io.c(193): error C2371: '_read': redefinition; different basic types\\n[00:12:31.966] C:\\\\Program Files (x86)\\\\Windows Kits\\\\10\\\\include\\\\10.0.20348.0\\\\ucrt\\\\corecrt_io.h(252): note: see declaration of '_read'\\n[00:12:31.966] ../src/bin/pg_dump/compress_io.c(210): error C2371: '_write': redefinition; different basic types\\n[00:12:31.966] C:\\\\Program Files (x86)\\\\Windows Kits\\\\10\\\\include\\\\10.0.20348.0\\\\ucrt\\\\corecrt_io.h(294): note: see declaration of '_write'\\n\\n> 002 breaks \\\"pg_dump -Fc -Z2\\\" because (I think) AllocateCompressor()\\n> doesn't store the passed-in compression_spec.\\n\\nHmm. This looks like a gap in the existing tests that we'd better fix\\nfirst. This CI is green on Linux.\\n\\n> 003 still uses <lz4.h> and not \\\"lz4.h\\\".\\n\\nThis should be <lz4.h>, not \\\"lz4.h\\\".\\n--\\nMichael\",\n \"msg_date\": \"Mon, 19 Dec 2022 13:06:00 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Monday, December 19th, 2022 at 5:06 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Sat, Dec 17, 2022 at 05:26:15PM -0600, Justin Pryzby wrote:\\n> \\n\\nThank you for the comments, please find v18 attached.\\n\\n> > 001: still refers to \\\"gzip\\\", which is correct for -Fp and -Fd but not\\n> > for -Fc, for which it's more correct to say \\\"zlib\\\".\\n> \\n> \\n> Or should we begin by changing all these existing \\\"not built with zlib\\n> support\\\" error strings to the more generic \\\"this build does not\\n> support compression with %s\\\" to reduce the number of messages to\\n> translate? That would bring consistency with the other tools dealing\\n> with compression.\\n\\nThis has been the approach from 0002 on-wards. In the attached it is also\\napplied on the remaining location in 0001. \\n\\n> \\n> > That affects the\\n> > name of the function, structures, comments, etc. I'm not sure if it's\\n> > an issue to re-use the basebackup compression routines here. Maybe we\\n> > should accept \\\"-Zn\\\" for zlib output (-Fc), but reject \\\"gzip:9\\\", which\\n> > I'm sure some will find confusing, as it does not output. Maybe 001\\n> > should be split into a patch to re-use the existing \\\"cfp\\\" interface\\n> > (which is a clear win), and 002 to re-use the basebackup interfaces for\\n> > user input and constants, etc.\\n> > \\n> > 001 still doesn't compile on freebsd, and 002 doesn't compile on\\n> > windows. Have you checked test results from cirrusci on your private\\n> > github account ?\\n\\nThere are still known gaps in 0002 and 0003, for example documentation,\\nand I have not been focusing too much on those. You are right, it is helpful\\nand kind to try to reduce the noise. The attached should have hopefully\\ntackled the ci errors.\\n\\n> \\n> FYI, I have re-added an entry to the CF app to get some automated\\n> coverage:\\n> https://commitfest.postgresql.org/41/3571/\\n\\nMuch obliged. Should I change the state to \\\"ready for review\\\" when post a\\nnew version or should I leave that to the senior personnel? \\n\\n> \\n> On MinGW, a complain about the open() callback, which I guess ought to\\n> be avoided with a rename:\\n> [00:16:37.254] compress_gzip.c:356:38: error: macro \\\"open\\\" passed 4 arguments, but takes just 3\\n> [00:16:37.254] 356 | ret = CFH->open(fname, -1, mode, CFH);\\n> \\n> [00:16:37.254] | ^\\n> [00:16:37.254] In file included from ../../../src/include/c.h:1309,\\n> [00:16:37.254] from ../../../src/include/postgres_fe.h:25,\\n> [00:16:37.254] from compress_gzip.c:15:\\n> \\n> On MSVC, some declaration conflicts, for a similar issue:\\n> [00:12:31.966] ../src/bin/pg_dump/compress_io.c(193): error C2371: '_read': redefinition; different basic types\\n> [00:12:31.966] C:\\\\Program Files (x86)\\\\Windows Kits\\\\10\\\\include\\\\10.0.20348.0\\\\ucrt\\\\corecrt_io.h(252): note: see declaration of '_read'\\n> [00:12:31.966] ../src/bin/pg_dump/compress_io.c(210): error C2371: '_write': redefinition; different basic types\\n> [00:12:31.966] C:\\\\Program Files (x86)\\\\Windows Kits\\\\10\\\\include\\\\10.0.20348.0\\\\ucrt\\\\corecrt_io.h(294): note: see declaration of '_write'\\n> \\n\\nA rename was enough.\\n\\n> > 002 breaks \\\"pg_dump -Fc -Z2\\\" because (I think) AllocateCompressor()\\n> > doesn't store the passed-in compression_spec.\\n> \\n\\nI am afraid I have not been able to reproduce this error. I tried both\\ndebian and freebsd after I addressed the compilation warnings. Which\\nerror did you get? Is it still present in the attached?\\n\\n> Hmm. This looks like a gap in the existing tests that we'd better fix\\n> first. This CI is green on Linux.\\n\\nAs the code stands, the compression level is not stored in the custom\\nformat's header as it is no longer relevant information. We can decide\\nto make it relevant for the tests only on the expense of increasing\\ndump size by four bytes. In either case this is not applicable in\\ncurrent head and can wait for 0002's turn.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> > 003 still uses <lz4.h> and not \\\"lz4.h\\\".\\n> \\n> \\n> This should be <lz4.h>, not \\\"lz4.h\\\".\\n> \\n> --\\n> Michael\",\n \"msg_date\": \"Mon, 19 Dec 2022 17:03:21 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\\n> > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\\n> > > windows. Have you checked test results from cirrusci on your private\\n> > > github account ?\\n> \\n> There are still known gaps in 0002 and 0003, for example documentation,\\n> and I have not been focusing too much on those. You are right, it is helpful\\n> and kind to try to reduce the noise. The attached should have hopefully\\n> tackled the ci errors.\\n\\nYep. Are you using cirrusci under your github account ?\\n\\n> > FYI, I have re-added an entry to the CF app to get some automated\\n> > coverage:\\n> > https://commitfest.postgresql.org/41/3571/\\n> \\n> Much obliged. Should I change the state to \\\"ready for review\\\" when post a\\n> new version or should I leave that to the senior personnel? \\n\\nIt's better to update it to reflect what you think its current status\\nis. If you think it's ready for review.\\n\\n> > > 002 breaks \\\"pg_dump -Fc -Z2\\\" because (I think) AllocateCompressor()\\n> > > doesn't store the passed-in compression_spec.\\n> \\n> I am afraid I have not been able to reproduce this error. I tried both\\n> debian and freebsd after I addressed the compilation warnings. Which\\n> error did you get? Is it still present in the attached?\\n\\nIt's not that there's an error - it's that compression isn't working.\\n\\n$ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\\n659956\\n$ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\\n637192\\n\\n$ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\\n1954890\\n$ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\\n1954890\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Mon, 19 Dec 2022 11:27:06 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Dec 19, 2022 at 01:06:00PM +0900, Michael Paquier wrote:\\n> On Sat, Dec 17, 2022 at 05:26:15PM -0600, Justin Pryzby wrote:\\n> > 001: still refers to \\\"gzip\\\", which is correct for -Fp and -Fd but not\\n> > for -Fc, for which it's more correct to say \\\"zlib\\\".\\n> \\n> Or should we begin by changing all these existing \\\"not built with zlib \\n> support\\\" error strings to the more generic \\\"this build does not\\n> support compression with %s\\\" to reduce the number of messages to\\n> translate? That would bring consistency with the other tools dealing\\n> with compression.\\n\\nThat's fine, but it doesn't touch on the issue I'm talking about, which\\nis that zlib != gzip.\\n\\nBTW I noticed that that also affects the pg_dump file itself; 002\\nchanges the file format to say \\\"gzip\\\", but that's wrong for -Fc, which\\ndoes not use gzip headers, which could be surprising to someone who\\nspecified \\\"gzip\\\".\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Mon, 19 Dec 2022 11:42:58 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Monday, December 19th, 2022 at 6:27 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n> \\n> \\n> On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\\n> \\n> > > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\\n> > > > windows. Have you checked test results from cirrusci on your private\\n> > > > github account ?\\n> > \\n> > There are still known gaps in 0002 and 0003, for example documentation,\\n> > and I have not been focusing too much on those. You are right, it is helpful\\n> > and kind to try to reduce the noise. The attached should have hopefully\\n> > tackled the ci errors.\\n> \\n> \\n> Yep. Are you using cirrusci under your github account ?\\n\\nThank you. To be very honest, I am not using github exclusively to post patches.\\nSometimes I do, sometimes I do not. Is github a requirement?\\n\\nTo answer your question, some of my github accounts are integrated with cirrusci,\\nothers are not.\\n\\nThe current cfbot build is green for what is worth.\\nhttps://cirrus-ci.com/build/5934319840002048\\n\\n> \\n> > > FYI, I have re-added an entry to the CF app to get some automated\\n> > > coverage:\\n> > > https://commitfest.postgresql.org/41/3571/\\n> > \\n> > Much obliged. Should I change the state to \\\"ready for review\\\" when post a\\n> > new version or should I leave that to the senior personnel?\\n> \\n> \\n> It's better to update it to reflect what you think its current status\\n> is. If you think it's ready for review.\\n\\nThank you.\\n\\n> \\n> > > > 002 breaks \\\"pg_dump -Fc -Z2\\\" because (I think) AllocateCompressor()\\n> > > > doesn't store the passed-in compression_spec.\\n> > \\n> > I am afraid I have not been able to reproduce this error. I tried both\\n> > debian and freebsd after I addressed the compilation warnings. Which\\n> > error did you get? Is it still present in the attached?\\n> \\n> \\n> It's not that there's an error - it's that compression isn't working.\\n> \\n> $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\\n> 659956\\n> $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\\n> 637192\\n> \\n> $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\\n> 1954890\\n> $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\\n> 1954890\\n> \\n\\nThank you. Now I understand what you mean. Trying the same on top of v18-0003\\non Ubuntu 22.04 yields:\\n\\n$ for compression in none gzip:1 gzip:6 gzip:9; do \\\\\\n pg_dump --format=custom --compress=\\\"$compression\\\" -f regression.\\\"$compression\\\".dump -d regression; \\\\\\n wc -c regression.\\\"$compression\\\".dump; \\\\\\n done;\\n14963753 regression.none.dump\\n3600183 regression.gzip:1.dump\\n3223755 regression.gzip:6.dump\\n3196903 regression.gzip:9.dump\\n\\nand on FreeBSD 13.1\\n\\n$ for compression in none gzip:1 gzip:6 gzip:9; do \\\\\\n pg_dump --format=custom --compress=\\\"$compression\\\" -f regression.\\\"$compression\\\".dump -d regression; \\\\\\n wc -c regression.\\\"$compression\\\".dump; \\\\\\n done;\\n14828822 regression.none.dump\\n3584304 regression.gzip:1.dump\\n3208548 regression.gzip:6.dump\\n3182044 regression.gzip:9.dump\\n\\nAlthough there are some variations between the installations, within the same\\ninstallation the size of the dump file is shrinking as expected.\\n\\nInvestigating a bit further on the issue, you are correct in identifying an\\nissue in v17. Up until v16, the compressor function looked like:\\n\\n+InitCompressorGzip(CompressorState *cs, int compressionLevel)\\n+{\\n+ GzipCompressorState *gzipcs;\\n+\\n+ cs->readData = ReadDataFromArchiveGzip;\\n+ cs->writeData = WriteDataToArchiveGzip;\\n+ cs->end = EndCompressorGzip;\\n+\\n+ gzipcs = (GzipCompressorState *) pg_malloc0(sizeof(GzipCompressorState));\\n+ gzipcs->compressionLevel = compressionLevel;\\n\\nV17 considered that more options could become available in the future\\nand changed the signature of the relevant Init functions to:\\n\\n+InitCompressorGzip(CompressorState *cs, const pg_compress_specification compression_spec)\\n+{\\n+ GzipCompressorState *gzipcs;\\n+\\n+ cs->readData = ReadDataFromArchiveGzip;\\n+ cs->writeData = WriteDataToArchiveGzip;\\n+ cs->end = EndCompressorGzip;\\n+\\n+ gzipcs = (GzipCompressorState *) pg_malloc0(sizeof(GzipCompressorState));\\n+\\n\\nV18 reinstated the assignment in similar fashion to InitCompressorNone and \\nInitCompressorLz4:\\n\\n+void\\n+InitCompressorGzip(CompressorState *cs, const pg_compress_specification compression_spec)\\n+{\\n+ GzipCompressorState *gzipcs;\\n+\\n+ cs->readData = ReadDataFromArchiveGzip;\\n+ cs->writeData = WriteDataToArchiveGzip;\\n+ cs->end = EndCompressorGzip;\\n+\\n+ cs->compression_spec = compression_spec;\\n+\\n+ gzipcs = (GzipCompressorState *) pg_malloc0(sizeof(GzipCompressorState));\\n\\nA test case can be added which performs a check similar to the loop above.\\nCreate a custom dump with the least and most compression for each method.\\nThen verify that the output sizes differ as expected. This addition could\\nbecome 0001 in the current series.\\n\\nThoughts?\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Justin\\n\\n\\n\",\n \"msg_date\": \"Tue, 20 Dec 2022 11:19:15 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Dec 20, 2022 at 11:19:15AM +0000, gkokolatos@pm.me wrote:\\n> ------- Original Message -------\\n> On Monday, December 19th, 2022 at 6:27 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\\n> > \\n> > > > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\\n> > > > > windows. Have you checked test results from cirrusci on your private\\n> > > > > github account ?\\n> > > \\n> > > There are still known gaps in 0002 and 0003, for example documentation,\\n> > > and I have not been focusing too much on those. You are right, it is helpful\\n> > > and kind to try to reduce the noise. The attached should have hopefully\\n> > > tackled the ci errors.\\n> > \\n> > \\n> > Yep. Are you using cirrusci under your github account ?\\n> \\n> Thank you. To be very honest, I am not using github exclusively to post patches.\\n> Sometimes I do, sometimes I do not. Is github a requirement?\\n\\nGithub isn't a requirement for postgres (but cirrusci only supports\\ngithub). I wasn't not trying to say that it's required, only trying to\\nmake sure that you (and others) know that it's available, since our\\ncirrus.yml is relatively new.\\n\\n> > > > > 002 breaks \\\"pg_dump -Fc -Z2\\\" because (I think) AllocateCompressor()\\n> > > > > doesn't store the passed-in compression_spec.\\n> > > \\n> > > I am afraid I have not been able to reproduce this error. I tried both\\n> > > debian and freebsd after I addressed the compilation warnings. Which\\n> > > error did you get? Is it still present in the attached?\\n> > \\n> > \\n> > It's not that there's an error - it's that compression isn't working.\\n> > \\n> > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\\n> > 659956\\n> > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\\n> > 637192\\n> > \\n> > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\\n> > 1954890\\n> > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\\n> > 1954890\\n> > \\n> \\n> Thank you. Now I understand what you mean. Trying the same on top of v18-0003\\n> on Ubuntu 22.04 yields:\\n\\nYou're right; this seems to be fixed in v18. Thanks.\\n\\nIt looks like I'd forgotten to run \\\"meson test tmp_install\\\", so had\\nretested v17...\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Tue, 20 Dec 2022 09:26:10 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Tuesday, December 20th, 2022 at 4:26 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n> \\n> \\n> On Tue, Dec 20, 2022 at 11:19:15AM +0000, gkokolatos@pm.me wrote:\\n> \\n> > ------- Original Message -------\\n> > On Monday, December 19th, 2022 at 6:27 PM, Justin Pryzby pryzby@telsasoft.com wrote:\\n> > \\n> > > On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\\n> > > \\n> > > > > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\\n> > > > > > windows. Have you checked test results from cirrusci on your private\\n> > > > > > github account ?\\n> > > > \\n> > > > There are still known gaps in 0002 and 0003, for example documentation,\\n> > > > and I have not been focusing too much on those. You are right, it is helpful\\n> > > > and kind to try to reduce the noise. The attached should have hopefully\\n> > > > tackled the ci errors.\\n> > > \\n> > > Yep. Are you using cirrusci under your github account ?\\n> > \\n> > Thank you. To be very honest, I am not using github exclusively to post patches.\\n> > Sometimes I do, sometimes I do not. Is github a requirement?\\n> \\n> \\n> Github isn't a requirement for postgres (but cirrusci only supports\\n> github). I wasn't not trying to say that it's required, only trying to\\n> make sure that you (and others) know that it's available, since our\\n> cirrus.yml is relatively new.\\n\\nGot it. Thank you very much for spreading the word. It is a useful feature which\\nshould be known. \\n\\n> \\n> > > > > > 002 breaks \\\"pg_dump -Fc -Z2\\\" because (I think) AllocateCompressor()\\n> > > > > > doesn't store the passed-in compression_spec.\\n> > > > \\n> > > > I am afraid I have not been able to reproduce this error. I tried both\\n> > > > debian and freebsd after I addressed the compilation warnings. Which\\n> > > > error did you get? Is it still present in the attached?\\n> > > \\n> > > It's not that there's an error - it's that compression isn't working.\\n> > > \\n> > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\\n> > > 659956\\n> > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\\n> > > 637192\\n> > > \\n> > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\\n> > > 1954890\\n> > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\\n> > > 1954890\\n> > \\n> > Thank you. Now I understand what you mean. Trying the same on top of v18-0003\\n> > on Ubuntu 22.04 yields:\\n> \\n> \\n> You're right; this seems to be fixed in v18. Thanks.\\n\\nGreat. Still there was a bug in v17 which you discovered. Thank you for the review\\neffort.\\n\\nPlease find in the attached v19 an extra check right before calling deflateInit().\\nThis check will verify that only compressed output will be generated for this\\nmethod.\\n\\nAlso v19 is rebased on top f450695e889 and applies cleanly.\\n\\nCheers.\\n//Georgios\\n\\n> --\\n> Justin\",\n \"msg_date\": \"Wed, 21 Dec 2022 10:09:38 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"There's a couple of lz4 bits which shouldn't be present in 002: file\\nextension and comments.\\n\\n\\n\",\n \"msg_date\": \"Thu, 22 Dec 2022 11:08:59 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Dec 22, 2022 at 11:08:59AM -0600, Justin Pryzby wrote:\\n> There's a couple of lz4 bits which shouldn't be present in 002: file\\n> extension and comments.\\n\\nThere were \\\"LZ4\\\" comments and file extension stuff in the preparatory\\ncommit. But now it seems like you *removed* them in the LZ4 commit\\n(where it actually belongs) rather than *moving* it from the\\nprior/parent commit *to* the lz4 commit. I recommend to run something\\nlike \\\"git diff @{1}\\\" whenever doing this kind of patch surgery.\\n\\n+ if (AH->compression_spec.algorithm != PG_COMPRESSION_NONE &&\\n+ AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\\n\\nThis looks wrong/redundant. The gzip part should be removed, right ?\\n\\nMaybe other places that check if (compression==PG_COMPRESSION_GZIP)\\nshould maybe change to say compression!=NONE?\\n\\n_PrepParallelRestore() references \\\".gz\\\", so I think it needs to be\\nretrofitted to handle .lz4. Ideally, that's built into a struct or list\\nof file extensions to try. Maybe compression.h should have a function\\nto return the file extension of a given algorithm. I'm planning to send\\na patch for zstd, and hoping its changes will be minimized by these\\npreparatory commits.\\n\\n+ errno = errno ? : ENOSPC;\\n\\n\\\"?:\\\" is a GNU extension (not the ternary operator, but the ternary\\noperator with only 2 args). It's not in use anywhere else in postgres.\\nYou could instead write it with 3 \\\"errno\\\"s or as \\\"if (errno==0):\\nerrno=ENOSPC\\\"\\n\\nYou wrote \\\"eol_flag == false\\\" and \\\"eol_flag == 0\\\" and true. But it's\\ncleaner to test it as a boolean: if (eol_flag) / if (!eol_flag).\\n\\nBoth LZ4File_init() and its callers check \\\"inited\\\". Better to do it in\\none place than 3. It's a static function, so I think there's no\\nperformance concern.\\n\\nGzip_close() still has a useless save_errno (or rebase issue?).\\n\\nI think it's confusing to have two functions, one named\\nInitCompressLZ4() and InitCompressorLZ4().\\n\\npg_compress_specification is being passed by value, but I think it\\nshould be passed as a pointer, as is done everywhere else.\\n\\npg_compress_algorithm is being writen directly into the pg_dump header.\\nCurrently, I think that's not an externally-visible value (it could be\\nrenumbered, theoretically even in a minor release). Maybe there should\\nbe a \\\"private\\\" enum for encoding the pg_dump header, similar to\\nWAL_COMPRESSION_LZ4 vs BKPIMAGE_COMPRESS_LZ4 ? Or else a comment there\\nshould warn that the values are encoded in pg_dump, and must never be\\nchanged.\\n\\n+ Verify that data files where compressed\\ntypo: s/where/were/\\n\\nAlso:\\ns/occurance/occurrence/\\ns/begining/beginning/\\ns/Verfiy/Verify/\\ns/nessary/necessary/\\n\\nBTW I noticed that cfdopen() was accidentally committed to compress_io.h\\nin master without being defined anywhere.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sun, 8 Jan 2023 13:45:25 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, 21 Dec 2022 at 15:40, <gkokolatos@pm.me> wrote:\\n>\\n>\\n>\\n>\\n>\\n>\\n> ------- Original Message -------\\n> On Tuesday, December 20th, 2022 at 4:26 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n>\\n> >\\n> >\\n> > On Tue, Dec 20, 2022 at 11:19:15AM +0000, gkokolatos@pm.me wrote:\\n> >\\n> > > ------- Original Message -------\\n> > > On Monday, December 19th, 2022 at 6:27 PM, Justin Pryzby pryzby@telsasoft.com wrote:\\n> > >\\n> > > > On Mon, Dec 19, 2022 at 05:03:21PM +0000, gkokolatos@pm.me wrote:\\n> > > >\\n> > > > > > > 001 still doesn't compile on freebsd, and 002 doesn't compile on\\n> > > > > > > windows. Have you checked test results from cirrusci on your private\\n> > > > > > > github account ?\\n> > > > >\\n> > > > > There are still known gaps in 0002 and 0003, for example documentation,\\n> > > > > and I have not been focusing too much on those. You are right, it is helpful\\n> > > > > and kind to try to reduce the noise. The attached should have hopefully\\n> > > > > tackled the ci errors.\\n> > > >\\n> > > > Yep. Are you using cirrusci under your github account ?\\n> > >\\n> > > Thank you. To be very honest, I am not using github exclusively to post patches.\\n> > > Sometimes I do, sometimes I do not. Is github a requirement?\\n> >\\n> >\\n> > Github isn't a requirement for postgres (but cirrusci only supports\\n> > github). I wasn't not trying to say that it's required, only trying to\\n> > make sure that you (and others) know that it's available, since our\\n> > cirrus.yml is relatively new.\\n>\\n> Got it. Thank you very much for spreading the word. It is a useful feature which\\n> should be known.\\n>\\n> >\\n> > > > > > > 002 breaks \\\"pg_dump -Fc -Z2\\\" because (I think) AllocateCompressor()\\n> > > > > > > doesn't store the passed-in compression_spec.\\n> > > > >\\n> > > > > I am afraid I have not been able to reproduce this error. I tried both\\n> > > > > debian and freebsd after I addressed the compilation warnings. Which\\n> > > > > error did you get? Is it still present in the attached?\\n> > > >\\n> > > > It's not that there's an error - it's that compression isn't working.\\n> > > >\\n> > > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fp regression |wc -c\\n> > > > 659956\\n> > > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fp regression |wc -c\\n> > > > 637192\\n> > > >\\n> > > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z1 -Fc regression |wc -c\\n> > > > 1954890\\n> > > > $ ./tmp_install/usr/local/pgsql/bin/pg_dump -h /tmp -Z2 -Fc regression |wc -c\\n> > > > 1954890\\n> > >\\n> > > Thank you. Now I understand what you mean. Trying the same on top of v18-0003\\n> > > on Ubuntu 22.04 yields:\\n> >\\n> >\\n> > You're right; this seems to be fixed in v18. Thanks.\\n>\\n> Great. Still there was a bug in v17 which you discovered. Thank you for the review\\n> effort.\\n>\\n> Please find in the attached v19 an extra check right before calling deflateInit().\\n> This check will verify that only compressed output will be generated for this\\n> method.\\n>\\n> Also v19 is rebased on top f450695e889 and applies cleanly.\\n\\nThe patch does not apply on top of HEAD as in [1], please post a rebased patch:\\n=== Applying patches on top of PostgreSQL commit ID\\nff23b592ad6621563d3128b26860bcb41daf9542 ===\\n=== applying patch ./v19-0002-Introduce-Compressor-API-in-pg_dump.patch\\npatching file src/bin/pg_dump/compress_io.h\\nHunk #1 FAILED at 37.\\n1 out of 1 hunk FAILED -- saving rejects to file\\nsrc/bin/pg_dump/compress_io.h.rej\\n\\n[1] - http://cfbot.cputube.org/patch_41_3571.log\\n\\nRegards,\\nVignesh\\n\\n\\n\",\n \"msg_date\": \"Sat, 14 Jan 2023 12:21:31 +0530\",\n \"msg_from\": \"vignesh C <vignesh21@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\\n> On Thu, Dec 22, 2022 at 11:08:59AM -0600, Justin Pryzby wrote:\\n> > There's a couple of lz4 bits which shouldn't be present in 002: file\\n> > extension and comments.\\n\\n> BTW I noticed that cfdopen() was accidentally committed to compress_io.h\\n> in master without being defined anywhere.\\n\\nThis was resolved in 69fb29d1a (so now needs to be re-added for this\\npatch series).\\n\\n> pg_compress_specification is being passed by value, but I think it\\n> should be passed as a pointer, as is done everywhere else.\\n\\nISTM that was an issue with 5e73a6048, affecting a few public and\\nprivate functions. I wrote a pre-preparatory patch which changes to\\npass by reference.\\n\\nAnd addressed a handful of other issues I reported as separate fixup\\ncommits. And changed to use LZ4 by default for CI.\\n\\nI also rebased my 2 year old patch to support zstd in pg_dump. I hope\\nit can finally added for v16. I'll send it for the next CF if these\\npatches progress.\\n\\nOne more thing: some comments still refer to the cfopen API, which this\\npatch removes.\\n\\n> There were \\\"LZ4\\\" comments and file extension stuff in the preparatory\\n> commit. But now it seems like you *removed* them in the LZ4 commit\\n> (where it actually belongs) rather than *moving* it from the\\n> prior/parent commit *to* the lz4 commit. I recommend to run something\\n> like \\\"git diff @{1}\\\" whenever doing this kind of patch surgery.\\n\\nTODO\\n\\n> Maybe other places that check if (compression==PG_COMPRESSION_GZIP)\\n> should maybe change to say compression!=NONE?\\n> \\n> _PrepParallelRestore() references \\\".gz\\\", so I think it needs to be\\n> retrofitted to handle .lz4. Ideally, that's built into a struct or list\\n> of file extensions to try. Maybe compression.h should have a function\\n> to return the file extension of a given algorithm. I'm planning to send\\n> a patch for zstd, and hoping its changes will be minimized by these\\n> preparatory commits.\\n\\nTODO\\n\\n> I think it's confusing to have two functions, one named\\n> InitCompressLZ4() and InitCompressorLZ4().\\n\\nTODO\\n\\n> pg_compress_algorithm is being writen directly into the pg_dump header.\\n> Currently, I think that's not an externally-visible value (it could be\\n> renumbered, theoretically even in a minor release). Maybe there should\\n> be a \\\"private\\\" enum for encoding the pg_dump header, similar to\\n> WAL_COMPRESSION_LZ4 vs BKPIMAGE_COMPRESS_LZ4 ? Or else a comment there\\n> should warn that the values are encoded in pg_dump, and must never be\\n> changed.\\n\\nMichael, WDYT ?\\n\\n-- \\nJustin\",\n \"msg_date\": \"Sat, 14 Jan 2023 15:43:09 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 14, 2023 at 03:43:08PM -0600, Justin Pryzby wrote:\\n> On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\\n> > pg_compress_specification is being passed by value, but I think it\\n> > should be passed as a pointer, as is done everywhere else.\\n> \\n> ISTM that was an issue with 5e73a6048, affecting a few public and\\n> private functions. I wrote a pre-preparatory patch which changes to\\n> pass by reference.\\n\\nI updated 001 to change SetOutput() to pass by reference, too (before,\\nthat ended up in the 002 patch).\\n\\nI can't see any issue in 002 other than the == GZIP change (the fix for\\nwhich I'd previously included in a later patch).\\n\\n> One more thing: some comments still refer to the cfopen API, which this\\n> patch removes.\\n> \\n> > There were \\\"LZ4\\\" comments and file extension stuff in the preparatory\\n> > commit. But now it seems like you *removed* them in the LZ4 commit\\n> > (where it actually belongs) rather than *moving* it from the\\n> > prior/parent commit *to* the lz4 commit. I recommend to run something\\n> > like \\\"git diff @{1}\\\" whenever doing this kind of patch surgery.\\n> \\n> TODO\\n\\nI addressed that in the fixup commits 005 and 007.\\n\\n-- \\nJustin\",\n \"msg_date\": \"Sun, 15 Jan 2023 16:32:11 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, Jan 14, 2023 at 03:43:09PM -0600, Justin Pryzby wrote:\\n> On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\\n>> pg_compress_specification is being passed by value, but I think it\\n>> should be passed as a pointer, as is done everywhere else.\\n> \\n> ISTM that was an issue with 5e73a6048, affecting a few public and\\n> private functions. I wrote a pre-preparatory patch which changes to\\n> pass by reference.\\n\\nThe functions changed by 0001 are cfopen[_write](),\\nAllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\\nto change these interfaces which basically exist to handle inputs? Is\\nthere some benefit in changing compression_spec within the internals\\nof these routines before going back one layer down to their callers?\\nChanging the compression_spec on-the-fly in these internal paths could\\nbe risky, actually, no?\\n\\n> And addressed a handful of other issues I reported as separate fixup\\n> commits. And changed to use LZ4 by default for CI.\\n\\nAre your slight changes shaped as of 0003-f.patch, 0005-f.patch and\\n0007-f.patch on top of the original patches sent by Georgios?\\n\\n> I also rebased my 2 year old patch to support zstd in pg_dump. I hope\\n> it can finally added for v16. I'll send it for the next CF if these\\n> patches progress.\\n\\nGood idea to see if what you have done for zstd fits with what's\\npresented here.\\n\\n>> pg_compress_algorithm is being writen directly into the pg_dump header.\\n\\nDo you mean that this is what happens once the patch series 0001~0008\\nsent upthread is applied on HEAD?\\n\\n>> Currently, I think that's not an externally-visible value (it could be\\n>> renumbered, theoretically even in a minor release). Maybe there should\\n>> be a \\\"private\\\" enum for encoding the pg_dump header, similar to\\n>> WAL_COMPRESSION_LZ4 vs BKPIMAGE_COMPRESS_LZ4 ? Or else a comment there\\n>> should warn that the values are encoded in pg_dump, and must never be\\n>> changed.\\n> \\n> Michael, WDYT ?\\n\\nChanging the order of the members in an enum would cause an ABI\\nbreakage, so that would not happen, and we tend to be very careful\\nabout that. Appending new members would be fine, though. FWIW, I'd\\nrather avoid adding more enums that would just be exact maps to\\npg_compress_algorithm.\\n\\n- /*\\n- * For now the compression type is implied by the level. This will need\\n- * to change once support for more compression algorithms is added,\\n- * requiring a format bump.\\n- */\\n- WriteInt(AH, AH->compression_spec.level);\\n+ AH->WriteBytePtr(AH, AH->compression_spec.algorithm);\\n\\nI may be missing something here, but it seems to me that you ought to\\nstore as well the level in the dump header, or it would not be\\npossible to report in the dump's description what was used? Hence,\\nK_VERS_1_15 should imply that we have both the method compression and\\nthe compression level.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 16 Jan 2023 10:28:50 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\\n> On Sat, Jan 14, 2023 at 03:43:09PM -0600, Justin Pryzby wrote:\\n> > On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\\n> >> pg_compress_specification is being passed by value, but I think it\\n> >> should be passed as a pointer, as is done everywhere else.\\n> > \\n> > ISTM that was an issue with 5e73a6048, affecting a few public and\\n> > private functions. I wrote a pre-preparatory patch which changes to\\n> > pass by reference.\\n> \\n> The functions changed by 0001 are cfopen[_write](),\\n> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\\n> to change these interfaces which basically exist to handle inputs?\\n\\nI changed to pass pg_compress_specification as a pointer, since that's\\nthe usual convention for structs, as followed by the existing uses of\\npg_compress_specification.\\n\\n> Is there some benefit in changing compression_spec within the\\n> internals of these routines before going back one layer down to their\\n> callers? Changing the compression_spec on-the-fly in these internal\\n> paths could be risky, actually, no?\\n\\nI think what you're saying is that if the spec is passed as a pointer,\\nthen the called functions shouldn't set spec->algorithm=something.\\n\\nI agree that if they need to do that, they should use a local variable.\\nWhich looks to be true for the functions that were changed in 001.\\n\\n> > And addressed a handful of other issues I reported as separate fixup\\n> > commits. And changed to use LZ4 by default for CI.\\n> \\n> Are your slight changes shaped as of 0003-f.patch, 0005-f.patch and\\n> 0007-f.patch on top of the original patches sent by Georgios?\\n\\nYes, the original patches, rebased as needed on top of HEAD and 001...\\n\\n> >> pg_compress_algorithm is being writen directly into the pg_dump header.\\n> \\n> Do you mean that this is what happens once the patch series 0001~0008\\n> sent upthread is applied on HEAD?\\n\\nYes\\n\\n> - /*\\n> - * For now the compression type is implied by the level. This will need\\n> - * to change once support for more compression algorithms is added,\\n> - * requiring a format bump.\\n> - */\\n> - WriteInt(AH, AH->compression_spec.level);\\n> + AH->WriteBytePtr(AH, AH->compression_spec.algorithm);\\n> \\n> I may be missing something here, but it seems to me that you ought to\\n> store as well the level in the dump header, or it would not be\\n> possible to report in the dump's description what was used? Hence,\\n> K_VERS_1_15 should imply that we have both the method compression and\\n> the compression level.\\n\\nMaybe. But the \\\"level\\\" isn't needed for decompression for any case I'm\\naware of.\\n\\nAlso, dumps with the default compression level currently say:\\n\\\"Compression: -1\\\", which does't seem valuable.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sun, 15 Jan 2023 19:56:25 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Oh, I didn’t realize you took over Justin? Why? After almost a year of work?\\n\\nThis is rather disheartening.\\n\\nOn Mon, Jan 16, 2023 at 02:56, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n> On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\\n>> On Sat, Jan 14, 2023 at 03:43:09PM -0600, Justin Pryzby wrote:\\n>> > On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:\\n>> >> pg_compress_specification is being passed by value, but I think it\\n>> >> should be passed as a pointer, as is done everywhere else.\\n>> >\\n>> > ISTM that was an issue with 5e73a6048, affecting a few public and\\n>> > private functions. I wrote a pre-preparatory patch which changes to\\n>> > pass by reference.\\n>>\\n>> The functions changed by 0001 are cfopen[_write](),\\n>> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\\n>> to change these interfaces which basically exist to handle inputs?\\n>\\n> I changed to pass pg_compress_specification as a pointer, since that's\\n> the usual convention for structs, as followed by the existing uses of\\n> pg_compress_specification.\\n>\\n>> Is there some benefit in changing compression_spec within the\\n>> internals of these routines before going back one layer down to their\\n>> callers? Changing the compression_spec on-the-fly in these internal\\n>> paths could be risky, actually, no?\\n>\\n> I think what you're saying is that if the spec is passed as a pointer,\\n> then the called functions shouldn't set spec->algorithm=something.\\n>\\n> I agree that if they need to do that, they should use a local variable.\\n> Which looks to be true for the functions that were changed in 001.\\n>\\n>> > And addressed a handful of other issues I reported as separate fixup\\n>> > commits. And changed to use LZ4 by default for CI.\\n>>\\n>> Are your slight changes shaped as of 0003-f.patch, 0005-f.patch and\\n>> 0007-f.patch on top of the original patches sent by Georgios?\\n>\\n> Yes, the original patches, rebased as needed on top of HEAD and 001...\\n>\\n>> >> pg_compress_algorithm is being writen directly into the pg_dump header.\\n>>\\n>> Do you mean that this is what happens once the patch series 0001~0008\\n>> sent upthread is applied on HEAD?\\n>\\n> Yes\\n>\\n>> - /*\\n>> - * For now the compression type is implied by the level. This will need\\n>> - * to change once support for more compression algorithms is added,\\n>> - * requiring a format bump.\\n>> - */\\n>> - WriteInt(AH, AH->compression_spec.level);\\n>> + AH->WriteBytePtr(AH, AH->compression_spec.algorithm);\\n>>\\n>> I may be missing something here, but it seems to me that you ought to\\n>> store as well the level in the dump header, or it would not be\\n>> possible to report in the dump's description what was used? Hence,\\n>> K_VERS_1_15 should imply that we have both the method compression and\\n>> the compression level.\\n>\\n> Maybe. But the \\\"level\\\" isn't needed for decompression for any case I'm\\n> aware of.\\n>\\n> Also, dumps with the default compression level currently say:\\n> \\\"Compression: -1\\\", which does't seem valuable.\\n>\\n> --\\n> Justin\\n Oh, I didn’t realize you took over Justin? Why? After almost a year of work?This is rather disheartening. On Mon, Jan 16, 2023 at 02:56, Justin Pryzby <pryzby@telsasoft.com> wrote: On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:> On Sat, Jan 14, 2023 at 03:43:09PM -0600, Justin Pryzby wrote:> > On Sun, Jan 08, 2023 at 01:45:25PM -0600, Justin Pryzby wrote:> >> pg_compress_specification is being passed by value, but I think it> >> should be passed as a pointer, as is done everywhere else.> >> > ISTM that was an issue with 5e73a6048, affecting a few public and> > private functions. I wrote a pre-preparatory patch which changes to> > pass by reference.>> The functions changed by 0001 are cfopen[_write](),> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea> to change these interfaces which basically exist to handle inputs?I changed to pass pg_compress_specification as a pointer, since that'sthe usual convention for structs, as followed by the existing uses ofpg_compress_specification.> Is there some benefit in changing compression_spec within the> internals of these routines before going back one layer down to their> callers? Changing the compression_spec on-the-fly in these internal> paths could be risky, actually, no?I think what you're saying is that if the spec is passed as a pointer,then the called functions shouldn't set spec->algorithm=something.I agree that if they need to do that, they should use a local variable.Which looks to be true for the functions that were changed in 001.> > And addressed a handful of other issues I reported as separate fixup> > commits. And changed to use LZ4 by default for CI.>> Are your slight changes shaped as of 0003-f.patch, 0005-f.patch and> 0007-f.patch on top of the original patches sent by Georgios?Yes, the original patches, rebased as needed on top of HEAD and 001...> >> pg_compress_algorithm is being writen directly into the pg_dump header.>> Do you mean that this is what happens once the patch series 0001~0008> sent upthread is applied on HEAD?Yes> - /*> - * For now the compression type is implied by the level. This will need> - * to change once support for more compression algorithms is added,> - * requiring a format bump.> - */> - WriteInt(AH, AH->compression_spec.level);> + AH->WriteBytePtr(AH, AH->compression_spec.algorithm);>> I may be missing something here, but it seems to me that you ought to> store as well the level in the dump header, or it would not be> possible to report in the dump's description what was used? Hence,> K_VERS_1_15 should imply that we have both the method compression and> the compression level.Maybe. But the \\\"level\\\" isn't needed for decompression for any case I'maware of.Also, dumps with the default compression level currently say:\\\"Compression: -1\\\", which does't seem valuable.--Justin\",\n \"msg_date\": \"Mon, 16 Jan 2023 02:27:56 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 16, 2023 at 02:27:56AM +0000, gkokolatos@pm.me wrote:\\n> Oh, I didn’t realize you took over Justin? Why? After almost a year of work?\\n> \\n> This is rather disheartening.\\n\\nI believe you've misunderstood my intent here. I sent rebased versions\\nof your patches with fixup commits implementing fixes that I'd\\npreviously sent. I don't think that's unusual. I hope your patches\\nwill be included in v16, and I hope to facilitate that. I don't mean\\nany offense. Actually, the fixups are provided as separate patches so\\nyou can adopt the changes easily into your branch. \\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sun, 15 Jan 2023 20:47:13 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sun, Jan 15, 2023 at 07:56:25PM -0600, Justin Pryzby wrote:\\n> On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\\n>> The functions changed by 0001 are cfopen[_write](),\\n>> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\\n>> to change these interfaces which basically exist to handle inputs?\\n> \\n> I changed to pass pg_compress_specification as a pointer, since that's\\n> the usual convention for structs, as followed by the existing uses of\\n> pg_compress_specification.\\n\\nOkay, but what do we gain here? It seems to me that this introduces\\nthe risk that a careless change in one of the internal routines if\\nthey change slight;ly compress_spec, hence impacting any of their\\ncallers? Or is that fixing an actual bug (except if I am missing your\\npoint, that does not seem to be the case)? \\n\\n>> Is there some benefit in changing compression_spec within the\\n>> internals of these routines before going back one layer down to their\\n>> callers? Changing the compression_spec on-the-fly in these internal\\n>> paths could be risky, actually, no?\\n> \\n> I think what you're saying is that if the spec is passed as a pointer,\\n> then the called functions shouldn't set spec->algorithm=something.\\n\\nYes. HEAD makes sure of that, 0001 would not prevent that. So I am a\\nbit confused in seeing how this is a benefit.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 16 Jan 2023 11:54:46 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nI admit I am completely at lost as to what is expected from me anymore.\\n\\nI had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\\nPlease find a rebased v20 attached.\\n\\nAlso please let me know if I should silently step away from it and let other people lead\\nit. I would be glad to comply either way.\\n\\nCheers,\\n//Georgios\\n\\n\\n------- Original Message -------\\nOn Monday, January 16th, 2023 at 3:54 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n>\\n>\\n> On Sun, Jan 15, 2023 at 07:56:25PM -0600, Justin Pryzby wrote:\\n>\\n> > On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\\n> >\\n> > > The functions changed by 0001 are cfopen_write,\\n> > > AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\\n> > > to change these interfaces which basically exist to handle inputs?\\n> >\\n> > I changed to pass pg_compress_specification as a pointer, since that's\\n> > the usual convention for structs, as followed by the existing uses of\\n> > pg_compress_specification.\\n>\\n>\\n> Okay, but what do we gain here? It seems to me that this introduces\\n> the risk that a careless change in one of the internal routines if\\n> they change slight;ly compress_spec, hence impacting any of their\\n> callers? Or is that fixing an actual bug (except if I am missing your\\n> point, that does not seem to be the case)?\\n>\\n> > > Is there some benefit in changing compression_spec within the\\n> > > internals of these routines before going back one layer down to their\\n> > > callers? Changing the compression_spec on-the-fly in these internal\\n> > > paths could be risky, actually, no?\\n> >\\n> > I think what you're saying is that if the spec is passed as a pointer,\\n> > then the called functions shouldn't set spec->algorithm=something.\\n>\\n>\\n> Yes. HEAD makes sure of that, 0001 would not prevent that. So I am a\\n> bit confused in seeing how this is a benefit.\\n> --\\n> Michael\",\n \"msg_date\": \"Mon, 16 Jan 2023 15:14:43 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 1/16/23 16:14, gkokolatos@pm.me wrote:\\n> Hi,\\n> \\n> I admit I am completely at lost as to what is expected from me anymore.\\n> \\n\\n:-(\\n\\nI understand it's frustrating not to know why a patch is not moving\\nforward. Particularly when is seems fairly straightforward ...\\n\\nLet me briefly explain my personal (and admittedly very subjective) view\\non picking what patches to review/commit. I'm sure other committers have\\nother criteria, but maybe this will help.\\n\\nThere are always more patches than I can review/commit, so I have to\\nprioritize, and pick which patches to look at. For me, it's mostly about\\ncost/benefit of the patch. The cost is e.g. the amount of time I need to\\nspend to review/commit the stuff, maybe read the thread, etc. Benefits\\nis mainly the new features/improvements.\\n\\nIt's oversimplified, we could talk about various bits that contribute to\\nthe costs and benefits, but this is what it boils down.\\n\\nThere's always the aspect of time - patches A and B have roughly the\\nsame benefits, but with A we get it \\\"immediately\\\" while B requires\\nadditional parts that we don't have ready yet (and if they don't make it\\nwe get no benefit), I'll probably pick A.\\n\\nUnfortunately, this plays against this patch - I'm certainly in favor of\\nadding lz4 (and other compression algos) into pg_dump, but if I commit\\n0001 we get little benefit, and the other parts actually adding lz4/zstd\\nare treated as \\\"WIP / for completeness\\\" so it's unclear when we'd get to\\ncommit them.\\n\\nSo if I could recommend one thing, it'd be to get at least one of those\\nWIP patches into a shape that's likely committable right after 0001.\\n\\n> I had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\\n> Please find a rebased v20 attached.\\n> \\n\\nI took a quick look at 0001, so a couple comments (sorry if some of this\\nwas already discussed in the thread):\\n\\n1) I don't think a \\\"refactoring\\\" patch should reference particular\\ncompression algorithms (lz4/zstd), and in particular I don't think we\\nshould have \\\"not yet implemented\\\" messages. We only have a couple other\\nplaces doing that, when we didn't have a better choice. But here we can\\nsimply reject the algorithm when parsing the options, we don't need to\\ndo that in a dozen other places.\\n\\n2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\\n\\\"none\\\" at the end. It might make backpatches harder.\\n\\n3) While building, I get bunch of warnings about missing cfdopen()\\nprototype and pg_backup_archiver.c not knowing about cfdopen() and\\nadding an implicit prototype (so I doubt it actually works).\\n\\n4) \\\"cfp\\\" struct no longer wraps gzFile, but the comment was not updated.\\nFWIW I'm not sure switching to \\\"void *\\\" is an improvement, maybe it'd be\\nbetter to have a \\\"union\\\" of correct types?\\n\\n5) cfopen/cfdopen are missing comments. cfopen_internal has an updated\\ncomment, but that's a static function while cfopen/cfdopen are the\\nactual API.\\n\\n> Also please let me know if I should silently step away from it and let other people lead\\n> it. I would be glad to comply either way.\\n> \\n\\nPlease don't. I promise to take a look at this patch again.\\n\\nThanks for doing all the work.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Wed, 18 Jan 2023 15:00:34 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Wednesday, January 18th, 2023 at 3:00 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> Hi,\\n> \\n> On 1/16/23 16:14, gkokolatos@pm.me wrote:\\n> \\n> > Hi,\\n> > \\n> > I admit I am completely at lost as to what is expected from me anymore.\\n> \\n<snip>\\n> \\n> Unfortunately, this plays against this patch - I'm certainly in favor of\\n> adding lz4 (and other compression algos) into pg_dump, but if I commit\\n> 0001 we get little benefit, and the other parts actually adding lz4/zstd\\n> are treated as \\\"WIP / for completeness\\\" so it's unclear when we'd get to\\n> commit them.\\n\\nThank you for your kindness and for taking the time to explain.\\n \\n> So if I could recommend one thing, it'd be to get at least one of those\\n> WIP patches into a shape that's likely committable right after 0001.\\n\\nThis was clearly my fault. I misunderstood a suggestion upthread to focus\\non the first patch of the series and ignore documentation and comments on\\nthe rest.\\n\\nPlease find v21 to contain 0002 and 0003 in a state which I no longer consider\\nas WIP but worthy of proper consideration. Some guidance on where is best to add\\ndocumentation in 0002 for the function pointers in CompressFileHandle will\\nbe welcomed.\\n\\n> \\n> > I had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\\n> > Please find a rebased v20 attached.\\n> \\n> \\n> I took a quick look at 0001, so a couple comments (sorry if some of this\\n> was already discussed in the thread):\\n\\nMuch appreciated!\\n\\n> \\n> 1) I don't think a \\\"refactoring\\\" patch should reference particular\\n> compression algorithms (lz4/zstd), and in particular I don't think we\\n> should have \\\"not yet implemented\\\" messages. We only have a couple other\\n> places doing that, when we didn't have a better choice. But here we can\\n> simply reject the algorithm when parsing the options, we don't need to\\n> do that in a dozen other places.\\n\\nI have now removed lz4/zstd from where they were present with the exception\\nof pg_dump.c which is responsible for parsing.\\n\\n> 2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\\n> \\\"none\\\" at the end. It might make backpatches harder.\\n\\nAgreed. However a 'default' is needed in order to avoid compilation warnings.\\nAlso note that 0002 completely does away with cases within WriteDataToArchive.\\n\\n> 3) While building, I get bunch of warnings about missing cfdopen()\\n> prototype and pg_backup_archiver.c not knowing about cfdopen() and\\n> adding an implicit prototype (so I doubt it actually works).\\n\\nFixed. cfdopen() got prematurely introduced in 5e73a6048 and then got removed\\nin 69fb29d1af. v20 failed to properly take 69fb29d1af in consideration. Note\\nthat cfdopen is removed in 0002 which explains why cfbot didn't complain.\\n \\n> 4) \\\"cfp\\\" struct no longer wraps gzFile, but the comment was not updated.\\n> FWIW I'm not sure switching to \\\"void *\\\" is an improvement, maybe it'd be\\n> better to have a \\\"union\\\" of correct types?\\n\\nPlease find and updated comment and a union in place of the void *. Also\\nnote that 0002 completely does away with cfp in favour of a new struct\\nCompressFileHandle. I maintained the void * there because it is used by\\nprivate methods of the compressors. 0003 contains such an example with\\nLZ4CompressorState.\\n\\n> 5) cfopen/cfdopen are missing comments. cfopen_internal has an updated\\n> comment, but that's a static function while cfopen/cfdopen are the\\n> actual API.\\n\\nAdded comments to cfopen/cfdopen.\\n\\n> \\n> > Also please let me know if I should silently step away from it and let other people lead\\n> > it. I would be glad to comply either way.\\n> \\n> \\n> Please don't. I promise to take a look at this patch again.\\n\\nThank you very much.\\n\\n> Thanks for doing all the work.\\n\\nThank you.\\n\\nCheers,\\n//Georgios\\n\\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Wed, 18 Jan 2023 19:05:43 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 1/18/23 20:05, gkokolatos@pm.me wrote:\\n> \\n> \\n> \\n> \\n> \\n> ------- Original Message -------\\n> On Wednesday, January 18th, 2023 at 3:00 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n> \\n> \\n>>\\n>>\\n>> Hi,\\n>>\\n>> On 1/16/23 16:14, gkokolatos@pm.me wrote:\\n>>\\n>>> Hi,\\n>>>\\n>>> I admit I am completely at lost as to what is expected from me anymore.\\n>>\\n> <snip>\\n>>\\n>> Unfortunately, this plays against this patch - I'm certainly in favor of\\n>> adding lz4 (and other compression algos) into pg_dump, but if I commit\\n>> 0001 we get little benefit, and the other parts actually adding lz4/zstd\\n>> are treated as \\\"WIP / for completeness\\\" so it's unclear when we'd get to\\n>> commit them.\\n> \\n> Thank you for your kindness and for taking the time to explain.\\n> \\n>> So if I could recommend one thing, it'd be to get at least one of those\\n>> WIP patches into a shape that's likely committable right after 0001.\\n> \\n> This was clearly my fault. I misunderstood a suggestion upthread to focus\\n> on the first patch of the series and ignore documentation and comments on\\n> the rest.\\n> \\n> Please find v21 to contain 0002 and 0003 in a state which I no longer consider\\n> as WIP but worthy of proper consideration. Some guidance on where is best to add\\n> documentation in 0002 for the function pointers in CompressFileHandle will\\n> be welcomed.\\n> \\n\\nThis is internal-only API, not meant for use by regular users and/or\\nextension authors, so I don't think we need sgml docs. I'd just add\\nregular code-level documentation to compress_io.h.\\n\\nFor inspiration see docs for \\\"struct ReorderBuffer\\\" in reorderbuffer.h,\\nor \\\"struct _archiveHandle\\\" in pg_backup_archiver.h.\\n\\nOr what other kind of documentation you had in mind?\\n\\n>>\\n>>> I had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\\n>>> Please find a rebased v20 attached.\\n>>\\n>>\\n>> I took a quick look at 0001, so a couple comments (sorry if some of this\\n>> was already discussed in the thread):\\n> \\n> Much appreciated!\\n> \\n>>\\n>> 1) I don't think a \\\"refactoring\\\" patch should reference particular\\n>> compression algorithms (lz4/zstd), and in particular I don't think we\\n>> should have \\\"not yet implemented\\\" messages. We only have a couple other\\n>> places doing that, when we didn't have a better choice. But here we can\\n>> simply reject the algorithm when parsing the options, we don't need to\\n>> do that in a dozen other places.\\n> \\n> I have now removed lz4/zstd from where they were present with the exception\\n> of pg_dump.c which is responsible for parsing.\\n> \\n\\nI'm not sure I understand why leave the lz4/zstd in this place?\\n\\n>> 2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\\n>> \\\"none\\\" at the end. It might make backpatches harder.\\n> \\n> Agreed. However a 'default' is needed in order to avoid compilation warnings.\\n> Also note that 0002 completely does away with cases within WriteDataToArchive.\\n> \\n\\nOK, although that's also a consequence of using a \\\"switch\\\" instead of\\nplan \\\"if\\\" branches.\\n\\nFurthermore, I'm not sure we really need the pg_fatal() about invalid\\ncompression method in these default blocks. I mean, how could we even\\nget to these places when the build does not support the algorithm? All\\nof this (ReadDataFromArchive, WriteDataToArchive, EndCompressor, ...)\\nhappens looong after the compressor was initialized and the method\\nchecked, no? So maybe either this should simply do Assert(false) or use\\na different error message.\\n\\n>> 3) While building, I get bunch of warnings about missing cfdopen()\\n>> prototype and pg_backup_archiver.c not knowing about cfdopen() and\\n>> adding an implicit prototype (so I doubt it actually works).\\n> \\n> Fixed. cfdopen() got prematurely introduced in 5e73a6048 and then got removed\\n> in 69fb29d1af. v20 failed to properly take 69fb29d1af in consideration. Note\\n> that cfdopen is removed in 0002 which explains why cfbot didn't complain.\\n> \\n\\nOK.\\n\\n>> 4) \\\"cfp\\\" struct no longer wraps gzFile, but the comment was not updated.\\n>> FWIW I'm not sure switching to \\\"void *\\\" is an improvement, maybe it'd be\\n>> better to have a \\\"union\\\" of correct types?\\n> \\n> Please find and updated comment and a union in place of the void *. Also\\n> note that 0002 completely does away with cfp in favour of a new struct\\n> CompressFileHandle. I maintained the void * there because it is used by\\n> private methods of the compressors. 0003 contains such an example with\\n> LZ4CompressorState.\\n> \\n\\nI wonder if this (and also the previous item) makes sense to keep 0001\\nand 0002 or to combine them. The \\\"intermediate\\\" state is a bit annoying.\\n\\n>> 5) cfopen/cfdopen are missing comments. cfopen_internal has an updated\\n>> comment, but that's a static function while cfopen/cfdopen are the\\n>> actual API.\\n> \\n> Added comments to cfopen/cfdopen.\\n> \\n\\nOK.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 19 Jan 2023 16:45:10 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Thursday, January 19th, 2023 at 4:45 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> On 1/18/23 20:05, gkokolatos@pm.me wrote:\\n> \\n> > ------- Original Message -------\\n> > On Wednesday, January 18th, 2023 at 3:00 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\\n> > \\n> > > Hi,\\n> > > \\n> > > On 1/16/23 16:14, gkokolatos@pm.me wrote:\\n> > > \\n> > > > Hi,\\n> > > > \\n> > > > I admit I am completely at lost as to what is expected from me anymore.\\n> > \\n> > <snip>\\n> > \\n> > > Unfortunately, this plays against this patch - I'm certainly in favor of\\n> > > adding lz4 (and other compression algos) into pg_dump, but if I commit\\n> > > 0001 we get little benefit, and the other parts actually adding lz4/zstd\\n> > > are treated as \\\"WIP / for completeness\\\" so it's unclear when we'd get to\\n> > > commit them.\\n> > \\n> > Thank you for your kindness and for taking the time to explain.\\n> > \\n> > > So if I could recommend one thing, it'd be to get at least one of those\\n> > > WIP patches into a shape that's likely committable right after 0001.\\n> > \\n> > This was clearly my fault. I misunderstood a suggestion upthread to focus\\n> > on the first patch of the series and ignore documentation and comments on\\n> > the rest.\\n> > \\n> > Please find v21 to contain 0002 and 0003 in a state which I no longer consider\\n> > as WIP but worthy of proper consideration. Some guidance on where is best to add\\n> > documentation in 0002 for the function pointers in CompressFileHandle will\\n> > be welcomed.\\n> \\n> \\n> This is internal-only API, not meant for use by regular users and/or\\n> extension authors, so I don't think we need sgml docs. I'd just add\\n> regular code-level documentation to compress_io.h.\\n> \\n> For inspiration see docs for \\\"struct ReorderBuffer\\\" in reorderbuffer.h,\\n> or \\\"struct _archiveHandle\\\" in pg_backup_archiver.h.\\n> \\n> Or what other kind of documentation you had in mind?\\n\\nThis is exactly what I was after. I was between compress_io.c and compress_io.h.\\nThank you.\\n\\n> > > > I had posted v19-0001 for a committer's consideration and v19-000{2,3} for completeness.\\n> > > > Please find a rebased v20 attached.\\n> > > \\n> > > I took a quick look at 0001, so a couple comments (sorry if some of this\\n> > > was already discussed in the thread):\\n> > \\n> > Much appreciated!\\n> > \\n> > > 1) I don't think a \\\"refactoring\\\" patch should reference particular\\n> > > compression algorithms (lz4/zstd), and in particular I don't think we\\n> > > should have \\\"not yet implemented\\\" messages. We only have a couple other\\n> > > places doing that, when we didn't have a better choice. But here we can\\n> > > simply reject the algorithm when parsing the options, we don't need to\\n> > > do that in a dozen other places.\\n> > \\n> > I have now removed lz4/zstd from where they were present with the exception\\n> > of pg_dump.c which is responsible for parsing.\\n> \\n> \\n> I'm not sure I understand why leave the lz4/zstd in this place?\\n\\nYou are right, it is not obvious. Those were added in 5e73a60488 which is\\nalready committed in master and I didn't want to backtrack. Of course, I am\\nnot opposing in doing so if you wish.\\n\\n> \\n> > > 2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\\n> > > \\\"none\\\" at the end. It might make backpatches harder.\\n> > \\n> > Agreed. However a 'default' is needed in order to avoid compilation warnings.\\n> > Also note that 0002 completely does away with cases within WriteDataToArchive.\\n> \\n> \\n> OK, although that's also a consequence of using a \\\"switch\\\" instead of\\n> plan \\\"if\\\" branches.\\n> \\n> Furthermore, I'm not sure we really need the pg_fatal() about invalid\\n> compression method in these default blocks. I mean, how could we even\\n> get to these places when the build does not support the algorithm? All\\n> of this (ReadDataFromArchive, WriteDataToArchive, EndCompressor, ...)\\n> happens looong after the compressor was initialized and the method\\n> checked, no? So maybe either this should simply do Assert(false) or use\\n> a different error message.\\n\\nI like Assert(false).\\n\\n> > > 3) While building, I get bunch of warnings about missing cfdopen()\\n> > > prototype and pg_backup_archiver.c not knowing about cfdopen() and\\n> > > adding an implicit prototype (so I doubt it actually works).\\n> > \\n> > Fixed. cfdopen() got prematurely introduced in 5e73a6048 and then got removed\\n> > in 69fb29d1af. v20 failed to properly take 69fb29d1af in consideration. Note\\n> > that cfdopen is removed in 0002 which explains why cfbot didn't complain.\\n> \\n> \\n> OK.\\n> \\n> > > 4) \\\"cfp\\\" struct no longer wraps gzFile, but the comment was not updated.\\n> > > FWIW I'm not sure switching to \\\"void *\\\" is an improvement, maybe it'd be\\n> > > better to have a \\\"union\\\" of correct types?\\n> > \\n> > Please find and updated comment and a union in place of the void *. Also\\n> > note that 0002 completely does away with cfp in favour of a new struct\\n> > CompressFileHandle. I maintained the void * there because it is used by\\n> > private methods of the compressors. 0003 contains such an example with\\n> > LZ4CompressorState.\\n> \\n> \\n> I wonder if this (and also the previous item) makes sense to keep 0001\\n> and 0002 or to combine them. The \\\"intermediate\\\" state is a bit annoying.\\n\\nAgreed. It was initially submitted as one patch. Then it was requested to be\\nsplit up in two parts, one to expand the use of the existing API and one to\\nreplace with the new interface. Unfortunately the expansion of usage of the\\nexisting API requires some tweaking, but that is not a very good reason for\\nthe current patch set. I should have done a better job there.\\n\\nPlease find v22 attach which combines back 0001 and 0002. It is missing the\\ndocumentation that was discussed above as I wanted to give a quick feedback.\\nLet me know if you think that the combined version is the one to move forward\\nwith.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> > > 5) cfopen/cfdopen are missing comments. cfopen_internal has an updated\\n> > > comment, but that's a static function while cfopen/cfdopen are the\\n> > > actual API.\\n> > \\n> > Added comments to cfopen/cfdopen.\\n> \\n> \\n> OK.\\n> \\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Thu, 19 Jan 2023 16:42:06 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 1/19/23 17:42, gkokolatos@pm.me wrote:\\n> \\n> ------- Original Message -------\\n> On Thursday, January 19th, 2023 at 4:45 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n>>\\n>> On 1/18/23 20:05, gkokolatos@pm.me wrote:\\n>>\\n>>> ------- Original Message -------\\n>>> On Wednesday, January 18th, 2023 at 3:00 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\\n>>\\n>> I'm not sure I understand why leave the lz4/zstd in this place?\\n> \\n> You are right, it is not obvious. Those were added in 5e73a60488 which is\\n> already committed in master and I didn't want to backtrack. Of course, I am\\n> not opposing in doing so if you wish.\\n> \\n\\nAh, I didn't realize it was already added by earlier commit. In that\\ncase let's not worry about it.\\n\\n>>\\n>>>> 2) I wouldn't reorder the cases in WriteDataToArchive, i.e. I'd keep\\n>>>> \\\"none\\\" at the end. It might make backpatches harder.\\n>>>\\n>>> Agreed. However a 'default' is needed in order to avoid compilation warnings.\\n>>> Also note that 0002 completely does away with cases within WriteDataToArchive.\\n>>\\n>>\\n>> OK, although that's also a consequence of using a \\\"switch\\\" instead of\\n>> plan \\\"if\\\" branches.\\n>>\\n>> Furthermore, I'm not sure we really need the pg_fatal() about invalid\\n>> compression method in these default blocks. I mean, how could we even\\n>> get to these places when the build does not support the algorithm? All\\n>> of this (ReadDataFromArchive, WriteDataToArchive, EndCompressor, ...)\\n>> happens looong after the compressor was initialized and the method\\n>> checked, no? So maybe either this should simply do Assert(false) or use\\n>> a different error message.\\n> \\n> I like Assert(false).\\n> \\n\\nOK, good. Do you agree we should never actually get there, if the\\nearlier checks work correctly?\\n\\n>>\\n>>>> 4) \\\"cfp\\\" struct no longer wraps gzFile, but the comment was not updated.\\n>>>> FWIW I'm not sure switching to \\\"void *\\\" is an improvement, maybe it'd be\\n>>>> better to have a \\\"union\\\" of correct types?\\n>>>\\n>>> Please find and updated comment and a union in place of the void *. Also\\n>>> note that 0002 completely does away with cfp in favour of a new struct\\n>>> CompressFileHandle. I maintained the void * there because it is used by\\n>>> private methods of the compressors. 0003 contains such an example with\\n>>> LZ4CompressorState.\\n>>\\n>>\\n>> I wonder if this (and also the previous item) makes sense to keep 0001\\n>> and 0002 or to combine them. The \\\"intermediate\\\" state is a bit annoying.\\n> \\n> Agreed. It was initially submitted as one patch. Then it was requested to be\\n> split up in two parts, one to expand the use of the existing API and one to\\n> replace with the new interface. Unfortunately the expansion of usage of the\\n> existing API requires some tweaking, but that is not a very good reason for\\n> the current patch set. I should have done a better job there.\\n> \\n> Please find v22 attach which combines back 0001 and 0002. It is missing the\\n> documentation that was discussed above as I wanted to give a quick feedback.\\n> Let me know if you think that the combined version is the one to move forward\\n> with.\\n> \\n\\nThanks, I'll take a look.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 19 Jan 2023 18:55:43 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 1/19/23 18:55, Tomas Vondra wrote:\\n> Hi,\\n> \\n> On 1/19/23 17:42, gkokolatos@pm.me wrote:\\n>>\\n>> ...\\n>>\\n>> Agreed. It was initially submitted as one patch. Then it was requested to be\\n>> split up in two parts, one to expand the use of the existing API and one to\\n>> replace with the new interface. Unfortunately the expansion of usage of the\\n>> existing API requires some tweaking, but that is not a very good reason for\\n>> the current patch set. I should have done a better job there.\\n>>\\n>> Please find v22 attach which combines back 0001 and 0002. It is missing the\\n>> documentation that was discussed above as I wanted to give a quick feedback.\\n>> Let me know if you think that the combined version is the one to move forward\\n>> with.\\n>>\\n> \\n> Thanks, I'll take a look.\\n> \\n\\nAfter taking a look and thinking about it a bit more, I think we should\\nkeep the two parts separate. I think Michael (or whoever proposed) the\\nsplit was right, it makes the patches easier to grok.\\n\\nSorry for the noise, hopefully we can just revert to the last version.\\n\\nWhile reading the thread, I also noticed this:\\n\\n> By the way, I think that this 0002 should drop all the default clauses\\n> in the switches for the compression method so as we'd catch any\\n> missing code paths with compiler warnings if a new compression method\\n> is added in the future.\\n\\nNow I realize why there were \\\"not yet implemented\\\" errors for lz4/zstd\\nin all the switches, and why after removing them you had to add a\\ndefault branch.\\n\\nWe DON'T want a default branch, because the idea is that after adding a\\nnew compression algorithm, we get warnings about switches not handling\\nit correctly.\\n\\nSo I guess we should walk back this change too :-( It's probably easier\\nto go back to v20 from January 16, and redo the couple remaining things\\nI commented on.\\n\\n\\nFWIW I think this is a hint that adding LZ4/ZSTD options, in 5e73a6048,\\nbut without implementation, was not a great idea. It mostly defeats the\\nidea of getting the compiler warnings - all the places already handle\\nPG_COMPRESSION_LZ4/PG_COMPRESSION_ZSTD by throwing a pg_fatal. So you'd\\nhave to grep for the options, inspect all the places or something like\\nthat anyway. The warnings would only work for entirely new methods.\\n\\nHowever, I now also realize the compressor API in 0002 replaces all of\\nthis with calls to a generic API callback, so trying to improve this was\\npretty silly from me.\\n\\n\\nPlease, fix the couple remaining details in v20, add the docs for the\\ncallbacks, and I'll try to polish it and get it committed.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 20 Jan 2023 00:34:34 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Friday, January 20th, 2023 at 12:34 AM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> On 1/19/23 18:55, Tomas Vondra wrote:\\n> \\n> > Hi,\\n> > \\n> > On 1/19/23 17:42, gkokolatos@pm.me wrote:\\n> > \\n> > > ...\\n> > > \\n> > > Agreed. It was initially submitted as one patch. Then it was requested to be\\n> > > split up in two parts, one to expand the use of the existing API and one to\\n> > > replace with the new interface. Unfortunately the expansion of usage of the\\n> > > existing API requires some tweaking, but that is not a very good reason for\\n> > > the current patch set. I should have done a better job there.\\n> > > \\n> > > Please find v22 attach which combines back 0001 and 0002. It is missing the\\n> > > documentation that was discussed above as I wanted to give a quick feedback.\\n> > > Let me know if you think that the combined version is the one to move forward\\n> > > with.\\n> > \\n> > Thanks, I'll take a look.\\n> \\n> \\n> After taking a look and thinking about it a bit more, I think we should\\n> keep the two parts separate. I think Michael (or whoever proposed) the\\n> split was right, it makes the patches easier to grok.\\n> \\n\\nExcellent. I will attempt a better split this time round.\\n\\n> \\n> While reading the thread, I also noticed this:\\n> \\n> > By the way, I think that this 0002 should drop all the default clauses\\n> > in the switches for the compression method so as we'd catch any\\n> > missing code paths with compiler warnings if a new compression method\\n> > is added in the future.\\n> \\n> \\n> Now I realize why there were \\\"not yet implemented\\\" errors for lz4/zstd\\n> in all the switches, and why after removing them you had to add a\\n> default branch.\\n> \\n> We DON'T want a default branch, because the idea is that after adding a\\n> new compression algorithm, we get warnings about switches not handling\\n> it correctly.\\n> \\n> So I guess we should walk back this change too :-( It's probably easier\\n> to go back to v20 from January 16, and redo the couple remaining things\\n> I commented on.\\n> \\n\\nSure.\\n\\n> \\n> FWIW I think this is a hint that adding LZ4/ZSTD options, in 5e73a6048,\\n> but without implementation, was not a great idea. It mostly defeats the\\n> idea of getting the compiler warnings - all the places already handle\\n> PG_COMPRESSION_LZ4/PG_COMPRESSION_ZSTD by throwing a pg_fatal. So you'd\\n> have to grep for the options, inspect all the places or something like\\n> that anyway. The warnings would only work for entirely new methods.\\n> \\n> However, I now also realize the compressor API in 0002 replaces all of\\n> this with calls to a generic API callback, so trying to improve this was\\n> pretty silly from me.\\n\\nI can try to do a better job at splitting things up.\\n\\n> \\n> Please, fix the couple remaining details in v20, add the docs for the\\n> callbacks, and I'll try to polish it and get it committed.\\n\\nExcellent. Allow me an attempt to polish and expect a new version soon.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 20 Jan 2023 15:03:30 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Friday, January 20th, 2023 at 12:34 AM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n>\\n>\\n> On 1/19/23 18:55, Tomas Vondra wrote:\\n>\\n> > Hi,\\n> >\\n> > On 1/19/23 17:42, gkokolatos@pm.me wrote:\\n> >\\n> > > ...\\n> > >\\n> > > Agreed. It was initially submitted as one patch. Then it was requested to be\\n> > > split up in two parts, one to expand the use of the existing API and one to\\n> > > replace with the new interface. Unfortunately the expansion of usage of the\\n> > > existing API requires some tweaking, but that is not a very good reason for\\n> > > the current patch set. I should have done a better job there.\\n> > >\\n> > > Please find v22 attach which combines back 0001 and 0002. It is missing the\\n> > > documentation that was discussed above as I wanted to give a quick feedback.\\n> > > Let me know if you think that the combined version is the one to move forward\\n> > > with.\\n> >\\n> > Thanks, I'll take a look.\\n>\\n>\\n> After taking a look and thinking about it a bit more, I think we should\\n> keep the two parts separate. I think Michael (or whoever proposed) the\\n> split was right, it makes the patches easier to grok.\\n\\nPlease find attached v23 which reintroduces the split.\\n\\n0001 is reworked to have a reduced footprint than before. Also in an attempt\\nto facilitate the readability, 0002 splits the API's and the uncompressed\\nimplementation in separate files.\\n\\n>\\n> While reading the thread, I also noticed this:\\n>\\n> > By the way, I think that this 0002 should drop all the default clauses\\n> > in the switches for the compression method so as we'd catch any\\n> > missing code paths with compiler warnings if a new compression method\\n> > is added in the future.\\n>\\n>\\n> Now I realize why there were \\\"not yet implemented\\\" errors for lz4/zstd\\n> in all the switches, and why after removing them you had to add a\\n> default branch.\\n>\\n> We DON'T want a default branch, because the idea is that after adding a\\n> new compression algorithm, we get warnings about switches not handling\\n> it correctly.\\n>\\n> So I guess we should walk back this change too :-( It's probably easier\\n> to go back to v20 from January 16, and redo the couple remaining things\\n> I commented on.\\n\\nNo problem.\\n\\n> FWIW I think this is a hint that adding LZ4/ZSTD options, in 5e73a6048,\\n> but without implementation, was not a great idea. It mostly defeats the\\n> idea of getting the compiler warnings - all the places already handle\\n> PG_COMPRESSION_LZ4/PG_COMPRESSION_ZSTD by throwing a pg_fatal. So you'd\\n> have to grep for the options, inspect all the places or something like\\n> that anyway. The warnings would only work for entirely new methods.\\n>\\n> However, I now also realize the compressor API in 0002 replaces all of\\n> this with calls to a generic API callback, so trying to improve this was\\n> pretty silly from me.\\n>\\n>\\n> Please, fix the couple remaining details in v20, add the docs for the\\n> callbacks, and I'll try to polish it and get it committed.\\n\\nThank you very much. Please find an attempt to comply with the requested\\nchanges in the attached.\\n\\n>\\n>\\n> regards\\n>\\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 23 Jan 2023 17:31:55 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\\n> Please find attached v23 which reintroduces the split.\\n> \\n> 0001 is reworked to have a reduced footprint than before. Also in an attempt\\n> to facilitate the readability, 0002 splits the API's and the uncompressed\\n> implementation in separate files.\\n\\nThanks for updating the patch. Could you address the review comments I\\nsent here ?\\nhttps://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\\n\\nThanks,\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Mon, 23 Jan 2023 12:00:14 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n> \\n> \\n> On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\\n> \\n> > Please find attached v23 which reintroduces the split.\\n> > \\n> > 0001 is reworked to have a reduced footprint than before. Also in an attempt\\n> > to facilitate the readability, 0002 splits the API's and the uncompressed\\n> > implementation in separate files.\\n> \\n> \\n> Thanks for updating the patch. Could you address the review comments I\\n> sent here ?\\n> https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\\n\\nPlease find v24 attached.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> Thanks,\\n> --\\n> Justin\",\n \"msg_date\": \"Tue, 24 Jan 2023 15:56:20 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 24, 2023 at 03:56:20PM +0000, gkokolatos@pm.me wrote:\\n> On Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\\n> > \\n> > > Please find attached v23 which reintroduces the split.\\n> > > \\n> > > 0001 is reworked to have a reduced footprint than before. Also in an attempt\\n> > > to facilitate the readability, 0002 splits the API's and the uncompressed\\n> > > implementation in separate files.\\n> > \\n> > Thanks for updating the patch. Could you address the review comments I\\n> > sent here ?\\n> > https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\\n> \\n> Please find v24 attached.\\n\\nThanks for updating the patch.\\n\\nIn 001, RestoreArchive() does:\\n\\n> -#ifndef HAVE_LIBZ\\n> - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\\n> - AH->PrintTocDataPtr != NULL)\\n> + supports_compression = false;\\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_NONE ||\\n> + AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> + supports_compression = true;\\n> +\\n> + if (AH->PrintTocDataPtr != NULL)\\n> {\\n> for (te = AH->toc->next; te != AH->toc; te = te->next)\\n> {\\n> if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\\n> - pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n> + {\\n> +#ifndef HAVE_LIBZ\\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> + supports_compression = false;\\n> +#endif\\n> + if (supports_compression == false)\\n> + pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n> + }\\n> }\\n> }\\n> -#endif\\n\\nThis first checks if the algorithm is implemented, and then checks if\\nthe algorithm is supported by the current build - that confused me for a\\nbit. It seems unnecessary to check for unimplemented algorithms before\\nlooping. That also requires referencing both GZIP and LZ4 in two\\nplaces.\\n\\nI think it could be written to avoid the need to change for added\\ncompression algorithms:\\n\\n+ if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\\n+ {\\n+ /* Check if the compression algorithm is supported */\\n+ pg_compress_specification spec;\\n+ parse_compress_specification(AH->compression_spec.algorithm, NULL, &spec);\\n+ if (spec->parse_error != NULL)\\n+ pg_fatal(spec->parse_error);\\n+ }\\n\\nOr maybe add a new function to compression.c to indicate whether a given\\nalgorithm is supported.\\n\\nThat would also indicate *which* compression library isn't supported.\\n\\nOther than that, I think 001 is ready.\\n\\n002/003 use these names, which I think are too similar - initially I\\ndidn't even realize there were two separate functions (each with a\\nsecond stub function to handle the case of unsupported compression):\\n\\n+extern void InitCompressorGzip(CompressorState *cs, const pg_compress_specification compression_spec); \\n+extern void InitCompressGzip(CompressFileHandle *CFH, const pg_compress_specification compression_spec); \\n\\n+extern void InitCompressorLZ4(CompressorState *cs, const pg_compress_specification compression_spec); \\n+extern void InitCompressLZ4(CompressFileHandle *CFH, const pg_compress_specification compression_spec); \\n\\ntypo:\\ns/not build with/not built with/\\n\\nShould AllocateCompressor() set cs->compression_spec, rather than doing\\nit in each compressor ?\\n\\nThanks for considering.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Tue, 24 Jan 2023 19:42:03 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Wednesday, January 25th, 2023 at 2:42 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n> \\n> \\n> On Tue, Jan 24, 2023 at 03:56:20PM +0000, gkokolatos@pm.me wrote:\\n> \\n> > On Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby pryzby@telsasoft.com wrote:\\n> > \\n> > > On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\\n> > > \\n> > > > Please find attached v23 which reintroduces the split.\\n> > > > \\n> > > > 0001 is reworked to have a reduced footprint than before. Also in an attempt\\n> > > > to facilitate the readability, 0002 splits the API's and the uncompressed\\n> > > > implementation in separate files.\\n> > > \\n> > > Thanks for updating the patch. Could you address the review comments I\\n> > > sent here ?\\n> > > https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\\n> > \\n> > Please find v24 attached.\\n> \\n> \\n> Thanks for updating the patch.\\n> \\n> In 001, RestoreArchive() does:\\n> \\n> > -#ifndef HAVE_LIBZ\\n> > - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\\n> > - AH->PrintTocDataPtr != NULL)\\n> > + supports_compression = false;\\n> > + if (AH->compression_spec.algorithm == PG_COMPRESSION_NONE ||\\n> > + AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> > + supports_compression = true;\\n> > +\\n> > + if (AH->PrintTocDataPtr != NULL)\\n> > {\\n> > for (te = AH->toc->next; te != AH->toc; te = te->next)\\n> > {\\n> > if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\\n> > - pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n> > + {\\n> > +#ifndef HAVE_LIBZ\\n> > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> > + supports_compression = false;\\n> > +#endif\\n> > + if (supports_compression == false)\\n> > + pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n> > + }\\n> > }\\n> > }\\n> > -#endif\\n> \\n> \\n> This first checks if the algorithm is implemented, and then checks if\\n> the algorithm is supported by the current build - that confused me for a\\n> bit. It seems unnecessary to check for unimplemented algorithms before\\n> looping. That also requires referencing both GZIP and LZ4 in two\\n> places.\\n\\nI am not certain that it is unnecessary, at least not in the way that is\\ndescribed. The idea is that new compression methods can be added, without\\nchanging the archive's version number. It is very possible that it is\\nrequested to restore an archive compressed with a method not implemented\\nin the current binary. The first check takes care of that and sets\\nsupports_compression only for the supported versions. It is possible to\\nenter the loop with supports_compression already set to false, for example\\nbecause the archive was compressed with ZSTD, triggering the fatal error.\\n\\nOf course, one can throw the error before entering the loop, yet I think\\nthat it does not help the readability of the code. IMHO it is easier to\\nfollow if the error is thrown once during that check.\\n\\n> \\n> I think it could be written to avoid the need to change for added\\n> compression algorithms:\\n> \\n> + if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\\n> \\n> + {\\n> + /* Check if the compression algorithm is supported */\\n> + pg_compress_specification spec;\\n> + parse_compress_specification(AH->compression_spec.algorithm, NULL, &spec);\\n> \\n> + if (spec->parse_error != NULL)\\n> \\n> + pg_fatal(spec->parse_error);\\n> \\n> + }\\n\\nI am not certain how that would work in the example with ZSTD above.\\nIf I am not wrong, parse_compress_specification() will not throw an error\\nif the codebase supports ZSTD, yet this specific pg_dump binary will not\\nsupport it because ZSTD is not implemented. parse_compress_specification()\\nis not aware of that and should not be aware of it, should it?\\n\\n> \\n> Or maybe add a new function to compression.c to indicate whether a given\\n> algorithm is supported.\\n\\nI am not certain how this would help, as compression.c is supposed to be\\nused by multiple binaries while this is a pg_dump specific detail.\\n\\n> That would also indicate which compression library isn't supported.\\n\\nIf anything, I can suggest to throw an error much earlier, i.e. in ReadHead(),\\nand remove altogether this check. On the other hand, I like the belts\\nand suspenders approach because there are no more checks after this point.\\n\\n> Other than that, I think 001 is ready.\\n\\nThank you.\\n\\n> 002/003 use these names, which I think are too similar - initially I\\n> didn't even realize there were two separate functions (each with a\\n> second stub function to handle the case of unsupported compression):\\n> \\n> +extern void InitCompressorGzip(CompressorState *cs, const pg_compress_specification compression_spec);\\n> +extern void InitCompressGzip(CompressFileHandle *CFH, const pg_compress_specification compression_spec);\\n> \\n> +extern void InitCompressorLZ4(CompressorState *cs, const pg_compress_specification compression_spec);\\n> +extern void InitCompressLZ4(CompressFileHandle *CFH, const pg_compress_specification compression_spec);\\n\\nFair enough. Names are now updated.\\n\\n> \\n> typo:\\n> s/not build with/not built with/\\n\\nThank you.\\n\\n> \\n> Should AllocateCompressor() set cs->compression_spec, rather than doing\\n> it in each compressor ?\\n\\nI think that compression_spec should be owned by each compressor. With that\\nin mind, it makes more sense to set it within each compressor. This is not\\na hill I am willing to die on though.\\n\\nPlease find v25 attached.\\n\\n> \\n> Thanks for considering.\\n> \\n> --\\n> Justin\",\n \"msg_date\": \"Wed, 25 Jan 2023 15:37:12 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\nOn 1/25/23 16:37, gkokolatos@pm.me wrote:\\n> \\n> \\n> \\n> \\n> \\n> ------- Original Message -------\\n> On Wednesday, January 25th, 2023 at 2:42 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> \\n> \\n>>\\n>>\\n>> On Tue, Jan 24, 2023 at 03:56:20PM +0000, gkokolatos@pm.me wrote:\\n>>\\n>>> On Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby pryzby@telsasoft.com wrote:\\n>>>\\n>>>> On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\\n>>>>\\n>>>>> Please find attached v23 which reintroduces the split.\\n>>>>>\\n>>>>> 0001 is reworked to have a reduced footprint than before. Also in an attempt\\n>>>>> to facilitate the readability, 0002 splits the API's and the uncompressed\\n>>>>> implementation in separate files.\\n>>>>\\n>>>> Thanks for updating the patch. Could you address the review comments I\\n>>>> sent here ?\\n>>>> https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\\n>>>\\n>>> Please find v24 attached.\\n>>\\n>>\\n>> Thanks for updating the patch.\\n>>\\n>> In 001, RestoreArchive() does:\\n>>\\n>>> -#ifndef HAVE_LIBZ\\n>>> - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\\n>>> - AH->PrintTocDataPtr != NULL)\\n>>> + supports_compression = false;\\n>>> + if (AH->compression_spec.algorithm == PG_COMPRESSION_NONE ||\\n>>> + AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n>>> + supports_compression = true;\\n>>> +\\n>>> + if (AH->PrintTocDataPtr != NULL)\\n>>> {\\n>>> for (te = AH->toc->next; te != AH->toc; te = te->next)\\n>>> {\\n>>> if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\\n>>> - pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n>>> + {\\n>>> +#ifndef HAVE_LIBZ\\n>>> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n>>> + supports_compression = false;\\n>>> +#endif\\n>>> + if (supports_compression == false)\\n>>> + pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n>>> + }\\n>>> }\\n>>> }\\n>>> -#endif\\n>>\\n>>\\n>> This first checks if the algorithm is implemented, and then checks if\\n>> the algorithm is supported by the current build - that confused me for a\\n>> bit. It seems unnecessary to check for unimplemented algorithms before\\n>> looping. That also requires referencing both GZIP and LZ4 in two\\n>> places.\\n> \\n> I am not certain that it is unnecessary, at least not in the way that is\\n> described. The idea is that new compression methods can be added, without\\n> changing the archive's version number. It is very possible that it is\\n> requested to restore an archive compressed with a method not implemented\\n> in the current binary. The first check takes care of that and sets\\n> supports_compression only for the supported versions. It is possible to\\n> enter the loop with supports_compression already set to false, for example\\n> because the archive was compressed with ZSTD, triggering the fatal error.\\n> \\n> Of course, one can throw the error before entering the loop, yet I think\\n> that it does not help the readability of the code. IMHO it is easier to\\n> follow if the error is thrown once during that check.\\n> \\n\\nActually, I don't understand why 0001 moves the check into the loop. I\\nmean, why not check HAVE_LIBZ before the loop?\\n\\n>>\\n>> I think it could be written to avoid the need to change for added\\n>> compression algorithms:\\n>>\\n>> + if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\\n>>\\n>> + {\\n>> + /* Check if the compression algorithm is supported */\\n>> + pg_compress_specification spec;\\n>> + parse_compress_specification(AH->compression_spec.algorithm, NULL, &spec);\\n>>\\n>> + if (spec->parse_error != NULL)\\n>>\\n>> + pg_fatal(spec->parse_error);\\n>>\\n>> + }\\n> \\n> I am not certain how that would work in the example with ZSTD above.\\n> If I am not wrong, parse_compress_specification() will not throw an error\\n> if the codebase supports ZSTD, yet this specific pg_dump binary will not\\n> support it because ZSTD is not implemented. parse_compress_specification()\\n> is not aware of that and should not be aware of it, should it?\\n> \\n\\nNot sure. What happens in a similar situation now? That is, when trying\\nto deal with an archive gzip-compressed in a build without libz?\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 25 Jan 2023 18:28:49 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 25, 2023 at 03:37:12PM +0000, gkokolatos@pm.me wrote:\\n> Of course, one can throw the error before entering the loop, yet I think\\n> that it does not help the readability of the code. IMHO it is easier to\\n> follow if the error is thrown once during that check.\\n\\n> If anything, I can suggest to throw an error much earlier, i.e. in ReadHead(),\\n> and remove altogether this check. On the other hand, I like the belts\\n> and suspenders approach because there are no more checks after this point.\\n\\nWhile looking at this, I realized that commit 5e73a6048 introduced a\\nregression:\\n\\n@@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\\n\\n- if (AH->compression != 0)\\n- pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n+ if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n+ pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n\\nBefore, it was possible to restore non-data chunks of a dump file, even\\nif the current build didn't support its compression. But that's now\\nimpossible - and it makes the code we're discussing in RestoreArchive()\\nunreachable.\\n\\nI don't think we can currently test for that, since it requires creating a dump\\nusing a build --with compression and then trying to restore using a build\\n--without compression. The coverage report disagrees with me, though...\\nhttps://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#3901\\n\\n> > I think it could be written to avoid the need to change for added\\n> > compression algorithms:\\n...\\n> \\n> I am not certain how that would work in the example with ZSTD above.\\n> If I am not wrong, parse_compress_specification() will not throw an error\\n> if the codebase supports ZSTD, yet this specific pg_dump binary will not\\n> support it because ZSTD is not implemented. parse_compress_specification()\\n> is not aware of that and should not be aware of it, should it?\\n\\nYou're right.\\n\\nI think the 001 patch should try to remove hardcoded references to\\nLIBZ/GZIP, such that the later patches don't need to update those same\\nplaces for LZ4. For example in ReadHead() and RestoreArchive(), and\\nmaybe other places dealing with file extensions. Maybe that could be\\ndone by adding a function specific to pg_dump indicating whether or not\\nan algorithm is implemented and supported.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Wed, 25 Jan 2023 12:00:20 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Wednesday, January 25th, 2023 at 6:28 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> \\n> On 1/25/23 16:37, gkokolatos@pm.me wrote:\\n> \\n> > ------- Original Message -------\\n> > On Wednesday, January 25th, 2023 at 2:42 AM, Justin Pryzby pryzby@telsasoft.com wrote:\\n> > \\n> > > On Tue, Jan 24, 2023 at 03:56:20PM +0000, gkokolatos@pm.me wrote:\\n> > > \\n> > > > On Monday, January 23rd, 2023 at 7:00 PM, Justin Pryzby pryzby@telsasoft.com wrote:\\n> > > > \\n> > > > > On Mon, Jan 23, 2023 at 05:31:55PM +0000, gkokolatos@pm.me wrote:\\n> > > > > \\n> > > > > > Please find attached v23 which reintroduces the split.\\n> > > > > > \\n> > > > > > 0001 is reworked to have a reduced footprint than before. Also in an attempt\\n> > > > > > to facilitate the readability, 0002 splits the API's and the uncompressed\\n> > > > > > implementation in separate files.\\n> > > > > \\n> > > > > Thanks for updating the patch. Could you address the review comments I\\n> > > > > sent here ?\\n> > > > > https://www.postgresql.org/message-id/20230108194524.GA27637%40telsasoft.com\\n> > > > \\n> > > > Please find v24 attached.\\n> > > \\n> > > Thanks for updating the patch.\\n> > > \\n> > > In 001, RestoreArchive() does:\\n> > > \\n> > > > -#ifndef HAVE_LIBZ\\n> > > > - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\\n> > > > - AH->PrintTocDataPtr != NULL)\\n> > > > + supports_compression = false;\\n> > > > + if (AH->compression_spec.algorithm == PG_COMPRESSION_NONE ||\\n> > > > + AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> > > > + supports_compression = true;\\n> > > > +\\n> > > > + if (AH->PrintTocDataPtr != NULL)\\n> > > > {\\n> > > > for (te = AH->toc->next; te != AH->toc; te = te->next)\\n> > > > {\\n> > > > if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\\n> > > > - pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n> > > > + {\\n> > > > +#ifndef HAVE_LIBZ\\n> > > > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> > > > + supports_compression = false;\\n> > > > +#endif\\n> > > > + if (supports_compression == false)\\n> > > > + pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n> > > > + }\\n> > > > }\\n> > > > }\\n> > > > -#endif\\n> > > \\n> > > This first checks if the algorithm is implemented, and then checks if\\n> > > the algorithm is supported by the current build - that confused me for a\\n> > > bit. It seems unnecessary to check for unimplemented algorithms before\\n> > > looping. That also requires referencing both GZIP and LZ4 in two\\n> > > places.\\n> > \\n> > I am not certain that it is unnecessary, at least not in the way that is\\n> > described. The idea is that new compression methods can be added, without\\n> > changing the archive's version number. It is very possible that it is\\n> > requested to restore an archive compressed with a method not implemented\\n> > in the current binary. The first check takes care of that and sets\\n> > supports_compression only for the supported versions. It is possible to\\n> > enter the loop with supports_compression already set to false, for example\\n> > because the archive was compressed with ZSTD, triggering the fatal error.\\n> > \\n> > Of course, one can throw the error before entering the loop, yet I think\\n> > that it does not help the readability of the code. IMHO it is easier to\\n> > follow if the error is thrown once during that check.\\n> \\n> \\n> Actually, I don't understand why 0001 moves the check into the loop. I\\n> mean, why not check HAVE_LIBZ before the loop?\\n\\nThe intention is to be able to restore archives that don't contain\\ndata. In that case compression becomes irrelevant as only the data in\\nan archive is compressed.\\n\\n> \\n> > > I think it could be written to avoid the need to change for added\\n> > > compression algorithms:\\n> > > \\n> > > + if (te->hadDumper && (te->reqs & REQ_DATA) != 0)\\n> > > \\n> > > + {\\n> > > + /* Check if the compression algorithm is supported */\\n> > > + pg_compress_specification spec;\\n> > > + parse_compress_specification(AH->compression_spec.algorithm, NULL, &spec);\\n> > > \\n> > > + if (spec->parse_error != NULL)\\n> > > \\n> > > + pg_fatal(spec->parse_error);\\n> > > \\n> > > + }\\n> > \\n> > I am not certain how that would work in the example with ZSTD above.\\n> > If I am not wrong, parse_compress_specification() will not throw an error\\n> > if the codebase supports ZSTD, yet this specific pg_dump binary will not\\n> > support it because ZSTD is not implemented. parse_compress_specification()\\n> > is not aware of that and should not be aware of it, should it?\\n> \\n> \\n> Not sure. What happens in a similar situation now? That is, when trying\\n> to deal with an archive gzip-compressed in a build without libz?\\n\\n\\nIn case that there are no data chunks, the archive will be restored.\\n\\nCheers,\\n//Georgios\\n\\n\\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 25 Jan 2023 19:50:49 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Wednesday, January 25th, 2023 at 7:00 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n> \\n> \\n> On Wed, Jan 25, 2023 at 03:37:12PM +0000, gkokolatos@pm.me wrote:\\n> \\n\\n> While looking at this, I realized that commit 5e73a6048 introduced a\\n> regression:\\n> \\n> @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\\n> \\n> - if (AH->compression != 0)\\n> \\n> - pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> \\n> + pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n> \\n> Before, it was possible to restore non-data chunks of a dump file, even\\n> if the current build didn't support its compression. But that's now\\n> impossible - and it makes the code we're discussing in RestoreArchive()\\n> unreachable.\\n\\nNice catch!\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Justin\\n\\n\\n\",\n \"msg_date\": \"Wed, 25 Jan 2023 19:57:18 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 25, 2023 at 07:57:18PM +0000, gkokolatos@pm.me wrote:\\n> Nice catch!\\n\\nLet me see..\\n--\\nMichael\",\n \"msg_date\": \"Thu, 26 Jan 2023 13:46:54 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 25, 2023 at 12:00:20PM -0600, Justin Pryzby wrote:\\n> While looking at this, I realized that commit 5e73a6048 introduced a\\n> regression:\\n> \\n> @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\\n> \\n> - if (AH->compression != 0)\\n> - pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n> + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> + pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n> \\n> Before, it was possible to restore non-data chunks of a dump file, even\\n> if the current build didn't support its compression. But that's now\\n> impossible - and it makes the code we're discussing in RestoreArchive()\\n> unreachable.\\n\\nRight. The impacts the possibility of looking at the header data,\\nwhich is useful with pg_restore -l for example. On a dump that's been\\ncompressed, pg_restore <= 15 would always print the TOC entries with\\nor without compression support. On HEAD, this code prevents the\\nheader lookup. All *nix or BSD platforms should have support for\\nzlib, I hope.. Still that could be an issue on Windows, and this\\nwould prevent folks to check the contents of the dumps after saving it\\non a WIN32 host, so let's undo that.\\n\\nSo, I have been testing the attached with four sets of binaries from\\n15/HEAD and with[out] zlib support, and this brings HEAD back to the\\npre-15 state (header information able to show up, still failure when\\nattempting to restore the dump's data without zlib).\\n\\n> I don't think we can currently test for that, since it requires creating a dump\\n> using a build --with compression and then trying to restore using a build\\n> --without compression.\\n\\nRight, the location of the data is in the header, and I don't see how\\nyou would be able to do that without two sets of binaries at hand, but\\nour tests run under the assumption that you have only one. Well,\\nthat's not entirely true as well, as you could create a TAP test like\\npg_upgrade that relies on a environment variable pointing to a second\\nset of binaries. That's not worth the complication involved, IMO.\\n\\n> The coverage report disagrees with me, though...\\n> https://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#3901\\n\\nIsn't that one of the tests like compression_gzip_plain?\\n\\nThoughts?\\n--\\nMichael\",\n \"msg_date\": \"Thu, 26 Jan 2023 14:49:27 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 26, 2023 at 02:49:27PM +0900, Michael Paquier wrote:\\n> On Wed, Jan 25, 2023 at 12:00:20PM -0600, Justin Pryzby wrote:\\n> > While looking at this, I realized that commit 5e73a6048 introduced a\\n> > regression:\\n> > \\n> > @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\\n> > \\n> > - if (AH->compression != 0)\\n> > - pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n> > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> > + pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n> > \\n> > Before, it was possible to restore non-data chunks of a dump file, even\\n> > if the current build didn't support its compression. But that's now\\n> > impossible - and it makes the code we're discussing in RestoreArchive()\\n> > unreachable.\\n> \\n> Right. The impacts the possibility of looking at the header data,\\n> which is useful with pg_restore -l for example.\\n\\nIt's not just header data - it's schema and (I think) everything other\\nthan table data.\\n\\n> > The coverage report disagrees with me, though...\\n> > https://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#3901\\n> \\n> Isn't that one of the tests like compression_gzip_plain?\\n\\nI'm not sure what you mean. Plain dump is restored with psql and not\\nwith pg_restore.\\n\\nMy line number was wrong:\\nhttps://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#390\\n\\nWhat test would hit that code without rebuilding ?\\n\\n394 : #ifndef HAVE_LIBZ\\n395 : if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\\n\\n> Thoughts?\\n> #ifndef HAVE_LIBZ\\n> \\tif (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> -\\t\\tpg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n> +\\t\\tpg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n\\nYour patch is fine for now, but these errors should eventually specify\\n*which* compression algorithm is unavailable. I think that should be a\\npart of the 001 patch, ideally in a way that minimizes the number of\\nplaces which need to be updated when adding an algorithm.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 Jan 2023 00:28:46 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Thursday, January 26th, 2023 at 7:28 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n> \\n> \\n> On Thu, Jan 26, 2023 at 02:49:27PM +0900, Michael Paquier wrote:\\n> \\n> > On Wed, Jan 25, 2023 at 12:00:20PM -0600, Justin Pryzby wrote:\\n> > \\n> > > While looking at this, I realized that commit 5e73a6048 introduced a\\n> > > regression:\\n> > > \\n> > > @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\\n> > > \\n> > > - if (AH->compression != 0)\\n> > > - pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n> > > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> > > + pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n> > > \\n> > > Before, it was possible to restore non-data chunks of a dump file, even\\n> > > if the current build didn't support its compression. But that's now\\n> > > impossible - and it makes the code we're discussing in RestoreArchive()\\n> > > unreachable.\\n> > \\n> > Right. The impacts the possibility of looking at the header data,\\n> > which is useful with pg_restore -l for example.\\n> \\n> \\n> It's not just header data - it's schema and (I think) everything other\\n> than table data.\\n> \\n> > > The coverage report disagrees with me, though...\\n> > > https://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#3901\\n> > \\n> > Isn't that one of the tests like compression_gzip_plain?\\n> \\n> \\n> I'm not sure what you mean. Plain dump is restored with psql and not\\n> with pg_restore.\\n> \\n> My line number was wrong:\\n> https://coverage.postgresql.org/src/bin/pg_dump/pg_backup_archiver.c.gcov.html#390\\n> \\n> What test would hit that code without rebuilding ?\\n> \\n> 394 : #ifndef HAVE_LIBZ\\n> 395 : if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP &&\\n> \\n> > Thoughts?\\n> > #ifndef HAVE_LIBZ\\n> > if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> > - pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n> > + pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n> \\n> \\n> Your patch is fine for now, but these errors should eventually specify\\n> which compression algorithm is unavailable. I think that should be a\\n> part of the 001 patch, ideally in a way that minimizes the number of\\n> places which need to be updated when adding an algorithm.\\n\\nI gave this a little bit of thought. I think that ReadHead should not\\nemit a warning, or at least not this warning as it is slightly misleading.\\nIt implies that it will automatically turn off data restoration, which is\\nfalse. Further ahead, the code will fail with a conflicting error message\\nstating that the compression is not available.\\n\\nInstead, it would be cleaner both for the user and the maintainer to\\nmove the check in RestoreArchive and make it the sole responsible for\\nthis logic.\\n\\nPlease find v26 attached. 0001 does the above and 0002 addresses Justin's\\ncomplaints regarding the code footprint.\\n\\n//Cheers,\\nGeorgios \\n\\n\\n> \\n> --\\n> Justin\",\n \"msg_date\": \"Thu, 26 Jan 2023 11:24:47 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 26, 2023 at 11:24:47AM +0000, gkokolatos@pm.me wrote:\\n> I gave this a little bit of thought. I think that ReadHead should not\\n> emit a warning, or at least not this warning as it is slightly misleading.\\n> It implies that it will automatically turn off data restoration, which is\\n> false. Further ahead, the code will fail with a conflicting error message\\n> stating that the compression is not available.\\n> \\n> Instead, it would be cleaner both for the user and the maintainer to\\n> move the check in RestoreArchive and make it the sole responsible for\\n> this logic.\\n\\n- pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n+ pg_fatal(\\\"cannot restore data from compressed archive (compression not supported in this installation)\\\");\\nHmm. I don't mind changing this part as you suggest.\\n\\n-#ifndef HAVE_LIBZ\\n- if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n- pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n-#endif\\nHowever I think that we'd better keep the warning, as it can offer a\\nhint when using pg_restore -l not built with compression support if\\nlooking at a dump that has been compressed.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 26 Jan 2023 20:53:28 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Thursday, January 26th, 2023 at 12:53 PM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Thu, Jan 26, 2023 at 11:24:47AM +0000, gkokolatos@pm.me wrote:\\n> \\n> > I gave this a little bit of thought. I think that ReadHead should not\\n> > emit a warning, or at least not this warning as it is slightly misleading.\\n> > It implies that it will automatically turn off data restoration, which is\\n> > false. Further ahead, the code will fail with a conflicting error message\\n> > stating that the compression is not available.\\n> > \\n> > Instead, it would be cleaner both for the user and the maintainer to\\n> > move the check in RestoreArchive and make it the sole responsible for\\n> > this logic.\\n> \\n> \\n> - pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n> + pg_fatal(\\\"cannot restore data from compressed archive (compression not supported in this installation)\\\");\\n> Hmm. I don't mind changing this part as you suggest.\\n> \\n> -#ifndef HAVE_LIBZ\\n> - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> \\n> - pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n> -#endif\\n> However I think that we'd better keep the warning, as it can offer a\\n> hint when using pg_restore -l not built with compression support if\\n> looking at a dump that has been compressed.\\n\\nFair enough. Please find v27 attached.\\n\\nCheers,\\n//Georgios\\n\\n\\n> --\\n> Michael\",\n \"msg_date\": \"Thu, 26 Jan 2023 18:03:59 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Jan 25, 2023 at 07:57:18PM +0000, gkokolatos@pm.me wrote:\\n> On Wednesday, January 25th, 2023 at 7:00 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > While looking at this, I realized that commit 5e73a6048 introduced a\\n> > regression:\\n> > \\n> > @@ -3740,19 +3762,24 @@ ReadHead(ArchiveHandle *AH)\\n> > \\n> > - if (AH->compression != 0)\\n> > \\n> > - pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n> > + if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> > \\n> > + pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n> > \\n> > Before, it was possible to restore non-data chunks of a dump file, even\\n> > if the current build didn't support its compression. But that's now\\n> > impossible - and it makes the code we're discussing in RestoreArchive()\\n> > unreachable.\\n\\nOn Thu, Jan 26, 2023 at 08:53:28PM +0900, Michael Paquier wrote:\\n> On Thu, Jan 26, 2023 at 11:24:47AM +0000, gkokolatos@pm.me wrote:\\n> > I gave this a little bit of thought. I think that ReadHead should not\\n> > emit a warning, or at least not this warning as it is slightly misleading.\\n> > It implies that it will automatically turn off data restoration, which is\\n> > false. Further ahead, the code will fail with a conflicting error message\\n> > stating that the compression is not available.\\n> > \\n> > Instead, it would be cleaner both for the user and the maintainer to\\n> > move the check in RestoreArchive and make it the sole responsible for\\n> > this logic.\\n> \\n> - pg_fatal(\\\"cannot restore from compressed archive (compression not supported in this installation)\\\");\\n> + pg_fatal(\\\"cannot restore data from compressed archive (compression not supported in this installation)\\\");\\n> Hmm. I don't mind changing this part as you suggest.\\n> \\n> -#ifndef HAVE_LIBZ\\n> - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> - pg_fatal(\\\"archive is compressed, but this installation does not support compression\\\");\\n> -#endif\\n> However I think that we'd better keep the warning, as it can offer a\\n> hint when using pg_restore -l not built with compression support if\\n> looking at a dump that has been compressed.\\n\\nYeah. But the original log_warning text was better, and should be\\nrestored:\\n\\n- if (AH->compression != 0)\\n- pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n\\nThat commit also added this to pg-dump.c:\\n\\n+ case PG_COMPRESSION_ZSTD:\\n+ pg_fatal(\\\"compression with %s is not yet supported\\\", \\\"ZSTD\\\");\\n+ break;\\n+ case PG_COMPRESSION_LZ4:\\n+ pg_fatal(\\\"compression with %s is not yet supported\\\", \\\"LZ4\\\");\\n+ break;\\n\\nIn 002, that could be simplified by re-using the supports_compression()\\nfunction. (And maybe the same in WriteDataToArchive()?)\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 Jan 2023 12:22:45 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 26, 2023 at 12:22:45PM -0600, Justin Pryzby wrote:\\n> Yeah. But the original log_warning text was better, and should be\\n> restored:\\n> \\n> - if (AH->compression != 0)\\n> - pg_log_warning(\\\"archive is compressed, but this installation does not support compression -- no data will be available\\\");\\n\\nYeah, this one's on me. So I have gone with the simplest solution and\\napplied a fix to restore the original behavior, with the same warning\\nshowing up.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 27 Jan 2023 10:23:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Jan 16, 2023 at 11:54:46AM +0900, Michael Paquier wrote:\\n> On Sun, Jan 15, 2023 at 07:56:25PM -0600, Justin Pryzby wrote:\\n> > On Mon, Jan 16, 2023 at 10:28:50AM +0900, Michael Paquier wrote:\\n> >> The functions changed by 0001 are cfopen[_write](),\\n> >> AllocateCompressor() and ReadDataFromArchive(). Why is it a good idea\\n> >> to change these interfaces which basically exist to handle inputs?\\n> > \\n> > I changed to pass pg_compress_specification as a pointer, since that's\\n> > the usual convention for structs, as followed by the existing uses of\\n> > pg_compress_specification.\\n> \\n> Okay, but what do we gain here? It seems to me that this introduces\\n> the risk that a careless change in one of the internal routines if\\n> they change slight;ly compress_spec, hence impacting any of their\\n> callers? Or is that fixing an actual bug (except if I am missing your\\n> point, that does not seem to be the case)? \\n\\nTo circle back to this: I was not saying there's any bug. The proposed\\nchange was only to follow normal and existing normal conventions for\\npassing structs. It could also be a pointer to const. It's fine with\\nme if you say that it's intentional how it's written already.\\n\\n> >> Is there some benefit in changing compression_spec within the\\n> >> internals of these routines before going back one layer down to their\\n> >> callers? Changing the compression_spec on-the-fly in these internal\\n> >> paths could be risky, actually, no?\\n> > \\n> > I think what you're saying is that if the spec is passed as a pointer,\\n> > then the called functions shouldn't set spec->algorithm=something.\\n> \\n> Yes. HEAD makes sure of that, 0001 would not prevent that. So I am a\\n> bit confused in seeing how this is a benefit.\\n\\n\\n\",\n \"msg_date\": \"Thu, 26 Jan 2023 21:56:32 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Jan 26, 2023 at 12:22:45PM -0600, Justin Pryzby wrote:\\n> That commit also added this to pg-dump.c:\\n> \\n> + case PG_COMPRESSION_ZSTD:\\n> + pg_fatal(\\\"compression with %s is not yet supported\\\", \\\"ZSTD\\\");\\n> + break;\\n> + case PG_COMPRESSION_LZ4:\\n> + pg_fatal(\\\"compression with %s is not yet supported\\\", \\\"LZ4\\\");\\n> + break;\\n> \\n> In 002, that could be simplified by re-using the supports_compression()\\n> function. (And maybe the same in WriteDataToArchive()?)\\n\\nThe first patch aims to minimize references to \\\".gz\\\" and \\\"GZIP\\\" and\\nZLIB. pg_backup_directory.c comments still refers to \\\".gz\\\". I think\\nthe patch should ideally change to refer to \\\"the compressed file\\nextension\\\" (similar to compress_io.c), avoiding the need to update it\\nlater.\\n\\nI think the file extension stuff could be generalized, so it doesn't\\nneed to be updated in multiple places (pg_backup_directory.c and\\ncompress_io.c). Maybe it's useful to add a function to return the\\nextension of a given compression method. It could go in compression.c,\\nand be useful in basebackup.\\n\\nFor the 2nd patch:\\n\\nI might be in the minority, but I still think some references to \\\"gzip\\\"\\nshould say \\\"zlib\\\":\\n\\n+} GzipCompressorState;\\n+\\n+/* Private routines that support gzip compressed data I/O */\\n+static void\\n+DeflateCompressorGzip(ArchiveHandle *AH, CompressorState *cs, bool flush)\\n\\nIn my mind, three things here are misleading, because it doesn't use\\ngzip headers:\\n\\n| GzipCompressorState, DeflateCompressorGzip, \\\"gzip compressed\\\".\\n\\nThis comment is about exactly that:\\n\\n * underlying stream. The second API is a wrapper around fopen/gzopen and\\n * friends, providing an interface similar to those, but abstracts away\\n * the possible compression. Both APIs use libz for the compression, but\\n * the second API uses gzip headers, so the resulting files can be easily\\n * manipulated with the gzip utility.\\n\\nAIUI, Michael says that it's fine that the user-facing command-line\\noptions use \\\"-Z gzip\\\" (even though the \\\"custom\\\" format doesn't use gzip\\nheaders). I'm okay with that, as long as that's discussed/understood.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Fri, 27 Jan 2023 11:23:20 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Friday, January 27th, 2023 at 6:23 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n> \\n> \\n> On Thu, Jan 26, 2023 at 12:22:45PM -0600, Justin Pryzby wrote:\\n> \\n> > That commit also added this to pg-dump.c:\\n> > \\n> > + case PG_COMPRESSION_ZSTD:\\n> > + pg_fatal(\\\"compression with %s is not yet supported\\\", \\\"ZSTD\\\");\\n> > + break;\\n> > + case PG_COMPRESSION_LZ4:\\n> > + pg_fatal(\\\"compression with %s is not yet supported\\\", \\\"LZ4\\\");\\n> > + break;\\n> > \\n> > In 002, that could be simplified by re-using the supports_compression()\\n> > function. (And maybe the same in WriteDataToArchive()?)\\n> \\n> \\n> The first patch aims to minimize references to \\\".gz\\\" and \\\"GZIP\\\" and\\n> ZLIB. pg_backup_directory.c comments still refers to \\\".gz\\\". I think\\n> the patch should ideally change to refer to \\\"the compressed file\\n> extension\\\" (similar to compress_io.c), avoiding the need to update it\\n> later.\\n> \\n> I think the file extension stuff could be generalized, so it doesn't\\n> need to be updated in multiple places (pg_backup_directory.c and\\n> compress_io.c). Maybe it's useful to add a function to return the\\n> extension of a given compression method. It could go in compression.c,\\n> and be useful in basebackup.\\n> \\n> For the 2nd patch:\\n> \\n> I might be in the minority, but I still think some references to \\\"gzip\\\"\\n> should say \\\"zlib\\\":\\n> \\n> +} GzipCompressorState;\\n> +\\n> +/* Private routines that support gzip compressed data I/O */\\n> +static void\\n> +DeflateCompressorGzip(ArchiveHandle *AH, CompressorState *cs, bool flush)\\n> \\n> In my mind, three things here are misleading, because it doesn't use\\n> gzip headers:\\n> \\n> | GzipCompressorState, DeflateCompressorGzip, \\\"gzip compressed\\\".\\n> \\n> This comment is about exactly that:\\n> \\n> * underlying stream. The second API is a wrapper around fopen/gzopen and\\n> * friends, providing an interface similar to those, but abstracts away\\n> * the possible compression. Both APIs use libz for the compression, but\\n> * the second API uses gzip headers, so the resulting files can be easily\\n> * manipulated with the gzip utility.\\n> \\n> AIUI, Michael says that it's fine that the user-facing command-line\\n> options use \\\"-Z gzip\\\" (even though the \\\"custom\\\" format doesn't use gzip\\n> headers). I'm okay with that, as long as that's discussed/understood.\\n> \\n\\nThank you for the input Justin. I am currently waiting for input from a\\nthird person to get some conclusion. I thought that it should be stated\\nbefore my inactiveness is considered as indifference, which is not.\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Justin\\n\\n\\n\",\n \"msg_date\": \"Tue, 31 Jan 2023 09:00:56 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Jan 31, 2023 at 09:00:56AM +0000, gkokolatos@pm.me wrote:\\n> > In my mind, three things here are misleading, because it doesn't use\\n> > gzip headers:\\n> > \\n> > | GzipCompressorState, DeflateCompressorGzip, \\\"gzip compressed\\\".\\n> > \\n> > This comment is about exactly that:\\n> > \\n> > * underlying stream. The second API is a wrapper around fopen/gzopen and\\n> > * friends, providing an interface similar to those, but abstracts away\\n> > * the possible compression. Both APIs use libz for the compression, but\\n> > * the second API uses gzip headers, so the resulting files can be easily\\n> > * manipulated with the gzip utility.\\n> > \\n> > AIUI, Michael says that it's fine that the user-facing command-line\\n> > options use \\\"-Z gzip\\\" (even though the \\\"custom\\\" format doesn't use gzip\\n> > headers). I'm okay with that, as long as that's discussed/understood.\\n> \\n> Thank you for the input Justin. I am currently waiting for input from a\\n> third person to get some conclusion. I thought that it should be stated\\n> before my inactiveness is considered as indifference, which is not.\\n\\nI'm not sure what there is to lose by making the names more accurate -\\nespecially since they're private/internal-only.\\n\\nTomas marked himself as a committer, so maybe could comment.\\n\\nIt'd be nice to also come to some conclusion about whether -Fc -Z gzip\\nis confusing (due to not actually using gzip).\\n\\nBTW, do you intend to merge this for v16 ? I verified in earlier patch\\nversions that tests all pass with lz4 as the default compression method.\\nAnd checked that gzip output is compatible with before, and that old\\ndumps restore correctly, and there's no memory leaks or other errors.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Tue, 7 Feb 2023 20:37:18 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Jan 27, 2023 2:04 AM gkokolatos@pm.me <gkokolatos@pm.me> wrote:\\r\\n> \\r\\n> ------- Original Message -------\\r\\n> On Thursday, January 26th, 2023 at 12:53 PM, Michael Paquier\\r\\n> <michael@paquier.xyz> wrote:\\r\\n> \\r\\n> \\r\\n> >\\r\\n> >\\r\\n> > On Thu, Jan 26, 2023 at 11:24:47AM +0000, gkokolatos@pm.me wrote:\\r\\n> >\\r\\n> > > I gave this a little bit of thought. I think that ReadHead should not\\r\\n> > > emit a warning, or at least not this warning as it is slightly misleading.\\r\\n> > > It implies that it will automatically turn off data restoration, which is\\r\\n> > > false. Further ahead, the code will fail with a conflicting error message\\r\\n> > > stating that the compression is not available.\\r\\n> > >\\r\\n> > > Instead, it would be cleaner both for the user and the maintainer to\\r\\n> > > move the check in RestoreArchive and make it the sole responsible for\\r\\n> > > this logic.\\r\\n> >\\r\\n> >\\r\\n> > - pg_fatal(\\\"cannot restore from compressed archive (compression not\\r\\n> supported in this installation)\\\");\\r\\n> > + pg_fatal(\\\"cannot restore data from compressed archive (compression not\\r\\n> supported in this installation)\\\");\\r\\n> > Hmm. I don't mind changing this part as you suggest.\\r\\n> >\\r\\n> > -#ifndef HAVE_LIBZ\\r\\n> > - if (AH->compression_spec.algorithm == PG_COMPRESSION_GZIP)\\r\\n> >\\r\\n> > - pg_fatal(\\\"archive is compressed, but this installation does not support\\r\\n> compression\\\");\\r\\n> > -#endif\\r\\n> > However I think that we'd better keep the warning, as it can offer a\\r\\n> > hint when using pg_restore -l not built with compression support if\\r\\n> > looking at a dump that has been compressed.\\r\\n> \\r\\n> Fair enough. Please find v27 attached.\\r\\n> \\r\\n\\r\\nHi,\\r\\n\\r\\nI am interested in this feature and tried the patch. While reading the comments,\\r\\nI noticed some minor things that could possibly be improved (in v27-0003 patch).\\r\\n\\r\\n1.\\r\\n+\\t/*\\r\\n+\\t * Open a file for writing.\\r\\n+\\t *\\r\\n+\\t * 'mode' can be one of ''w', 'wb', 'a', and 'ab'. Requrires an already\\r\\n+\\t * initialized CompressFileHandle.\\r\\n+\\t */\\r\\n+\\tint\\t\\t\\t(*open_write_func) (const char *path, const char *mode,\\r\\n+\\t\\t\\t\\t\\t\\t\\t\\t\\tCompressFileHandle *CFH);\\r\\n\\r\\nThere is a redundant single quote in front of 'w'.\\r\\n\\r\\n2.\\r\\n/*\\r\\n * Callback function for WriteDataToArchive. Writes one block of (compressed)\\r\\n * data to the archive.\\r\\n */\\r\\n/*\\r\\n * Callback function for ReadDataFromArchive. To keep things simple, we\\r\\n * always read one compressed block at a time.\\r\\n */\\r\\n\\r\\nShould the function names in the comments be updated?\\r\\n\\r\\nWriteDataToArchive\\r\\n->\\r\\nwriteData\\r\\n\\r\\nReadDataFromArchive\\r\\n->\\r\\nreadData\\r\\n\\r\\n3.\\r\\n+\\tAssert(strcmp(mode, \\\"r\\\") == 0 || strcmp(mode, \\\"rb\\\") == 0);\\r\\n\\r\\nCould we use PG_BINARY_R instead of \\\"r\\\" and \\\"rb\\\" here?\\r\\n\\r\\nRegards,\\r\\nShi Yu\\r\\n\\r\\n\",\n \"msg_date\": \"Wed, 15 Feb 2023 13:51:21 +0000\",\n \"msg_from\": \"\\\"shiy.fnst@fujitsu.com\\\" <shiy.fnst@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Wednesday, February 15th, 2023 at 2:51 PM, shiy.fnst@fujitsu.com <shiy.fnst@fujitsu.com> wrote:\\n\\n\\n> \\n> Hi,\\n> \\n> I am interested in this feature and tried the patch. While reading the comments,\\n> I noticed some minor things that could possibly be improved (in v27-0003 patch).\\n\\nThank you very much for the interest. Please find a rebased v28 attached. Due to\\nthe rebase, 0001 of v27 is no longer relevant and has been removed. Your comments\\nare applied on v28-0002.\\n\\n> \\n> 1.\\n> + /*\\n> + * Open a file for writing.\\n> + *\\n> + * 'mode' can be one of ''w', 'wb', 'a', and 'ab'. Requrires an already\\n> + * initialized CompressFileHandle.\\n> + */\\n> + int (*open_write_func) (const char *path, const char *mode,\\n> + CompressFileHandle CFH);\\n> \\n> There is a redundant single quote in front of 'w'.\\n\\nFixed.\\n\\n> \\n> 2.\\n> /\\n> * Callback function for WriteDataToArchive. Writes one block of (compressed)\\n> * data to the archive.\\n> /\\n> /\\n> * Callback function for ReadDataFromArchive. To keep things simple, we\\n> * always read one compressed block at a time.\\n> */\\n> \\n> Should the function names in the comments be updated?\\n\\nAgreed. Fixed.\\n\\n> \\n> 3.\\n> + Assert(strcmp(mode, \\\"r\\\") == 0 || strcmp(mode, \\\"rb\\\") == 0);\\n> \\n> Could we use PG_BINARY_R instead of \\\"r\\\" and \\\"rb\\\" here?\\n\\nWe could and we should. Using PG_BINARY_R has the added benefit\\nof needing only one strcmp() call. Fixed.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> Regards,\\n> Shi Yu\",\n \"msg_date\": \"Thu, 16 Feb 2023 11:38:00 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hi Georgios,\\n\\nI spent some time looking at the patch again, and IMO it's RFC. But I\\nneed some help with the commit messages - I updated 0001 and 0002 but I\\nwasn't quite sure what some of the stuff meant to say and/or it seemed\\nmaybe coming from an earlier patch version and obsolete.\\n\\nCould you go over them and check if I got it right? Also feel free to\\nupdate the list of reviewers (I compiled that from substantial reviews\\non the thread).\\n\\nThe 0003 commit message seems somewhat confusing - I added some XXX\\nlines asking about unclear stuff.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Sun, 19 Feb 2023 18:10:13 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Some little updates since I last checked:\\n\\n+ * This file also includes the implementation when compression is none for\\n+ * both API's.\\n\\n=> this comment is obsolete.\\n\\ns/deffer/infer/ ?\\nor determine ?\\nThis typo occurs multiple times.\\n\\ncurrently this includes only \\\".gz\\\"\\n=> remove this phase from the 002 patch (or at least update it in 003).\\n\\ndeferred by iteratively\\n=> inferred?\\n\\ns/Requrires/Requires/\\ntwice.\\n\\ns/occured/occurred/\\n\\ns/disc/disk/ ?\\nProbably unimportant, but \\\"disc\\\" isn't used anywhere else.\\n\\n\\\"compress file handle\\\"\\n=> maybe these should say \\\"compressed\\\"\\n\\nsupports_compression():\\nSince this is an exported function, it should probably be called\\npgdump_supports_compresion.\\n\\n\\n\",\n \"msg_date\": \"Sun, 19 Feb 2023 13:43:45 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Sunday, February 19th, 2023 at 6:10 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n>\\n>\\n> Hi Georgios,\\n>\\n> I spent some time looking at the patch again, and IMO it's RFC. But I\\n> need some help with the commit messages - I updated 0001 and 0002 but I\\n> wasn't quite sure what some of the stuff meant to say and/or it seemed\\n> maybe coming from an earlier patch version and obsolete.\\n\\nThank you very much Tomas! Indeed I have not being paying any attention\\nat the commit messages.\\n\\n> Could you go over them and check if I got it right? Also feel free to\\n> update the list of reviewers (I compiled that from substantial reviews\\n> on the thread).\\n\\nDone. Rachel has been correctly identified as author in the relevant parts\\nup to commit 98fe74218d. After that, she had offered review comments and I\\nhave taken the liberty to add her as a reviewer through out.\\n\\nAlso I think that Shi Yu should be credited as a reviewer of 0003.\\n\\n>\\n> The 0003 commit message seems somewhat confusing - I added some XXX\\n> lines asking about unclear stuff.\\n\\nPlease find in the attached v30 an updated message, as well as an amended\\nreviewer list. Also v30 addresses the final comments raised by Justin.\\n\\nCheers,\\n//Georgios\\n\\n> regards\\n>\\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 20 Feb 2023 13:09:22 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Thanks for v30 with the updated commit messages. I've pushed 0001 after\\nfixing a comment typo and removing (I think) an unnecessary change in an\\nerror message.\\n\\nI'll give the buildfarm a bit of time before pushing 0002 and 0003.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 23 Feb 2023 16:26:35 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 2/23/23 16:26, Tomas Vondra wrote:\\n> Thanks for v30 with the updated commit messages. I've pushed 0001 after\\n> fixing a comment typo and removing (I think) an unnecessary change in an\\n> error message.\\n> \\n> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\\n> \\n\\nI've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\nand marked the CF entry as committed. Thanks for the patch!\\n\\nI wonder how difficult would it be to add the zstd compression, so that\\nwe don't have the annoying \\\"unsupported\\\" cases.\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 23 Feb 2023 21:24:46 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 23, 2023 at 09:24:46PM +0100, Tomas Vondra wrote:\\n> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\n> and marked the CF entry as committed. Thanks for the patch!\\n\\nA big thanks from me to everyone involved.\\n\\n> I wonder how difficult would it be to add the zstd compression, so that\\n> we don't have the annoying \\\"unsupported\\\" cases.\\n\\nI'll send a patch soon. I first submitted patches for that 2 years ago\\n(before PGDG was ready to add zstd).\\nhttps://commitfest.postgresql.org/31/2888/\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Thu, 23 Feb 2023 19:51:16 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 23, 2023 at 07:51:16PM -0600, Justin Pryzby wrote:\\n> On Thu, Feb 23, 2023 at 09:24:46PM +0100, Tomas Vondra wrote:\\n>> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\n>> and marked the CF entry as committed. Thanks for the patch!\\n> \\n> A big thanks from me to everyone involved.\\n\\nWow, nice! The APIs are clear to follow.\\n\\n> I'll send a patch soon. I first submitted patches for that 2 years ago\\n> (before PGDG was ready to add zstd).\\n> https://commitfest.postgresql.org/31/2888/\\n\\nThanks. It should be straight-forward to see that in 16, I guess.\\n--\\nMichael\",\n \"msg_date\": \"Fri, 24 Feb 2023 13:35:14 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Friday, February 24th, 2023 at 5:35 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Thu, Feb 23, 2023 at 07:51:16PM -0600, Justin Pryzby wrote:\\n> \\n> > On Thu, Feb 23, 2023 at 09:24:46PM +0100, Tomas Vondra wrote:\\n> > \\n> > > I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\n> > > and marked the CF entry as committed. Thanks for the patch!\\n> > \\n> > A big thanks from me to everyone involved.\\n> \\n> \\n> Wow, nice! The APIs are clear to follow.\\n\\nI am out of words, thank you all so very much. I learned a lot. \\n\\n> \\n> > I'll send a patch soon. I first submitted patches for that 2 years ago\\n> > (before PGDG was ready to add zstd).\\n> > https://commitfest.postgresql.org/31/2888/\\n> \\n> \\n> Thanks. It should be straight-forward to see that in 16, I guess.\\n> --\\n> Michael\\n\\n\\n\",\n \"msg_date\": \"Fri, 24 Feb 2023 09:21:32 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"I have some fixes (attached) and questions while polishing the patch for\\nzstd compression. The fixes are small and could be integrated with the\\npatch for zstd, but could be applied independently.\\n\\n- I'm unclear about get_error_func(). That's called in three places\\n from pg_backup_directory.c, after failures from write_func(), to\\n supply an compression-specific error message to pg_fatal(). But it's\\n not being used outside of directory format, nor for errors for other\\n function pointers, or even for all errors in write_func(). Is there\\n some reason why each compression method's write_func() shouldn't call\\n pg_fatal() directly, with its compression-specific message ?\\n\\n- I still think supports_compression() should be renamed, or made into a\\n static function in the necessary file. The main reason is that it's\\n more clear what it indicates - whether compression is \\\"implemented by\\n pgdump\\\" and not whether compression is \\\"supported by this postgres\\n build\\\". It also seems possible that we'd want to add a function\\n called something like supports_compression(), indicating whether the\\n algorithm is supported by the current build. It'd be better if pgdump\\n didn't subjugate that name.\\n\\n- Finally, the \\\"Nothing to do in the default case\\\" comment comes from\\n Michael's commit 5e73a6048:\\n\\n+ /*\\n+ * Custom and directory formats are compressed by default with gzip when\\n+ * available, not the others.\\n+ */\\n+ if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\\n+ !user_compression_defined)\\n {\\n #ifdef HAVE_LIBZ\\n- if (archiveFormat == archCustom || archiveFormat == archDirectory)\\n- compressLevel = Z_DEFAULT_COMPRESSION;\\n- else\\n+ parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\\n+ &compression_spec);\\n+#else\\n+ /* Nothing to do in the default case */\\n #endif\\n- compressLevel = 0;\\n }\\n\\n\\nAs the comment says: for -Fc and -Fd, the compression is set to zlib, if\\nenabled, and when not otherwise specified by the user.\\n\\nBefore 5e73a6048, this set compressLevel=0 for -Fp and -Ft, *and* when\\nzlib was unavailable.\\n\\nBut I'm not sure why there's now an empty \\\"#else\\\". I also don't know\\nwhat \\\"the default case\\\" refers to.\\n\\nMaybe the best thing here is to move the preprocessor #if, since it's no\\nlonger in the middle of a runtime conditional:\\n\\n #ifdef HAVE_LIBZ\\n+ if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\\n+ !user_compression_defined)\\n+ parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\\n+ &compression_spec);\\n #endif\\n\\n...but that elicits a warning about \\\"variable set but not used\\\"...\\n\\n-- \\nJustin\",\n \"msg_date\": \"Fri, 24 Feb 2023 23:02:14 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n> I have some fixes (attached) and questions while polishing the patch for\\n> zstd compression. The fixes are small and could be integrated with the\\n> patch for zstd, but could be applied independently.\\n\\nOne more - WriteDataToArchiveGzip() says:\\n\\n+ if (cs->compression_spec.level == 0)\\n+ pg_fatal(\\\"requested to compress the archive yet no level was specified\\\");\\n\\nThat was added at e9960732a. \\n\\nBut if you specify gzip:0, the compression level is already enforced by\\nvalidate_compress_specification(), before hitting gzip.c:\\n\\n| pg_dump: error: invalid compression specification: compression algorithm \\\"gzip\\\" expects a compression level between 1 and 9 (default at -1)\\n\\n5e73a6048 intended that to work as before, and you *can* specify -Z0:\\n\\n The change is backward-compatible, hence specifying only an integer\\n leads to no compression for a level of 0 and gzip compression when the\\n level is greater than 0.\\n\\n $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\\n /dev/stdin: ASCII text\\n\\nRight now, I think that pg_fatal in gzip.c is dead code - that was first\\nadded in the patch version sent on 21 Dec 2022.\\n\\n-- \\nJustin\",\n \"msg_date\": \"Sat, 25 Feb 2023 08:05:53 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2/25/23 06:02, Justin Pryzby wrote:\\n> I have some fixes (attached) and questions while polishing the patch for\\n> zstd compression. The fixes are small and could be integrated with the\\n> patch for zstd, but could be applied independently.\\n> \\n> - I'm unclear about get_error_func(). That's called in three places\\n> from pg_backup_directory.c, after failures from write_func(), to\\n> supply an compression-specific error message to pg_fatal(). But it's\\n> not being used outside of directory format, nor for errors for other\\n> function pointers, or even for all errors in write_func(). Is there\\n> some reason why each compression method's write_func() shouldn't call\\n> pg_fatal() directly, with its compression-specific message ?\\n> \\n\\nI think there are a couple more places that might/should call\\nget_error_func(). For example ahwrite() in pg_backup_archiver.c now\\nsimply does\\n\\n if (bytes_written != size * nmemb)\\n WRITE_ERROR_EXIT;\\n\\nbut perhaps it should call get_error_func() too. There are probably\\nother places that call write_func() and should use get_error_func().\\n\\n> - I still think supports_compression() should be renamed, or made into a\\n> static function in the necessary file. The main reason is that it's\\n> more clear what it indicates - whether compression is \\\"implemented by\\n> pgdump\\\" and not whether compression is \\\"supported by this postgres\\n> build\\\". It also seems possible that we'd want to add a function\\n> called something like supports_compression(), indicating whether the\\n> algorithm is supported by the current build. It'd be better if pgdump\\n> didn't subjugate that name.\\n> \\n\\nIf we choose to rename this to have pgdump_ prefix, fine with me. But I\\ndon't think there's a realistic chance of conflict, as it's restricted\\nto pgdump header etc. And it's not part of an API, so I guess we could\\nrename that in the future if needed.\\n\\n> - Finally, the \\\"Nothing to do in the default case\\\" comment comes from\\n> Michael's commit 5e73a6048:\\n> \\n> + /*\\n> + * Custom and directory formats are compressed by default with gzip when\\n> + * available, not the others.\\n> + */\\n> + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\\n> + !user_compression_defined)\\n> {\\n> #ifdef HAVE_LIBZ\\n> - if (archiveFormat == archCustom || archiveFormat == archDirectory)\\n> - compressLevel = Z_DEFAULT_COMPRESSION;\\n> - else\\n> + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\\n> + &compression_spec);\\n> +#else\\n> + /* Nothing to do in the default case */\\n> #endif\\n> - compressLevel = 0;\\n> }\\n> \\n> \\n> As the comment says: for -Fc and -Fd, the compression is set to zlib, if\\n> enabled, and when not otherwise specified by the user.\\n> \\n> Before 5e73a6048, this set compressLevel=0 for -Fp and -Ft, *and* when\\n> zlib was unavailable.\\n> \\n> But I'm not sure why there's now an empty \\\"#else\\\". I also don't know\\n> what \\\"the default case\\\" refers to.\\n> \\n> Maybe the best thing here is to move the preprocessor #if, since it's no\\n> longer in the middle of a runtime conditional:\\n> \\n> #ifdef HAVE_LIBZ\\n> + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\\n> + !user_compression_defined)\\n> + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\\n> + &compression_spec);\\n> #endif\\n> \\n> ...but that elicits a warning about \\\"variable set but not used\\\"...\\n> \\n\\nNot sure, I need to think about this a bit.\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sun, 26 Feb 2023 15:59:41 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 25, 2023 at 08:05:53AM -0600, Justin Pryzby wrote:\\n> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n> > I have some fixes (attached) and questions while polishing the patch for\\n> > zstd compression. The fixes are small and could be integrated with the\\n> > patch for zstd, but could be applied independently.\\n> \\n> One more - WriteDataToArchiveGzip() says:\\n\\nOne more again.\\n\\nThe LZ4 path is using non-streaming mode, which compresses each block\\nwithout persistent state, giving poor compression for -Fc compared with\\n-Fp. If the data is highly compressible, the difference can be orders\\nof magnitude.\\n\\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fp |wc -c\\n12351763\\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\\n21890708\\n\\nThat's not true for gzip:\\n\\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fc |wc -c\\n2118869\\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fp |wc -c\\n2115832\\n\\nThe function ought to at least use streaming mode, so each block/row\\nisn't compressioned in isolation. 003 is a simple patch to use\\nstreaming mode, which improves the -Fc case:\\n\\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\\n15178283\\n\\nHowever, that still flushes the compression buffer, writing a block\\nheader, for every row. With a single-column table, pg_dump -Fc -Z lz4\\nstill outputs ~10% *more* data than with no compression at all. And\\nthat's for compressible data.\\n\\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z lz4 |wc -c\\n12890296\\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z none |wc -c\\n11890296\\n\\nI think this should use the LZ4F API with frames, which are buffered to\\navoid outputting a header for every single row. The LZ4F format isn't\\ncompatible with the LZ4 format, so (unlike changing to the streaming\\nAPI) that's not something we can change in a bugfix release. I consider\\nthis an Opened Item.\\n\\nWith the LZ4F API in 004, -Fp and -Fc are essentially the same size\\n(like gzip). (Oh, and the output is three times smaller, too.)\\n\\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fp |wc -c\\n4155448\\n$ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fc |wc -c\\n4156548\\n\\n-- \\nJustin\",\n \"msg_date\": \"Sun, 26 Feb 2023 22:49:10 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Sunday, February 26th, 2023 at 3:59 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> \\n> \\n> On 2/25/23 06:02, Justin Pryzby wrote:\\n> \\n> > I have some fixes (attached) and questions while polishing the patch for\\n> > zstd compression. The fixes are small and could be integrated with the\\n> > patch for zstd, but could be applied independently.\\n> > \\n> > - I'm unclear about get_error_func(). That's called in three places\\n> > from pg_backup_directory.c, after failures from write_func(), to\\n> > supply an compression-specific error message to pg_fatal(). But it's\\n> > not being used outside of directory format, nor for errors for other\\n> > function pointers, or even for all errors in write_func(). Is there\\n> > some reason why each compression method's write_func() shouldn't call\\n> > pg_fatal() directly, with its compression-specific message ?\\n> \\n> \\n> I think there are a couple more places that might/should call\\n> get_error_func(). For example ahwrite() in pg_backup_archiver.c now\\n> simply does\\n> \\n> if (bytes_written != size * nmemb)\\n> WRITE_ERROR_EXIT;\\n> \\n> but perhaps it should call get_error_func() too. There are probably\\n> other places that call write_func() and should use get_error_func().\\n\\nAgreed, calling get_error_func() would be preferable to a fatal error. It\\nshould be the caller of the api who decides how to proceed.\\n\\n> \\n> > - I still think supports_compression() should be renamed, or made into a\\n> > static function in the necessary file. The main reason is that it's\\n> > more clear what it indicates - whether compression is \\\"implemented by\\n> > pgdump\\\" and not whether compression is \\\"supported by this postgres\\n> > build\\\". It also seems possible that we'd want to add a function\\n> > called something like supports_compression(), indicating whether the\\n> > algorithm is supported by the current build. It'd be better if pgdump\\n> > didn't subjugate that name.\\n> \\n> \\n> If we choose to rename this to have pgdump_ prefix, fine with me. But I\\n> don't think there's a realistic chance of conflict, as it's restricted\\n> to pgdump header etc. And it's not part of an API, so I guess we could\\n> rename that in the future if needed.\\n\\nAgreed, it is very unrealistic that one will include that header file anywhere\\nbut within pg_dump. Also. I think that adding a prefix, \\\"pgdump\\\", \\\"pg_dump\\\",\\nor similar does not add value and subtracts readability. \\n\\n> \\n> > - Finally, the \\\"Nothing to do in the default case\\\" comment comes from\\n> > Michael's commit 5e73a6048:\\n> > \\n> > + /*\\n> > + * Custom and directory formats are compressed by default with gzip when\\n> > + * available, not the others.\\n> > + /\\n> > + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\\n> > + !user_compression_defined)\\n> > {\\n> > #ifdef HAVE_LIBZ\\n> > - if (archiveFormat == archCustom || archiveFormat == archDirectory)\\n> > - compressLevel = Z_DEFAULT_COMPRESSION;\\n> > - else\\n> > + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\\n> > + &compression_spec);\\n> > +#else\\n> > + / Nothing to do in the default case */\\n> > #endif\\n> > - compressLevel = 0;\\n> > }\\n> > \\n> > As the comment says: for -Fc and -Fd, the compression is set to zlib, if\\n> > enabled, and when not otherwise specified by the user.\\n> > \\n> > Before 5e73a6048, this set compressLevel=0 for -Fp and -Ft, and when\\n> > zlib was unavailable.\\n> > \\n> > But I'm not sure why there's now an empty \\\"#else\\\". I also don't know\\n> > what \\\"the default case\\\" refers to.\\n> > \\n> > Maybe the best thing here is to move the preprocessor #if, since it's no\\n> > longer in the middle of a runtime conditional:\\n> > \\n> > #ifdef HAVE_LIBZ\\n> > + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\\n> > + !user_compression_defined)\\n> > + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\\n> > + &compression_spec);\\n> > #endif\\n> > \\n> > ...but that elicits a warning about \\\"variable set but not used\\\"...\\n> \\n> \\n> Not sure, I need to think about this a bit.\\n\\nNot having warnings is preferable, isn't it? I can understand the confusion\\non the message though. Maybe a phrasing like:\\n/* Nothing to do for the default case when LIBZ is not available */\\nis easier to understand.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Mon, 27 Feb 2023 14:33:04 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Saturday, February 25th, 2023 at 3:05 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n> \\n> \\n> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n> \\n> > I have some fixes (attached) and questions while polishing the patch for\\n> > zstd compression. The fixes are small and could be integrated with the\\n> > patch for zstd, but could be applied independently.\\n\\n\\nPlease find some comments on the rest of the fixes patch that Tomas has not\\ncommented on.\\n\\n can be compressed with the <application>gzip</application> or\\n- <application>lz4</application>tool.\\n+ <application>lz4</application> tools.\\n\\n+1\\n\\n The compression method can be set to <literal>gzip</literal> or\\n- <literal>lz4</literal> or <literal>none</literal> for no compression.\\n+ <literal>lz4</literal>, or <literal>none</literal> for no compression.\\n\\nI am not a native English speaker. Yet I think that if one adds commas\\nin one of the options, then one should add commas to all the options.\\nNamely, the aboveis missing a comma between gzip and lz4. However I\\nthink that not having any commas still works grammatically and\\nsyntactically.\\n\\n- /*\\n- * A level of zero simply copies the input one block at the time. This\\n- * is probably not what the user wanted when calling this interface.\\n- */\\n- if (cs->compression_spec.level == 0)\\n- pg_fatal(\\\"requested to compress the archive yet no level was specified\\\");\\n\\n\\nI disagree with change. WriteDataToArchiveGzip() is far away from\\nwhat ever the code in pg_dump.c is doing. Any non valid values for\\nlevel will emit an error in when the proper gzip/zlib code is\\ncalled. A zero value however, will not emit such error. Having the\\nextra check there is a future proof guarantee in a very low cost.\\nFurthermore, it quickly informs the reader of the code about that\\nspecific value helping with readability and comprehension.\\n\\nIf any change is required, something for which I vote strongly\\nagainst, I would at least recommend to replace it with an\\nassertion.\\n\\n- * Initialize a compress file stream. Deffer the compression algorithm\\n+ * Initialize a compress file stream. Infer the compression algorithm\\n\\n:+1:\\n\\n- # Skip command-level tests for gzip if there is no support for it.\\n+ # Skip command-level tests for gzip/lz4 if they're not supported.\\n\\nWe will be back at that again soon. Maybe change to:\\n\\nSkip command-level test for unsupported compression methods\\n\\nTo include everything.\\n\\n\\n- ($pgdump_runs{$run}->{compile_option} eq 'gzip' && !$supports_gzip) ||\\n- ($pgdump_runs{$run}->{compile_option} eq 'lz4' && !$supports_lz4))\\n+ (($pgdump_runs{$run}->{compile_option} eq 'gzip' && !$supports_gzip) ||\\n+ ($pgdump_runs{$run}->{compile_option} eq 'lz4' && !$supports_lz4)))\\n\\nGood catch, :+1:\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Justin\\n\\n\\n\",\n \"msg_date\": \"Mon, 27 Feb 2023 14:56:08 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Feb 23, 2023 at 09:24:46PM +0100, Tomas Vondra wrote:\\n> On 2/23/23 16:26, Tomas Vondra wrote:\\n> > Thanks for v30 with the updated commit messages. I've pushed 0001 after\\n> > fixing a comment typo and removing (I think) an unnecessary change in an\\n> > error message.\\n> > \\n> > I'll give the buildfarm a bit of time before pushing 0002 and 0003.\\n> > \\n> \\n> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\n> and marked the CF entry as committed. Thanks for the patch!\\n\\nI found that e9960732a broke writing of empty gzip-compressed data,\\nspecifically LOs. pg_dump succeeds, but then the restore fails:\\n\\npostgres=# SELECT lo_create(1234);\\nlo_create | 1234\\n\\n$ time ./src/bin/pg_dump/pg_dump -h /tmp -d postgres -Fc |./src/bin/pg_dump/pg_restore -f /dev/null -v \\npg_restore: implied data-only restore\\npg_restore: executing BLOB 1234\\npg_restore: processing BLOBS\\npg_restore: restoring large object with OID 1234\\npg_restore: error: could not uncompress data: (null)\\n\\nThe inline patch below fixes it, but you won't be able to apply it\\ndirectly, as it's on top of other patches which rename the functions\\nback to \\\"Zlib\\\" and rearranges the functions to their original order, to\\nallow running:\\n\\ngit diff --diff-algorithm=minimal -w e9960732a~:./src/bin/pg_dump/compress_io.c ./src/bin/pg_dump/compress_gzip.c\\n\\nThe current function order avoids 3 lines of declarations, but it's\\nobviously pretty useful to be able to run that diff command. I already\\nargued for not calling the functions \\\"Gzip\\\" on the grounds that the name\\nwas inaccurate.\\n\\nI'd want to create an empty large object in src/test/sql/largeobject.sql\\nto exercise this tested during pgupgrade. But unfortunately that\\ndoesn't use -Fc, so this isn't hit. Empty input is an important enough\\ntest case to justify a tap test, if there's no better way.\\n\\ndiff --git a/src/bin/pg_dump/compress_gzip.c b/src/bin/pg_dump/compress_gzip.c\\nindex f3f5e87c9a8..68f3111b2fe 100644\\n--- a/src/bin/pg_dump/compress_gzip.c\\n+++ b/src/bin/pg_dump/compress_gzip.c\\n@@ -55,6 +55,32 @@ InitCompressorZlib(CompressorState *cs,\\n \\tgzipcs = (ZlibCompressorState *) pg_malloc0(sizeof(ZlibCompressorState));\\n \\n \\tcs->private_data = gzipcs;\\n+\\n+\\tif (cs->writeF)\\n+\\t{\\n+\\t\\tz_streamp\\tzp;\\n+\\t\\tzp = gzipcs->zp = (z_streamp) pg_malloc0(sizeof(z_stream));\\n+\\t\\tzp->zalloc = Z_NULL;\\n+\\t\\tzp->zfree = Z_NULL;\\n+\\t\\tzp->opaque = Z_NULL;\\n+\\n+\\t\\t/*\\n+\\t\\t * outsize is the buffer size we tell zlib it can output to. We\\n+\\t\\t * actually allocate one extra byte because some routines want to append a\\n+\\t\\t * trailing zero byte to the zlib output.\\n+\\t\\t */\\n+\\n+\\t\\tgzipcs->outbuf = pg_malloc(ZLIB_OUT_SIZE + 1);\\n+\\t\\tgzipcs->outsize = ZLIB_OUT_SIZE;\\n+\\n+\\t\\tif (deflateInit(gzipcs->zp, cs->compression_spec.level) != Z_OK)\\n+\\t\\t\\tpg_fatal(\\\"could not initialize compression library: %s\\\",\\n+\\t\\t\\t\\t\\tzp->msg);\\n+\\n+\\t\\t/* Just be paranoid - maybe End is called after Start, with no Write */\\n+\\t\\tzp->next_out = gzipcs->outbuf;\\n+\\t\\tzp->avail_out = gzipcs->outsize;\\n+\\t}\\n }\\n \\n static void\\n@@ -63,7 +89,7 @@ EndCompressorZlib(ArchiveHandle *AH, CompressorState *cs)\\n \\tZlibCompressorState *gzipcs = (ZlibCompressorState *) cs->private_data;\\n \\tz_streamp\\tzp;\\n \\n-\\tif (gzipcs->zp)\\n+\\tif (cs->writeF != NULL)\\n \\t{\\n \\t\\tzp = gzipcs->zp;\\n \\t\\tzp->next_in = NULL;\\n@@ -131,29 +157,6 @@ WriteDataToArchiveZlib(ArchiveHandle *AH, CompressorState *cs,\\n \\t\\t\\t\\t\\t const void *data, size_t dLen)\\n {\\n \\tZlibCompressorState *gzipcs = (ZlibCompressorState *) cs->private_data;\\n-\\tz_streamp\\tzp;\\n-\\n-\\tif (!gzipcs->zp)\\n-\\t{\\n-\\t\\tzp = gzipcs->zp = (z_streamp) pg_malloc(sizeof(z_stream));\\n-\\t\\tzp->zalloc = Z_NULL;\\n-\\t\\tzp->zfree = Z_NULL;\\n-\\t\\tzp->opaque = Z_NULL;\\n-\\n-\\t\\t/*\\n-\\t\\t * outsize is the buffer size we tell zlib it can output to. We\\n-\\t\\t * actually allocate one extra byte because some routines want to\\n-\\t\\t * append a trailing zero byte to the zlib output.\\n-\\t\\t */\\n-\\t\\tgzipcs->outbuf = pg_malloc(ZLIB_OUT_SIZE + 1);\\n-\\t\\tgzipcs->outsize = ZLIB_OUT_SIZE;\\n-\\n-\\t\\tif (deflateInit(zp, cs->compression_spec.level) != Z_OK)\\n-\\t\\t\\tpg_fatal(\\\"could not initialize compression library: %s\\\", zp->msg);\\n-\\n-\\t\\tzp->next_out = gzipcs->outbuf;\\n-\\t\\tzp->avail_out = gzipcs->outsize;\\n-\\t}\\n \\n \\tgzipcs->zp->next_in = (void *) unconstify(void *, data);\\n \\tgzipcs->zp->avail_in = dLen;\",\n \"msg_date\": \"Tue, 28 Feb 2023 17:58:34 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Feb 28, 2023 at 05:58:34PM -0600, Justin Pryzby wrote:\\n> I found that e9960732a broke writing of empty gzip-compressed data,\\n> specifically LOs. pg_dump succeeds, but then the restore fails:\\n\\nThe number of issues you have been reporting here begins to worries\\nme.. How many of them have you found? Is it right to assume that all\\nof them have basically 03d02f5 as oldest origin point?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 1 Mar 2023 16:24:59 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Wednesday, March 1st, 2023 at 12:58 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n\\n\\n\\n> I found that e9960732a broke writing of empty gzip-compressed data,\\n> specifically LOs. pg_dump succeeds, but then the restore fails:\\n> \\n> postgres=# SELECT lo_create(1234);\\n> lo_create | 1234\\n> \\n> $ time ./src/bin/pg_dump/pg_dump -h /tmp -d postgres -Fc |./src/bin/pg_dump/pg_restore -f /dev/null -v\\n> pg_restore: implied data-only restore\\n> pg_restore: executing BLOB 1234\\n> pg_restore: processing BLOBS\\n> pg_restore: restoring large object with OID 1234\\n> pg_restore: error: could not uncompress data: (null)\\n> \\n\\nThank you for looking. This was an untested case.\\n\\n> The inline patch below fixes it, but you won't be able to apply it\\n> directly, as it's on top of other patches which rename the functions\\n> back to \\\"Zlib\\\" and rearranges the functions to their original order, to\\n> allow running:\\n> \\n> git diff --diff-algorithm=minimal -w e9960732a~:./src/bin/pg_dump/compress_io.c ./src/bin/pg_dump/compress_gzip.c\\n> \\n\\nPlease find a patch attached that can be applied directly.\\n\\n> The current function order avoids 3 lines of declarations, but it's\\n> obviously pretty useful to be able to run that diff command. I already\\n> argued for not calling the functions \\\"Gzip\\\" on the grounds that the name\\n> was inaccurate.\\n\\nI have no idea why we are back on the naming issue. I stand by the name\\nbecause in my humble opinion helps the code reader. There is a certain\\nuniformity when the compression_spec.algorithm and the compressor\\nfunctions match as the following code sample shows.\\n\\n if (compression_spec.algorithm == PG_COMPRESSION_NONE) \\n InitCompressorNone(cs, compression_spec);\\n else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n InitCompressorGzip(cs, compression_spec);\\n else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\\n InitCompressorLZ4(cs, compression_spec); \\n \\nWhen the reader wants to see what happens when the PG_COMPRESSION_XXX\\nis set, has to simply search for the XXX part. I think that this is\\njustification enough for the use of the names.\\n\\n> \\n> I'd want to create an empty large object in src/test/sql/largeobject.sql\\n> to exercise this tested during pgupgrade. But unfortunately that\\n> doesn't use -Fc, so this isn't hit. Empty input is an important enough\\n> test case to justify a tap test, if there's no better way.\\n\\nPlease find in the attached a test case that exercises this codepath.\\n\\nCheers,\\n//Georgios\",\n \"msg_date\": \"Wed, 01 Mar 2023 13:39:14 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 3/1/23 08:24, Michael Paquier wrote:\\n> On Tue, Feb 28, 2023 at 05:58:34PM -0600, Justin Pryzby wrote:\\n>> I found that e9960732a broke writing of empty gzip-compressed data,\\n>> specifically LOs. pg_dump succeeds, but then the restore fails:\\n> \\n> The number of issues you have been reporting here begins to worries\\n> me.. How many of them have you found? Is it right to assume that all\\n> of them have basically 03d02f5 as oldest origin point?\\n\\nAFAICS a lot of the issues are more a discussion about wording in a\\ncouple places, whether it's nicer to do A or B, name the functions\\ndifferently or what.\\n\\nI'm aware of three genuine issues that I intend to fix shortly:\\n\\n1) incorrect \\\"if\\\" condition in a TAP test\\n\\n2) failure when compressing empty LO (which we had no test for)\\n\\n3) change in handling \\\"compression level = 0\\\" (which I believe should be\\nmade to behave like before)\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Mar 2023 15:33:00 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 3/1/23 14:39, gkokolatos@pm.me wrote:\\n> \\n> \\n> \\n> \\n> \\n> ------- Original Message -------\\n> On Wednesday, March 1st, 2023 at 12:58 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> \\n> \\n> \\n>> I found that e9960732a broke writing of empty gzip-compressed data,\\n>> specifically LOs. pg_dump succeeds, but then the restore fails:\\n>>\\n>> postgres=# SELECT lo_create(1234);\\n>> lo_create | 1234\\n>>\\n>> $ time ./src/bin/pg_dump/pg_dump -h /tmp -d postgres -Fc |./src/bin/pg_dump/pg_restore -f /dev/null -v\\n>> pg_restore: implied data-only restore\\n>> pg_restore: executing BLOB 1234\\n>> pg_restore: processing BLOBS\\n>> pg_restore: restoring large object with OID 1234\\n>> pg_restore: error: could not uncompress data: (null)\\n>>\\n> \\n> Thank you for looking. This was an untested case.\\n> \\n\\nYeah :-(\\n\\n>> The inline patch below fixes it, but you won't be able to apply it\\n>> directly, as it's on top of other patches which rename the functions\\n>> back to \\\"Zlib\\\" and rearranges the functions to their original order, to\\n>> allow running:\\n>>\\n>> git diff --diff-algorithm=minimal -w e9960732a~:./src/bin/pg_dump/compress_io.c ./src/bin/pg_dump/compress_gzip.c\\n>>\\n> \\n> Please find a patch attached that can be applied directly.\\n> \\n>> The current function order avoids 3 lines of declarations, but it's\\n>> obviously pretty useful to be able to run that diff command. I already\\n>> argued for not calling the functions \\\"Gzip\\\" on the grounds that the name\\n>> was inaccurate.\\n> \\n> I have no idea why we are back on the naming issue. I stand by the name\\n> because in my humble opinion helps the code reader. There is a certain\\n> uniformity when the compression_spec.algorithm and the compressor\\n> functions match as the following code sample shows.\\n> \\n> if (compression_spec.algorithm == PG_COMPRESSION_NONE) \\n> InitCompressorNone(cs, compression_spec);\\n> else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> InitCompressorGzip(cs, compression_spec);\\n> else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\\n> InitCompressorLZ4(cs, compression_spec); \\n> \\n> When the reader wants to see what happens when the PG_COMPRESSION_XXX\\n> is set, has to simply search for the XXX part. I think that this is\\n> justification enough for the use of the names.\\n> \\n\\nI don't recall the previous discussion about the naming, but I'm not\\nsure why would it be inaccurate. We call it 'gzip' pretty much\\neverywhere, and I agree with Georgios there's it helps to make this\\nconsistent with the PG_COMPRESSION_ stuff.\\n\\nThe one thing that concerned me while reviewing it earlier was that it\\nmight make the backpatcheing harder. But that's mostly irrelevant due to\\nall the other changes I think.\\n\\n>>\\n>> I'd want to create an empty large object in src/test/sql/largeobject.sql\\n>> to exercise this tested during pgupgrade. But unfortunately that\\n>> doesn't use -Fc, so this isn't hit. Empty input is an important enough\\n>> test case to justify a tap test, if there's no better way.\\n> \\n> Please find in the attached a test case that exercises this codepath.\\n> \\n\\nThanks. That seems correct to me, but I find it somewhat confusing,\\nbecause we now have\\n\\n DeflateCompressorInit vs. InitCompressorGzip\\n\\n DeflateCompressorEnd vs. EndCompressorGzip\\n\\n DeflateCompressorData - The name doesn't really say what it does (would\\n be better to have a verb in there, I think).\\n\\nI wonder if we can make this somehow clearer?\\n\\nAlso, InitCompressorGzip says this:\\n\\n /*\\n * If the caller has defined a write function, prepare the necessary\\n * state. Avoid initializing during the first write call, because End\\n * may be called without ever writing any data.\\n */\\n if (cs->writeF)\\n DeflateCompressorInit(cs);\\n\\nDoes it actually make sense to not have writeF defined in some cases?\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Mar 2023 16:52:49 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2/27/23 15:56, gkokolatos@pm.me wrote:\\n> \\n> \\n> \\n> \\n> \\n> ------- Original Message -------\\n> On Saturday, February 25th, 2023 at 3:05 PM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> \\n> \\n>>\\n>>\\n>> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n>>\\n>>> I have some fixes (attached) and questions while polishing the patch for\\n>>> zstd compression. The fixes are small and could be integrated with the\\n>>> patch for zstd, but could be applied independently.\\n> \\n> \\n> Please find some comments on the rest of the fixes patch that Tomas has not\\n> commented on.\\n> \\n> can be compressed with the <application>gzip</application> or\\n> - <application>lz4</application>tool.\\n> + <application>lz4</application> tools.\\n> \\n> +1\\n> \\n> The compression method can be set to <literal>gzip</literal> or\\n> - <literal>lz4</literal> or <literal>none</literal> for no compression.\\n> + <literal>lz4</literal>, or <literal>none</literal> for no compression.\\n> \\n> I am not a native English speaker. Yet I think that if one adds commas\\n> in one of the options, then one should add commas to all the options.\\n> Namely, the aboveis missing a comma between gzip and lz4. However I\\n> think that not having any commas still works grammatically and\\n> syntactically.\\n> \\n\\nI pushed a fix with most of these wording changes. As for this comma, I\\nbelieve the correct style is\\n\\n a, b, or c\\n\\nAt least that's what the other places in the pg_dump.sgml file do.\\n\\n> - ($pgdump_runs{$run}->{compile_option} eq 'gzip' && !$supports_gzip) ||\\n> - ($pgdump_runs{$run}->{compile_option} eq 'lz4' && !$supports_lz4))\\n> + (($pgdump_runs{$run}->{compile_option} eq 'gzip' && !$supports_gzip) ||\\n> + ($pgdump_runs{$run}->{compile_option} eq 'lz4' && !$supports_lz4)))\\n> \\n\\nPushed a fix for this too.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Mar 2023 17:14:08 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2/25/23 15:05, Justin Pryzby wrote:\\n> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n>> I have some fixes (attached) and questions while polishing the patch for\\n>> zstd compression. The fixes are small and could be integrated with the\\n>> patch for zstd, but could be applied independently.\\n> \\n> One more - WriteDataToArchiveGzip() says:\\n> \\n> + if (cs->compression_spec.level == 0)\\n> + pg_fatal(\\\"requested to compress the archive yet no level was specified\\\");\\n> \\n> That was added at e9960732a. \\n> \\n> But if you specify gzip:0, the compression level is already enforced by\\n> validate_compress_specification(), before hitting gzip.c:\\n> \\n> | pg_dump: error: invalid compression specification: compression algorithm \\\"gzip\\\" expects a compression level between 1 and 9 (default at -1)\\n> \\n> 5e73a6048 intended that to work as before, and you *can* specify -Z0:\\n> \\n> The change is backward-compatible, hence specifying only an integer\\n> leads to no compression for a level of 0 and gzip compression when the\\n> level is greater than 0.\\n> \\n> $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\\n> /dev/stdin: ASCII text\\n> \\n\\nFWIW I agree we should make this backwards-compatible - accept \\\"0\\\" and\\ntreat it as no compression.\\n\\nGeorgios, can you prepare a patch doing that?\\n\\n\\nregards\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Mar 2023 17:20:05 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 2/27/23 05:49, Justin Pryzby wrote:\\n> On Sat, Feb 25, 2023 at 08:05:53AM -0600, Justin Pryzby wrote:\\n>> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n>>> I have some fixes (attached) and questions while polishing the patch for\\n>>> zstd compression. The fixes are small and could be integrated with the\\n>>> patch for zstd, but could be applied independently.\\n>>\\n>> One more - WriteDataToArchiveGzip() says:\\n> \\n> One more again.\\n> \\n> The LZ4 path is using non-streaming mode, which compresses each block\\n> without persistent state, giving poor compression for -Fc compared with\\n> -Fp. If the data is highly compressible, the difference can be orders\\n> of magnitude.\\n> \\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fp |wc -c\\n> 12351763\\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\\n> 21890708\\n> \\n> That's not true for gzip:\\n> \\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fc |wc -c\\n> 2118869\\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fp |wc -c\\n> 2115832\\n> \\n> The function ought to at least use streaming mode, so each block/row\\n> isn't compressioned in isolation. 003 is a simple patch to use\\n> streaming mode, which improves the -Fc case:\\n> \\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\\n> 15178283\\n> \\n> However, that still flushes the compression buffer, writing a block\\n> header, for every row. With a single-column table, pg_dump -Fc -Z lz4\\n> still outputs ~10% *more* data than with no compression at all. And\\n> that's for compressible data.\\n> \\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z lz4 |wc -c\\n> 12890296\\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z none |wc -c\\n> 11890296\\n> \\n> I think this should use the LZ4F API with frames, which are buffered to\\n> avoid outputting a header for every single row. The LZ4F format isn't\\n> compatible with the LZ4 format, so (unlike changing to the streaming\\n> API) that's not something we can change in a bugfix release. I consider\\n> this an Opened Item.\\n> \\n> With the LZ4F API in 004, -Fp and -Fc are essentially the same size\\n> (like gzip). (Oh, and the output is three times smaller, too.)\\n> \\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fp |wc -c\\n> 4155448\\n> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fc |wc -c\\n> 4156548\\n> \\n\\nThanks. Those are definitely interesting improvements/optimizations!\\n\\nI suggest we track them as a separate patch series - please add them to\\nthe CF app (I guess you'll have to add them to 2023-07 at this point,\\nbut we can get them in, I think).\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Mar 2023 17:39:54 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 01, 2023 at 01:39:14PM +0000, gkokolatos@pm.me wrote:\\n> On Wednesday, March 1st, 2023 at 12:58 AM, Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> \\n> > The current function order avoids 3 lines of declarations, but it's\\n> > obviously pretty useful to be able to run that diff command. I already\\n> > argued for not calling the functions \\\"Gzip\\\" on the grounds that the name\\n> > was inaccurate.\\n> \\n> I have no idea why we are back on the naming issue. I stand by the name\\n> because in my humble opinion helps the code reader. There is a certain\\n> uniformity when the compression_spec.algorithm and the compressor\\n> functions match as the following code sample shows.\\n\\nI mentioned that it's because this allows usefully running \\\"diff\\\"\\nagainst the previous commits.\\n\\n> if (compression_spec.algorithm == PG_COMPRESSION_NONE) \\n> InitCompressorNone(cs, compression_spec);\\n> else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> InitCompressorGzip(cs, compression_spec);\\n> else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\\n> InitCompressorLZ4(cs, compression_spec); \\n> \\n> When the reader wants to see what happens when the PG_COMPRESSION_XXX\\n> is set, has to simply search for the XXX part. I think that this is\\n> justification enough for the use of the names.\\n\\nYou're right about that.\\n\\nBut (with the exception of InitCompressorGzip), I'm referring to the\\nnaming of internal functions, static to gzip.c, so renaming can't be\\nsaid to cause a loss of clarity.\\n\\n> > I'd want to create an empty large object in src/test/sql/largeobject.sql\\n> > to exercise this tested during pgupgrade. But unfortunately that\\n> > doesn't use -Fc, so this isn't hit. Empty input is an important enough\\n> > test case to justify a tap test, if there's no better way.\\n> \\n> Please find in the attached a test case that exercises this codepath.\\n\\nThanks for writing it.\\n\\nThis patch could be an opportunity to improve the \\\"diff\\\" output, without\\nrenaming anything.\\n\\nThe old order of functions was:\\n-InitCompressorZlib\\n-EndCompressorZlib\\n-DeflateCompressorZlib\\n-WriteDataToArchiveZlib\\n-ReadDataFromArchiveZlib\\n\\nIf you put DeflateCompressorEnd immediately after DeflateCompressorInit,\\ndiff works nicely. You'll have to add at least one declaration, which\\nseems very worth it.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Wed, 1 Mar 2023 11:08:17 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Wednesday, March 1st, 2023 at 5:20 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> \\n> \\n> On 2/25/23 15:05, Justin Pryzby wrote:\\n> \\n> > On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n> > \\n> > > I have some fixes (attached) and questions while polishing the patch for\\n> > > zstd compression. The fixes are small and could be integrated with the\\n> > > patch for zstd, but could be applied independently.\\n> > \\n> > One more - WriteDataToArchiveGzip() says:\\n> > \\n> > + if (cs->compression_spec.level == 0)\\n> > + pg_fatal(\\\"requested to compress the archive yet no level was specified\\\");\\n> > \\n> > That was added at e9960732a.\\n> > \\n> > But if you specify gzip:0, the compression level is already enforced by\\n> > validate_compress_specification(), before hitting gzip.c:\\n> > \\n> > | pg_dump: error: invalid compression specification: compression algorithm \\\"gzip\\\" expects a compression level between 1 and 9 (default at -1)\\n> > \\n> > 5e73a6048 intended that to work as before, and you can specify -Z0:\\n> > \\n> > The change is backward-compatible, hence specifying only an integer\\n> > leads to no compression for a level of 0 and gzip compression when the\\n> > level is greater than 0.\\n> > \\n> > $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\\n> > /dev/stdin: ASCII text\\n> \\n> \\n> FWIW I agree we should make this backwards-compatible - accept \\\"0\\\" and\\n> treat it as no compression.\\n> \\n> Georgios, can you prepare a patch doing that?\\n\\nPlease find a patch attached. However I am a bit at a loss, the backwards\\ncompatible behaviour has not changed. Passing -Z0/--compress=0 does produce\\na non compressed output. So I am not really certain as to what broke and\\nneeds fixing.\\n\\nWhat commit 5e73a6048 did fail to do, is test the backwards compatible\\nbehaviour. The attached amends it.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> \\n> regards\\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Thu, 02 Mar 2023 16:34:23 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 01, 2023 at 05:20:05PM +0100, Tomas Vondra wrote:\\n> On 2/25/23 15:05, Justin Pryzby wrote:\\n> > On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n> >> I have some fixes (attached) and questions while polishing the patch for\\n> >> zstd compression. The fixes are small and could be integrated with the\\n> >> patch for zstd, but could be applied independently.\\n> > \\n> > One more - WriteDataToArchiveGzip() says:\\n> > \\n> > + if (cs->compression_spec.level == 0)\\n> > + pg_fatal(\\\"requested to compress the archive yet no level was specified\\\");\\n> > \\n> > That was added at e9960732a. \\n> > \\n> > But if you specify gzip:0, the compression level is already enforced by\\n> > validate_compress_specification(), before hitting gzip.c:\\n> > \\n> > | pg_dump: error: invalid compression specification: compression algorithm \\\"gzip\\\" expects a compression level between 1 and 9 (default at -1)\\n> > \\n> > 5e73a6048 intended that to work as before, and you *can* specify -Z0:\\n> > \\n> > The change is backward-compatible, hence specifying only an integer\\n> > leads to no compression for a level of 0 and gzip compression when the\\n> > level is greater than 0.\\n> > \\n> > $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\\n> > /dev/stdin: ASCII text\\n> \\n> FWIW I agree we should make this backwards-compatible - accept \\\"0\\\" and\\n> treat it as no compression.\\n> \\n> Georgios, can you prepare a patch doing that?\\n\\nI think maybe Tomas misunderstood. What I was trying to say is that -Z\\n0 *is* accepted to mean no compression. This part wasn't quoted, but I\\nsaid:\\n\\n> Right now, I think that pg_fatal in gzip.c is dead code - that was first\\n> added in the patch version sent on 21 Dec 2022.\\n\\nIf you run the diff command that I've been talking about, you'll see\\nthat InitCompressorZlib was almost unchanged - e9960732 is essentially a\\nrefactoring. I don't think it's desirable to add a pg_fatal() in a\\nfunction that's otherwise nearly-unchanged. The fact that it's\\nnearly-unchanged is a good thing: it simplifies reading of what changed.\\nIf someone wants to add a pg_fatal() in that code path, it'd be better\\ndone in its own commit, with a separate message explaining the change.\\n\\nIf you insist on changing anything here, you might add an assertion (as\\nyou said earlier) along with a comment like\\n/* -Z 0 uses the \\\"None\\\" compressor rather than zlib with no compression */\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Mar 2023 11:18:24 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 3/2/23 18:18, Justin Pryzby wrote:\\n> On Wed, Mar 01, 2023 at 05:20:05PM +0100, Tomas Vondra wrote:\\n>> On 2/25/23 15:05, Justin Pryzby wrote:\\n>>> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n>>>> I have some fixes (attached) and questions while polishing the patch for\\n>>>> zstd compression. The fixes are small and could be integrated with the\\n>>>> patch for zstd, but could be applied independently.\\n>>>\\n>>> One more - WriteDataToArchiveGzip() says:\\n>>>\\n>>> + if (cs->compression_spec.level == 0)\\n>>> + pg_fatal(\\\"requested to compress the archive yet no level was specified\\\");\\n>>>\\n>>> That was added at e9960732a. \\n>>>\\n>>> But if you specify gzip:0, the compression level is already enforced by\\n>>> validate_compress_specification(), before hitting gzip.c:\\n>>>\\n>>> | pg_dump: error: invalid compression specification: compression algorithm \\\"gzip\\\" expects a compression level between 1 and 9 (default at -1)\\n>>>\\n>>> 5e73a6048 intended that to work as before, and you *can* specify -Z0:\\n>>>\\n>>> The change is backward-compatible, hence specifying only an integer\\n>>> leads to no compression for a level of 0 and gzip compression when the\\n>>> level is greater than 0.\\n>>>\\n>>> $ time ./src/bin/pg_dump/pg_dump -h /tmp regression -t int8_tbl -Fp --compress 0 |file -\\n>>> /dev/stdin: ASCII text\\n>>\\n>> FWIW I agree we should make this backwards-compatible - accept \\\"0\\\" and\\n>> treat it as no compression.\\n>>\\n>> Georgios, can you prepare a patch doing that?\\n> \\n> I think maybe Tomas misunderstood. What I was trying to say is that -Z\\n> 0 *is* accepted to mean no compression. This part wasn't quoted, but I\\n> said:\\n> \\n\\nAh, I see. Well, I also tried but with \\\"-Z gzip:0\\\" (and not -Z 0), and\\nthat does fail:\\n\\n error: invalid compression specification: compression algorithm \\\"gzip\\\"\\n expects a compression level between 1 and 9 (default at -1)\\n\\nIt's a bit weird these two cases behave differently, when both translate\\nto the same default compression method (gzip).\\n\\n>> Right now, I think that pg_fatal in gzip.c is dead code - that was first\\n>> added in the patch version sent on 21 Dec 2022.\\n> \\n> If you run the diff command that I've been talking about, you'll see\\n> that InitCompressorZlib was almost unchanged - e9960732 is essentially a\\n> refactoring. I don't think it's desirable to add a pg_fatal() in a\\n> function that's otherwise nearly-unchanged. The fact that it's\\n> nearly-unchanged is a good thing: it simplifies reading of what changed.\\n> If someone wants to add a pg_fatal() in that code path, it'd be better\\n> done in its own commit, with a separate message explaining the change.\\n> \\n> If you insist on changing anything here, you might add an assertion (as\\n> you said earlier) along with a comment like\\n> /* -Z 0 uses the \\\"None\\\" compressor rather than zlib with no compression */\\n> \\n\\nYeah, a comment would be helpful.\\n\\nAlso, after thinking about it a bit more maybe having the unreachable\\npg_fatal() is not a good thing, as it will just confuse people (I'd\\ncertainly assume having such check means there's a way in which case it\\nmight trigger.). Maybe an assert would be better?\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 2 Mar 2023 19:27:00 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 01, 2023 at 04:52:49PM +0100, Tomas Vondra wrote:\\n> Thanks. That seems correct to me, but I find it somewhat confusing,\\n> because we now have\\n> \\n> DeflateCompressorInit vs. InitCompressorGzip\\n> \\n> DeflateCompressorEnd vs. EndCompressorGzip\\n> \\n> DeflateCompressorData - The name doesn't really say what it does (would\\n> be better to have a verb in there, I think).\\n> \\n> I wonder if we can make this somehow clearer?\\n\\nTo move things along, I updated Georgios' patch:\\n\\nRename DeflateCompressorData() to DeflateCompressorCommon();\\nRearrange functions to their original order allowing a cleaner diff to the prior code;\\nChange pg_fatal() to an assertion+comment;\\nUpdate the commit message and fix a few typos;\\n\\n> Also, InitCompressorGzip says this:\\n> \\n> /*\\n> * If the caller has defined a write function, prepare the necessary\\n> * state. Avoid initializing during the first write call, because End\\n> * may be called without ever writing any data.\\n> */\\n> if (cs->writeF)\\n> DeflateCompressorInit(cs);\\n>\\n> Does it actually make sense to not have writeF defined in some cases?\\n\\nInitCompressor is being called for either reading or writing, either of\\nwhich could be null:\\n\\nsrc/bin/pg_dump/pg_backup_custom.c: ctx->cs = AllocateCompressor(AH->compression_spec,\\nsrc/bin/pg_dump/pg_backup_custom.c- NULL,\\nsrc/bin/pg_dump/pg_backup_custom.c- _CustomWriteFunc);\\n--\\nsrc/bin/pg_dump/pg_backup_custom.c: cs = AllocateCompressor(AH->compression_spec,\\nsrc/bin/pg_dump/pg_backup_custom.c- _CustomReadFunc, NULL);\\n\\nIt's confusing that the comment says \\\"Avoid initializing...\\\". What it\\nreally means is \\\"Initialize eagerly...\\\". But that makes more sense in\\nthe context of the commit message for this bugfix than in a comment. So\\nI changed that too.\\n\\n+ /* If deflation was initialized, finalize it */ \\n+ if (cs->private_data) \\n+ DeflateCompressorEnd(AH, cs); \\n\\nMaybe it'd be more clear if this used \\\"if (cs->writeF)\\\", like in the\\ninit function ?\\n\\n-- \\nJustin\",\n \"msg_date\": \"Tue, 7 Mar 2023 19:45:20 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 01, 2023 at 05:39:54PM +0100, Tomas Vondra wrote:\\n> On 2/27/23 05:49, Justin Pryzby wrote:\\n> > On Sat, Feb 25, 2023 at 08:05:53AM -0600, Justin Pryzby wrote:\\n> >> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n> >>> I have some fixes (attached) and questions while polishing the patch for\\n> >>> zstd compression. The fixes are small and could be integrated with the\\n> >>> patch for zstd, but could be applied independently.\\n> >>\\n> >> One more - WriteDataToArchiveGzip() says:\\n> > \\n> > One more again.\\n> > \\n> > The LZ4 path is using non-streaming mode, which compresses each block\\n> > without persistent state, giving poor compression for -Fc compared with\\n> > -Fp. If the data is highly compressible, the difference can be orders\\n> > of magnitude.\\n> > \\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fp |wc -c\\n> > 12351763\\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\\n> > 21890708\\n> > \\n> > That's not true for gzip:\\n> > \\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fc |wc -c\\n> > 2118869\\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fp |wc -c\\n> > 2115832\\n> > \\n> > The function ought to at least use streaming mode, so each block/row\\n> > isn't compressioned in isolation. 003 is a simple patch to use\\n> > streaming mode, which improves the -Fc case:\\n> > \\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\\n> > 15178283\\n> > \\n> > However, that still flushes the compression buffer, writing a block\\n> > header, for every row. With a single-column table, pg_dump -Fc -Z lz4\\n> > still outputs ~10% *more* data than with no compression at all. And\\n> > that's for compressible data.\\n> > \\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z lz4 |wc -c\\n> > 12890296\\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z none |wc -c\\n> > 11890296\\n> > \\n> > I think this should use the LZ4F API with frames, which are buffered to\\n> > avoid outputting a header for every single row. The LZ4F format isn't\\n> > compatible with the LZ4 format, so (unlike changing to the streaming\\n> > API) that's not something we can change in a bugfix release. I consider\\n> > this an Opened Item.\\n> > \\n> > With the LZ4F API in 004, -Fp and -Fc are essentially the same size\\n> > (like gzip). (Oh, and the output is three times smaller, too.)\\n> > \\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fp |wc -c\\n> > 4155448\\n> > $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fc |wc -c\\n> > 4156548\\n> \\n> Thanks. Those are definitely interesting improvements/optimizations!\\n> \\n> I suggest we track them as a separate patch series - please add them to\\n> the CF app (I guess you'll have to add them to 2023-07 at this point,\\n> but we can get them in, I think).\\n\\nThanks for looking. I'm not sure if I'm the best person to write/submit\\nthe patch to implement that for LZ4. Georgios, would you want to take\\non this change ?\\n\\nI think that needs to be changed for v16, since 1) LZ4F works so much\\nbetter like this, and 2) we can't change it later without breaking\\ncompatibility of the dumpfiles by changing the header with another name\\nother than \\\"lz4\\\". Also, I imagine we'd want to continue supporting the\\nability to *restore* a dumpfile using the old(current) format, which\\nwould be untestable code unless we also preserved the ability to write\\nit somehow (like -Z lz4-old).\\n\\nOne issue is that LZ4F_createCompressionContext() and\\nLZ4F_compressBegin() ought to be called in InitCompressorLZ4(). It\\nseems like it might *need* to be called there to avoid exactly the kind\\nof issue that I reported with empty LOs with gzip. But\\nInitCompressorLZ4() isn't currently passed the ArchiveHandle, so can't\\nwrite the header. And LZ4CompressorState has a simple char *buf, and\\nnot an more elaborate data structure like zlib. You could work around\\nthat by keeping storing the \\\"len\\\" of the existing buffer, and flushing\\nit in EndCompressorLZ4(), but that adds needless complexity to the Write\\nand End functions. Maybe the Init function should be passed the AH.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Mar 2023 10:15:55 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 3/9/23 17:15, Justin Pryzby wrote:\\n> On Wed, Mar 01, 2023 at 05:39:54PM +0100, Tomas Vondra wrote:\\n>> On 2/27/23 05:49, Justin Pryzby wrote:\\n>>> On Sat, Feb 25, 2023 at 08:05:53AM -0600, Justin Pryzby wrote:\\n>>>> On Fri, Feb 24, 2023 at 11:02:14PM -0600, Justin Pryzby wrote:\\n>>>>> I have some fixes (attached) and questions while polishing the patch for\\n>>>>> zstd compression. The fixes are small and could be integrated with the\\n>>>>> patch for zstd, but could be applied independently.\\n>>>>\\n>>>> One more - WriteDataToArchiveGzip() says:\\n>>>\\n>>> One more again.\\n>>>\\n>>> The LZ4 path is using non-streaming mode, which compresses each block\\n>>> without persistent state, giving poor compression for -Fc compared with\\n>>> -Fp. If the data is highly compressible, the difference can be orders\\n>>> of magnitude.\\n>>>\\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fp |wc -c\\n>>> 12351763\\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\\n>>> 21890708\\n>>>\\n>>> That's not true for gzip:\\n>>>\\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fc |wc -c\\n>>> 2118869\\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z gzip -Fp |wc -c\\n>>> 2115832\\n>>>\\n>>> The function ought to at least use streaming mode, so each block/row\\n>>> isn't compressioned in isolation. 003 is a simple patch to use\\n>>> streaming mode, which improves the -Fc case:\\n>>>\\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -Z lz4 -Fc |wc -c\\n>>> 15178283\\n>>>\\n>>> However, that still flushes the compression buffer, writing a block\\n>>> header, for every row. With a single-column table, pg_dump -Fc -Z lz4\\n>>> still outputs ~10% *more* data than with no compression at all. And\\n>>> that's for compressible data.\\n>>>\\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z lz4 |wc -c\\n>>> 12890296\\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Fc -Z none |wc -c\\n>>> 11890296\\n>>>\\n>>> I think this should use the LZ4F API with frames, which are buffered to\\n>>> avoid outputting a header for every single row. The LZ4F format isn't\\n>>> compatible with the LZ4 format, so (unlike changing to the streaming\\n>>> API) that's not something we can change in a bugfix release. I consider\\n>>> this an Opened Item.\\n>>>\\n>>> With the LZ4F API in 004, -Fp and -Fc are essentially the same size\\n>>> (like gzip). (Oh, and the output is three times smaller, too.)\\n>>>\\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fp |wc -c\\n>>> 4155448\\n>>> $ ./src/bin/pg_dump/pg_dump -h /tmp postgres -t t1 -Z lz4 -Fc |wc -c\\n>>> 4156548\\n>>\\n>> Thanks. Those are definitely interesting improvements/optimizations!\\n>>\\n>> I suggest we track them as a separate patch series - please add them to\\n>> the CF app (I guess you'll have to add them to 2023-07 at this point,\\n>> but we can get them in, I think).\\n> \\n> Thanks for looking. I'm not sure if I'm the best person to write/submit\\n> the patch to implement that for LZ4. Georgios, would you want to take\\n> on this change ?\\n> \\n> I think that needs to be changed for v16, since 1) LZ4F works so much\\n> better like this, and 2) we can't change it later without breaking\\n> compatibility of the dumpfiles by changing the header with another name\\n> other than \\\"lz4\\\". Also, I imagine we'd want to continue supporting the\\n> ability to *restore* a dumpfile using the old(current) format, which\\n> would be untestable code unless we also preserved the ability to write\\n> it somehow (like -Z lz4-old).\\n> \\n\\nI'm a bit confused about the lz4 vs. lz4f stuff, TBH. If we switch to\\nlz4f, doesn't that mean it (e.g. restore) won't work on systems that\\nonly have older lz4 version? What would/should happen if we take backup\\ncompressed with lz4f, an then try restoring it on an older system where\\nlz4 does not support lz4f?\\n\\nMaybe if lz4f format is incompatible with regular lz4, we should treat\\nit as a separate compression method 'lz4f'?\\n\\nI'm mostly afk until the end of the week, but I tried searching for lz4f\\ninfo - the results are not particularly enlightening, unfortunately.\\n\\nAFAICS this only applies to lz4f stuff. Or would the streaming mode be a\\nbreaking change too?\\n\\n> One issue is that LZ4F_createCompressionContext() and\\n> LZ4F_compressBegin() ought to be called in InitCompressorLZ4(). It\\n> seems like it might *need* to be called there to avoid exactly the kind\\n> of issue that I reported with empty LOs with gzip. But\\n> InitCompressorLZ4() isn't currently passed the ArchiveHandle, so can't\\n> write the header. And LZ4CompressorState has a simple char *buf, and\\n> not an more elaborate data structure like zlib. You could work around\\n> that by keeping storing the \\\"len\\\" of the existing buffer, and flushing\\n> it in EndCompressorLZ4(), but that adds needless complexity to the Write\\n> and End functions. Maybe the Init function should be passed the AH.\\n> \\n\\nNot sure, but looking at GzipCompressorState I see the only extra thing\\nit has (compared to LZ4CompressorState) is \\\"z_streamp\\\". I can't\\nexperiment with this until the end of this week, so perhaps that's not\\nworkable, but wouldn't it be better to add a similar field into\\nLZ4CompressorState? Passing AH to the init function seems like a\\nviolation of abstraction.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 9 Mar 2023 18:58:20 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 09, 2023 at 06:58:20PM +0100, Tomas Vondra wrote:\\n> I'm a bit confused about the lz4 vs. lz4f stuff, TBH. If we switch to\\n> lz4f, doesn't that mean it (e.g. restore) won't work on systems that\\n> only have older lz4 version? What would/should happen if we take backup\\n> compressed with lz4f, an then try restoring it on an older system where\\n> lz4 does not support lz4f?\\n\\nYou seem to be thinking about LZ4F as a weird, new innovation I'm\\nexperimenting with, but compress_lz4.c already uses LZ4F for its \\\"file\\\"\\nAPI. LZ4F is also what's written by the lz4 CLI tool, and I found that\\nLZ4F has been included in the library for ~8 years:\\n\\nhttps://github.com/lz4/lz4/releases?page=2\\nr126 Dec 24, 2014\\nNew : lz4frame API is now integrated into liblz4\\n\\n> Maybe if lz4f format is incompatible with regular lz4, we should treat\\n> it as a separate compression method 'lz4f'?\\n> \\n> I'm mostly afk until the end of the week, but I tried searching for lz4f\\n> info - the results are not particularly enlightening, unfortunately.\\n> \\n> AFAICS this only applies to lz4f stuff. Or would the streaming mode be a\\n> breaking change too?\\n\\nStreaming mode outputs the same format as the existing code, but gives\\nbetter compression. We could (theoretically) change it in a bugfix\\nrelease, and old output would still be restorable (I think new output\\nwould even be restorable with the old versions of pg_restore).\\n\\nBut that's not true for LZ4F. The benefit there is that it avoids\\noutputing a separate block for each row. That's essential for narrow\\ntables, for which the block header currently being written has an\\noverhead several times larger than the data.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Fri, 10 Mar 2023 07:05:49 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 10, 2023 at 07:05:49AM -0600, Justin Pryzby wrote:\\n> On Thu, Mar 09, 2023 at 06:58:20PM +0100, Tomas Vondra wrote:\\n>> I'm a bit confused about the lz4 vs. lz4f stuff, TBH. If we switch to\\n>> lz4f, doesn't that mean it (e.g. restore) won't work on systems that\\n>> only have older lz4 version? What would/should happen if we take backup\\n>> compressed with lz4f, an then try restoring it on an older system where\\n>> lz4 does not support lz4f?\\n> \\n> You seem to be thinking about LZ4F as a weird, new innovation I'm\\n> experimenting with, but compress_lz4.c already uses LZ4F for its \\\"file\\\"\\n> API.\\n\\nNote: we already use lz4 frames in pg_receivewal (for WAL) and\\npg_basebackup (bbstreamer).\\n--\\nMichael\",\n \"msg_date\": \"Sat, 11 Mar 2023 08:00:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hello,\\n23.02.2023 23:24, Tomas Vondra wrote:\\n> On 2/23/23 16:26, Tomas Vondra wrote:\\n>> Thanks for v30 with the updated commit messages. I've pushed 0001 after\\n>> fixing a comment typo and removing (I think) an unnecessary change in an\\n>> error message.\\n>>\\n>> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\\n>>\\n> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\n> and marked the CF entry as committed. Thanks for the patch!\\n>\\n> I wonder how difficult would it be to add the zstd compression, so that\\n> we don't have the annoying \\\"unsupported\\\" cases.\\n\\nWith the patch 0003 committed, a single warning -Wtype-limits appeared in the\\nmaster branch:\\n$ CPPFLAGS=\\\"-Og -Wtype-limits\\\" ./configure --with-lz4 -q && make -s -j8\\ncompress_lz4.c: In function ‘LZ4File_gets’:\\ncompress_lz4.c:492:19: warning: comparison of unsigned expression in ‘< 0’ is \\nalways false [-Wtype-limits]\\n 492 | if (dsize < 0)\\n |\\n(I wonder, is it accidental that there no other places that triggers\\nthe warning, or some buildfarm animals had this check enabled before?)\\n\\nIt is not a false positive as can be proved by the 002_pg_dump.pl modified as\\nfollows:\\n- program => $ENV{'LZ4'},\\n+ program => 'mv',\\n args => [\\n- '-z', '-f', '--rm',\\n\\\"$tempdir/compression_lz4_dir/blobs.toc\\\",\\n\\\"$tempdir/compression_lz4_dir/blobs.toc.lz4\\\",\\n ],\\n },\\nA diagnostic logging added shows:\\nLZ4File_gets() after LZ4File_read_internal; dsize: 18446744073709551615\\n\\nand pg_restore fails with:\\nerror: invalid line in large object TOC file \\n\\\".../src/bin/pg_dump/tmp_check/tmp_test_22ri/compression_lz4_dir/blobs.toc\\\": \\\"????\\\"\\n\\nBest regards,\\nAlexander\\n\\n\\n\",\n \"msg_date\": \"Sat, 11 Mar 2023 09:00:00 +0300\",\n \"msg_from\": \"Alexander Lakhin <exclusion@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Saturday, March 11th, 2023 at 7:00 AM, Alexander Lakhin <exclusion@gmail.com> wrote:\\n\\n> Hello,\\n> 23.02.2023 23:24, Tomas Vondra wrote:\\n> \\n> > On 2/23/23 16:26, Tomas Vondra wrote:\\n> > \\n> > > Thanks for v30 with the updated commit messages. I've pushed 0001 after\\n> > > fixing a comment typo and removing (I think) an unnecessary change in an\\n> > > error message.\\n> > > \\n> > > I'll give the buildfarm a bit of time before pushing 0002 and 0003.\\n> > \\n> > I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\n> > and marked the CF entry as committed. Thanks for the patch!\\n> > \\n> > I wonder how difficult would it be to add the zstd compression, so that\\n> > we don't have the annoying \\\"unsupported\\\" cases.\\n> \\n> \\n> With the patch 0003 committed, a single warning -Wtype-limits appeared in the\\n> master branch:\\n> $ CPPFLAGS=\\\"-Og -Wtype-limits\\\" ./configure --with-lz4 -q && make -s -j8\\n> compress_lz4.c: In function ‘LZ4File_gets’:\\n> compress_lz4.c:492:19: warning: comparison of unsigned expression in ‘< 0’ is\\n> always false [-Wtype-limits]\\n> 492 | if (dsize < 0)\\n> |\\n\\nThank you Alexander. Please find attached an attempt to address it.\\n\\n> (I wonder, is it accidental that there no other places that triggers\\n> the warning, or some buildfarm animals had this check enabled before?)\\n\\nI can not answer about the buildfarms. Do you think that adding an explicit\\ncheck for this warning in meson would help? I am a bit uncertain as I think\\nthat type-limits are included in extra.\\n\\n@@ -1748,6 +1748,7 @@ common_warning_flags = [\\n '-Wshadow=compatible-local',\\n # This was included in -Wall/-Wformat in older GCC versions\\n '-Wformat-security',\\n+ '-Wtype-limits',\\n ]\\n\\n> \\n> It is not a false positive as can be proved by the 002_pg_dump.pl modified as\\n> follows:\\n> - program => $ENV{'LZ4'},\\n> \\n> + program => 'mv',\\n> \\n> args => [\\n> \\n> - '-z', '-f', '--rm',\\n> \\\"$tempdir/compression_lz4_dir/blobs.toc\\\",\\n> \\\"$tempdir/compression_lz4_dir/blobs.toc.lz4\\\",\\n> ],\\n> },\\n\\nCorrect, it is not a false positive. The existing testing framework provides\\nlimited support for exercising error branches. Especially so when those are\\ndependent on generated output. \\n\\n> A diagnostic logging added shows:\\n> LZ4File_gets() after LZ4File_read_internal; dsize: 18446744073709551615\\n> \\n> and pg_restore fails with:\\n> error: invalid line in large object TOC file\\n> \\\".../src/bin/pg_dump/tmp_check/tmp_test_22ri/compression_lz4_dir/blobs.toc\\\": \\\"????\\\"\\n\\nIt is a good thing that the restore fails with bad input. Yet it should\\nhave failed earlier. The attached makes certain it does fail earlier. \\n\\nCheers,\\n//Georgios\\n\\n> \\n> Best regards,\\n> Alexander\",\n \"msg_date\": \"Sat, 11 Mar 2023 10:50:06 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hi Georgios,\\n\\n11.03.2023 13:50, gkokolatos@pm.me wrote:\\n> I can not answer about the buildfarms. Do you think that adding an explicit\\n> check for this warning in meson would help? I am a bit uncertain as I think\\n> that type-limits are included in extra.\\n>\\n> @@ -1748,6 +1748,7 @@ common_warning_flags = [\\n> '-Wshadow=compatible-local',\\n> # This was included in -Wall/-Wformat in older GCC versions\\n> '-Wformat-security',\\n> + '-Wtype-limits',\\n> ]\\nI'm not sure that I can promote additional checks (or determine where\\nto put them), but if some patch introduces a warning of a type that wasn't\\npresent before, I think it's worth to eliminate the warning (if it is\\nsensible) to keep the source code check baseline at the same level\\nor even lift it up gradually.\\nI've also found that the same commit introduced a single instance of\\nthe analyzer-possible-null-argument warning:\\nCPPFLAGS=\\\"-Og -fanalyzer -Wno-analyzer-malloc-leak -Wno-analyzer-file-leak \\n-Wno-analyzer-null-dereference -Wno-analyzer-shift-count-overflow \\n-Wno-analyzer-free-of-non-heap -Wno-analyzer-null-argument \\n-Wno-analyzer-double-free -Wanalyzer-possible-null-argument\\\" ./configure \\n--with-lz4 -q && make -s -j8\\ncompress_io.c: In function ‘hasSuffix’:\\ncompress_io.c:158:47: warning: use of possibly-NULL ‘filename’ where non-null \\nexpected [CWE-690] [-Wanalyzer-possible-null-argument]\\n 158 | int filenamelen = strlen(filename);\\n | ^~~~~~~~~~~~~~~~\\n ‘InitDiscoverCompressFileHandle’: events 1-3\\n...\\n\\n(I use gcc-11.3.)\\nAs I can see, many existing uses of strdup() are followed by a check for\\nnull result, so maybe it's a common practice and a similar check should\\nbe added in InitDiscoverCompressFileHandle().\\n(There also a couple of other warnings introduced with the lz4 compression\\npatches, but those ones are not unique, so I maybe they aren't worth fixing.)\\n\\n>> It is a good thing that the restore fails with bad input. Yet it should\\n>> have failed earlier. The attached makes certain it does fail earlier.\\n>>\\nThanks! Your patch definitely fixes the issue.\\n\\nBest regards,\\nAlexander\\n\\n\\n\",\n \"msg_date\": \"Sat, 11 Mar 2023 21:00:00 +0300\",\n \"msg_from\": \"Alexander Lakhin <exclusion@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 11.03.23 07:00, Alexander Lakhin wrote:\\n> Hello,\\n> 23.02.2023 23:24, Tomas Vondra wrote:\\n>> On 2/23/23 16:26, Tomas Vondra wrote:\\n>>> Thanks for v30 with the updated commit messages. I've pushed 0001 after\\n>>> fixing a comment typo and removing (I think) an unnecessary change in an\\n>>> error message.\\n>>>\\n>>> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\\n>>>\\n>> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\n>> and marked the CF entry as committed. Thanks for the patch!\\n>>\\n>> I wonder how difficult would it be to add the zstd compression, so that\\n>> we don't have the annoying \\\"unsupported\\\" cases.\\n> \\n> With the patch 0003 committed, a single warning -Wtype-limits appeared \\n> in the\\n> master branch:\\n> $ CPPFLAGS=\\\"-Og -Wtype-limits\\\" ./configure --with-lz4 -q && make -s -j8\\n> compress_lz4.c: In function ‘LZ4File_gets’:\\n> compress_lz4.c:492:19: warning: comparison of unsigned expression in ‘< \\n> 0’ is always false [-Wtype-limits]\\n> 492 | if (dsize < 0)\\n> |\\n> (I wonder, is it accidental that there no other places that triggers\\n> the warning, or some buildfarm animals had this check enabled before?)\\n\\nI think there is an underlying problem in this code that it dances back \\nand forth between size_t and int in an unprincipled way.\\n\\nIn the code that triggers the warning, dsize is size_t. dsize is the \\nreturn from LZ4File_read_internal(), which is declared to return int. \\nThe variable that LZ4File_read_internal() returns in the success case is \\nsize_t, but in case of an error it returns -1. (So the code that is \\nwarning is meaning to catch this error case, but it won't ever work.) \\nFurther below LZ4File_read_internal() calls LZ4File_read_overflow(), \\nwhich is declared to return int, but in some cases it returns \\nfs->overflowlen, which is size_t.\\n\\nThis should be cleaned up.\\n\\nAFAICT, the upstream API in lz4.h uses int for size values, but \\nlz4frame.h uses size_t, so I don't know what the correct approach is.\\n\\n\\n\",\n \"msg_date\": \"Sun, 12 Mar 2023 11:07:33 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 3/12/23 11:07, Peter Eisentraut wrote:\\n> On 11.03.23 07:00, Alexander Lakhin wrote:\\n>> Hello,\\n>> 23.02.2023 23:24, Tomas Vondra wrote:\\n>>> On 2/23/23 16:26, Tomas Vondra wrote:\\n>>>> Thanks for v30 with the updated commit messages. I've pushed 0001 after\\n>>>> fixing a comment typo and removing (I think) an unnecessary change\\n>>>> in an\\n>>>> error message.\\n>>>>\\n>>>> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\\n>>>>\\n>>> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\n>>> and marked the CF entry as committed. Thanks for the patch!\\n>>>\\n>>> I wonder how difficult would it be to add the zstd compression, so that\\n>>> we don't have the annoying \\\"unsupported\\\" cases.\\n>>\\n>> With the patch 0003 committed, a single warning -Wtype-limits appeared\\n>> in the\\n>> master branch:\\n>> $ CPPFLAGS=\\\"-Og -Wtype-limits\\\" ./configure --with-lz4 -q && make -s -j8\\n>> compress_lz4.c: In function ‘LZ4File_gets’:\\n>> compress_lz4.c:492:19: warning: comparison of unsigned expression in\\n>> ‘< 0’ is always false [-Wtype-limits]\\n>> 492 | if (dsize < 0)\\n>> |\\n>> (I wonder, is it accidental that there no other places that triggers\\n>> the warning, or some buildfarm animals had this check enabled before?)\\n> \\n> I think there is an underlying problem in this code that it dances back\\n> and forth between size_t and int in an unprincipled way.\\n> \\n> In the code that triggers the warning, dsize is size_t. dsize is the\\n> return from LZ4File_read_internal(), which is declared to return int.\\n> The variable that LZ4File_read_internal() returns in the success case is\\n> size_t, but in case of an error it returns -1. (So the code that is\\n> warning is meaning to catch this error case, but it won't ever work.)\\n> Further below LZ4File_read_internal() calls LZ4File_read_overflow(),\\n> which is declared to return int, but in some cases it returns\\n> fs->overflowlen, which is size_t.\\n> \\n\\nI agree. I just got home so I looked at this only very briefly, but I\\nthink it's clearly wrong to assign the LZ4File_read_internal() result to\\na size_t variable (and it seems to me LZ4File_gets does the same mistake\\nwith LZ4File_read_internal() result).\\n\\nI'll get this fixed early next week, I'm too tired to do that now\\nwithout likely causing further issues.\\n\\n> This should be cleaned up.\\n> \\n> AFAICT, the upstream API in lz4.h uses int for size values, but\\n> lz4frame.h uses size_t, so I don't know what the correct approach is.\\n\\nYeah, that's a good point. I think Justin is right we should be using\\nthe LZ4F stuff, so ultimately we'll probably switch to size_t. But IMO\\nit's definitely better to correct the current code first, and only then\\nswitch to LZ4F (from one correct state to another).\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sun, 12 Mar 2023 19:34:02 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 3/11/23 11:50, gkokolatos@pm.me wrote:\\n> ------- Original Message -------\\n> On Saturday, March 11th, 2023 at 7:00 AM, Alexander Lakhin <exclusion@gmail.com> wrote:\\n> \\n>> Hello,\\n>> 23.02.2023 23:24, Tomas Vondra wrote:\\n>>\\n>>> On 2/23/23 16:26, Tomas Vondra wrote:\\n>>>\\n>>>> Thanks for v30 with the updated commit messages. I've pushed 0001 after\\n>>>> fixing a comment typo and removing (I think) an unnecessary change in an\\n>>>> error message.\\n>>>>\\n>>>> I'll give the buildfarm a bit of time before pushing 0002 and 0003.\\n>>>\\n>>> I've now pushed 0002 and 0003, after minor tweaks (a couple typos etc.),\\n>>> and marked the CF entry as committed. Thanks for the patch!\\n>>>\\n>>> I wonder how difficult would it be to add the zstd compression, so that\\n>>> we don't have the annoying \\\"unsupported\\\" cases.\\n>>\\n>>\\n>> With the patch 0003 committed, a single warning -Wtype-limits appeared in the\\n>> master branch:\\n>> $ CPPFLAGS=\\\"-Og -Wtype-limits\\\" ./configure --with-lz4 -q && make -s -j8\\n>> compress_lz4.c: In function ‘LZ4File_gets’:\\n>> compress_lz4.c:492:19: warning: comparison of unsigned expression in ‘< 0’ is\\n>> always false [-Wtype-limits]\\n>> 492 | if (dsize < 0)\\n>> |\\n> \\n> Thank you Alexander. Please find attached an attempt to address it.\\n> \\n>> (I wonder, is it accidental that there no other places that triggers\\n>> the warning, or some buildfarm animals had this check enabled before?)\\n> \\n> I can not answer about the buildfarms. Do you think that adding an explicit\\n> check for this warning in meson would help? I am a bit uncertain as I think\\n> that type-limits are included in extra.\\n> \\n> @@ -1748,6 +1748,7 @@ common_warning_flags = [\\n> '-Wshadow=compatible-local',\\n> # This was included in -Wall/-Wformat in older GCC versions\\n> '-Wformat-security',\\n> + '-Wtype-limits',\\n> ]\\n> \\n>>\\n>> It is not a false positive as can be proved by the 002_pg_dump.pl modified as\\n>> follows:\\n>> - program => $ENV{'LZ4'},\\n>>\\n>> + program => 'mv',\\n>>\\n>> args => [\\n>>\\n>> - '-z', '-f', '--rm',\\n>> \\\"$tempdir/compression_lz4_dir/blobs.toc\\\",\\n>> \\\"$tempdir/compression_lz4_dir/blobs.toc.lz4\\\",\\n>> ],\\n>> },\\n> \\n> Correct, it is not a false positive. The existing testing framework provides\\n> limited support for exercising error branches. Especially so when those are\\n> dependent on generated output. \\n> \\n>> A diagnostic logging added shows:\\n>> LZ4File_gets() after LZ4File_read_internal; dsize: 18446744073709551615\\n>>\\n>> and pg_restore fails with:\\n>> error: invalid line in large object TOC file\\n>> \\\".../src/bin/pg_dump/tmp_check/tmp_test_22ri/compression_lz4_dir/blobs.toc\\\": \\\"????\\\"\\n> \\n> It is a good thing that the restore fails with bad input. Yet it should\\n> have failed earlier. The attached makes certain it does fail earlier. \\n> \\n\\nThanks for the patch.\\n\\nI did look if there are other places that might have the same issue, and\\nI think there are - see attached 0002. For example LZ4File_write is\\ndeclared to return size_t, but then it also does\\n\\n if (LZ4F_isError(status))\\n {\\n fs->errcode = status;\\n return -1;\\n }\\n\\nThat won't work :-(\\n\\nAnd these issues may not be restricted to lz4 code - Gzip_write is\\ndeclared to return size_t, but it does\\n\\n return gzwrite(gzfp, ptr, size);\\n\\nand gzwrite returns int. Although, maybe that's correct, because\\ngzwrite() is \\\"0 on error\\\" so maybe this is fine ...\\n\\nHowever, Gzip_read assigns gzread() to size_t, and that does not seem\\ngreat. It probably will still trigger the following pg_fatal() because\\nit'd be very lucky to match the expected 'size', but it's confusing.\\n\\n\\nI wonder whether CompressorState should use int or size_t for the\\nread_func/write_func callbacks. I guess no option is perfect, i.e. no\\ndata type will work for all compression libraries we might use (lz4 uses\\nint while lz4f uses size_t, to there's that).\\n\\nIt's a bit weird the \\\"open\\\" functions return int and the read/write\\nsize_t. Maybe we should stick to int, which is what the old functions\\n(cfwrite etc.) did.\\n\\n\\nBut I think the actual problem here is that the API does not clearly\\ndefine how errors are communicated. I mean, it's nice to return the\\nvalue returned by the library function without \\\"mangling\\\" it by\\nconversion to size_t, but what if the libraries communicate errors in\\ndifferent way? Some may return \\\"0\\\" while others may return \\\"-1\\\".\\n\\nI think the right approach is to handle all library errors and not just\\nlet them through. So Gzip_write() needs to check the return value, and\\neither call pg_fatal() or translate it to an error defined by the API.\\n\\nFor example we might say \\\"returns 0 on error\\\" and then translate all\\nlibrary-specific errors to that.\\n\\n\\nWhile looking at the code I realized a couple function comments don't\\nsay what's returned in case of error, etc. So 0004 adds those.\\n\\n0003 is a couple minor assorted comments/questions:\\n\\n- Should we move ZLIB_OUT_SIZE/ZLIB_IN_SIZE to compress_gzip.c?\\n\\n- Why are LZ4 buffer sizes different (ZLIB has both 4kB)?\\n\\n- I wonder if we actually need LZ4F_HEADER_SIZE_MAX? Is it even possible\\nfor LZ4F_compressBound to return value this small (especially for 16kB\\ninput buffer)?\\n\\n\\n\\nregards\\n\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 13 Mar 2023 21:21:11 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hi Justin,\\n\\nThanks for the patch.\\n\\nOn 3/8/23 02:45, Justin Pryzby wrote:\\n> On Wed, Mar 01, 2023 at 04:52:49PM +0100, Tomas Vondra wrote:\\n>> Thanks. That seems correct to me, but I find it somewhat confusing,\\n>> because we now have\\n>>\\n>> DeflateCompressorInit vs. InitCompressorGzip\\n>>\\n>> DeflateCompressorEnd vs. EndCompressorGzip\\n>>\\n>> DeflateCompressorData - The name doesn't really say what it does (would\\n>> be better to have a verb in there, I think).\\n>>\\n>> I wonder if we can make this somehow clearer?\\n> \\n> To move things along, I updated Georgios' patch:\\n> \\n> Rename DeflateCompressorData() to DeflateCompressorCommon();\\n\\nHmmm, I don't find \\\"common\\\" any clearer than \\\"data\\\" :-( There needs to\\nat least be a comment explaining what \\\"common\\\" does.\\n\\n> Rearrange functions to their original order allowing a cleaner diff to the prior code;\\n\\nOK. I wasn't very enthusiastic about this initially, but after thinking\\nabout it a bit I think it's meaningful to make diffs clearer. But I\\ndon't see much difference with/without the patch. The\\n\\ngit diff --diff-algorithm=minimal -w\\ne9960732a~:src/bin/pg_dump/compress_io.c src/bin/pg_dump/compress_gzip.c\\n\\nProduces ~25k diff with/without the patch. What am I doing wrong?\\n\\n> Change pg_fatal() to an assertion+comment;\\n\\nYeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\\ncould add such protections against \\\"impossible\\\" stuff to a zillion other\\nplaces and the confusion likely outweighs the benefits.\\n\\n> Update the commit message and fix a few typos;\\n> \\n\\nThanks. I don't want to annoy you too much, but could you split the\\npatch into the \\\"empty-data\\\" fix and all the other changes (rearranging\\nfunctions etc.)? I'd rather not mix those in the same commit.\\n\\n\\n>> Also, InitCompressorGzip says this:\\n>>\\n>> /*\\n>> * If the caller has defined a write function, prepare the necessary\\n>> * state. Avoid initializing during the first write call, because End\\n>> * may be called without ever writing any data.\\n>> */\\n>> if (cs->writeF)\\n>> DeflateCompressorInit(cs);\\n>>\\n>> Does it actually make sense to not have writeF defined in some cases?\\n> \\n> InitCompressor is being called for either reading or writing, either of\\n> which could be null:\\n> \\n> src/bin/pg_dump/pg_backup_custom.c: ctx->cs = AllocateCompressor(AH->compression_spec,\\n> src/bin/pg_dump/pg_backup_custom.c- NULL,\\n> src/bin/pg_dump/pg_backup_custom.c- _CustomWriteFunc);\\n> --\\n> src/bin/pg_dump/pg_backup_custom.c: cs = AllocateCompressor(AH->compression_spec,\\n> src/bin/pg_dump/pg_backup_custom.c- _CustomReadFunc, NULL);\\n> \\n> It's confusing that the comment says \\\"Avoid initializing...\\\". What it\\n> really means is \\\"Initialize eagerly...\\\". But that makes more sense in\\n> the context of the commit message for this bugfix than in a comment. So\\n> I changed that too.\\n> \\n> + /* If deflation was initialized, finalize it */ \\n> + if (cs->private_data) \\n> + DeflateCompressorEnd(AH, cs); \\n> \\n> Maybe it'd be more clear if this used \\\"if (cs->writeF)\\\", like in the\\n> init function ?\\n> \\n\\nYeah, if the two checks are equivalent, it'd be better to stick to the\\nsame check everywhere.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Mon, 13 Mar 2023 22:47:12 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n------- Original Message -------\\nOn Monday, March 13th, 2023 at 10:47 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n\\n> \\n> > Change pg_fatal() to an assertion+comment;\\n> \\n> \\n> Yeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\\n> could add such protections against \\\"impossible\\\" stuff to a zillion other\\n> places and the confusion likely outweighs the benefits.\\n> \\n\\nA minor note to add is to not ignore the lessons learned from a7885c9bb.\\n\\nFor example, as the testing framework stands, one can not test that the\\ncontents of the custom format are indeed compressed. One can infer it by\\nexamining the header of the produced dump and searching for the\\ncompression flag. The code responsible for writing the header and the\\ncode responsible for actually dealing with data, is not the same. Also,\\nthe compression library itself will happily read and write uncompressed\\ndata.\\n\\nA pg_fatal, assertion, or similar, is the only guard rail against this\\nkind of error. Without it, the tests will continue passing even after\\ne.g. a wrong initialization of the API. It was such a case that lead to\\na7885c9bb in the first place. I do think that we wish it to be an\\n\\\"impossible\\\" case. Also it will be an untested case with some history\\nwithout such a guard rail.\\n\\nOf course I will not object to removing it, if you think that is more\\nconfusing than useful.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Tue, 14 Mar 2023 11:07:25 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Monday, March 13th, 2023 at 9:21 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n>\\n>\\n>\\n>\\n> On 3/11/23 11:50, gkokolatos@pm.me wrote:\\n>\\n> > ------- Original Message -------\\n> > On Saturday, March 11th, 2023 at 7:00 AM, Alexander Lakhin exclusion@gmail.com wrote:\\n> >\\n> > > Hello,\\n> > > 23.02.2023 23:24, Tomas Vondra wrote:\\n\\n>\\n>\\n> Thanks for the patch.\\n>\\n> I did look if there are other places that might have the same issue, and\\n> I think there are - see attached 0002. For example LZ4File_write is\\n> declared to return size_t, but then it also does\\n>\\n> if (LZ4F_isError(status))\\n> {\\n> fs->errcode = status;\\n>\\n> return -1;\\n> }\\n>\\n> That won't work :-(\\n\\nYou are right. It is confusing.\\n\\n>\\n> And these issues may not be restricted to lz4 code - Gzip_write is\\n> declared to return size_t, but it does\\n>\\n> return gzwrite(gzfp, ptr, size);\\n>\\n> and gzwrite returns int. Although, maybe that's correct, because\\n> gzwrite() is \\\"0 on error\\\" so maybe this is fine ...\\n>\\n> However, Gzip_read assigns gzread() to size_t, and that does not seem\\n> great. It probably will still trigger the following pg_fatal() because\\n> it'd be very lucky to match the expected 'size', but it's confusing.\\n\\nAgreed.\\n\\n>\\n>\\n> I wonder whether CompressorState should use int or size_t for the\\n> read_func/write_func callbacks. I guess no option is perfect, i.e. no\\n> data type will work for all compression libraries we might use (lz4 uses\\n> int while lz4f uses size_t, to there's that).\\n>\\n> It's a bit weird the \\\"open\\\" functions return int and the read/write\\n> size_t. Maybe we should stick to int, which is what the old functions\\n> (cfwrite etc.) did.\\n>\\nYou are right. These functions are modeled by the open/fread/\\nfwrite etc, and they have kept the return types of these ones. Their\\ncallers do check the return value of read_func and write_func against\\nthe requested size of bytes to be transferred.\\n\\n>\\n> But I think the actual problem here is that the API does not clearly\\n> define how errors are communicated. I mean, it's nice to return the\\n> value returned by the library function without \\\"mangling\\\" it by\\n> conversion to size_t, but what if the libraries communicate errors in\\n> different way? Some may return \\\"0\\\" while others may return \\\"-1\\\".\\n\\nAgreed.\\n\\n>\\n> I think the right approach is to handle all library errors and not just\\n> let them through. So Gzip_write() needs to check the return value, and\\n> either call pg_fatal() or translate it to an error defined by the API.\\n\\nIt makes sense. It will change some of the behaviour of the callers,\\nmostly on what constitutes an error, and what error message is emitted.\\nThis is a reasonable change though.\\n\\n>\\n> For example we might say \\\"returns 0 on error\\\" and then translate all\\n> library-specific errors to that.\\n\\nOk.\\n\\n> While looking at the code I realized a couple function comments don't\\n> say what's returned in case of error, etc. So 0004 adds those.\\n>\\n> 0003 is a couple minor assorted comments/questions:\\n>\\n> - Should we move ZLIB_OUT_SIZE/ZLIB_IN_SIZE to compress_gzip.c?\\n\\nIt would make things clearer.\\n\\n> - Why are LZ4 buffer sizes different (ZLIB has both 4kB)?\\n\\nClearly some comments are needed, if the difference makes sense.\\n\\n> - I wonder if we actually need LZ4F_HEADER_SIZE_MAX? Is it even possible\\n> for LZ4F_compressBound to return value this small (especially for 16kB\\n> input buffer)?\\n>\\n\\nI would recommend to keep it. Earlier versions of LZ4F_HEADER_SIZE_MAX\\ndo not have it. Later versions do advise to use it.\\n\\nWould you mind me trying to come with a patch to address your points?\\n\\nCheers,\\n//Georgios\\n\\n>\\n>\\n> regards\\n>\\n>\\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Tue, 14 Mar 2023 15:18:33 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 3/14/23 16:18, gkokolatos@pm.me wrote:\\n> ...> Would you mind me trying to come with a patch to address your points?\\n> \\n\\nThat'd be great, thanks. Please keep it split into smaller patches - two\\nmight work, with one patch for \\\"cosmetic\\\" changes and the other tweaking\\nthe API error-handling stuff.\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Tue, 14 Mar 2023 16:32:56 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 3/14/23 12:07, gkokolatos@pm.me wrote:\\n> \\n> \\n> ------- Original Message -------\\n> On Monday, March 13th, 2023 at 10:47 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n> \\n> \\n> \\n>>\\n>>> Change pg_fatal() to an assertion+comment;\\n>>\\n>>\\n>> Yeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\\n>> could add such protections against \\\"impossible\\\" stuff to a zillion other\\n>> places and the confusion likely outweighs the benefits.\\n>>\\n> \\n> A minor note to add is to not ignore the lessons learned from a7885c9bb.\\n> \\n> For example, as the testing framework stands, one can not test that the\\n> contents of the custom format are indeed compressed. One can infer it by\\n> examining the header of the produced dump and searching for the\\n> compression flag. The code responsible for writing the header and the\\n> code responsible for actually dealing with data, is not the same. Also,\\n> the compression library itself will happily read and write uncompressed\\n> data.\\n> \\n> A pg_fatal, assertion, or similar, is the only guard rail against this\\n> kind of error. Without it, the tests will continue passing even after\\n> e.g. a wrong initialization of the API. It was such a case that lead to\\n> a7885c9bb in the first place. I do think that we wish it to be an\\n> \\\"impossible\\\" case. Also it will be an untested case with some history\\n> without such a guard rail.\\n> \\n\\nSo is the pg_fatal() a dead code or not? My understanding was it's not\\nreally reachable, and the main purpose is to remind people this is not\\npossible. Or am I mistaken/confused?\\n\\nIf it's reachable, can we test it? AFAICS we don't, per the coverage\\nreports.\\n\\nIf it's just a protection against incorrect API initialization, then an\\nassert is the right solution, I think. With proper comment. But can't we\\nactually verify that *during* the initialization?\\n\\nAlso, how come WriteDataToArchiveLZ4() doesn't need this protection too?\\nOr is that due to gzip being the default compression method?\\n\\n> Of course I will not object to removing it, if you think that is more\\n> confusing than useful.\\n> \\n\\nNot sure, I have a feeling I don't quite understand in what situation\\nthis actually helps.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Tue, 14 Mar 2023 16:52:20 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 13, 2023 at 10:47:12PM +0100, Tomas Vondra wrote:\\n> > Rearrange functions to their original order allowing a cleaner diff to the prior code;\\n> \\n> OK. I wasn't very enthusiastic about this initially, but after thinking\\n> about it a bit I think it's meaningful to make diffs clearer. But I\\n> don't see much difference with/without the patch. The\\n> \\n> git diff --diff-algorithm=minimal -w e9960732a~:src/bin/pg_dump/compress_io.c src/bin/pg_dump/compress_gzip.c\\n> \\n> Produces ~25k diff with/without the patch. What am I doing wrong?\\n\\nDo you mean 25 kB of diff ? I agree that the statistics of the diff\\noutput don't change a lot:\\n\\n 1 file changed, 201 insertions(+), 570 deletions(-)\\n 1 file changed, 198 insertions(+), 548 deletions(-)\\n\\nBut try reading the diff while looking for the cause of a bug. It's the\\ndifference between reading 50, two-line changes, and reading a hunk that\\nreplaces 100 lines with a different 100 lines, with empty/unrelated\\nlines randomly thrown in as context.\\n\\nWhen the diff is readable, the pg_fatal() also stands out.\\n\\n> > Change pg_fatal() to an assertion+comment;\\n> \\n> Yeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\\n> could add such protections against \\\"impossible\\\" stuff to a zillion other\\n> places and the confusion likely outweighs the benefits.\\n> \\n> > Update the commit message and fix a few typos;\\n> \\n> Thanks. I don't want to annoy you too much, but could you split the\\n> patch into the \\\"empty-data\\\" fix and all the other changes (rearranging\\n> functions etc.)? I'd rather not mix those in the same commit.\\n\\nI don't know if that makes sense? The \\\"empty-data\\\" fix creates a new\\nfunction called DeflateCompressorInit(). My proposal was to add the new\\nfunction in the same place in the file as it used to be.\\n\\nThe patch also moves the pg_fatal() that's being removed. I don't think\\nit's going to look any cleaner to read a history involving the\\npg_fatal() first being added, then moved, then removed. Anyway, I'll\\nwait while the community continues discussion about the pg_fatal().\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Wed, 15 Mar 2023 19:20:59 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Tuesday, March 14th, 2023 at 4:32 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> \\n> \\n> On 3/14/23 16:18, gkokolatos@pm.me wrote:\\n> \\n> > ...> Would you mind me trying to come with a patch to address your points?\\n> \\n> \\n> That'd be great, thanks. Please keep it split into smaller patches - two\\n> might work, with one patch for \\\"cosmetic\\\" changes and the other tweaking\\n> the API error-handling stuff.\\n\\nPlease find attached a set for it. I will admit that the splitting in the\\nseries might not be ideal and what you requested. It is split on what\\nseemed as a logical units. Please advice how a better split can look like.\\n\\n0001 is unifying types and return values on the API\\n0002 is addressing the constant definitions\\n0003 is your previous 0004 adding comments\\n\\nAs far as the error handling is concerned, you had said upthread:\\n\\n> I think the right approach is to handle all library errors and not just\\n> let them through. So Gzip_write() needs to check the return value, and\\n> either call pg_fatal() or translate it to an error defined by the API.\\n\\nWhile working on it, I thought it would be clearer and more consistent\\nfor the pg_fatal() to be called by the caller of the individual functions.\\nEach individual function can keep track of the specifics of the error\\ninternally. Then the caller upon detecting that there was an error by\\nchecking the return value, can call pg_fatal() with a uniform error\\nmessage and then add the specifics by calling the get_error_func().\\n\\nThoughts?\\n\\nCheers,\\n//Georgios\\n\\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Thu, 16 Mar 2023 17:04:16 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 3/16/23 18:04, gkokolatos@pm.me wrote:\\n> \\n> ------- Original Message -------\\n> On Tuesday, March 14th, 2023 at 4:32 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n>>\\n>> On 3/14/23 16:18, gkokolatos@pm.me wrote:\\n>>\\n>>> ...> Would you mind me trying to come with a patch to address your points?\\n>>\\n>>\\n>> That'd be great, thanks. Please keep it split into smaller patches - two\\n>> might work, with one patch for \\\"cosmetic\\\" changes and the other tweaking\\n>> the API error-handling stuff.\\n> \\n> Please find attached a set for it. I will admit that the splitting in the\\n> series might not be ideal and what you requested. It is split on what\\n> seemed as a logical units. Please advice how a better split can look like.\\n> \\n> 0001 is unifying types and return values on the API\\n> 0002 is addressing the constant definitions\\n> 0003 is your previous 0004 adding comments\\n> \\n\\nThanks. I think the split seems reasonable - the goal was to not mix\\ndifferent changes, and from that POV it works.\\n\\nI'm not sure I understand the Gzip_read/Gzip_write changes in 0001. I\\nmean, gzread/gzwrite returns int, so how does renaming the size_t\\nvariable solve the issue of negative values for errors? I mean, this\\n\\n- size_t ret;\\n+ size_t gzret;\\n\\n- ret = gzread(gzfp, ptr, size);\\n+ gzret = gzread(gzfp, ptr, size);\\n\\nmeans we still lost the information gzread() returned a negative value,\\nno? We'll still probably trigger an error, but it's a bit weird.\\n\\n\\nISTM all this kinda assumes we're processing chunks of memory small\\nenough that we'll never actually overflow int - I did check what the\\ncode in 15 does, and it seems use int and size_t quite arbitrarily.\\n\\nFor example cfread() seems quite sane:\\n\\n int\\n cfread(void *ptr, int size, cfp *fp)\\n {\\n int ret;\\n ...\\n ret = gzread(fp->compressedfp, ptr, size);\\n ...\\n return ret;\\n }\\n\\nbut then _PrintFileData() happily stashes it into a size_t, ignoring the\\nsignedness. Surely, if\\n\\n static void\\n _PrintFileData(ArchiveHandle *AH, char *filename)\\n {\\n size_t cnt;\\n ...\\n while ((cnt = cfread(buf, buflen, cfp)))\\n {\\n ahwrite(buf, 1, cnt, AH);\\n }\\n ...\\n }\\n\\nUnless I'm missing something, if gzread() ever returns -1 or some other\\nnegative error value, we'll cast it to size_t, while condition will\\nevaluate to \\\"true\\\" and we'll happily chew on some random chunk of data.\\n\\nSo the confusion is (at least partially) a preexisting issue ...\\n\\nFor gzwrite() it seems to be fine, because that only returns 0 on error.\\nOTOH it's defined to take 'int size' but then we happily pass size_t\\nvalues to it.\\n\\nAs I wrote earlier, this apparently assumes we never need to deal with\\nbuffers larger than int, and I don't think we have the ambition to relax\\nthat (I'm not sure it's even needed / possible).\\n\\n\\nI see the read/write functions are now defined as int, but we only ever\\nreturn 0/1 from them, and then interpret that as bool. Why not to define\\nit like that? I don't think we need to adhere to the custom that\\neverything returns \\\"int\\\". This is an internal API. Or if we want to\\nstick to int, I'd define meaningful \\\"nice\\\" constants for 0/1.\\n\\n\\n\\n0002 seems fine to me. I see you've ditched the idea of having two\\nseparate buffers, and replaced them with DEFAULT_IO_BUFFER_SIZE. Fine\\nwith me, although I wonder if this might have negative impact on\\nperformance or something (but I doubt that).\\n\\n0003 seems fine too.\\n\\n\\n> As far as the error handling is concerned, you had said upthread:\\n> \\n>> I think the right approach is to handle all library errors and not just\\n>> let them through. So Gzip_write() needs to check the return value, and\\n>> either call pg_fatal() or translate it to an error defined by the API.\\n> \\n> While working on it, I thought it would be clearer and more consistent\\n> for the pg_fatal() to be called by the caller of the individual functions.\\n> Each individual function can keep track of the specifics of the error\\n> internally. Then the caller upon detecting that there was an error by\\n> checking the return value, can call pg_fatal() with a uniform error\\n> message and then add the specifics by calling the get_error_func().\\n> \\n\\nI agree it's cleaner the way you did it.\\n\\nI was thinking that with each compression function handling error\\ninternally, the callers would not need to do that. But I haven't\\nrealized there's logic to detect ENOSPC and so on, and we'd need to\\nduplicate that in every compression func.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Mar 2023 22:20:10 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 3/16/23 01:20, Justin Pryzby wrote:\\n> On Mon, Mar 13, 2023 at 10:47:12PM +0100, Tomas Vondra wrote:\\n>>> Rearrange functions to their original order allowing a cleaner diff to the prior code;\\n>>\\n>> OK. I wasn't very enthusiastic about this initially, but after thinking\\n>> about it a bit I think it's meaningful to make diffs clearer. But I\\n>> don't see much difference with/without the patch. The\\n>>\\n>> git diff --diff-algorithm=minimal -w e9960732a~:src/bin/pg_dump/compress_io.c src/bin/pg_dump/compress_gzip.c\\n>>\\n>> Produces ~25k diff with/without the patch. What am I doing wrong?\\n> \\n> Do you mean 25 kB of diff ?\\n\\nYes, if you redirect the git-diff to a file, it's a 25kB file.\\n\\n> I agree that the statistics of the diff output don't change a lot:\\n> \\n> 1 file changed, 201 insertions(+), 570 deletions(-)\\n> 1 file changed, 198 insertions(+), 548 deletions(-)\\n> \\n> But try reading the diff while looking for the cause of a bug. It's the\\n> difference between reading 50, two-line changes, and reading a hunk that\\n> replaces 100 lines with a different 100 lines, with empty/unrelated\\n> lines randomly thrown in as context.\\n> \\n> When the diff is readable, the pg_fatal() also stands out.\\n> \\n\\nI don't know, maybe I'm doing something wrong or maybe I just am bad at\\nlooking at diffs, but if I apply the patch you submitted on 8/3 and do\\nthe git-diff above (output attached), it seems pretty incomprehensible\\nto me :-( I don't see 50 two-line changes (I certainly wouldn't be able\\nto identify the root cause of the bug based on that).\\n\\n>>> Change pg_fatal() to an assertion+comment;\\n>>\\n>> Yeah, that's reasonable. I'd even ditch the assert/comment, TBH. We\\n>> could add such protections against \\\"impossible\\\" stuff to a zillion other\\n>> places and the confusion likely outweighs the benefits.\\n>>\\n>>> Update the commit message and fix a few typos;\\n>>\\n>> Thanks. I don't want to annoy you too much, but could you split the\\n>> patch into the \\\"empty-data\\\" fix and all the other changes (rearranging\\n>> functions etc.)? I'd rather not mix those in the same commit.\\n> \\n> I don't know if that makes sense? The \\\"empty-data\\\" fix creates a new\\n> function called DeflateCompressorInit(). My proposal was to add the new\\n> function in the same place in the file as it used to be.\\n> \\n\\nGot it. In that case I agree it's fine to do that in a single commit.\\n\\n> The patch also moves the pg_fatal() that's being removed. I don't think\\n> it's going to look any cleaner to read a history involving the\\n> pg_fatal() first being added, then moved, then removed. Anyway, I'll\\n> wait while the community continues discussion about the pg_fatal().\\n> \\n\\nI think the agreement was to replace the pg_fatal with and assert, and I\\nsee your patch already does that.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Thu, 16 Mar 2023 23:30:50 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 16, 2023 at 11:30:50PM +0100, Tomas Vondra wrote:\\n> On 3/16/23 01:20, Justin Pryzby wrote:\\n> > But try reading the diff while looking for the cause of a bug. It's the\\n> > difference between reading 50, two-line changes, and reading a hunk that\\n> > replaces 100 lines with a different 100 lines, with empty/unrelated\\n> > lines randomly thrown in as context.\\n>\\n> I don't know, maybe I'm doing something wrong or maybe I just am bad at\\n> looking at diffs, but if I apply the patch you submitted on 8/3 and do\\n> the git-diff above (output attached), it seems pretty incomprehensible\\n> to me :-( I don't see 50 two-line changes (I certainly wouldn't be able\\n> to identify the root cause of the bug based on that).\\n\\nIt's true that most of the diff is still incomprehensible...\\n\\nBut look at the part relevant to the \\\"empty-data\\\" bug:\\n\\n[... incomprehensible changes elided ...]\\n> static void\\n> -InitCompressorZlib(CompressorState *cs, int level)\\n> +DeflateCompressorInit(CompressorState *cs)\\n> {\\n> +\\tGzipCompressorState *gzipcs;\\n> \\tz_streamp\\tzp;\\n> \\n> -\\tzp = cs->zp = (z_streamp) pg_malloc(sizeof(z_stream));\\n> +\\tgzipcs = (GzipCompressorState *) pg_malloc0(sizeof(GzipCompressorState));\\n> +\\tzp = gzipcs->zp = (z_streamp) pg_malloc(sizeof(z_stream));\\n> \\tzp->zalloc = Z_NULL;\\n> \\tzp->zfree = Z_NULL;\\n> \\tzp->opaque = Z_NULL;\\n> \\n> \\t/*\\n> -\\t * zlibOutSize is the buffer size we tell zlib it can output to. We\\n> -\\t * actually allocate one extra byte because some routines want to append a\\n> -\\t * trailing zero byte to the zlib output.\\n> +\\t * outsize is the buffer size we tell zlib it can output to. We actually\\n> +\\t * allocate one extra byte because some routines want to append a trailing\\n> +\\t * zero byte to the zlib output.\\n> \\t */\\n> -\\tcs->zlibOut = (char *) pg_malloc(ZLIB_OUT_SIZE + 1);\\n> -\\tcs->zlibOutSize = ZLIB_OUT_SIZE;\\n> +\\tgzipcs->outbuf = pg_malloc(ZLIB_OUT_SIZE + 1);\\n> +\\tgzipcs->outsize = ZLIB_OUT_SIZE;\\n> \\n> -\\tif (deflateInit(zp, level) != Z_OK)\\n> -\\t\\tpg_fatal(\\\"could not initialize compression library: %s\\\",\\n> -\\t\\t\\t\\t zp->msg);\\n> +\\t/* -Z 0 uses the \\\"None\\\" compressor -- not zlib with no compression */\\n> +\\tAssert(cs->compression_spec.level != 0);\\n> +\\n> +\\tif (deflateInit(zp, cs->compression_spec.level) != Z_OK)\\n> +\\t\\tpg_fatal(\\\"could not initialize compression library: %s\\\", zp->msg);\\n> \\n> \\t/* Just be paranoid - maybe End is called after Start, with no Write */\\n> -\\tzp->next_out = (void *) cs->zlibOut;\\n> -\\tzp->avail_out = cs->zlibOutSize;\\n> +\\tzp->next_out = gzipcs->outbuf;\\n> +\\tzp->avail_out = gzipcs->outsize;\\n> +\\n> +\\t/* Keep track of gzipcs */\\n> +\\tcs->private_data = gzipcs;\\n> }\\n> \\n> static void\\n> -EndCompressorZlib(ArchiveHandle *AH, CompressorState *cs)\\n> +DeflateCompressorEnd(ArchiveHandle *AH, CompressorState *cs)\\n> {\\n> -\\tz_streamp\\tzp = cs->zp;\\n> +\\tGzipCompressorState *gzipcs = (GzipCompressorState *) cs->private_data;\\n> +\\tz_streamp\\tzp;\\n> \\n> +\\tzp = gzipcs->zp;\\n> \\tzp->next_in = NULL;\\n> \\tzp->avail_in = 0;\\n> \\n> \\t/* Flush any remaining data from zlib buffer */\\n> -\\tDeflateCompressorZlib(AH, cs, true);\\n> +\\tDeflateCompressorCommon(AH, cs, true);\\n> \\n> \\tif (deflateEnd(zp) != Z_OK)\\n> \\t\\tpg_fatal(\\\"could not close compression stream: %s\\\", zp->msg);\\n> \\n> -\\tfree(cs->zlibOut);\\n> -\\tfree(cs->zp);\\n> +\\tpg_free(gzipcs->outbuf);\\n> +\\tpg_free(gzipcs->zp);\\n> +\\tpg_free(gzipcs);\\n> +\\tcs->private_data = NULL;\\n> }\\n> \\n> static void\\n> -DeflateCompressorZlib(ArchiveHandle *AH, CompressorState *cs, bool flush)\\n> +DeflateCompressorCommon(ArchiveHandle *AH, CompressorState *cs, bool flush)\\n> {\\n> -\\tz_streamp\\tzp = cs->zp;\\n> -\\tchar\\t *out = cs->zlibOut;\\n> +\\tGzipCompressorState *gzipcs = (GzipCompressorState *) cs->private_data;\\n> +\\tz_streamp\\tzp = gzipcs->zp;\\n> +\\tvoid\\t *out = gzipcs->outbuf;\\n> \\tint\\t\\t\\tres = Z_OK;\\n> \\n> -\\twhile (cs->zp->avail_in != 0 || flush)\\n> +\\twhile (gzipcs->zp->avail_in != 0 || flush)\\n> \\t{\\n> \\t\\tres = deflate(zp, flush ? Z_FINISH : Z_NO_FLUSH);\\n> \\t\\tif (res == Z_STREAM_ERROR)\\n> \\t\\t\\tpg_fatal(\\\"could not compress data: %s\\\", zp->msg);\\n> -\\t\\tif ((flush && (zp->avail_out < cs->zlibOutSize))\\n> +\\t\\tif ((flush && (zp->avail_out < gzipcs->outsize))\\n> \\t\\t\\t|| (zp->avail_out == 0)\\n> \\t\\t\\t|| (zp->avail_in != 0)\\n> \\t\\t\\t)\\n> @@ -289,18 +122,18 @@ DeflateCompressorZlib(ArchiveHandle *AH, CompressorState *cs, bool flush)\\n> \\t\\t\\t * chunk is the EOF marker in the custom format. This should never\\n> \\t\\t\\t * happen but ...\\n> \\t\\t\\t */\\n> -\\t\\t\\tif (zp->avail_out < cs->zlibOutSize)\\n> +\\t\\t\\tif (zp->avail_out < gzipcs->outsize)\\n> \\t\\t\\t{\\n> \\t\\t\\t\\t/*\\n> \\t\\t\\t\\t * Any write function should do its own error checking but to\\n> \\t\\t\\t\\t * make sure we do a check here as well ...\\n> \\t\\t\\t\\t */\\n> -\\t\\t\\t\\tsize_t\\t\\tlen = cs->zlibOutSize - zp->avail_out;\\n> +\\t\\t\\t\\tsize_t\\t\\tlen = gzipcs->outsize - zp->avail_out;\\n> \\n> -\\t\\t\\t\\tcs->writeF(AH, out, len);\\n> +\\t\\t\\t\\tcs->writeF(AH, (char *) out, len);\\n> \\t\\t\\t}\\n> -\\t\\t\\tzp->next_out = (void *) out;\\n> -\\t\\t\\tzp->avail_out = cs->zlibOutSize;\\n> +\\t\\t\\tzp->next_out = out;\\n> +\\t\\t\\tzp->avail_out = gzipcs->outsize;\\n> \\t\\t}\\n> \\n> \\t\\tif (res == Z_STREAM_END)\\n> @@ -309,16 +142,26 @@ DeflateCompressorZlib(ArchiveHandle *AH, CompressorState *cs, bool flush)\\n> }\\n> \\n> static void\\n> -WriteDataToArchiveZlib(ArchiveHandle *AH, CompressorState *cs,\\n> -\\t\\t\\t\\t\\t const char *data, size_t dLen)\\n> +EndCompressorGzip(ArchiveHandle *AH, CompressorState *cs)\\n> {\\n> -\\tcs->zp->next_in = (void *) unconstify(char *, data);\\n> -\\tcs->zp->avail_in = dLen;\\n> -\\tDeflateCompressorZlib(AH, cs, false);\\n> +\\t/* If deflation was initialized, finalize it */\\n> +\\tif (cs->private_data)\\n> +\\t\\tDeflateCompressorEnd(AH, cs);\\n> }\\n> \\n> static void\\n> -ReadDataFromArchiveZlib(ArchiveHandle *AH, ReadFunc readF)\\n> +WriteDataToArchiveGzip(ArchiveHandle *AH, CompressorState *cs,\\n> +\\t\\t\\t\\t\\t const void *data, size_t dLen)\\n> +{\\n> +\\tGzipCompressorState *gzipcs = (GzipCompressorState *) cs->private_data;\\n> +\\n> +\\tgzipcs->zp->next_in = (void *) unconstify(void *, data);\\n> +\\tgzipcs->zp->avail_in = dLen;\\n> +\\tDeflateCompressorCommon(AH, cs, false);\\n> +}\\n> +\\n> +static void\\n> +ReadDataFromArchiveGzip(ArchiveHandle *AH, CompressorState *cs)\\n> {\\n> \\tz_streamp\\tzp;\\n> \\tchar\\t *out;\\n> @@ -342,7 +185,7 @@ ReadDataFromArchiveZlib(ArchiveHandle *AH, ReadFunc readF)\\n> \\t\\t\\t\\t zp->msg);\\n> \\n> \\t/* no minimal chunk size for zlib */\\n> -\\twhile ((cnt = readF(AH, &buf, &buflen)))\\n> +\\twhile ((cnt = cs->readF(AH, &buf, &buflen)))\\n> \\t{\\n> \\t\\tzp->next_in = (void *) buf;\\n> \\t\\tzp->avail_in = cnt;\\n> @@ -382,389 +225,196 @@ ReadDataFromArchiveZlib(ArchiveHandle *AH, ReadFunc readF)\\n> \\tfree(out);\\n> \\tfree(zp);\\n> }\\n[... more incomprehensible changes elided ...]\\n\\n\\n\",\n \"msg_date\": \"Thu, 16 Mar 2023 17:58:42 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 3/16/23 23:58, Justin Pryzby wrote:\\n> On Thu, Mar 16, 2023 at 11:30:50PM +0100, Tomas Vondra wrote:\\n>> On 3/16/23 01:20, Justin Pryzby wrote:\\n>>> But try reading the diff while looking for the cause of a bug. It's the\\n>>> difference between reading 50, two-line changes, and reading a hunk that\\n>>> replaces 100 lines with a different 100 lines, with empty/unrelated\\n>>> lines randomly thrown in as context.\\n>>\\n>> I don't know, maybe I'm doing something wrong or maybe I just am bad at\\n>> looking at diffs, but if I apply the patch you submitted on 8/3 and do\\n>> the git-diff above (output attached), it seems pretty incomprehensible\\n>> to me :-( I don't see 50 two-line changes (I certainly wouldn't be able\\n>> to identify the root cause of the bug based on that).\\n> \\n> It's true that most of the diff is still incomprehensible...\\n> \\n> But look at the part relevant to the \\\"empty-data\\\" bug:\\n> \\n\\nWell, yeah. If you know where to look, and if you squint just the right\\nway, then you can see any bug. I don't think I'd be able to spot the bug\\nin the diff unless I knew in advance what the bug is.\\n\\nThat being said, I don't object to moving the function etc. Unless there\\nare alternative ideas how to fix the empty-data issue, I'll get this\\ncommitted after playing with it a bit more.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 17 Mar 2023 03:02:22 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Thursday, March 16th, 2023 at 10:20 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> \\n> \\n> On 3/16/23 18:04, gkokolatos@pm.me wrote:\\n> \\n> > ------- Original Message -------\\n> > On Tuesday, March 14th, 2023 at 4:32 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\\n> > \\n> > > On 3/14/23 16:18, gkokolatos@pm.me wrote:\\n> > > \\n> > > > ...> Would you mind me trying to come with a patch to address your points?\\n> > > \\n> > > That'd be great, thanks. Please keep it split into smaller patches - two\\n> > > might work, with one patch for \\\"cosmetic\\\" changes and the other tweaking\\n> > > the API error-handling stuff.\\n> > \\n> > Please find attached a set for it. I will admit that the splitting in the\\n> > series might not be ideal and what you requested. It is split on what\\n> > seemed as a logical units. Please advice how a better split can look like.\\n> > \\n> > 0001 is unifying types and return values on the API\\n> > 0002 is addressing the constant definitions\\n> > 0003 is your previous 0004 adding comments\\n> \\n> \\n> Thanks. I think the split seems reasonable - the goal was to not mix\\n> different changes, and from that POV it works.\\n> \\n> I'm not sure I understand the Gzip_read/Gzip_write changes in 0001. I\\n> mean, gzread/gzwrite returns int, so how does renaming the size_t\\n> variable solve the issue of negative values for errors? I mean, this\\n> \\n> - size_t ret;\\n> + size_t gzret;\\n> \\n> - ret = gzread(gzfp, ptr, size);\\n> + gzret = gzread(gzfp, ptr, size);\\n> \\n> means we still lost the information gzread() returned a negative value,\\n> no? We'll still probably trigger an error, but it's a bit weird.\\n\\nYou are obviously correct. My bad, I miss-read the return type of gzread().\\n\\nPlease find an amended version attached.\\n\\n> Unless I'm missing something, if gzread() ever returns -1 or some other\\n> negative error value, we'll cast it to size_t, while condition will\\n> evaluate to \\\"true\\\" and we'll happily chew on some random chunk of data.\\n> \\n> So the confusion is (at least partially) a preexisting issue ...\\n> \\n> For gzwrite() it seems to be fine, because that only returns 0 on error.\\n> OTOH it's defined to take 'int size' but then we happily pass size_t\\n> values to it.\\n> \\n> As I wrote earlier, this apparently assumes we never need to deal with\\n> buffers larger than int, and I don't think we have the ambition to relax\\n> that (I'm not sure it's even needed / possible).\\n\\nAgreed.\\n\\n\\n> I see the read/write functions are now defined as int, but we only ever\\n> return 0/1 from them, and then interpret that as bool. Why not to define\\n> it like that? I don't think we need to adhere to the custom that\\n> everything returns \\\"int\\\". This is an internal API. Or if we want to\\n> stick to int, I'd define meaningful \\\"nice\\\" constants for 0/1.\\n\\nThe return types are now booleans and the callers have been made aware.\\n\\n\\n> 0002 seems fine to me. I see you've ditched the idea of having two\\n> separate buffers, and replaced them with DEFAULT_IO_BUFFER_SIZE. Fine\\n> with me, although I wonder if this might have negative impact on\\n> performance or something (but I doubt that).\\n> \\n\\nI doubt that too. Thank you.\\n\\n> 0003 seems fine too.\\n\\nThank you.\\n\\n\\n> > As far as the error handling is concerned, you had said upthread:\\n> > \\n> > > I think the right approach is to handle all library errors and not just\\n> > > let them through. So Gzip_write() needs to check the return value, and\\n> > > either call pg_fatal() or translate it to an error defined by the API.\\n> > \\n> > While working on it, I thought it would be clearer and more consistent\\n> > for the pg_fatal() to be called by the caller of the individual functions.\\n> > Each individual function can keep track of the specifics of the error\\n> > internally. Then the caller upon detecting that there was an error by\\n> > checking the return value, can call pg_fatal() with a uniform error\\n> > message and then add the specifics by calling the get_error_func().\\n> \\n> \\n> I agree it's cleaner the way you did it.\\n> \\n> I was thinking that with each compression function handling error\\n> internally, the callers would not need to do that. But I haven't\\n> realized there's logic to detect ENOSPC and so on, and we'd need to\\n> duplicate that in every compression func.\\n> \\n\\nIf you agree, I can prepare a patch to improve on the error handling\\naspect of the API as a separate thread, since here we are trying to\\nfocus on correctness.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Fri, 17 Mar 2023 15:43:58 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 3/17/23 16:43, gkokolatos@pm.me wrote:\\n>>\\n>> ...\\n>>\\n>> I agree it's cleaner the way you did it.\\n>>\\n>> I was thinking that with each compression function handling error\\n>> internally, the callers would not need to do that. But I haven't\\n>> realized there's logic to detect ENOSPC and so on, and we'd need to\\n>> duplicate that in every compression func.\\n>>\\n> \\n> If you agree, I can prepare a patch to improve on the error handling\\n> aspect of the API as a separate thread, since here we are trying to\\n> focus on correctness.\\n> \\n\\nYes, that makes sense. There are far too many patches in this thread\\nalready ...\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Fri, 17 Mar 2023 17:56:58 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nI was preparing to get the 3 cleanup patches pushed, so I\\nupdated/reworded the commit messages a bit (attached, please check).\\n\\nBut I noticed the commit message for 0001 said:\\n\\n In passing save the appropriate errno in LZ4File_open_write in\\n case that the caller is not using the API's get_error_func.\\n\\nI think that's far too low-level for a commit message, it'd be much more\\nappropriate for a comment at the function.\\n\\nHowever, do we even need this behavior? I was looking for code calling\\nthis function without using get_error_func(), but no luck. And if there\\nis such caller, shouldn't we fix it to use get_error_func()?\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 20 Mar 2023 21:48:58 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 17, 2023 at 03:43:58PM +0000, gkokolatos@pm.me wrote:\\n> From a174cdff4ec8aad59f5bcc7e8d52218a14fe56fc Mon Sep 17 00:00:00 2001\\n> From: Georgios Kokolatos <gkokolatos@pm.me>\\n> Date: Fri, 17 Mar 2023 14:45:58 +0000\\n> Subject: [PATCH v3 1/3] Improve type handling in pg_dump's compress file API\\n\\n> -int\\n> +bool\\n> EndCompressFileHandle(CompressFileHandle *CFH)\\n> {\\n> -\\tint\\t\\t\\tret = 0;\\n> +\\tbool\\t\\tret = 0;\\n\\nShould say \\\"= false\\\" ?\\n\\n> \\t/*\\n> \\t * Write 'size' bytes of data into the file from 'ptr'.\\n> +\\t *\\n> +\\t * Returns true on success and false on error.\\n> +\\t */\\n> + bool (*write_func) (const void *ptr, size_t size,\\n\\n> - * Get a pointer to a string that describes an error that occurred during a\\n> - * compress file handle operation.\\n> + * Get a pointer to a string that describes an error that occurred during\\n> + * a compress file handle operation.\\n> */\\n> const char *(*get_error_func) (CompressFileHandle *CFH);\\n\\nThis should mention that the error accessible in error_func() applies (only) to\\nwrite_func() ?\\n\\nAs long as this touches pg_backup_directory.c you could update the\\nheader comment to refer to \\\"compressed extensions\\\", not just .gz.\\n\\nI noticed that EndCompressorLZ4() tests \\\"if (LZ4cs)\\\", but that should\\nalways be true.\\n\\nI was able to convert the zstd patch to this new API with no issue.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Mon, 20 Mar 2023 17:40:32 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nI looked at this again, and I realized I misunderstood the bit about\\nerrno in LZ4File_open_write a bit. I now see it simply just brings the\\nfunction in line with Gzip_open_write(), so that the callers can just do\\npg_fatal(\\\"%m\\\"). I still think the special \\\"errno\\\" handling in this one\\nplace feels a bit random, and handling it by get_error_func() would be\\nnicer, but we can leave that for a separate patch - no need to block\\nthese changes because of that.\\n\\nSo pushed all three parts, after updating the commit messages a bit.\\n\\nThis leaves the empty-data issue (which we have a fix for) and the\\nswitch to LZ4F. And then the zstd part.\\n\\n\\nOn 3/20/23 23:40, Justin Pryzby wrote:\\n> On Fri, Mar 17, 2023 at 03:43:58PM +0000, gkokolatos@pm.me wrote:\\n>> From a174cdff4ec8aad59f5bcc7e8d52218a14fe56fc Mon Sep 17 00:00:00 2001\\n>> From: Georgios Kokolatos <gkokolatos@pm.me>\\n>> Date: Fri, 17 Mar 2023 14:45:58 +0000\\n>> Subject: [PATCH v3 1/3] Improve type handling in pg_dump's compress file API\\n> \\n>> -int\\n>> +bool\\n>> EndCompressFileHandle(CompressFileHandle *CFH)\\n>> {\\n>> -\\tint\\t\\t\\tret = 0;\\n>> +\\tbool\\t\\tret = 0;\\n> \\n> Should say \\\"= false\\\" ?\\n> \\n\\nRight, fixed.\\n\\n>> \\t/*\\n>> \\t * Write 'size' bytes of data into the file from 'ptr'.\\n>> +\\t *\\n>> +\\t * Returns true on success and false on error.\\n>> +\\t */\\n>> + bool (*write_func) (const void *ptr, size_t size,\\n> \\n>> - * Get a pointer to a string that describes an error that occurred during a\\n>> - * compress file handle operation.\\n>> + * Get a pointer to a string that describes an error that occurred during\\n>> + * a compress file handle operation.\\n>> */\\n>> const char *(*get_error_func) (CompressFileHandle *CFH);\\n> \\n> This should mention that the error accessible in error_func() applies (only) to\\n> write_func() ?\\n> \\n> As long as this touches pg_backup_directory.c you could update the\\n> header comment to refer to \\\"compressed extensions\\\", not just .gz.\\n> \\n> I noticed that EndCompressorLZ4() tests \\\"if (LZ4cs)\\\", but that should\\n> always be true.\\n> \\n\\nI haven't done these two things. We can/should do that, but it didn't\\nfit into the three patches.\\n\\n> I was able to convert the zstd patch to this new API with no issue.\\n> \\n\\nGood to hear.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Thu, 23 Mar 2023 18:10:27 +0100\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n\\n> \\n> So pushed all three parts, after updating the commit messages a bit.\\n\\nThank you very much.\\n\\n> \\n> This leaves the empty-data issue (which we have a fix for) and the\\n> switch to LZ4F. And then the zstd part.\\n\\nPlease expect promptly a patch for the switch to frames.\\n\\nCheers,\\n//Georgios\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 24 Mar 2023 09:30:28 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Thursday, March 16th, 2023 at 11:30 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> \\n> \\n> On 3/16/23 01:20, Justin Pryzby wrote:\\n> \\n> > On Mon, Mar 13, 2023 at 10:47:12PM +0100, Tomas Vondra wrote:\\n> > \\n> > > \\n> > > Thanks. I don't want to annoy you too much, but could you split the\\n> > > patch into the \\\"empty-data\\\" fix and all the other changes (rearranging\\n> > > functions etc.)? I'd rather not mix those in the same commit.\\n> > \\n> > I don't know if that makes sense? The \\\"empty-data\\\" fix creates a new\\n> > function called DeflateCompressorInit(). My proposal was to add the new\\n> > function in the same place in the file as it used to be.\\n> \\n> \\n> Got it. In that case I agree it's fine to do that in a single commit.\\n\\nFor what is worth, I think that this patch should get a +1 and get in. It\\nsolves the empty writes problem and includes a test to a previous untested\\ncase.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> > The patch also moves the pg_fatal() that's being removed. I don't think\\n> > it's going to look any cleaner to read a history involving the\\n> > pg_fatal() first being added, then moved, then removed. Anyway, I'll\\n> > wait while the community continues discussion about the pg_fatal().\\n> \\n> \\n> I think the agreement was to replace the pg_fatal with and assert, and I\\n> see your patch already does that.\\n> \\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Mon, 27 Mar 2023 22:34:52 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me <gkokolatos@pm.me> wrote:\\n\\n> \\n> ------- Original Message -------\\n> On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\\n> \\n> > This leaves the empty-data issue (which we have a fix for) and the\\n> > switch to LZ4F. And then the zstd part.\\n> \\n> Please expect promptly a patch for the switch to frames.\\n\\nPlease find the expected patch attached. Note that the bulk of the\\npatch is code unification, variable renaming to something more\\nappropriate, and comment addition. These are changes that are not\\nstrictly necessary to switch to LZ4F. I do believe that are\\nessential for code hygiene after the switch and they do belong\\non the same commit. \\n\\nCheers,\\n//Georgios\\n\\n> \\n> Cheers,\\n> //Georgios\",\n \"msg_date\": \"Tue, 28 Mar 2023 16:07:30 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 3/28/23 18:07, gkokolatos@pm.me wrote:\\n> \\n> \\n> \\n> \\n> \\n> ------- Original Message -------\\n> On Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me <gkokolatos@pm.me> wrote:\\n> \\n>>\\n>> ------- Original Message -------\\n>> On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\\n>>\\n>>> This leaves the empty-data issue (which we have a fix for) and the\\n>>> switch to LZ4F. And then the zstd part.\\n>>\\n>> Please expect promptly a patch for the switch to frames.\\n> \\n> Please find the expected patch attached. Note that the bulk of the\\n> patch is code unification, variable renaming to something more\\n> appropriate, and comment addition. These are changes that are not\\n> strictly necessary to switch to LZ4F. I do believe that are\\n> essential for code hygiene after the switch and they do belong\\n> on the same commit. \\n> \\n\\nThanks!\\n\\nI agree the renames & cleanup are appropriate - it'd be silly to stick\\nto misleading naming etc. Would it make sense to split the patch into\\ntwo, to separate the renames and the switch to lz4f?\\n\\nThat'd make it the changes necessary for lz4f switch clearer.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Tue, 28 Mar 2023 18:40:03 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 28, 2023 at 06:40:03PM +0200, Tomas Vondra wrote:\\n> On 3/28/23 18:07, gkokolatos@pm.me wrote:\\n> > ------- Original Message -------\\n> > On Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me <gkokolatos@pm.me> wrote:\\n> > \\n> >> ------- Original Message -------\\n> >> On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\\n> >>\\n> >>> This leaves the empty-data issue (which we have a fix for) and the\\n> >>> switch to LZ4F. And then the zstd part.\\n> >>\\n> >> Please expect promptly a patch for the switch to frames.\\n> > \\n> > Please find the expected patch attached. Note that the bulk of the\\n> > patch is code unification, variable renaming to something more\\n> > appropriate, and comment addition. These are changes that are not\\n> > strictly necessary to switch to LZ4F. I do believe that are\\n> > essential for code hygiene after the switch and they do belong\\n> > on the same commit. \\n> \\n> Thanks!\\n> \\n> I agree the renames & cleanup are appropriate - it'd be silly to stick\\n> to misleading naming etc. Would it make sense to split the patch into\\n> two, to separate the renames and the switch to lz4f?\\n> That'd make it the changes necessary for lz4f switch clearer.\\n\\nI don't think so. Did you mean separate commits only for review ?\\n\\nThe patch is pretty readable - the File API has just some renames, and\\nthe compressor API is what's being replaced, which isn't going to be any\\nmore clear.\\n\\n@Georgeos: did you consider using a C union in LZ4State, to separate the\\nparts used by the different APIs ?\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Tue, 28 Mar 2023 13:47:21 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 3/28/23 18:07, gkokolatos@pm.me wrote:\\n> \\n> ------- Original Message -------\\n> On Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me <gkokolatos@pm.me> wrote:\\n> \\n>>\\n>> ------- Original Message -------\\n>> On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\\n>>\\n>>> This leaves the empty-data issue (which we have a fix for) and the\\n>>> switch to LZ4F. And then the zstd part.\\n>>\\n>> Please expect promptly a patch for the switch to frames.\\n> \\n> Please find the expected patch attached. Note that the bulk of the\\n> patch is code unification, variable renaming to something more\\n> appropriate, and comment addition. These are changes that are not\\n> strictly necessary to switch to LZ4F. I do believe that are\\n> essential for code hygiene after the switch and they do belong\\n> on the same commit. \\n> \\n\\n\\nI think the patch is fine, but I'm wondering if the renames shouldn't go\\na bit further. It removes references to LZ4File struct, but there's a\\nbunch of functions with LZ4File_ prefix. Why not to simply use LZ4_\\nprefix? We don't have GzipFile either.\\n\\nSure, it might be a bit confusing because lz4.h uses LZ4_ prefix, but\\nthen we probably should not define LZ4_compressor_init ...\\n\\nAlso, maybe the comments shouldn't use \\\"File API\\\" when compress_io.c\\ncalls that \\\"Compressed stream API\\\".\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 29 Mar 2023 00:02:00 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 3/28/23 00:34, gkokolatos@pm.me wrote:\\n> \\n> ...\\n>\\n>> Got it. In that case I agree it's fine to do that in a single commit.\\n> \\n> For what is worth, I think that this patch should get a +1 and get in. It\\n> solves the empty writes problem and includes a test to a previous untested\\n> case.\\n> \\n\\nPushed, after updating / rewording the commit message a little bit.\\n\\nThanks!\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 29 Mar 2023 02:40:15 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Wednesday, March 29th, 2023 at 12:02 AM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> On 3/28/23 18:07, gkokolatos@pm.me wrote:\\n> \\n> > ------- Original Message -------\\n> > On Friday, March 24th, 2023 at 10:30 AM, gkokolatos@pm.me gkokolatos@pm.me wrote:\\n> > \\n> > > ------- Original Message -------\\n> > > On Thursday, March 23rd, 2023 at 6:10 PM, Tomas Vondra tomas.vondra@enterprisedb.com wrote:\\n> > > \\n> > > > This leaves the empty-data issue (which we have a fix for) and the\\n> > > > switch to LZ4F. And then the zstd part.\\n> > > \\n> > > Please expect promptly a patch for the switch to frames.\\n> > \\n> > Please find the expected patch attached. Note that the bulk of the\\n> > patch is code unification, variable renaming to something more\\n> > appropriate, and comment addition. These are changes that are not\\n> > strictly necessary to switch to LZ4F. I do believe that are\\n> > essential for code hygiene after the switch and they do belong\\n> > on the same commit.\\n> \\n> \\n> I think the patch is fine, but I'm wondering if the renames shouldn't go\\n> a bit further. It removes references to LZ4File struct, but there's a\\n> bunch of functions with LZ4File_ prefix. Why not to simply use LZ4_\\n> prefix? We don't have GzipFile either.\\n> \\n> Sure, it might be a bit confusing because lz4.h uses LZ4_ prefix, but\\n> then we probably should not define LZ4_compressor_init ...\\n\\nThis is a good point. The initial thought was that since lz4.h is now\\nremoved, such ambiguity will not be present. In v2 of the patch the\\nfunction is renamed to `LZ4State_compression_init` since this name\\ndescribes better its purpose. It initializes the LZ4State for\\ncompression.\\n\\nAs for the LZ4File_ prefix, I have no objections. Please find the\\nprefix changed to LZ4Stream_. For the record, the word 'File' is not\\nunique to the lz4 implementation. The common data structure used by\\nthe API in compress_io.h:\\n\\n typedef struct CompressFileHandle CompressFileHandle; \\n\\nThe public functions for this API are named:\\n\\n InitCompressFileHandle\\n InitDiscoverCompressFileHandle\\n EndCompressFileHandle\\n\\nAnd within InitCompressFileHandle the pattern is:\\n\\n if (compression_spec.algorithm == PG_COMPRESSION_NONE)\\n InitCompressFileHandleNone(CFH, compression_spec);\\n else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n InitCompressFileHandleGzip(CFH, compression_spec);\\n else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\\n InitCompressFileHandleLZ4(CFH, compression_spec);\\n\\nIt was felt that a prefix was required due to the inclusion 'lz4.h'\\nheader where naming functions as 'LZ4_' would be wrong. The prefix\\n'LZ4File_' seemed to be in line with the naming of the rest of\\nthe relevant functions and structures. Other compressions, gzip and\\nnone, did not face the same issue.\\n\\nTo conclude, I think that having a prefix is slightly preferred\\nover not having one. Since the prefix `LZ4File_` is not desired,\\nI propose `LZ4Stream_` in v2.\\n\\nI will not object to dismissing the argument and drop `File` from\\nthe prefix, if so requested.\\n\\n> \\n> Also, maybe the comments shouldn't use \\\"File API\\\" when compress_io.c\\n> calls that \\\"Compressed stream API\\\".\\n\\nDone.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Fri, 31 Mar 2023 09:19:47 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 3/31/23 11:19, gkokolatos@pm.me wrote:\\n> \\n>> ...\\n>>\\n>>\\n>> I think the patch is fine, but I'm wondering if the renames shouldn't go\\n>> a bit further. It removes references to LZ4File struct, but there's a\\n>> bunch of functions with LZ4File_ prefix. Why not to simply use LZ4_\\n>> prefix? We don't have GzipFile either.\\n>>\\n>> Sure, it might be a bit confusing because lz4.h uses LZ4_ prefix, but\\n>> then we probably should not define LZ4_compressor_init ...\\n> \\n> This is a good point. The initial thought was that since lz4.h is now\\n> removed, such ambiguity will not be present. In v2 of the patch the\\n> function is renamed to `LZ4State_compression_init` since this name\\n> describes better its purpose. It initializes the LZ4State for\\n> compression.\\n> \\n> As for the LZ4File_ prefix, I have no objections. Please find the\\n> prefix changed to LZ4Stream_. For the record, the word 'File' is not\\n> unique to the lz4 implementation. The common data structure used by\\n> the API in compress_io.h:\\n> \\n> typedef struct CompressFileHandle CompressFileHandle; \\n> \\n> The public functions for this API are named:\\n> \\n> InitCompressFileHandle\\n> InitDiscoverCompressFileHandle\\n> EndCompressFileHandle\\n> \\n> And within InitCompressFileHandle the pattern is:\\n> \\n> if (compression_spec.algorithm == PG_COMPRESSION_NONE)\\n> InitCompressFileHandleNone(CFH, compression_spec);\\n> else if (compression_spec.algorithm == PG_COMPRESSION_GZIP)\\n> InitCompressFileHandleGzip(CFH, compression_spec);\\n> else if (compression_spec.algorithm == PG_COMPRESSION_LZ4)\\n> InitCompressFileHandleLZ4(CFH, compression_spec);\\n> \\n> It was felt that a prefix was required due to the inclusion 'lz4.h'\\n> header where naming functions as 'LZ4_' would be wrong. The prefix\\n> 'LZ4File_' seemed to be in line with the naming of the rest of\\n> the relevant functions and structures. Other compressions, gzip and\\n> none, did not face the same issue.\\n> \\n> To conclude, I think that having a prefix is slightly preferred\\n> over not having one. Since the prefix `LZ4File_` is not desired,\\n> I propose `LZ4Stream_` in v2.\\n> \\n> I will not object to dismissing the argument and drop `File` from\\n> the prefix, if so requested.\\n> \\n\\nThanks.\\n\\nI think the LZ4Stream prefix is reasonable, so let's roll with that. I\\ncleaned up the patch a little bit (mostly comment tweaks, etc.), updated\\nthe commit message and pushed it.\\n\\nThe main tweak I did is renaming all the LZ4State variables from \\\"fs\\\" to\\nstate. The old name referred to the now abandoned \\\"file state\\\", but\\nafter the rename to LZ4State that seems confusing. Some of the places\\nalready used \\\"state\\\"and it's easier to know \\\"state\\\" is always LZ4State,\\nso let's keep it consistent.\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Sat, 1 Apr 2023 01:34:16 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 27, 2023 at 02:33:04PM +0000, gkokolatos@pm.me wrote:\\n> > > - Finally, the \\\"Nothing to do in the default case\\\" comment comes from\\n> > > Michael's commit 5e73a6048:\\n> > > \\n> > > + /*\\n> > > + * Custom and directory formats are compressed by default with gzip when\\n> > > + * available, not the others.\\n> > > + /\\n> > > + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\\n> > > + !user_compression_defined)\\n> > > {\\n> > > #ifdef HAVE_LIBZ\\n> > > - if (archiveFormat == archCustom || archiveFormat == archDirectory)\\n> > > - compressLevel = Z_DEFAULT_COMPRESSION;\\n> > > - else\\n> > > + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\\n> > > + &compression_spec);\\n> > > +#else\\n> > > + / Nothing to do in the default case */\\n> > > #endif\\n> > > - compressLevel = 0;\\n> > > }\\n> > > \\n> > > As the comment says: for -Fc and -Fd, the compression is set to zlib, if\\n> > > enabled, and when not otherwise specified by the user.\\n> > > \\n> > > Before 5e73a6048, this set compressLevel=0 for -Fp and -Ft, and when\\n> > > zlib was unavailable.\\n> > > \\n> > > But I'm not sure why there's now an empty \\\"#else\\\". I also don't know\\n> > > what \\\"the default case\\\" refers to.\\n> > > \\n> > > Maybe the best thing here is to move the preprocessor #if, since it's no\\n> > > longer in the middle of a runtime conditional:\\n> > > \\n> > > #ifdef HAVE_LIBZ\\n> > > + if ((archiveFormat == archCustom || archiveFormat == archDirectory) &&\\n> > > + !user_compression_defined)\\n> > > + parse_compress_specification(PG_COMPRESSION_GZIP, NULL,\\n> > > + &compression_spec);\\n> > > #endif\\n> > > \\n> > > ...but that elicits a warning about \\\"variable set but not used\\\"...\\n> > \\n> > \\n> > Not sure, I need to think about this a bit.\\n\\n> /* Nothing to do for the default case when LIBZ is not available */\\n> is easier to understand.\\n\\nMaybe I would write it as: \\\"if zlib is unavailable, default to no\\ncompression\\\". But I think that's best done in the leading comment, and\\nnot inside an empty preprocessor #else.\\n\\nI was hoping Michael would comment on this.\\nThe placement and phrasing of the comment makes no sense to me.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Tue, 11 Apr 2023 19:41:11 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 11, 2023 at 07:41:11PM -0500, Justin Pryzby wrote:\\n> Maybe I would write it as: \\\"if zlib is unavailable, default to no\\n> compression\\\". But I think that's best done in the leading comment, and\\n> not inside an empty preprocessor #else.\\n> \\n> I was hoping Michael would comment on this.\\n\\n(Sorry for the late reply, somewhat missed that.)\\n\\n> The placement and phrasing of the comment makes no sense to me.\\n\\nYes, this comment gives no value as it stands. I would be tempted to\\nfollow the suggestion to group the whole code block in a single ifdef,\\nincluding the check, and remove this comment. Like the attached\\nperhaps?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 12 Apr 2023 10:07:08 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 12, 2023 at 10:07:08AM +0900, Michael Paquier wrote:\\n> On Tue, Apr 11, 2023 at 07:41:11PM -0500, Justin Pryzby wrote:\\n> > Maybe I would write it as: \\\"if zlib is unavailable, default to no\\n> > compression\\\". But I think that's best done in the leading comment, and\\n> > not inside an empty preprocessor #else.\\n> > \\n> > I was hoping Michael would comment on this.\\n> \\n> (Sorry for the late reply, somewhat missed that.)\\n> \\n> > The placement and phrasing of the comment makes no sense to me.\\n> \\n> Yes, this comment gives no value as it stands. I would be tempted to\\n> follow the suggestion to group the whole code block in a single ifdef,\\n> including the check, and remove this comment. Like the attached\\n> perhaps?\\n\\n+1\\n\\n\\n\",\n \"msg_date\": \"Tue, 11 Apr 2023 20:19:59 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 11, 2023 at 08:19:59PM -0500, Justin Pryzby wrote:\\n> On Wed, Apr 12, 2023 at 10:07:08AM +0900, Michael Paquier wrote:\\n>> Yes, this comment gives no value as it stands. I would be tempted to\\n>> follow the suggestion to group the whole code block in a single ifdef,\\n>> including the check, and remove this comment. Like the attached\\n>> perhaps?\\n> \\n> +1\\n\\nLet me try this one again, as the previous patch would cause a warning\\nunder --without:-zlib as user_compression_defined would be unused. We\\ncould do something like the attached instead. It means doing twice\\nparse_compress_specification() for the non-zlib path, still we are\\nalready doing so for the zlib path.\\n\\nIf there are other ideas, feel free.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 13 Apr 2023 07:23:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 13, 2023 at 07:23:48AM +0900, Michael Paquier wrote:\\n> On Tue, Apr 11, 2023 at 08:19:59PM -0500, Justin Pryzby wrote:\\n> > On Wed, Apr 12, 2023 at 10:07:08AM +0900, Michael Paquier wrote:\\n> >> Yes, this comment gives no value as it stands. I would be tempted to\\n> >> follow the suggestion to group the whole code block in a single ifdef,\\n> >> including the check, and remove this comment. Like the attached\\n> >> perhaps?\\n> > \\n> > +1\\n> \\n> Let me try this one again, as the previous patch would cause a warning\\n> under --without:-zlib as user_compression_defined would be unused. We\\n> could do something like the attached instead. It means doing twice\\n> parse_compress_specification() for the non-zlib path, still we are\\n> already doing so for the zlib path.\\n> \\n> If there are other ideas, feel free.\\n\\nI don't think you need to call parse_compress_specification(NONE).\\nAs you wrote it, if zlib is unavailable, there's no parse(NONE) call,\\neven for directory and custom formats. And there's no parse(NONE) call\\nfor plan format when zlib is available.\\n\\nThe old way had preprocessor #if around both the \\\"if\\\" and \\\"else\\\" - is\\nthat what you meant ?\\n\\nIf you don't insist on calling parse(NONE), the only change is to remove\\nthe empty #else, which was my original patch.\\n\\n\\\"if no compression specification has been specified\\\" is redundant with\\n\\\"by default\\\", and causes \\\"not the others\\\" to dangle.\\n\\nIf I were to rewrite the comment, it'd say:\\n\\n+ * When gzip is available, custom and directory formats are compressed by\\n+ * default\\n\\n\\n\",\n \"msg_date\": \"Wed, 12 Apr 2023 17:52:40 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 12, 2023 at 05:52:40PM -0500, Justin Pryzby wrote:\\n> I don't think you need to call parse_compress_specification(NONE).\\n> As you wrote it, if zlib is unavailable, there's no parse(NONE) call,\\n> even for directory and custom formats. And there's no parse(NONE) call\\n> for plan format when zlib is available.\\n\\nYeah, that's not necessary, but I was wondering if it made the code a\\nbit cleaner, or else the non-zlib path would rely on the default\\ncompression method string.\\n\\n> The old way had preprocessor #if around both the \\\"if\\\" and \\\"else\\\" - is\\n> that what you meant?\\n> \\n> If you don't insist on calling parse(NONE), the only change is to remove\\n> the empty #else, which was my original patch.\\n\\nRemoving the empty else has as problem to create an empty if block,\\nwhich could be itself a cause of warnings?\\n\\n> If I were to rewrite the comment, it'd say:\\n> \\n> + * When gzip is available, custom and directory formats are compressed by\\n> + * default\\n\\nOkay.\\n--\\nMichael\",\n \"msg_date\": \"Thu, 13 Apr 2023 09:37:06 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 13, 2023 at 09:37:06AM +0900, Michael Paquier wrote:\\n> > If you don't insist on calling parse(NONE), the only change is to remove\\n> > the empty #else, which was my original patch.\\n> \\n> Removing the empty else has as problem to create an empty if block,\\n> which could be itself a cause of warnings?\\n\\nI doubt it - in the !HAVE_LIBZ case, it's currently an \\\"if\\\" statement\\nwith nothing but a comment, which isn't a problem.\\n\\nI think the only issue with an empty \\\"if\\\" is when you have no braces,\\nlike:\\n\\n\\tif (...)\\n#if ...\\n\\t\\tsomething;\\n#endif\\n\\n\\t// problem here //\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Wed, 12 Apr 2023 19:53:53 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 12, 2023 at 07:53:53PM -0500, Justin Pryzby wrote:\\n> I doubt it - in the !HAVE_LIBZ case, it's currently an \\\"if\\\" statement\\n> with nothing but a comment, which isn't a problem.\\n> \\n> I think the only issue with an empty \\\"if\\\" is when you have no braces,\\n> like:\\n> \\n> \\tif (...)\\n> #if ...\\n> \\t\\tsomething;\\n> #endif\\n> \\n> \\t// problem here //\\n\\n(My apologies for the late reply.)\\n\\nStill it could be easily messed up, and that's not a style that\\nreally exists in the tree, either, because there are always #else\\nblocks set up in such cases. Another part that makes me a bit\\nuncomfortable is that we would still call twice\\nparse_compress_specification(), something that should not happen but\\nwe are doing so on HEAD because the default compression_algorithm_str\\nis \\\"none\\\" and we want to enforce \\\"gzip\\\" for custom and directory\\nformats when building with zlib.\\n\\nWhat about just moving this block a bit up, just before the\\ncompression spec parsing, then? If we set compression_algorithm_str,\\nthe specification is compiled with the expected default, once instead\\nof twice.\\n--\\nMichael\",\n \"msg_date\": \"Tue, 25 Apr 2023 15:02:18 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\n\\n\\n\\n------- Original Message -------\\nOn Tuesday, April 25th, 2023 at 8:02 AM, Michael Paquier <michael@paquier.xyz> wrote:\\n\\n\\n> \\n> \\n> On Wed, Apr 12, 2023 at 07:53:53PM -0500, Justin Pryzby wrote:\\n> \\n> > I doubt it - in the !HAVE_LIBZ case, it's currently an \\\"if\\\" statement\\n> > with nothing but a comment, which isn't a problem.\\n> > \\n> > I think the only issue with an empty \\\"if\\\" is when you have no braces,\\n> > like:\\n> > \\n> > if (...)\\n> > #if ...\\n> > something;\\n> > #endif\\n> > \\n> > // problem here //\\n> \\n> \\n> (My apologies for the late reply.)\\n> \\n> Still it could be easily messed up, and that's not a style that\\n> really exists in the tree, either, because there are always #else\\n> blocks set up in such cases. Another part that makes me a bit\\n> uncomfortable is that we would still call twice\\n> parse_compress_specification(), something that should not happen but\\n> we are doing so on HEAD because the default compression_algorithm_str\\n> is \\\"none\\\" and we want to enforce \\\"gzip\\\" for custom and directory\\n> formats when building with zlib.\\n> \\n> What about just moving this block a bit up, just before the\\n> compression spec parsing, then? If we set compression_algorithm_str,\\n> the specification is compiled with the expected default, once instead\\n> of twice.\\n\\nFor what is worth, I think this would be the best approach. +1\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\\n\\n\\n\",\n \"msg_date\": \"Wed, 26 Apr 2023 08:50:46 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Wed, Apr 26, 2023 at 08:50:46AM +0000, gkokolatos@pm.me wrote:\\n> For what is worth, I think this would be the best approach. +1\\n\\nThanks. I have gone with that, then!\\n--\\nMichael\",\n \"msg_date\": \"Thu, 27 Apr 2023 13:38:48 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"23.03.2023 20:10, Tomas Vondra wrote:\\n> So pushed all three parts, after updating the commit messages a bit.\\n>\\n> This leaves the empty-data issue (which we have a fix for) and the\\n> switch to LZ4F. And then the zstd part.\\n>\\n\\nI'm sorry that I haven't noticed/checked that before, but when trying to\\nperform check-world with Valgrind I've discovered another issue presumably\\nrelated to LZ4File_gets().\\nWhen running under Valgrind:\\nPROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\\nI get:\\n...\\n[07:07:11.683](0.000s) ok 1939 - compression_lz4_dir: glob check for \\n.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\\n# Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql \\n.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\\n\\n==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\\n==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\\n==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\\n==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\\n==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\\n==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\\n==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\\n==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\\n==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\\n==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\\n==00:00:00:00.579 2811926==\\n...\\n\\nIt looks like the line variable returned by gets_func() here is not\\nnull-terminated:\\n while ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\\n {\\n...\\n if (sscanf(line, \\\"%u %\\\" CppAsString2(MAXPGPATH) \\\"s\\\\n\\\", &oid, lofname) != 2)\\n...\\nAnd Valgrind doesn't like it.\\n\\nBest regards,\\nAlexander\\n\\n\\n\",\n \"msg_date\": \"Fri, 5 May 2023 09:00:01 +0300\",\n \"msg_from\": \"Alexander Lakhin <exclusion@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Friday, May 5th, 2023 at 8:00 AM, Alexander Lakhin <exclusion@gmail.com> wrote:\\n\\n\\n> \\n> \\n> 23.03.2023 20:10, Tomas Vondra wrote:\\n> \\n> > So pushed all three parts, after updating the commit messages a bit.\\n> > \\n> > This leaves the empty-data issue (which we have a fix for) and the\\n> > switch to LZ4F. And then the zstd part.\\n> \\n> \\n> I'm sorry that I haven't noticed/checked that before, but when trying to\\n> perform check-world with Valgrind I've discovered another issue presumably\\n> related to LZ4File_gets().\\n> When running under Valgrind:\\n> PROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\\n> I get:\\n> ...\\n> 07:07:11.683 ok 1939 - compression_lz4_dir: glob check for\\n> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\\n> # Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql\\n> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\\n> \\n> ==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\\n> ==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\\n> ==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\\n> ==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\\n> ==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\\n> ==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\\n> ==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\\n> ==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\\n> ==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\\n> ==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\\n> ==00:00:00:00.579 2811926==\\n> ...\\n> \\n> It looks like the line variable returned by gets_func() here is not\\n> null-terminated:\\n> while ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\\n> \\n> {\\n> ...\\n> if (sscanf(line, \\\"%u %\\\" CppAsString2(MAXPGPATH) \\\"s\\\\n\\\", &oid, lofname) != 2)\\n> ...\\n> And Valgrind doesn't like it.\\n> \\n\\nValgrind is correct to not like it. LZ4Stream_gets() got modeled after\\ngets() when it should have been modeled after fgets().\\n\\nPlease find a patch attached to address it.\\n\\nCheers,\\n//Georgios\\n\\n> Best regards,\\n> Alexander\",\n \"msg_date\": \"Fri, 05 May 2023 10:02:12 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 2023-05-05 Fr 06:02, gkokolatos@pm.me wrote:\\n>\\n>\\n>\\n>\\n> ------- Original Message -------\\n> On Friday, May 5th, 2023 at 8:00 AM, Alexander Lakhin<exclusion@gmail.com> wrote:\\n>\\n>\\n>>\\n>> 23.03.2023 20:10, Tomas Vondra wrote:\\n>>\\n>>> So pushed all three parts, after updating the commit messages a bit.\\n>>>\\n>>> This leaves the empty-data issue (which we have a fix for) and the\\n>>> switch to LZ4F. And then the zstd part.\\n>>\\n>> I'm sorry that I haven't noticed/checked that before, but when trying to\\n>> perform check-world with Valgrind I've discovered another issue presumably\\n>> related to LZ4File_gets().\\n>> When running under Valgrind:\\n>> PROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\\n>> I get:\\n>> ...\\n>> 07:07:11.683 ok 1939 - compression_lz4_dir: glob check for\\n>> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\\n>> # Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql\\n>> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\\n>>\\n>> ==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\\n>> ==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\\n>> ==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\\n>> ==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\\n>> ==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\\n>> ==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\\n>> ==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\\n>> ==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\\n>> ==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\\n>> ==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\\n>> ==00:00:00:00.579 2811926==\\n>> ...\\n>>\\n>> It looks like the line variable returned by gets_func() here is not\\n>> null-terminated:\\n>> while ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\\n>>\\n>> {\\n>> ...\\n>> if (sscanf(line, \\\"%u %\\\" CppAsString2(MAXPGPATH) \\\"s\\\\n\\\", &oid, lofname) != 2)\\n>> ...\\n>> And Valgrind doesn't like it.\\n>>\\n> Valgrind is correct to not like it. LZ4Stream_gets() got modeled after\\n> gets() when it should have been modeled after fgets().\\n>\\n> Please find a patch attached to address it.\\n>\\n>\\n\\nIsn't using memset here a bit wasteful? Why not just put a null at the \\nend after calling LZ4Stream_read_internal(), which tells you how many \\nbytes it has written?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB:https://www.enterprisedb.com\\n\\n\\n\\n\\n\\n\\n\\n\\nOn 2023-05-05 Fr 06:02,\\n gkokolatos@pm.me wrote:\\n\\n\\n\\n\\n\\n\\n\\n------- Original Message -------\\nOn Friday, May 5th, 2023 at 8:00 AM, Alexander Lakhin <exclusion@gmail.com> wrote:\\n\\n\\n\\n\\n\\n\\n23.03.2023 20:10, Tomas Vondra wrote:\\n\\n\\n\\nSo pushed all three parts, after updating the commit messages a bit.\\n\\nThis leaves the empty-data issue (which we have a fix for) and the\\nswitch to LZ4F. And then the zstd part.\\n\\n\\n\\n\\nI'm sorry that I haven't noticed/checked that before, but when trying to\\nperform check-world with Valgrind I've discovered another issue presumably\\nrelated to LZ4File_gets().\\nWhen running under Valgrind:\\nPROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\\nI get:\\n...\\n07:07:11.683 ok 1939 - compression_lz4_dir: glob check for\\n.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\\n# Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql\\n.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\\n\\n==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\\n==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\\n==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\\n==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\\n==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\\n==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\\n==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\\n==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\\n==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\\n==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\\n==00:00:00:00.579 2811926==\\n...\\n\\nIt looks like the line variable returned by gets_func() here is not\\nnull-terminated:\\nwhile ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\\n\\n{\\n...\\nif (sscanf(line, \\\"%u %\\\" CppAsString2(MAXPGPATH) \\\"s\\\\n\\\", &oid, lofname) != 2)\\n...\\nAnd Valgrind doesn't like it.\\n\\n\\n\\n\\nValgrind is correct to not like it. LZ4Stream_gets() got modeled after\\ngets() when it should have been modeled after fgets().\\n\\nPlease find a patch attached to address it.\\n\\n\\n\\n\\n\\n\\nIsn't using memset here a bit wasteful? Why not just put a null\\n at the end after calling LZ4Stream_read_internal(), which tells\\n you how many bytes it has written?\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\",\n \"msg_date\": \"Fri, 5 May 2023 09:23:37 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Friday, May 5th, 2023 at 3:23 PM, Andrew Dunstan <andrew@dunslane.net> wrote:\\n\\n> On 2023-05-05 Fr 06:02, gkokolatos@pm.me wrote:\\n>\\n>> ------- Original Message -------\\n>> On Friday, May 5th, 2023 at 8:00 AM, Alexander Lakhin\\n>> [<exclusion@gmail.com>](mailto:exclusion@gmail.com)\\n>> wrote:\\n>>\\n>>> 23.03.2023 20:10, Tomas Vondra wrote:\\n>>>\\n>>>> So pushed all three parts, after updating the commit messages a bit.\\n>>>>\\n>>>> This leaves the empty-data issue (which we have a fix for) and the\\n>>>> switch to LZ4F. And then the zstd part.\\n>>>\\n>>> I'm sorry that I haven't noticed/checked that before, but when trying to\\n>>> perform check-world with Valgrind I've discovered another issue presumably\\n>>> related to LZ4File_gets().\\n>>> When running under Valgrind:\\n>>> PROVE_TESTS=t/002_pg_dump.pl make check -C src/bin/pg_dump/\\n>>> I get:\\n>>> ...\\n>>> 07:07:11.683 ok 1939 - compression_lz4_dir: glob check for\\n>>> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir/*.dat.lz4\\n>>> # Running: pg_restore --jobs=2 --file=.../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir.sql\\n>>> .../src/bin/pg_dump/tmp_check/tmp_test_HB6A/compression_lz4_dir\\n>>>\\n>>> ==00:00:00:00.579 2811926== Conditional jump or move depends on uninitialised value(s)\\n>>> ==00:00:00:00.579 2811926== at 0x4853376: rawmemchr (vg_replace_strmem.c:1548)\\n>>> ==00:00:00:00.579 2811926== by 0x4C96A67: _IO_str_init_static_internal (strops.c:41)\\n>>> ==00:00:00:00.579 2811926== by 0x4C693A2: _IO_strfile_read (strfile.h:95)\\n>>> ==00:00:00:00.579 2811926== by 0x4C693A2: __isoc99_sscanf (isoc99_sscanf.c:28)\\n>>> ==00:00:00:00.579 2811926== by 0x11DB6F: _LoadLOs (pg_backup_directory.c:458)\\n>>> ==00:00:00:00.579 2811926== by 0x11DD1E: _PrintTocData (pg_backup_directory.c:422)\\n>>> ==00:00:00:00.579 2811926== by 0x118484: restore_toc_entry (pg_backup_archiver.c:882)\\n>>> ==00:00:00:00.579 2811926== by 0x1190CC: RestoreArchive (pg_backup_archiver.c:699)\\n>>> ==00:00:00:00.579 2811926== by 0x10F25D: main (pg_restore.c:414)\\n>>> ==00:00:00:00.579 2811926==\\n>>> ...\\n>>>\\n>>> It looks like the line variable returned by gets_func() here is not\\n>>> null-terminated:\\n>>> while ((CFH->gets_func(line, MAXPGPATH, CFH)) != NULL)\\n>>>\\n>>> {\\n>>> ...\\n>>> if (sscanf(line, \\\"%u %\\\" CppAsString2(MAXPGPATH) \\\"s\\\\n\\\", &oid, lofname) != 2)\\n>>> ...\\n>>> And Valgrind doesn't like it.\\n>>\\n>> Valgrind is correct to not like it. LZ4Stream_gets() got modeled after\\n>> gets() when it should have been modeled after fgets().\\n>>\\n>> Please find a patch attached to address it.\\n>\\n> Isn't using memset here a bit wasteful? Why not just put a null at the end after calling LZ4Stream_read_internal(), which tells you how many bytes it has written?\\n\\nGood point. I thought about it before submitting the patch. I concluded that given the complexity and operations involved in LZ4Stream_read_internal() and the rest of the pg_dump/pg_restore code, the memset() call will be negligible. However from the readability point of view, the function is a bit cleaner with the memset().\\n\\nI will not object to any suggestion though, as this is a very trivial point. Please find attached a v2 of the patch following the suggested approach.\\n\\nCheers,\\n\\n//Georgios\\n\\n> cheers\\n>\\n> andrew\\n>\\n> --\\n> Andrew Dunstan\\n> EDB:\\n> https://www.enterprisedb.com\",\n \"msg_date\": \"Fri, 05 May 2023 14:13:28 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:\\n> Good point. I thought about it before submitting the patch. I\\n> concluded that given the complexity and operations involved in\\n> LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore\\n> code, the memset() call will be negligible. However from the\\n> readability point of view, the function is a bit cleaner with the\\n> memset(). \\n> \\n> I will not object to any suggestion though, as this is a very\\n> trivial point. Please find attached a v2 of the patch following the\\n> suggested approach.\\n\\nPlease note that an open item has been added for this stuff.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 6 May 2023 11:51:19 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sat, May 6, 2023 at 04:51, Michael Paquier <[michael@paquier.xyz](mailto:On Sat, May 6, 2023 at 04:51, Michael Paquier <<a href=)> wrote:\\n\\n> On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:\\n>> Good point. I thought about it before submitting the patch. I\\n>> concluded that given the complexity and operations involved in\\n>> LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore\\n>> code, the memset() call will be negligible. However from the\\n>> readability point of view, the function is a bit cleaner with the\\n>> memset().\\n>>\\n>> I will not object to any suggestion though, as this is a very\\n>> trivial point. Please find attached a v2 of the patch following the\\n>> suggested approach.\\n>\\n> Please note that an open item has been added for this stuff.\\n\\nThank you but I am not certain I know what that means. Can you please explain?\\n\\nCheers,\\n//Georgios\\n\\n> --\\n> Michael\\n On Sat, May 6, 2023 at 04:51, Michael Paquier <michael@paquier.xyz> wrote: On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:> Good point. I thought about it before submitting the patch. I> concluded that given the complexity and operations involved in> LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore> code, the memset() call will be negligible. However from the> readability point of view, the function is a bit cleaner with the> memset(). > > I will not object to any suggestion though, as this is a very> trivial point. Please find attached a v2 of the patch following the> suggested approach.Please note that an open item has been added for this stuff.Thank you but I am not certain I know what that means. Can you please explain?Cheers,//Georgios--Michael\",\n \"msg_date\": \"Sun, 07 May 2023 15:01:52 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sun, May 07, 2023 at 03:01:52PM +0000, gkokolatos@pm.me wrote:\\n> Thank you but I am not certain I know what that means. Can you please explain?\\n\\nIt means that this thread has been added to the following list:\\nhttps://wiki.postgresql.org/wiki/PostgreSQL_16_Open_Items#Open_Issues\\n\\npg_dump/compress_lz4.c is new as of PostgreSQL 16, and this patch is\\nfixing a deficiency. That's just a way outside of the commit fest to\\ntrack any problems and make sure these are fixed before the release\\nhappens.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 8 May 2023 07:26:23 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:\\n> Good point. I thought about it before submitting the patch. I\\n> concluded that given the complexity and operations involved in\\n> LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore\\n> code, the memset() call will be negligible. However from the\\n> readability point of view, the function is a bit cleaner with the\\n> memset(). \\n> \\n> I will not object to any suggestion though, as this is a very\\n> trivial point. Please find attached a v2 of the patch following the\\n> suggested approach.\\n\\nHmm. I was looking at this patch, and what you are trying to do\\nsounds rather right to keep a parallel with the gzip and zstd code\\npaths.\\n\\nLooking at the code of gzread.c, gzgets() enforces a null-termination\\non the string read. Still, isn't that something we'd better enforce\\nin read_none() as well? compress_io.h lists this as a requirement of\\nthe callback, and Zstd_gets() does so already. read_none() does not\\nenforce that, unfortunately.\\n\\n+ /* No work needs to be done for a zero-sized output buffer */\\n+ if (size <= 0)\\n+ return 0;\\n\\nIndeed. This should be OK.\\n\\n- ret = LZ4Stream_read_internal(state, ptr, size, true);\\n+ Assert(size > 1);\\n\\nThe addition of this assertion is a bit surprising, and this is\\ninconsistent with Zstd_gets where a length of 1 is authorized. We\\nshould be more consistent across all the callbacks, IMO, not less, so\\nas we apply the same API contract across all the compression methods.\\n\\nWhile testing this patch, I have triggered an error pointing out that\\nthe decompression path of LZ4 is broken for table data. I can\\nreproduce that with a dump of the regression database, as of:\\nmake installcheck\\npg_dump --format=d --file=dump_lz4 --compress=lz4 regression\\ncreatedb regress_lz4\\npg_restore --format=d -d regress_lz4 dump_lz4\\npg_restore: error: COPY failed for table \\\"clstr_tst\\\": ERROR: extra data after last expected column\\nCONTEXT: COPY clstr_tst, line 15: \\\"32\\t6\\tseis\\txyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzy...\\\"\\npg_restore: warning: errors ignored on restore: 1\\n\\nThis does not show up with gzip or zstd, and the patch does not\\ninfluence the result. In short it shows up with and without the\\npatch, on HEAD. That does not look really stable :/\\n--\\nMichael\",\n \"msg_date\": \"Mon, 8 May 2023 09:31:06 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"Michael Paquier <michael@paquier.xyz> writes:\\n> While testing this patch, I have triggered an error pointing out that\\n> the decompression path of LZ4 is broken for table data. I can\\n> reproduce that with a dump of the regression database, as of:\\n> make installcheck\\n> pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\\n> createdb regress_lz4\\n> pg_restore --format=d -d regress_lz4 dump_lz4\\n> pg_restore: error: COPY failed for table \\\"clstr_tst\\\": ERROR: extra data after last expected column\\n> CONTEXT: COPY clstr_tst, line 15: \\\"32\\t6\\tseis\\txyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzyxyzzy...\\\"\\n> pg_restore: warning: errors ignored on restore: 1\\n\\nUgh. Reproduced here ... so we need an open item for this.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 07 May 2023 21:09:25 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Sun, May 07, 2023 at 09:09:25PM -0400, Tom Lane wrote:\\n> Ugh. Reproduced here ... so we need an open item for this.\\n\\nYep. Already added.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 8 May 2023 10:13:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> Michael Paquier <michael@paquier.xyz> writes:\\n>> While testing this patch, I have triggered an error pointing out that\\n>> the decompression path of LZ4 is broken for table data. I can\\n>> reproduce that with a dump of the regression database, as of:\\n>> make installcheck\\n>> pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\\n\\n> Ugh. Reproduced here ... so we need an open item for this.\\n\\nBTW, it seems to work with --format=c.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 07 May 2023 21:16:51 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 5/7/23 17:01, gkokolatos@pm.me wrote:\\n> \\n> \\n> \\n> On Sat, May 6, 2023 at 04:51, Michael Paquier <michael@paquier.xyz\\n> <mailto:On Sat, May 6, 2023 at 04:51, Michael Paquier <<a href=>> wrote:\\n>> On Fri, May 05, 2023 at 02:13:28PM +0000, gkokolatos@pm.me wrote:\\n>> > Good point. I thought about it before submitting the patch. I\\n>> > concluded that given the complexity and operations involved in\\n>> > LZ4Stream_read_internal() and the rest of t he pg_dump/pg_restore\\n>> > code, the memset() call will be negligible. However from the\\n>> > readability point of view, the function is a bit cleaner with the\\n>> > memset().\\n>> >\\n>> > I will not object to any suggestion though, as this is a very\\n>> > trivial point. Please find attached a v2 of the patch following the\\n>> > suggested approach.\\n>>\\n>> Please note that an open item has been added for this stuff.\\n> Thank you but I am not certain I know what that means. Can you please\\n> explain?\\n> \\n\\nIt means it was added to the list of items we need to fix before PG16\\ngets out:\\n\\nhttps://wiki.postgresql.org/wiki/PostgreSQL_16_Open_Items\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 May 2023 03:34:15 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Monday, May 8th, 2023 at 3:16 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n\\n\\n> \\n> \\n> I wrote:\\n> \\n> > Michael Paquier michael@paquier.xyz writes:\\n> > \\n> > > While testing this patch, I have triggered an error pointing out that\\n> > > the decompression path of LZ4 is broken for table data. I can\\n> > > reproduce that with a dump of the regression database, as of:\\n> > > make installcheck\\n> > > pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\\n> \\n> > Ugh. Reproduced here ... so we need an open item for this.\\n> \\n> \\n> BTW, it seems to work with --format=c.\\n> \\n\\nThank you for the extra tests. It seems that exists a gap in the test\\ncoverage. Please find a patch attached that is addressing the issue\\nand attempt to provide tests for it.\\n\\nCheers,\\n//Georgios\\n\\n> regards, tom lane\",\n \"msg_date\": \"Mon, 08 May 2023 16:19:43 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 5/8/23 03:16, Tom Lane wrote:\\n> I wrote:\\n>> Michael Paquier <michael@paquier.xyz> writes:\\n>>> While testing this patch, I have triggered an error pointing out that\\n>>> the decompression path of LZ4 is broken for table data. I can\\n>>> reproduce that with a dump of the regression database, as of:\\n>>> make installcheck\\n>>> pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\\n> \\n>> Ugh. Reproduced here ... so we need an open item for this.\\n> \\n> BTW, it seems to work with --format=c.\\n> \\n\\nThe LZ4Stream_write() forgot to move the pointer to the next chunk, so\\nit was happily decompressing the initial chunk over and over. A bit\\nembarrassing oversight :-(\\n\\nThe custom format calls WriteDataToArchiveLZ4(), which was correct.\\n\\nThe attached patch fixes this for me.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 8 May 2023 20:00:39 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 5/8/23 18:19, gkokolatos@pm.me wrote:\\n> \\n> \\n> \\n> \\n> \\n> ------- Original Message -------\\n> On Monday, May 8th, 2023 at 3:16 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> \\n> \\n>>\\n>>\\n>> I wrote:\\n>>\\n>>> Michael Paquier michael@paquier.xyz writes:\\n>>>\\n>>>> While testing this patch, I have triggered an error pointing out that\\n>>>> the decompression path of LZ4 is broken for table data. I can\\n>>>> reproduce that with a dump of the regression database, as of:\\n>>>> make installcheck\\n>>>> pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\\n>>\\n>>> Ugh. Reproduced here ... so we need an open item for this.\\n>>\\n>>\\n>> BTW, it seems to work with --format=c.\\n>>\\n> \\n> Thank you for the extra tests. It seems that exists a gap in the test\\n> coverage. Please find a patch attached that is addressing the issue\\n> and attempt to provide tests for it.\\n> \\n\\nSeems I'm getting messages with a delay - this is mostly the same fix I\\nended up with, not realizing you already posted a fix.\\n\\nI don't think we need the local \\\"in\\\" variable - the pointer parameter is\\nlocal in the function, so we can modify it directly (with a cast).\\nWriteDataToArchiveLZ4 does it that way too.\\n\\nThe tests are definitely a good idea. I wonder if we should add a\\ncomment to DEFAULT_IO_BUFFER_SIZE mentioning that if we choose to\\nincrease the value in the future, we needs to tweak the tests too to use\\nmore data in order to exercise the buffering etc. Maybe it's obvious?\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Mon, 8 May 2023 20:20:42 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Monday, May 8th, 2023 at 8:20 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> \\n> \\n> On 5/8/23 18:19, gkokolatos@pm.me wrote:\\n> \\n> > ------- Original Message -------\\n> > On Monday, May 8th, 2023 at 3:16 AM, Tom Lane tgl@sss.pgh.pa.us wrote:\\n> > \\n> > > I wrote:\\n> > > \\n> > > > Michael Paquier michael@paquier.xyz writes:\\n> > > > \\n> > > > > While testing this patch, I have triggered an error pointing out that\\n> > > > > the decompression path of LZ4 is broken for table data. I can\\n> > > > > reproduce that with a dump of the regression database, as of:\\n> > > > > make installcheck\\n> > > > > pg_dump --format=d --file=dump_lz4 --compress=lz4 regression\\n> > > \\n> > > > Ugh. Reproduced here ... so we need an open item for this.\\n> > > \\n> > > BTW, it seems to work with --format=c.\\n> > \\n> > Thank you for the extra tests. It seems that exists a gap in the test\\n> > coverage. Please find a patch attached that is addressing the issue\\n> > and attempt to provide tests for it.\\n> \\n> \\n> Seems I'm getting messages with a delay - this is mostly the same fix I\\n> ended up with, not realizing you already posted a fix.\\n\\nThank you very much for looking.\\n\\n> I don't think we need the local \\\"in\\\" variable - the pointer parameter is\\n> local in the function, so we can modify it directly (with a cast).\\n> WriteDataToArchiveLZ4 does it that way too.\\n\\nSure, patch updated.\\n \\n> The tests are definitely a good idea.\\n\\nThank you.\\n\\n> I wonder if we should add a\\n> comment to DEFAULT_IO_BUFFER_SIZE mentioning that if we choose to\\n> increase the value in the future, we needs to tweak the tests too to use\\n> more data in order to exercise the buffering etc. Maybe it's obvious?\\n> \\n\\nYou are right. Added a comment both in the header and in the test.\\n\\nI hope v2 gets closer to closing the open item for this.\\n\\nCheers,\\n//Georgios\\n\\n\\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Mon, 08 May 2023 19:53:28 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Mon, May 08, 2023 at 08:00:39PM +0200, Tomas Vondra wrote:\\n> The LZ4Stream_write() forgot to move the pointer to the next chunk, so\\n> it was happily decompressing the initial chunk over and over. A bit\\n> embarrassing oversight :-(\\n> \\n> The custom format calls WriteDataToArchiveLZ4(), which was correct.\\n> \\n> The attached patch fixes this for me.\\n\\nOuch. So this was corrupting the dumps and the compression when\\ntrying to write more than two chunks at once, not the decompression\\nsteps. That addresses the issue here as well, thanks!\\n--\\nMichael\",\n \"msg_date\": \"Tue, 9 May 2023 07:10:55 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 5/9/23 00:10, Michael Paquier wrote:\\n> On Mon, May 08, 2023 at 08:00:39PM +0200, Tomas Vondra wrote:\\n>> The LZ4Stream_write() forgot to move the pointer to the next chunk, so\\n>> it was happily decompressing the initial chunk over and over. A bit\\n>> embarrassing oversight :-(\\n>>\\n>> The custom format calls WriteDataToArchiveLZ4(), which was correct.\\n>>\\n>> The attached patch fixes this for me.\\n> \\n> Ouch. So this was corrupting the dumps and the compression when\\n> trying to write more than two chunks at once, not the decompression\\n> steps. That addresses the issue here as well, thanks!\\n\\nYeah. Thanks for the report, should have been found during review.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Tue, 9 May 2023 14:54:31 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"------- Original Message -------\\nOn Tuesday, May 9th, 2023 at 2:54 PM, Tomas Vondra <tomas.vondra@enterprisedb.com> wrote:\\n\\n\\n> \\n> \\n> On 5/9/23 00:10, Michael Paquier wrote:\\n> \\n> > On Mon, May 08, 2023 at 08:00:39PM +0200, Tomas Vondra wrote:\\n> > \\n> > > The LZ4Stream_write() forgot to move the pointer to the next chunk, so\\n> > > it was happily decompressing the initial chunk over and over. A bit\\n> > > embarrassing oversight :-(\\n> > > \\n> > > The custom format calls WriteDataToArchiveLZ4(), which was correct.\\n> > > \\n> > > The attached patch fixes this for me.\\n> > \\n> > Ouch. So this was corrupting the dumps and the compression when\\n> > trying to write more than two chunks at once, not the decompression\\n> > steps. That addresses the issue here as well, thanks!\\n> \\n> \\n> Yeah. Thanks for the report, should have been found during review.\\n\\nThank you both for looking. A small consolation is that now there are\\ntests for this case.\\n\\nMoving on to the other open item for this, please find attached v2\\nof the patch as requested.\\n\\nCheers,\\n//Georgios\\n\\n> \\n> \\n> regards\\n> \\n> --\\n> Tomas Vondra\\n> EnterpriseDB: http://www.enterprisedb.com\\n> The Enterprise PostgreSQL Company\",\n \"msg_date\": \"Tue, 09 May 2023 14:12:44 +0000\",\n \"msg_from\": \"gkokolatos@pm.me\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, May 09, 2023 at 02:12:44PM +0000, gkokolatos@pm.me wrote:\\n> Thank you both for looking. A small consolation is that now there are\\n> tests for this case.\\n\\n+1, noticing that was pure luck ;)\\n\\nWorth noting that the patch posted in [1] has these tests, not the\\nversion posted in [2].\\n\\n+ create_sql => 'INSERT INTO dump_test.test_compression_method (col1) '\\n+ . 'SELECT string_agg(a::text, \\\\'\\\\') FROM generate_series(1,4096) a;',\\n\\nYep, good and cheap idea to check for longer chunks. That should be\\nenough to loop twice.\\n\\n[1]: https://www.postgresql.org/message-id/SYTRcNgtAbzyn3y3IInh1x-UfNTKMNpnFvI3mr6SyqyVf3PkaDsMy_cpKKgsl3_HdLy2MFAH4zwjxDmFfiLO8rWtSiJWBtqT06OMjeNo4GA=@pm.me\\n[2]: https://www.postgresql.org/message-id/f735df01-0bb4-2fbc-1297-73a520cfc534@enterprisedb.com\\n\\n> Moving on to the other open item for this, please find attached v2\\n> of the patch as requested.\\n\\nDid you notice the comments of [3] about the second patch that aims to\\nadd the null termination in the line from the LZ4 fgets() callback?\\n\\n[3]: https://www.postgresql.org/message-id/ZFhCyn4Gm2eu60rB@paquier.xyz\\n--\\nMichael\",\n \"msg_date\": \"Wed, 10 May 2023 11:22:10 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On Tue, May 09, 2023 at 02:54:31PM +0200, Tomas Vondra wrote:\\n> Yeah. Thanks for the report, should have been found during review.\\n\\nTomas, are you planning to do something by the end of this week for\\nbeta1? Or do you need some help of any kind?\\n--\\nMichael\",\n \"msg_date\": \"Wed, 17 May 2023 15:18:58 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"On 5/17/23 08:18, Michael Paquier wrote:\\n> On Tue, May 09, 2023 at 02:54:31PM +0200, Tomas Vondra wrote:\\n>> Yeah. Thanks for the report, should have been found during review.\\n> \\n> Tomas, are you planning to do something by the end of this week for\\n> beta1? Or do you need some help of any kind?\\n\\nI'll take care of it.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 May 2023 10:59:02 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n },\n {\n \"msg_contents\": \"\\n\\nOn 5/17/23 10:59, Tomas Vondra wrote:\\n> On 5/17/23 08:18, Michael Paquier wrote:\\n>> On Tue, May 09, 2023 at 02:54:31PM +0200, Tomas Vondra wrote:\\n>>> Yeah. Thanks for the report, should have been found during review.\\n>>\\n>> Tomas, are you planning to do something by the end of this week for\\n>> beta1? Or do you need some help of any kind?\\n> \\n> I'll take care of it.\\n> \\n\\nFWIW I've pushed fixes for both open issues associated with the pg_dump\\ncompression. I'll keep an eye on the buildfarm, but hopefully that'll do\\nit for beta1.\\n\\n\\nregards\\n\\n-- \\nTomas Vondra\\nEnterpriseDB: http://www.enterprisedb.com\\nThe Enterprise PostgreSQL Company\\n\\n\\n\",\n \"msg_date\": \"Wed, 17 May 2023 17:02:22 +0200\",\n \"msg_from\": \"Tomas Vondra <tomas.vondra@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add LZ4 compression in pg_dump\"\n }\n]","isConversational":false}}},{"rowIdx":53581,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Here's a patch to add the sum of timings for JIT counters to\npg_stat_statements, as a way to follow-up on if JIT is doing a good or\na bad job in a configuration.\n\nI decided to only store the total time for the timings, since there\nare 4 different timings and storing max/min/etc for each one of them\nwould lead to a bit too much data. This can of course be reconsidered,\nbut I think that's a reasonable tradeoff.\n\nAnother option I guess could be to store the max/min/etc, but only\nstore for the total jit time instead of for each individual one. Maybe\nthat'd actually be more useful?\n\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/","msg_date":"Fri, 25 Feb 2022 14:06:29 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"Hi,\n\nOn Fri, Feb 25, 2022 at 02:06:29PM +0100, Magnus Hagander wrote:\n> Here's a patch to add the sum of timings for JIT counters to\n> pg_stat_statements, as a way to follow-up on if JIT is doing a good or\n> a bad job in a configuration.\n\n+1, it seems like something quite useful.\n\n> I decided to only store the total time for the timings, since there\n> are 4 different timings and storing max/min/etc for each one of them\n> would lead to a bit too much data. This can of course be reconsidered,\n> but I think that's a reasonable tradeoff.\n\nI think the cumulated total time is enough. Looking at the patch, I think we\nshould also cumulate the number of time jit was triggered, and\nprobably the same for each other main operation (optimization and inlining).\nOtherwise the values may be wrong and look artificially low.\n\n\n","msg_date":"Fri, 25 Feb 2022 21:33:22 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Fri, Feb 25, 2022 at 2:33 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n>\n> Hi,\n>\n> On Fri, Feb 25, 2022 at 02:06:29PM +0100, Magnus Hagander wrote:\n> > Here's a patch to add the sum of timings for JIT counters to\n> > pg_stat_statements, as a way to follow-up on if JIT is doing a good or\n> > a bad job in a configuration.\n>\n> +1, it seems like something quite useful.\n\nGiven the amount of time often spent debugging JIT -- getting more\ninsight is going to make it easier to tune it instead of like what\nunfortunately many people do and just turn it off..\n\n\n> > I decided to only store the total time for the timings, since there\n> > are 4 different timings and storing max/min/etc for each one of them\n> > would lead to a bit too much data. This can of course be reconsidered,\n> > but I think that's a reasonable tradeoff.\n>\n> I think the cumulated total time is enough. Looking at the patch, I think we\n> should also cumulate the number of time jit was triggered, and\n> probably the same for each other main operation (optimization and inlining).\n> Otherwise the values may be wrong and look artificially low.\n\nSo just to be clear, you're basically thinking:\n\njit_count = count of entries where jit_functions>0\njit_functions = <same as now>\njit_optimizatinos = count of entries where time spent on jit_optimizations > 0\n\netc?\n\nSo we count the times with min/max like other times for the total one,\nbut instead add a counter for each of the details?\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/\n\n\n","msg_date":"Fri, 25 Feb 2022 16:19:27 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\n> \n> So just to be clear, you're basically thinking:\n> \n> jit_count = count of entries where jit_functions>0\n> jit_functions = <same as now>\n> jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\n> \n> etc?\n\nYes exactly, so 3 new fields on top of the one you already added.\n\n> So we count the times with min/max like other times for the total one,\n> but instead add a counter for each of the details?\n\nI don't understand this one. Did you mean we *don't* count times with min/max?\nIf that's the case then yes :)\n\n\n","msg_date":"Fri, 25 Feb 2022 23:40:01 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Fri, Feb 25, 2022 at 4:40 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n>\n> On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\n> >\n> > So just to be clear, you're basically thinking:\n> >\n> > jit_count = count of entries where jit_functions>0\n> > jit_functions = <same as now>\n> > jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\n> >\n> > etc?\n>\n> Yes exactly, so 3 new fields on top of the one you already added.\n>\n> > So we count the times with min/max like other times for the total one,\n> > but instead add a counter for each of the details?\n>\n> I don't understand this one. Did you mean we *don't* count times with min/max?\n> If that's the case then yes :)\n\nHmm. Yeah, that was a bit unclear. I mean we track total time with\nmin/max, and detailed time not at all. For details, we only track\ncount, not time.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/\n\n\n","msg_date":"Fri, 25 Feb 2022 16:41:22 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Fri, Feb 25, 2022 at 4:41 PM Magnus Hagander <magnus@hagander.net> wrote:\n>\n> On Fri, Feb 25, 2022 at 4:40 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> >\n> > On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\n> > >\n> > > So just to be clear, you're basically thinking:\n> > >\n> > > jit_count = count of entries where jit_functions>0\n> > > jit_functions = <same as now>\n> > > jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\n> > >\n> > > etc?\n> >\n> > Yes exactly, so 3 new fields on top of the one you already added.\n> >\n> > > So we count the times with min/max like other times for the total one,\n> > > but instead add a counter for each of the details?\n> >\n> > I don't understand this one. Did you mean we *don't* count times with min/max?\n> > If that's the case then yes :)\n>\n> Hmm. Yeah, that was a bit unclear. I mean we track total time with\n> min/max, and detailed time not at all. For details, we only track\n> count, not time.\n\nPer some off-list discussion with Julien, we have clearly been talking\nin slightly different terms. So let's summarize the options into what\ntheÿ́d actually be:\n\nOption 0: what is int he patch now\n\nOption 1:\njit_count - number of executions using jit\ntotal_jit_time - for sum of functions+inlining+optimization+emission time\nmin_jit_time - for sum of functions+inlining+optimization+emission time\nmax_jit_time - for sum of functions+inlining+optimization+emission time\nmean_jit_time - for sum of functions+inlining+optimization+emission time\nstddev_jit_time - for sum of functions+inlining+optimization+emission time\njit_functions - number of functions\njit_inlining_count - number of executions where inlining happened\njit_optimization_count - number of executions where optimization happened\njit_emission_count - number of executions where emission happened\n\nOption 2:\njit_count\njit_functions\njit_generation_time\njit_inlining_count\njit_inlining_time\njit_optimization_count\njit_optimization_time\njit_emission_count\njit_emission_time\n\n\n(We can bikeshed on the exact names of the fields once we have decided\nwhich option is preferred)\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/\n\n\n","msg_date":"Fri, 25 Feb 2022 17:38:45 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"> On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\n> On Fri, Feb 25, 2022 at 2:33 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> >\n> > Hi,\n> >\n> > On Fri, Feb 25, 2022 at 02:06:29PM +0100, Magnus Hagander wrote:\n> > > Here's a patch to add the sum of timings for JIT counters to\n> > > pg_stat_statements, as a way to follow-up on if JIT is doing a good or\n> > > a bad job in a configuration.\n> >\n> > +1, it seems like something quite useful.\n>\n> Given the amount of time often spent debugging JIT -- getting more\n> insight is going to make it easier to tune it instead of like what\n> unfortunately many people do and just turn it off..\n\nIndeed, sounds convenient, although I wonder how exactly one would use it\nto tune JIT? I'm curious, because I got used to situations when one\nsingle long query takes much longer than expected due to JIT issues --\nbut it seems the target of this patch are situations when there are a\nlot of long queries using JIT, and it's easier to analyze them via pgss?\n\n> > > I decided to only store the total time for the timings, since there\n> > > are 4 different timings and storing max/min/etc for each one of them\n> > > would lead to a bit too much data. This can of course be reconsidered,\n> > > but I think that's a reasonable tradeoff.\n> >\n> > I think the cumulated total time is enough. Looking at the patch, I think we\n> > should also cumulate the number of time jit was triggered, and\n> > probably the same for each other main operation (optimization and inlining).\n> > Otherwise the values may be wrong and look artificially low.\n>\n> So just to be clear, you're basically thinking:\n>\n> jit_count = count of entries where jit_functions>0\n> jit_functions = <same as now>\n> jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\n\nOne interesting not-very-relevant for the patch thing I've noticed while\nreading it, is that there seems to be no way to find out what fraction\nof time jit_tuple_deforming is taking alone, it's sort of merged\ntogether with jit_expressions in generation_counter.\n\n\n","msg_date":"Fri, 25 Feb 2022 17:40:08 +0100","msg_from":"Dmitry Dolgov <9erthalion6@gmail.com>","msg_from_op":false,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Fri, Feb 25, 2022 at 05:38:45PM +0100, Magnus Hagander wrote:\n>\n> Per some off-list discussion with Julien, we have clearly been talking\n> in slightly different terms. So let's summarize the options into what\n> theÿ́d actually be:\n>\n> Option 0: what is int he patch now\n>\n> Option 1:\n> jit_count - number of executions using jit\n> total_jit_time - for sum of functions+inlining+optimization+emission time\n> min_jit_time - for sum of functions+inlining+optimization+emission time\n> max_jit_time - for sum of functions+inlining+optimization+emission time\n> mean_jit_time - for sum of functions+inlining+optimization+emission time\n> stddev_jit_time - for sum of functions+inlining+optimization+emission time\n> jit_functions - number of functions\n> jit_inlining_count - number of executions where inlining happened\n> jit_optimization_count - number of executions where optimization happened\n> jit_emission_count - number of executions where emission happened\n>\n> Option 2:\n> jit_count\n> jit_functions\n> jit_generation_time\n> jit_inlining_count\n> jit_inlining_time\n> jit_optimization_count\n> jit_optimization_time\n> jit_emission_count\n> jit_emission_time\n\nI'm for option 2, I think it's important to have the timing details for\ninlining and optimization and be able to compute correct stats.\n\n\n","msg_date":"Sat, 26 Feb 2022 00:43:15 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On 25.02.22 14:06, Magnus Hagander wrote:\n> + OUT jit_generation_time float8,\n> + OUT jit_inlining_time float8,\n> + OUT jit_optimization_time float8,\n> + OUT jit_emission_time float8\n\nPerhaps those should be of type interval?\n\n\n","msg_date":"Mon, 28 Feb 2022 17:00:05 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Mon, Feb 28, 2022 at 05:00:05PM +0100, Peter Eisentraut wrote:\n> On 25.02.22 14:06, Magnus Hagander wrote:\n> > + OUT jit_generation_time float8,\n> > + OUT jit_inlining_time float8,\n> > + OUT jit_optimization_time float8,\n> > + OUT jit_emission_time float8\n> \n> Perhaps those should be of type interval?\n\nThe rest of the \"time\" fields are declared as float8, so I think it's better to\nkeep things consistent. And changing the type of existing fields, even in a\nmajor version, seems like a bit too much trouble.\n\n\n","msg_date":"Tue, 1 Mar 2022 00:24:49 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Fri, Feb 25, 2022 at 5:40 PM Dmitry Dolgov <9erthalion6@gmail.com> wrote:\n>\n> > On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\n> > On Fri, Feb 25, 2022 at 2:33 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> > >\n> > > Hi,\n> > >\n> > > On Fri, Feb 25, 2022 at 02:06:29PM +0100, Magnus Hagander wrote:\n> > > > Here's a patch to add the sum of timings for JIT counters to\n> > > > pg_stat_statements, as a way to follow-up on if JIT is doing a good or\n> > > > a bad job in a configuration.\n> > >\n> > > +1, it seems like something quite useful.\n> >\n> > Given the amount of time often spent debugging JIT -- getting more\n> > insight is going to make it easier to tune it instead of like what\n> > unfortunately many people do and just turn it off..\n>\n> Indeed, sounds convenient, although I wonder how exactly one would use it\n> to tune JIT? I'm curious, because I got used to situations when one\n> single long query takes much longer than expected due to JIT issues --\n> but it seems the target of this patch are situations when there are a\n> lot of long queries using JIT, and it's easier to analyze them via pgss?\n\n\"tune jit\" might be a bad wording. But tune the values of\njit_above_cost for example, and in particular figure specific queries\nwhere it needs adapting.\n\nAnd like most things with pgss, just getting proper insight into what\nyour system is actually doing and spending it's time on. It might be\nspending a huge amount of time on JITting many relatively simple\nqueries, and it may still be a win -- but this lets you know that this\nwas the case.\n\n\n> > > > I decided to only store the total time for the timings, since there\n> > > > are 4 different timings and storing max/min/etc for each one of them\n> > > > would lead to a bit too much data. This can of course be reconsidered,\n> > > > but I think that's a reasonable tradeoff.\n> > >\n> > > I think the cumulated total time is enough. Looking at the patch, I think we\n> > > should also cumulate the number of time jit was triggered, and\n> > > probably the same for each other main operation (optimization and inlining).\n> > > Otherwise the values may be wrong and look artificially low.\n> >\n> > So just to be clear, you're basically thinking:\n> >\n> > jit_count = count of entries where jit_functions>0\n> > jit_functions = <same as now>\n> > jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\n>\n> One interesting not-very-relevant for the patch thing I've noticed while\n> reading it, is that there seems to be no way to find out what fraction\n> of time jit_tuple_deforming is taking alone, it's sort of merged\n> together with jit_expressions in generation_counter.\n\nThat's missing att a deeper level though, right? We don't have it in\nEXPLAIN ANALYZE either. So while I agree that's useful, I think that's\nthe job of another patch, and these two sets of counters should be the\nsame.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/\n\n\n","msg_date":"Mon, 7 Mar 2022 13:27:02 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Fri, Feb 25, 2022 at 5:43 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n>\n> On Fri, Feb 25, 2022 at 05:38:45PM +0100, Magnus Hagander wrote:\n> >\n> > Per some off-list discussion with Julien, we have clearly been talking\n> > in slightly different terms. So let's summarize the options into what\n> > theÿ́d actually be:\n> >\n> > Option 0: what is int he patch now\n> >\n> > Option 1:\n> > jit_count - number of executions using jit\n> > total_jit_time - for sum of functions+inlining+optimization+emission time\n> > min_jit_time - for sum of functions+inlining+optimization+emission time\n> > max_jit_time - for sum of functions+inlining+optimization+emission time\n> > mean_jit_time - for sum of functions+inlining+optimization+emission time\n> > stddev_jit_time - for sum of functions+inlining+optimization+emission time\n> > jit_functions - number of functions\n> > jit_inlining_count - number of executions where inlining happened\n> > jit_optimization_count - number of executions where optimization happened\n> > jit_emission_count - number of executions where emission happened\n> >\n> > Option 2:\n> > jit_count\n> > jit_functions\n> > jit_generation_time\n> > jit_inlining_count\n> > jit_inlining_time\n> > jit_optimization_count\n> > jit_optimization_time\n> > jit_emission_count\n> > jit_emission_time\n>\n> I'm for option 2, I think it's important to have the timing details for\n> inlining and optimization and be able to compute correct stats.\n\nI wonder if there might be an interesting middle ground, or if that is\nmaking it too much. That is, we could have an\nOption 3:\njit_count\ntotal_jit_time - for sum of functions+inlining+optimization+emission time\nmin_jit_time - for sum of functions+inlining+optimization+emission time\nmax_jit_time - for sum of functions+inlining+optimization+emission time\nmean_jit_time - for sum of functions+inlining+optimization+emission time\nstddev_jit_time - for sum of functions+inlining+optimization+emission time\njit_functions\njit_generation_time\njit_inlining_count\njit_inlining_time\njit_optimization_count\njit_optimization_time\njit_emission_count\njit_emission_time\n\nThat is, we'd get the more detailed timings across the total time, but\nnot on the details. But that might be overkill.\n\nBut -- here's an updated patched based on Option 2.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/","msg_date":"Mon, 7 Mar 2022 13:40:34 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"> On Mon, Mar 07, 2022 at 01:27:02PM +0100, Magnus Hagander wrote:\n> On Fri, Feb 25, 2022 at 5:40 PM Dmitry Dolgov <9erthalion6@gmail.com> wrote:\n> > One interesting not-very-relevant for the patch thing I've noticed while\n> > reading it, is that there seems to be no way to find out what fraction\n> > of time jit_tuple_deforming is taking alone, it's sort of merged\n> > together with jit_expressions in generation_counter.\n>\n> That's missing att a deeper level though, right? We don't have it in\n> EXPLAIN ANALYZE either. So while I agree that's useful, I think that's\n> the job of another patch, and these two sets of counters should be the\n> same.\n\nRight, it's missing on the instrumentation level, I was just surprised\nto notice that.\n\n\n","msg_date":"Mon, 7 Mar 2022 14:01:08 +0100","msg_from":"Dmitry Dolgov <9erthalion6@gmail.com>","msg_from_op":false,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Mon, Mar 07, 2022 at 01:40:34PM +0100, Magnus Hagander wrote:\n>\n> I wonder if there might be an interesting middle ground, or if that is\n> making it too much. That is, we could have an\n> Option 3:\n> jit_count\n> total_jit_time - for sum of functions+inlining+optimization+emission time\n> min_jit_time - for sum of functions+inlining+optimization+emission time\n> max_jit_time - for sum of functions+inlining+optimization+emission time\n> mean_jit_time - for sum of functions+inlining+optimization+emission time\n> stddev_jit_time - for sum of functions+inlining+optimization+emission time\n> jit_functions\n> jit_generation_time\n> jit_inlining_count\n> jit_inlining_time\n> jit_optimization_count\n> jit_optimization_time\n> jit_emission_count\n> jit_emission_time\n>\n> That is, we'd get the more detailed timings across the total time, but\n> not on the details. But that might be overkill.\n\nI also thought about it but it seems overkill. pg_stat_statements view is\nalready very big, and I think that the JIT time should be somewhat stable, at\nleast compared to how much a query execution time can vary depending on the\nparameters. This approach would also be a bit useless if you change the\ncosting of underlying JIT operation.\n\n> But -- here's an updated patched based on Option 2.\n\nThanks!\n\nCode-wide, the patch looks good. For the doc, it seems that you documented\njit_inlining_count three times rather than documenting jit_optimization_count\nand jit_emission_count.\n\nI don't think we can add tests there, and having a test for every new counter\nbeing >= 0 seems entirely useless, however there should be a new test added for\nthe \"oldextversions\" test to make sure that there's no issue with old SQL / new\nshlib compatibility. And looking at it I see that it was already missed for\nversion 1.9 :(\n\n\n","msg_date":"Tue, 8 Mar 2022 11:08:01 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"On Tue, Mar 8, 2022 at 4:08 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\n\n> On Mon, Mar 07, 2022 at 01:40:34PM +0100, Magnus Hagander wrote:\n> >\n> > I wonder if there might be an interesting middle ground, or if that is\n> > making it too much. That is, we could have an\n> > Option 3:\n> > jit_count\n> > total_jit_time - for sum of functions+inlining+optimization+emission time\n> > min_jit_time - for sum of functions+inlining+optimization+emission time\n> > max_jit_time - for sum of functions+inlining+optimization+emission time\n> > mean_jit_time - for sum of functions+inlining+optimization+emission time\n> > stddev_jit_time - for sum of functions+inlining+optimization+emission\n> time\n> > jit_functions\n> > jit_generation_time\n> > jit_inlining_count\n> > jit_inlining_time\n> > jit_optimization_count\n> > jit_optimization_time\n> > jit_emission_count\n> > jit_emission_time\n> >\n> > That is, we'd get the more detailed timings across the total time, but\n> > not on the details. But that might be overkill.\n>\n> I also thought about it but it seems overkill. pg_stat_statements view is\n> already very big, and I think that the JIT time should be somewhat stable,\n> at\n> least compared to how much a query execution time can vary depending on the\n> parameters. This approach would also be a bit useless if you change the\n> costing of underlying JIT operation.\n>\n> > But -- here's an updated patched based on Option 2.\n>\n> Thanks!\n>\n> Code-wide, the patch looks good. For the doc, it seems that you documented\n> jit_inlining_count three times rather than documenting\n> jit_optimization_count\n> and jit_emission_count.\n>\n\nOops, thanks and fixed.\n\n\nI don't think we can add tests there, and having a test for every new\n> counter\n> being >= 0 seems entirely useless, however there should be a new test\n> added for\n> the \"oldextversions\" test to make sure that there's no issue with old SQL\n> / new\n> shlib compatibility. And looking at it I see that it was already missed\n> for\n> version 1.9 :(\n>\n\nIndeed. Fixed here.\n\nMichael had already applied a patch that took us to 1.10 and added that\ntest, so I've just updated it here. I don't think we normally bump the\nversion twice int he same day, so I just mergd the SQL script changes as\nwell.\n\nPFA a \"final\" version for the CI to run.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>","msg_date":"Fri, 8 Apr 2022 13:47:20 +0200","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"Hi,\r\nthank you for the great features.\r\n\r\nThe attached small patch changes the data type in the document.\r\nThe following columns are actually double precision but bigint in the docs.\r\njit_generation_time\r\njit_inlining_time\r\njit_optimization_time\r\njit_emission_count\r\n\r\nRegards,\r\nNoriyoshi Shinoda\r\n\r\nFrom: Magnus Hagander <magnus@hagander.net>\r\nSent: Friday, April 8, 2022 8:47 PM\r\nTo: Julien Rouhaud <rjuju123@gmail.com>\r\nCc: PostgreSQL Developers <pgsql-hackers@lists.postgresql.org>\r\nSubject: Re: Expose JIT counters/timing in pg_stat_statements\r\n\r\n\r\n\r\nOn Tue, Mar 8, 2022 at 4:08 AM Julien Rouhaud <rjuju123@gmail.com<mailto:rjuju123@gmail.com>> wrote:\r\nOn Mon, Mar 07, 2022 at 01:40:34PM +0100, Magnus Hagander wrote:\r\n>\r\n> I wonder if there might be an interesting middle ground, or if that is\r\n> making it too much. That is, we could have an\r\n> Option 3:\r\n> jit_count\r\n> total_jit_time - for sum of functions+inlining+optimization+emission time\r\n> min_jit_time - for sum of functions+inlining+optimization+emission time\r\n> max_jit_time - for sum of functions+inlining+optimization+emission time\r\n> mean_jit_time - for sum of functions+inlining+optimization+emission time\r\n> stddev_jit_time - for sum of functions+inlining+optimization+emission time\r\n> jit_functions\r\n> jit_generation_time\r\n> jit_inlining_count\r\n> jit_inlining_time\r\n> jit_optimization_count\r\n> jit_optimization_time\r\n> jit_emission_count\r\n> jit_emission_time\r\n>\r\n> That is, we'd get the more detailed timings across the total time, but\r\n> not on the details. But that might be overkill.\r\n\r\nI also thought about it but it seems overkill. pg_stat_statements view is\r\nalready very big, and I think that the JIT time should be somewhat stable, at\r\nleast compared to how much a query execution time can vary depending on the\r\nparameters. This approach would also be a bit useless if you change the\r\ncosting of underlying JIT operation.\r\n\r\n> But -- here's an updated patched based on Option 2.\r\n\r\nThanks!\r\n\r\nCode-wide, the patch looks good. For the doc, it seems that you documented\r\njit_inlining_count three times rather than documenting jit_optimization_count\r\nand jit_emission_count.\r\n\r\nOops, thanks and fixed.\r\n\r\n\r\nI don't think we can add tests there, and having a test for every new counter\r\nbeing >= 0 seems entirely useless, however there should be a new test added for\r\nthe \"oldextversions\" test to make sure that there's no issue with old SQL / new\r\nshlib compatibility. And looking at it I see that it was already missed for\r\nversion 1.9 :(\r\n\r\nIndeed. Fixed here.\r\n\r\nMichael had already applied a patch that took us to 1.10 and added that test, so I've just updated it here. I don't think we normally bump the version twice int he same day, so I just mergd the SQL script changes as well.\r\n\r\nPFA a \"final\" version for the CI to run.\r\n\r\n--\r\n Magnus Hagander\r\n Me: https://www.hagander.net/<http://www.hagander.net/>\r\n Work: https://www.redpill-linpro.com/<http://www.redpill-linpro.com/>","msg_date":"Sat, 9 Apr 2022 01:51:21 +0000","msg_from":"\"Shinoda, Noriyoshi (PN Japan FSIP)\" <noriyoshi.shinoda@hpe.com>","msg_from_op":false,"msg_subject":"RE: Expose JIT counters/timing in pg_stat_statements"},{"msg_contents":"Hi,\n\nOn Sat, Apr 09, 2022 at 01:51:21AM +0000, Shinoda, Noriyoshi (PN Japan FSIP) wrote:\n> Hi,\n> thank you for the great features.\n>\n> The attached small patch changes the data type in the document.\n> The following columns are actually double precision but bigint in the docs.\n> jit_generation_time\n> jit_inlining_time\n> jit_optimization_time\n> jit_emission_count\n\nIndeed! The patch looks good to me, I didn't find any other discrepancy.\n\n\n","msg_date":"Sat, 9 Apr 2022 11:47:57 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Expose JIT counters/timing in pg_stat_statements"}],"string":"[\n {\n \"msg_contents\": \"Here's a patch to add the sum of timings for JIT counters to\\npg_stat_statements, as a way to follow-up on if JIT is doing a good or\\na bad job in a configuration.\\n\\nI decided to only store the total time for the timings, since there\\nare 4 different timings and storing max/min/etc for each one of them\\nwould lead to a bit too much data. This can of course be reconsidered,\\nbut I think that's a reasonable tradeoff.\\n\\nAnother option I guess could be to store the max/min/etc, but only\\nstore for the total jit time instead of for each individual one. Maybe\\nthat'd actually be more useful?\\n\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Fri, 25 Feb 2022 14:06:29 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Fri, Feb 25, 2022 at 02:06:29PM +0100, Magnus Hagander wrote:\\n> Here's a patch to add the sum of timings for JIT counters to\\n> pg_stat_statements, as a way to follow-up on if JIT is doing a good or\\n> a bad job in a configuration.\\n\\n+1, it seems like something quite useful.\\n\\n> I decided to only store the total time for the timings, since there\\n> are 4 different timings and storing max/min/etc for each one of them\\n> would lead to a bit too much data. This can of course be reconsidered,\\n> but I think that's a reasonable tradeoff.\\n\\nI think the cumulated total time is enough. Looking at the patch, I think we\\nshould also cumulate the number of time jit was triggered, and\\nprobably the same for each other main operation (optimization and inlining).\\nOtherwise the values may be wrong and look artificially low.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 21:33:22 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 2:33 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n>\\n> Hi,\\n>\\n> On Fri, Feb 25, 2022 at 02:06:29PM +0100, Magnus Hagander wrote:\\n> > Here's a patch to add the sum of timings for JIT counters to\\n> > pg_stat_statements, as a way to follow-up on if JIT is doing a good or\\n> > a bad job in a configuration.\\n>\\n> +1, it seems like something quite useful.\\n\\nGiven the amount of time often spent debugging JIT -- getting more\\ninsight is going to make it easier to tune it instead of like what\\nunfortunately many people do and just turn it off..\\n\\n\\n> > I decided to only store the total time for the timings, since there\\n> > are 4 different timings and storing max/min/etc for each one of them\\n> > would lead to a bit too much data. This can of course be reconsidered,\\n> > but I think that's a reasonable tradeoff.\\n>\\n> I think the cumulated total time is enough. Looking at the patch, I think we\\n> should also cumulate the number of time jit was triggered, and\\n> probably the same for each other main operation (optimization and inlining).\\n> Otherwise the values may be wrong and look artificially low.\\n\\nSo just to be clear, you're basically thinking:\\n\\njit_count = count of entries where jit_functions>0\\njit_functions = <same as now>\\njit_optimizatinos = count of entries where time spent on jit_optimizations > 0\\n\\netc?\\n\\nSo we count the times with min/max like other times for the total one,\\nbut instead add a counter for each of the details?\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 16:19:27 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\\n> \\n> So just to be clear, you're basically thinking:\\n> \\n> jit_count = count of entries where jit_functions>0\\n> jit_functions = <same as now>\\n> jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\\n> \\n> etc?\\n\\nYes exactly, so 3 new fields on top of the one you already added.\\n\\n> So we count the times with min/max like other times for the total one,\\n> but instead add a counter for each of the details?\\n\\nI don't understand this one. Did you mean we *don't* count times with min/max?\\nIf that's the case then yes :)\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 23:40:01 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 4:40 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n>\\n> On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\\n> >\\n> > So just to be clear, you're basically thinking:\\n> >\\n> > jit_count = count of entries where jit_functions>0\\n> > jit_functions = <same as now>\\n> > jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\\n> >\\n> > etc?\\n>\\n> Yes exactly, so 3 new fields on top of the one you already added.\\n>\\n> > So we count the times with min/max like other times for the total one,\\n> > but instead add a counter for each of the details?\\n>\\n> I don't understand this one. Did you mean we *don't* count times with min/max?\\n> If that's the case then yes :)\\n\\nHmm. Yeah, that was a bit unclear. I mean we track total time with\\nmin/max, and detailed time not at all. For details, we only track\\ncount, not time.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 16:41:22 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 4:41 PM Magnus Hagander <magnus@hagander.net> wrote:\\n>\\n> On Fri, Feb 25, 2022 at 4:40 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> >\\n> > On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\\n> > >\\n> > > So just to be clear, you're basically thinking:\\n> > >\\n> > > jit_count = count of entries where jit_functions>0\\n> > > jit_functions = <same as now>\\n> > > jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\\n> > >\\n> > > etc?\\n> >\\n> > Yes exactly, so 3 new fields on top of the one you already added.\\n> >\\n> > > So we count the times with min/max like other times for the total one,\\n> > > but instead add a counter for each of the details?\\n> >\\n> > I don't understand this one. Did you mean we *don't* count times with min/max?\\n> > If that's the case then yes :)\\n>\\n> Hmm. Yeah, that was a bit unclear. I mean we track total time with\\n> min/max, and detailed time not at all. For details, we only track\\n> count, not time.\\n\\nPer some off-list discussion with Julien, we have clearly been talking\\nin slightly different terms. So let's summarize the options into what\\ntheÿ́d actually be:\\n\\nOption 0: what is int he patch now\\n\\nOption 1:\\njit_count - number of executions using jit\\ntotal_jit_time - for sum of functions+inlining+optimization+emission time\\nmin_jit_time - for sum of functions+inlining+optimization+emission time\\nmax_jit_time - for sum of functions+inlining+optimization+emission time\\nmean_jit_time - for sum of functions+inlining+optimization+emission time\\nstddev_jit_time - for sum of functions+inlining+optimization+emission time\\njit_functions - number of functions\\njit_inlining_count - number of executions where inlining happened\\njit_optimization_count - number of executions where optimization happened\\njit_emission_count - number of executions where emission happened\\n\\nOption 2:\\njit_count\\njit_functions\\njit_generation_time\\njit_inlining_count\\njit_inlining_time\\njit_optimization_count\\njit_optimization_time\\njit_emission_count\\njit_emission_time\\n\\n\\n(We can bikeshed on the exact names of the fields once we have decided\\nwhich option is preferred)\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 17:38:45 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"> On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\\n> On Fri, Feb 25, 2022 at 2:33 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> >\\n> > Hi,\\n> >\\n> > On Fri, Feb 25, 2022 at 02:06:29PM +0100, Magnus Hagander wrote:\\n> > > Here's a patch to add the sum of timings for JIT counters to\\n> > > pg_stat_statements, as a way to follow-up on if JIT is doing a good or\\n> > > a bad job in a configuration.\\n> >\\n> > +1, it seems like something quite useful.\\n>\\n> Given the amount of time often spent debugging JIT -- getting more\\n> insight is going to make it easier to tune it instead of like what\\n> unfortunately many people do and just turn it off..\\n\\nIndeed, sounds convenient, although I wonder how exactly one would use it\\nto tune JIT? I'm curious, because I got used to situations when one\\nsingle long query takes much longer than expected due to JIT issues --\\nbut it seems the target of this patch are situations when there are a\\nlot of long queries using JIT, and it's easier to analyze them via pgss?\\n\\n> > > I decided to only store the total time for the timings, since there\\n> > > are 4 different timings and storing max/min/etc for each one of them\\n> > > would lead to a bit too much data. This can of course be reconsidered,\\n> > > but I think that's a reasonable tradeoff.\\n> >\\n> > I think the cumulated total time is enough. Looking at the patch, I think we\\n> > should also cumulate the number of time jit was triggered, and\\n> > probably the same for each other main operation (optimization and inlining).\\n> > Otherwise the values may be wrong and look artificially low.\\n>\\n> So just to be clear, you're basically thinking:\\n>\\n> jit_count = count of entries where jit_functions>0\\n> jit_functions = <same as now>\\n> jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\\n\\nOne interesting not-very-relevant for the patch thing I've noticed while\\nreading it, is that there seems to be no way to find out what fraction\\nof time jit_tuple_deforming is taking alone, it's sort of merged\\ntogether with jit_expressions in generation_counter.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 17:40:08 +0100\",\n \"msg_from\": \"Dmitry Dolgov <9erthalion6@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 05:38:45PM +0100, Magnus Hagander wrote:\\n>\\n> Per some off-list discussion with Julien, we have clearly been talking\\n> in slightly different terms. So let's summarize the options into what\\n> theÿ́d actually be:\\n>\\n> Option 0: what is int he patch now\\n>\\n> Option 1:\\n> jit_count - number of executions using jit\\n> total_jit_time - for sum of functions+inlining+optimization+emission time\\n> min_jit_time - for sum of functions+inlining+optimization+emission time\\n> max_jit_time - for sum of functions+inlining+optimization+emission time\\n> mean_jit_time - for sum of functions+inlining+optimization+emission time\\n> stddev_jit_time - for sum of functions+inlining+optimization+emission time\\n> jit_functions - number of functions\\n> jit_inlining_count - number of executions where inlining happened\\n> jit_optimization_count - number of executions where optimization happened\\n> jit_emission_count - number of executions where emission happened\\n>\\n> Option 2:\\n> jit_count\\n> jit_functions\\n> jit_generation_time\\n> jit_inlining_count\\n> jit_inlining_time\\n> jit_optimization_count\\n> jit_optimization_time\\n> jit_emission_count\\n> jit_emission_time\\n\\nI'm for option 2, I think it's important to have the timing details for\\ninlining and optimization and be able to compute correct stats.\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 00:43:15 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On 25.02.22 14:06, Magnus Hagander wrote:\\n> + OUT jit_generation_time float8,\\n> + OUT jit_inlining_time float8,\\n> + OUT jit_optimization_time float8,\\n> + OUT jit_emission_time float8\\n\\nPerhaps those should be of type interval?\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 17:00:05 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 05:00:05PM +0100, Peter Eisentraut wrote:\\n> On 25.02.22 14:06, Magnus Hagander wrote:\\n> > + OUT jit_generation_time float8,\\n> > + OUT jit_inlining_time float8,\\n> > + OUT jit_optimization_time float8,\\n> > + OUT jit_emission_time float8\\n> \\n> Perhaps those should be of type interval?\\n\\nThe rest of the \\\"time\\\" fields are declared as float8, so I think it's better to\\nkeep things consistent. And changing the type of existing fields, even in a\\nmajor version, seems like a bit too much trouble.\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 00:24:49 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 5:40 PM Dmitry Dolgov <9erthalion6@gmail.com> wrote:\\n>\\n> > On Fri, Feb 25, 2022 at 04:19:27PM +0100, Magnus Hagander wrote:\\n> > On Fri, Feb 25, 2022 at 2:33 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> > >\\n> > > Hi,\\n> > >\\n> > > On Fri, Feb 25, 2022 at 02:06:29PM +0100, Magnus Hagander wrote:\\n> > > > Here's a patch to add the sum of timings for JIT counters to\\n> > > > pg_stat_statements, as a way to follow-up on if JIT is doing a good or\\n> > > > a bad job in a configuration.\\n> > >\\n> > > +1, it seems like something quite useful.\\n> >\\n> > Given the amount of time often spent debugging JIT -- getting more\\n> > insight is going to make it easier to tune it instead of like what\\n> > unfortunately many people do and just turn it off..\\n>\\n> Indeed, sounds convenient, although I wonder how exactly one would use it\\n> to tune JIT? I'm curious, because I got used to situations when one\\n> single long query takes much longer than expected due to JIT issues --\\n> but it seems the target of this patch are situations when there are a\\n> lot of long queries using JIT, and it's easier to analyze them via pgss?\\n\\n\\\"tune jit\\\" might be a bad wording. But tune the values of\\njit_above_cost for example, and in particular figure specific queries\\nwhere it needs adapting.\\n\\nAnd like most things with pgss, just getting proper insight into what\\nyour system is actually doing and spending it's time on. It might be\\nspending a huge amount of time on JITting many relatively simple\\nqueries, and it may still be a win -- but this lets you know that this\\nwas the case.\\n\\n\\n> > > > I decided to only store the total time for the timings, since there\\n> > > > are 4 different timings and storing max/min/etc for each one of them\\n> > > > would lead to a bit too much data. This can of course be reconsidered,\\n> > > > but I think that's a reasonable tradeoff.\\n> > >\\n> > > I think the cumulated total time is enough. Looking at the patch, I think we\\n> > > should also cumulate the number of time jit was triggered, and\\n> > > probably the same for each other main operation (optimization and inlining).\\n> > > Otherwise the values may be wrong and look artificially low.\\n> >\\n> > So just to be clear, you're basically thinking:\\n> >\\n> > jit_count = count of entries where jit_functions>0\\n> > jit_functions = <same as now>\\n> > jit_optimizatinos = count of entries where time spent on jit_optimizations > 0\\n>\\n> One interesting not-very-relevant for the patch thing I've noticed while\\n> reading it, is that there seems to be no way to find out what fraction\\n> of time jit_tuple_deforming is taking alone, it's sort of merged\\n> together with jit_expressions in generation_counter.\\n\\nThat's missing att a deeper level though, right? We don't have it in\\nEXPLAIN ANALYZE either. So while I agree that's useful, I think that's\\nthe job of another patch, and these two sets of counters should be the\\nsame.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 13:27:02 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 5:43 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n>\\n> On Fri, Feb 25, 2022 at 05:38:45PM +0100, Magnus Hagander wrote:\\n> >\\n> > Per some off-list discussion with Julien, we have clearly been talking\\n> > in slightly different terms. So let's summarize the options into what\\n> > theÿ́d actually be:\\n> >\\n> > Option 0: what is int he patch now\\n> >\\n> > Option 1:\\n> > jit_count - number of executions using jit\\n> > total_jit_time - for sum of functions+inlining+optimization+emission time\\n> > min_jit_time - for sum of functions+inlining+optimization+emission time\\n> > max_jit_time - for sum of functions+inlining+optimization+emission time\\n> > mean_jit_time - for sum of functions+inlining+optimization+emission time\\n> > stddev_jit_time - for sum of functions+inlining+optimization+emission time\\n> > jit_functions - number of functions\\n> > jit_inlining_count - number of executions where inlining happened\\n> > jit_optimization_count - number of executions where optimization happened\\n> > jit_emission_count - number of executions where emission happened\\n> >\\n> > Option 2:\\n> > jit_count\\n> > jit_functions\\n> > jit_generation_time\\n> > jit_inlining_count\\n> > jit_inlining_time\\n> > jit_optimization_count\\n> > jit_optimization_time\\n> > jit_emission_count\\n> > jit_emission_time\\n>\\n> I'm for option 2, I think it's important to have the timing details for\\n> inlining and optimization and be able to compute correct stats.\\n\\nI wonder if there might be an interesting middle ground, or if that is\\nmaking it too much. That is, we could have an\\nOption 3:\\njit_count\\ntotal_jit_time - for sum of functions+inlining+optimization+emission time\\nmin_jit_time - for sum of functions+inlining+optimization+emission time\\nmax_jit_time - for sum of functions+inlining+optimization+emission time\\nmean_jit_time - for sum of functions+inlining+optimization+emission time\\nstddev_jit_time - for sum of functions+inlining+optimization+emission time\\njit_functions\\njit_generation_time\\njit_inlining_count\\njit_inlining_time\\njit_optimization_count\\njit_optimization_time\\njit_emission_count\\njit_emission_time\\n\\nThat is, we'd get the more detailed timings across the total time, but\\nnot on the details. But that might be overkill.\\n\\nBut -- here's an updated patched based on Option 2.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Mon, 7 Mar 2022 13:40:34 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"> On Mon, Mar 07, 2022 at 01:27:02PM +0100, Magnus Hagander wrote:\\n> On Fri, Feb 25, 2022 at 5:40 PM Dmitry Dolgov <9erthalion6@gmail.com> wrote:\\n> > One interesting not-very-relevant for the patch thing I've noticed while\\n> > reading it, is that there seems to be no way to find out what fraction\\n> > of time jit_tuple_deforming is taking alone, it's sort of merged\\n> > together with jit_expressions in generation_counter.\\n>\\n> That's missing att a deeper level though, right? We don't have it in\\n> EXPLAIN ANALYZE either. So while I agree that's useful, I think that's\\n> the job of another patch, and these two sets of counters should be the\\n> same.\\n\\nRight, it's missing on the instrumentation level, I was just surprised\\nto notice that.\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 14:01:08 +0100\",\n \"msg_from\": \"Dmitry Dolgov <9erthalion6@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 07, 2022 at 01:40:34PM +0100, Magnus Hagander wrote:\\n>\\n> I wonder if there might be an interesting middle ground, or if that is\\n> making it too much. That is, we could have an\\n> Option 3:\\n> jit_count\\n> total_jit_time - for sum of functions+inlining+optimization+emission time\\n> min_jit_time - for sum of functions+inlining+optimization+emission time\\n> max_jit_time - for sum of functions+inlining+optimization+emission time\\n> mean_jit_time - for sum of functions+inlining+optimization+emission time\\n> stddev_jit_time - for sum of functions+inlining+optimization+emission time\\n> jit_functions\\n> jit_generation_time\\n> jit_inlining_count\\n> jit_inlining_time\\n> jit_optimization_count\\n> jit_optimization_time\\n> jit_emission_count\\n> jit_emission_time\\n>\\n> That is, we'd get the more detailed timings across the total time, but\\n> not on the details. But that might be overkill.\\n\\nI also thought about it but it seems overkill. pg_stat_statements view is\\nalready very big, and I think that the JIT time should be somewhat stable, at\\nleast compared to how much a query execution time can vary depending on the\\nparameters. This approach would also be a bit useless if you change the\\ncosting of underlying JIT operation.\\n\\n> But -- here's an updated patched based on Option 2.\\n\\nThanks!\\n\\nCode-wide, the patch looks good. For the doc, it seems that you documented\\njit_inlining_count three times rather than documenting jit_optimization_count\\nand jit_emission_count.\\n\\nI don't think we can add tests there, and having a test for every new counter\\nbeing >= 0 seems entirely useless, however there should be a new test added for\\nthe \\\"oldextversions\\\" test to make sure that there's no issue with old SQL / new\\nshlib compatibility. And looking at it I see that it was already missed for\\nversion 1.9 :(\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Mar 2022 11:08:01 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 8, 2022 at 4:08 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n\\n> On Mon, Mar 07, 2022 at 01:40:34PM +0100, Magnus Hagander wrote:\\n> >\\n> > I wonder if there might be an interesting middle ground, or if that is\\n> > making it too much. That is, we could have an\\n> > Option 3:\\n> > jit_count\\n> > total_jit_time - for sum of functions+inlining+optimization+emission time\\n> > min_jit_time - for sum of functions+inlining+optimization+emission time\\n> > max_jit_time - for sum of functions+inlining+optimization+emission time\\n> > mean_jit_time - for sum of functions+inlining+optimization+emission time\\n> > stddev_jit_time - for sum of functions+inlining+optimization+emission\\n> time\\n> > jit_functions\\n> > jit_generation_time\\n> > jit_inlining_count\\n> > jit_inlining_time\\n> > jit_optimization_count\\n> > jit_optimization_time\\n> > jit_emission_count\\n> > jit_emission_time\\n> >\\n> > That is, we'd get the more detailed timings across the total time, but\\n> > not on the details. But that might be overkill.\\n>\\n> I also thought about it but it seems overkill. pg_stat_statements view is\\n> already very big, and I think that the JIT time should be somewhat stable,\\n> at\\n> least compared to how much a query execution time can vary depending on the\\n> parameters. This approach would also be a bit useless if you change the\\n> costing of underlying JIT operation.\\n>\\n> > But -- here's an updated patched based on Option 2.\\n>\\n> Thanks!\\n>\\n> Code-wide, the patch looks good. For the doc, it seems that you documented\\n> jit_inlining_count three times rather than documenting\\n> jit_optimization_count\\n> and jit_emission_count.\\n>\\n\\nOops, thanks and fixed.\\n\\n\\nI don't think we can add tests there, and having a test for every new\\n> counter\\n> being >= 0 seems entirely useless, however there should be a new test\\n> added for\\n> the \\\"oldextversions\\\" test to make sure that there's no issue with old SQL\\n> / new\\n> shlib compatibility. And looking at it I see that it was already missed\\n> for\\n> version 1.9 :(\\n>\\n\\nIndeed. Fixed here.\\n\\nMichael had already applied a patch that took us to 1.10 and added that\\ntest, so I've just updated it here. I don't think we normally bump the\\nversion twice int he same day, so I just mergd the SQL script changes as\\nwell.\\n\\nPFA a \\\"final\\\" version for the CI to run.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\",\n \"msg_date\": \"Fri, 8 Apr 2022 13:47:20 +0200\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"Hi,\\r\\nthank you for the great features.\\r\\n\\r\\nThe attached small patch changes the data type in the document.\\r\\nThe following columns are actually double precision but bigint in the docs.\\r\\njit_generation_time\\r\\njit_inlining_time\\r\\njit_optimization_time\\r\\njit_emission_count\\r\\n\\r\\nRegards,\\r\\nNoriyoshi Shinoda\\r\\n\\r\\nFrom: Magnus Hagander <magnus@hagander.net>\\r\\nSent: Friday, April 8, 2022 8:47 PM\\r\\nTo: Julien Rouhaud <rjuju123@gmail.com>\\r\\nCc: PostgreSQL Developers <pgsql-hackers@lists.postgresql.org>\\r\\nSubject: Re: Expose JIT counters/timing in pg_stat_statements\\r\\n\\r\\n\\r\\n\\r\\nOn Tue, Mar 8, 2022 at 4:08 AM Julien Rouhaud <rjuju123@gmail.com<mailto:rjuju123@gmail.com>> wrote:\\r\\nOn Mon, Mar 07, 2022 at 01:40:34PM +0100, Magnus Hagander wrote:\\r\\n>\\r\\n> I wonder if there might be an interesting middle ground, or if that is\\r\\n> making it too much. That is, we could have an\\r\\n> Option 3:\\r\\n> jit_count\\r\\n> total_jit_time - for sum of functions+inlining+optimization+emission time\\r\\n> min_jit_time - for sum of functions+inlining+optimization+emission time\\r\\n> max_jit_time - for sum of functions+inlining+optimization+emission time\\r\\n> mean_jit_time - for sum of functions+inlining+optimization+emission time\\r\\n> stddev_jit_time - for sum of functions+inlining+optimization+emission time\\r\\n> jit_functions\\r\\n> jit_generation_time\\r\\n> jit_inlining_count\\r\\n> jit_inlining_time\\r\\n> jit_optimization_count\\r\\n> jit_optimization_time\\r\\n> jit_emission_count\\r\\n> jit_emission_time\\r\\n>\\r\\n> That is, we'd get the more detailed timings across the total time, but\\r\\n> not on the details. But that might be overkill.\\r\\n\\r\\nI also thought about it but it seems overkill. pg_stat_statements view is\\r\\nalready very big, and I think that the JIT time should be somewhat stable, at\\r\\nleast compared to how much a query execution time can vary depending on the\\r\\nparameters. This approach would also be a bit useless if you change the\\r\\ncosting of underlying JIT operation.\\r\\n\\r\\n> But -- here's an updated patched based on Option 2.\\r\\n\\r\\nThanks!\\r\\n\\r\\nCode-wide, the patch looks good. For the doc, it seems that you documented\\r\\njit_inlining_count three times rather than documenting jit_optimization_count\\r\\nand jit_emission_count.\\r\\n\\r\\nOops, thanks and fixed.\\r\\n\\r\\n\\r\\nI don't think we can add tests there, and having a test for every new counter\\r\\nbeing >= 0 seems entirely useless, however there should be a new test added for\\r\\nthe \\\"oldextversions\\\" test to make sure that there's no issue with old SQL / new\\r\\nshlib compatibility. And looking at it I see that it was already missed for\\r\\nversion 1.9 :(\\r\\n\\r\\nIndeed. Fixed here.\\r\\n\\r\\nMichael had already applied a patch that took us to 1.10 and added that test, so I've just updated it here. I don't think we normally bump the version twice int he same day, so I just mergd the SQL script changes as well.\\r\\n\\r\\nPFA a \\\"final\\\" version for the CI to run.\\r\\n\\r\\n--\\r\\n Magnus Hagander\\r\\n Me: https://www.hagander.net/<http://www.hagander.net/>\\r\\n Work: https://www.redpill-linpro.com/<http://www.redpill-linpro.com/>\",\n \"msg_date\": \"Sat, 9 Apr 2022 01:51:21 +0000\",\n \"msg_from\": \"\\\"Shinoda, Noriyoshi (PN Japan FSIP)\\\" <noriyoshi.shinoda@hpe.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Expose JIT counters/timing in pg_stat_statements\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Sat, Apr 09, 2022 at 01:51:21AM +0000, Shinoda, Noriyoshi (PN Japan FSIP) wrote:\\n> Hi,\\n> thank you for the great features.\\n>\\n> The attached small patch changes the data type in the document.\\n> The following columns are actually double precision but bigint in the docs.\\n> jit_generation_time\\n> jit_inlining_time\\n> jit_optimization_time\\n> jit_emission_count\\n\\nIndeed! The patch looks good to me, I didn't find any other discrepancy.\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Apr 2022 11:47:57 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Expose JIT counters/timing in pg_stat_statements\"\n }\n]","isConversational":false}}},{"rowIdx":53582,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nWhile doing some work using our functions [1] for calculate relations size\nI noticed an inconsistency between pg_total_relation_size and calculate\neverything separately, have a look in this example:\n\nfabrizio=# create table test_size (id bigserial primary key, toast_column\ntext);\nCREATE TABLE\n\nfabrizio=# insert into test_size (toast_column)\n\n select repeat('X'::text, pg_size_bytes('1MB')::integer)\n from generate_series(1,1000);\nINSERT 0 1000\n\nfabrizio=# with relations as (\n select schemaname, relname, relid\n from pg_stat_user_tables\n where relname = 'test_size'\n),\nsizes as (\n select\n schemaname,\n r.relname,\n\n pg_total_relation_size(relid) AS total_bytes,\n\n pg_relation_size(relid, 'main') +\n pg_relation_size(relid, 'init') +\n pg_relation_size(relid, 'fsm') +\n pg_relation_size(relid, 'vm') AS heap_bytes,\n pg_indexes_size(relid) AS index_bytes,\n pg_table_size(reltoastrelid) AS toast_bytes\n from relations r\n join pg_class on pg_class.oid = r.relid\n)\nselect\n total_bytes, heap_bytes, index_bytes, toast_bytes,\n (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \"Equal?\",\n (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \"Diff\"\nfrom sizes;\n total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff\n-------------+------------+-------------+-------------+--------+--------\n 14000128 | 90112 | 40960 | 13688832 | f | 180224\n(1 row)\n\nI want to calculate separately HEAP, INDEXES and TOAST (including indexes)\nsizes but it seems it's a bit inconsistent with pg_total_relation_size.\n\nIs it correct or am I missing something?\n\nRegards,\n\n[1]\nhttps://www.postgresql.org/docs/current/functions-admin.html#FUNCTIONS-ADMIN-DBSIZE\n\n-- \nFabrízio de Royes Mello\n\nHi all,While doing some work using our functions [1] for calculate relations size I noticed an inconsistency between pg_total_relation_size and calculate everything separately, have a look in this example:fabrizio=# create table test_size (id bigserial primary key, toast_column text);CREATE TABLEfabrizio=# insert into test_size (toast_column) select repeat('X'::text, pg_size_bytes('1MB')::integer) from generate_series(1,1000);INSERT 0 1000fabrizio=# with relations as ( select schemaname, relname, relid from pg_stat_user_tables where relname = 'test_size'),sizes as ( select schemaname, r.relname, pg_total_relation_size(relid) AS total_bytes, pg_relation_size(relid, 'main') + pg_relation_size(relid, 'init') + pg_relation_size(relid, 'fsm') + pg_relation_size(relid, 'vm') AS heap_bytes, pg_indexes_size(relid) AS index_bytes, pg_table_size(reltoastrelid) AS toast_bytes from relations r join pg_class on pg_class.oid = r.relid)select total_bytes, heap_bytes, index_bytes, toast_bytes, (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \"Equal?\", (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \"Diff\"from sizes; total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff -------------+------------+-------------+-------------+--------+-------- 14000128 | 90112 | 40960 | 13688832 | f | 180224(1 row)I want to calculate separately HEAP, INDEXES and TOAST (including indexes) sizes but it seems it's a bit inconsistent with pg_total_relation_size. Is it correct or am I missing something?Regards,[1] https://www.postgresql.org/docs/current/functions-admin.html#FUNCTIONS-ADMIN-DBSIZE-- Fabrízio de Royes Mello","msg_date":"Fri, 25 Feb 2022 11:58:48 -0300","msg_from":"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>","msg_from_op":true,"msg_subject":"Size functions inconsistent results"},{"msg_contents":"\nOn Fri, 25 Feb 2022 at 22:58, Fabrízio de Royes Mello <fabriziomello@gmail.com> wrote:\n> Hi all,\n>\n> While doing some work using our functions [1] for calculate relations size\n> I noticed an inconsistency between pg_total_relation_size and calculate\n> everything separately, have a look in this example:\n>\n> fabrizio=# create table test_size (id bigserial primary key, toast_column\n> text);\n> CREATE TABLE\n>\n> fabrizio=# insert into test_size (toast_column)\n>\n> select repeat('X'::text, pg_size_bytes('1MB')::integer)\n> from generate_series(1,1000);\n> INSERT 0 1000\n>\n> fabrizio=# with relations as (\n> select schemaname, relname, relid\n> from pg_stat_user_tables\n> where relname = 'test_size'\n> ),\n> sizes as (\n> select\n> schemaname,\n> r.relname,\n>\n> pg_total_relation_size(relid) AS total_bytes,\n>\n> pg_relation_size(relid, 'main') +\n> pg_relation_size(relid, 'init') +\n> pg_relation_size(relid, 'fsm') +\n> pg_relation_size(relid, 'vm') AS heap_bytes,\n> pg_indexes_size(relid) AS index_bytes,\n> pg_table_size(reltoastrelid) AS toast_bytes\n> from relations r\n> join pg_class on pg_class.oid = r.relid\n> )\n> select\n> total_bytes, heap_bytes, index_bytes, toast_bytes,\n> (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \"Equal?\",\n> (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \"Diff\"\n> from sizes;\n> total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff\n> -------------+------------+-------------+-------------+--------+--------\n> 14000128 | 90112 | 40960 | 13688832 | f | 180224\n> (1 row)\n>\n> I want to calculate separately HEAP, INDEXES and TOAST (including indexes)\n> sizes but it seems it's a bit inconsistent with pg_total_relation_size.\n>\n> Is it correct or am I missing something?\n>\n\nI think, you forget the index size of toast table.\n\nwith relations as (\n select schemaname, relname, relid\n from pg_stat_user_tables\n where relname = 'test_size'\n),\nsizes as (\n select\n schemaname,\n r.relname,\n\n pg_total_relation_size(relid) AS total_bytes,\n\n pg_relation_size(relid, 'main') +\n pg_relation_size(relid, 'init') +\n pg_relation_size(relid, 'fsm') +\n pg_relation_size(relid, 'vm') AS heap_bytes,\n pg_indexes_size(relid) AS index_bytes,\n pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS toast_bytes\n from relations r\n join pg_class on pg_class.oid = r.relid\n)\nselect\n total_bytes, heap_bytes, index_bytes, toast_bytes,\n (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \"Equal?\",\n (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \"Diff\"\nfrom sizes;\n\n-- \nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Fri, 25 Feb 2022 23:10:32 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":false,"msg_subject":"Re: Size functions inconsistent results"},{"msg_contents":"On Fri, Feb 25, 2022 at 12:10 PM Japin Li <japinli@hotmail.com> wrote:\n>\n>\n> I think, you forget the index size of toast table.\n>\n> with relations as (\n> select schemaname, relname, relid\n> from pg_stat_user_tables\n> where relname = 'test_size'\n> ),\n> sizes as (\n> select\n> schemaname,\n> r.relname,\n>\n> pg_total_relation_size(relid) AS total_bytes,\n>\n> pg_relation_size(relid, 'main') +\n> pg_relation_size(relid, 'init') +\n> pg_relation_size(relid, 'fsm') +\n> pg_relation_size(relid, 'vm') AS heap_bytes,\n> pg_indexes_size(relid) AS index_bytes,\n> pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS\ntoast_bytes\n> from relations r\n> join pg_class on pg_class.oid = r.relid\n> )\n> select\n> total_bytes, heap_bytes, index_bytes, toast_bytes,\n> (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \"Equal?\",\n> (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \"Diff\"\n> from sizes;\n>\n\nAhh perfect... thanks... make sense because pg_table_size don't compute the\nindexes size, now it worked:\n\nfabrizio=# with relations as (\n select schemaname, relname, relid\n from pg_stat_user_tables\n where relname = 'test_size'\n),\nsizes as (\n select\n schemaname,\n r.relname,\n\n pg_total_relation_size(relid) AS total_bytes,\n\n pg_relation_size(relid, 'main') +\n pg_relation_size(relid, 'init') +\n pg_relation_size(relid, 'fsm') +\n pg_relation_size(relid, 'vm') AS heap_bytes,\n pg_indexes_size(relid) AS index_bytes,\n pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS\ntoast_bytes\n from relations r\n join pg_class on pg_class.oid = r.relid\n)\nselect\n total_bytes, heap_bytes, index_bytes, toast_bytes,\n (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \"Equal?\",\n (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \"Diff\"\nfrom sizes;\n total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff\n-------------+------------+-------------+-------------+--------+------\n 14622720 | 65536 | 40960 | 14516224 | t | 0\n(1 row)\n\nRegards,\n\n--\nFabrízio de Royes Mello\n\nOn Fri, Feb 25, 2022 at 12:10 PM Japin Li <japinli@hotmail.com> wrote:>>> I think, you forget the index size of toast table.>> with relations as (> select schemaname, relname, relid> from pg_stat_user_tables> where relname = 'test_size'> ),> sizes as (> select> schemaname,> r.relname,>> pg_total_relation_size(relid) AS total_bytes,>> pg_relation_size(relid, 'main') +> pg_relation_size(relid, 'init') +> pg_relation_size(relid, 'fsm') +> pg_relation_size(relid, 'vm') AS heap_bytes,> pg_indexes_size(relid) AS index_bytes,> pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS toast_bytes> from relations r> join pg_class on pg_class.oid = r.relid> )> select> total_bytes, heap_bytes, index_bytes, toast_bytes,> (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \"Equal?\",> (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \"Diff\"> from sizes;>Ahh perfect... thanks... make sense because pg_table_size don't compute the indexes size, now it worked:fabrizio=# with relations as ( select schemaname, relname, relid from pg_stat_user_tables where relname = 'test_size'),sizes as ( select schemaname, r.relname, pg_total_relation_size(relid) AS total_bytes, pg_relation_size(relid, 'main') + pg_relation_size(relid, 'init') + pg_relation_size(relid, 'fsm') + pg_relation_size(relid, 'vm') AS heap_bytes, pg_indexes_size(relid) AS index_bytes, pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS toast_bytes from relations r join pg_class on pg_class.oid = r.relid)select total_bytes, heap_bytes, index_bytes, toast_bytes, (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \"Equal?\", (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \"Diff\"from sizes; total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff -------------+------------+-------------+-------------+--------+------ 14622720 | 65536 | 40960 | 14516224 | t | 0(1 row)Regards,--Fabrízio de Royes Mello","msg_date":"Fri, 25 Feb 2022 12:16:48 -0300","msg_from":"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>","msg_from_op":true,"msg_subject":"Re: Size functions inconsistent results"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nWhile doing some work using our functions [1] for calculate relations size\\nI noticed an inconsistency between pg_total_relation_size and calculate\\neverything separately, have a look in this example:\\n\\nfabrizio=# create table test_size (id bigserial primary key, toast_column\\ntext);\\nCREATE TABLE\\n\\nfabrizio=# insert into test_size (toast_column)\\n\\n select repeat('X'::text, pg_size_bytes('1MB')::integer)\\n from generate_series(1,1000);\\nINSERT 0 1000\\n\\nfabrizio=# with relations as (\\n select schemaname, relname, relid\\n from pg_stat_user_tables\\n where relname = 'test_size'\\n),\\nsizes as (\\n select\\n schemaname,\\n r.relname,\\n\\n pg_total_relation_size(relid) AS total_bytes,\\n\\n pg_relation_size(relid, 'main') +\\n pg_relation_size(relid, 'init') +\\n pg_relation_size(relid, 'fsm') +\\n pg_relation_size(relid, 'vm') AS heap_bytes,\\n pg_indexes_size(relid) AS index_bytes,\\n pg_table_size(reltoastrelid) AS toast_bytes\\n from relations r\\n join pg_class on pg_class.oid = r.relid\\n)\\nselect\\n total_bytes, heap_bytes, index_bytes, toast_bytes,\\n (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \\\"Equal?\\\",\\n (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \\\"Diff\\\"\\nfrom sizes;\\n total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff\\n-------------+------------+-------------+-------------+--------+--------\\n 14000128 | 90112 | 40960 | 13688832 | f | 180224\\n(1 row)\\n\\nI want to calculate separately HEAP, INDEXES and TOAST (including indexes)\\nsizes but it seems it's a bit inconsistent with pg_total_relation_size.\\n\\nIs it correct or am I missing something?\\n\\nRegards,\\n\\n[1]\\nhttps://www.postgresql.org/docs/current/functions-admin.html#FUNCTIONS-ADMIN-DBSIZE\\n\\n-- \\nFabrízio de Royes Mello\\n\\nHi all,While doing some work using our functions [1] for calculate relations size I noticed an inconsistency between pg_total_relation_size and calculate everything separately, have a look in this example:fabrizio=# create table test_size (id bigserial primary key, toast_column text);CREATE TABLEfabrizio=# insert into test_size (toast_column) select repeat('X'::text, pg_size_bytes('1MB')::integer) from generate_series(1,1000);INSERT 0 1000fabrizio=# with relations as ( select schemaname, relname, relid from pg_stat_user_tables where relname = 'test_size'),sizes as ( select schemaname, r.relname, pg_total_relation_size(relid) AS total_bytes, pg_relation_size(relid, 'main') + pg_relation_size(relid, 'init') + pg_relation_size(relid, 'fsm') + pg_relation_size(relid, 'vm') AS heap_bytes, pg_indexes_size(relid) AS index_bytes, pg_table_size(reltoastrelid) AS toast_bytes from relations r join pg_class on pg_class.oid = r.relid)select total_bytes, heap_bytes, index_bytes, toast_bytes, (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \\\"Equal?\\\", (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \\\"Diff\\\"from sizes; total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff -------------+------------+-------------+-------------+--------+-------- 14000128 | 90112 | 40960 | 13688832 | f | 180224(1 row)I want to calculate separately HEAP, INDEXES and TOAST (including indexes) sizes but it seems it's a bit inconsistent with pg_total_relation_size. Is it correct or am I missing something?Regards,[1] https://www.postgresql.org/docs/current/functions-admin.html#FUNCTIONS-ADMIN-DBSIZE-- Fabrízio de Royes Mello\",\n \"msg_date\": \"Fri, 25 Feb 2022 11:58:48 -0300\",\n \"msg_from\": \"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Size functions inconsistent results\"\n },\n {\n \"msg_contents\": \"\\nOn Fri, 25 Feb 2022 at 22:58, Fabrízio de Royes Mello <fabriziomello@gmail.com> wrote:\\n> Hi all,\\n>\\n> While doing some work using our functions [1] for calculate relations size\\n> I noticed an inconsistency between pg_total_relation_size and calculate\\n> everything separately, have a look in this example:\\n>\\n> fabrizio=# create table test_size (id bigserial primary key, toast_column\\n> text);\\n> CREATE TABLE\\n>\\n> fabrizio=# insert into test_size (toast_column)\\n>\\n> select repeat('X'::text, pg_size_bytes('1MB')::integer)\\n> from generate_series(1,1000);\\n> INSERT 0 1000\\n>\\n> fabrizio=# with relations as (\\n> select schemaname, relname, relid\\n> from pg_stat_user_tables\\n> where relname = 'test_size'\\n> ),\\n> sizes as (\\n> select\\n> schemaname,\\n> r.relname,\\n>\\n> pg_total_relation_size(relid) AS total_bytes,\\n>\\n> pg_relation_size(relid, 'main') +\\n> pg_relation_size(relid, 'init') +\\n> pg_relation_size(relid, 'fsm') +\\n> pg_relation_size(relid, 'vm') AS heap_bytes,\\n> pg_indexes_size(relid) AS index_bytes,\\n> pg_table_size(reltoastrelid) AS toast_bytes\\n> from relations r\\n> join pg_class on pg_class.oid = r.relid\\n> )\\n> select\\n> total_bytes, heap_bytes, index_bytes, toast_bytes,\\n> (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \\\"Equal?\\\",\\n> (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \\\"Diff\\\"\\n> from sizes;\\n> total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff\\n> -------------+------------+-------------+-------------+--------+--------\\n> 14000128 | 90112 | 40960 | 13688832 | f | 180224\\n> (1 row)\\n>\\n> I want to calculate separately HEAP, INDEXES and TOAST (including indexes)\\n> sizes but it seems it's a bit inconsistent with pg_total_relation_size.\\n>\\n> Is it correct or am I missing something?\\n>\\n\\nI think, you forget the index size of toast table.\\n\\nwith relations as (\\n select schemaname, relname, relid\\n from pg_stat_user_tables\\n where relname = 'test_size'\\n),\\nsizes as (\\n select\\n schemaname,\\n r.relname,\\n\\n pg_total_relation_size(relid) AS total_bytes,\\n\\n pg_relation_size(relid, 'main') +\\n pg_relation_size(relid, 'init') +\\n pg_relation_size(relid, 'fsm') +\\n pg_relation_size(relid, 'vm') AS heap_bytes,\\n pg_indexes_size(relid) AS index_bytes,\\n pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS toast_bytes\\n from relations r\\n join pg_class on pg_class.oid = r.relid\\n)\\nselect\\n total_bytes, heap_bytes, index_bytes, toast_bytes,\\n (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \\\"Equal?\\\",\\n (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \\\"Diff\\\"\\nfrom sizes;\\n\\n-- \\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 23:10:32 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Size functions inconsistent results\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 12:10 PM Japin Li <japinli@hotmail.com> wrote:\\n>\\n>\\n> I think, you forget the index size of toast table.\\n>\\n> with relations as (\\n> select schemaname, relname, relid\\n> from pg_stat_user_tables\\n> where relname = 'test_size'\\n> ),\\n> sizes as (\\n> select\\n> schemaname,\\n> r.relname,\\n>\\n> pg_total_relation_size(relid) AS total_bytes,\\n>\\n> pg_relation_size(relid, 'main') +\\n> pg_relation_size(relid, 'init') +\\n> pg_relation_size(relid, 'fsm') +\\n> pg_relation_size(relid, 'vm') AS heap_bytes,\\n> pg_indexes_size(relid) AS index_bytes,\\n> pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS\\ntoast_bytes\\n> from relations r\\n> join pg_class on pg_class.oid = r.relid\\n> )\\n> select\\n> total_bytes, heap_bytes, index_bytes, toast_bytes,\\n> (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \\\"Equal?\\\",\\n> (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \\\"Diff\\\"\\n> from sizes;\\n>\\n\\nAhh perfect... thanks... make sense because pg_table_size don't compute the\\nindexes size, now it worked:\\n\\nfabrizio=# with relations as (\\n select schemaname, relname, relid\\n from pg_stat_user_tables\\n where relname = 'test_size'\\n),\\nsizes as (\\n select\\n schemaname,\\n r.relname,\\n\\n pg_total_relation_size(relid) AS total_bytes,\\n\\n pg_relation_size(relid, 'main') +\\n pg_relation_size(relid, 'init') +\\n pg_relation_size(relid, 'fsm') +\\n pg_relation_size(relid, 'vm') AS heap_bytes,\\n pg_indexes_size(relid) AS index_bytes,\\n pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS\\ntoast_bytes\\n from relations r\\n join pg_class on pg_class.oid = r.relid\\n)\\nselect\\n total_bytes, heap_bytes, index_bytes, toast_bytes,\\n (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \\\"Equal?\\\",\\n (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \\\"Diff\\\"\\nfrom sizes;\\n total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff\\n-------------+------------+-------------+-------------+--------+------\\n 14622720 | 65536 | 40960 | 14516224 | t | 0\\n(1 row)\\n\\nRegards,\\n\\n--\\nFabrízio de Royes Mello\\n\\nOn Fri, Feb 25, 2022 at 12:10 PM Japin Li <japinli@hotmail.com> wrote:>>> I think, you forget the index size of toast table.>> with relations as (> select schemaname, relname, relid> from pg_stat_user_tables> where relname = 'test_size'> ),> sizes as (> select> schemaname,> r.relname,>> pg_total_relation_size(relid) AS total_bytes,>> pg_relation_size(relid, 'main') +> pg_relation_size(relid, 'init') +> pg_relation_size(relid, 'fsm') +> pg_relation_size(relid, 'vm') AS heap_bytes,> pg_indexes_size(relid) AS index_bytes,> pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS toast_bytes> from relations r> join pg_class on pg_class.oid = r.relid> )> select> total_bytes, heap_bytes, index_bytes, toast_bytes,> (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \\\"Equal?\\\",> (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \\\"Diff\\\"> from sizes;>Ahh perfect... thanks... make sense because pg_table_size don't compute the indexes size, now it worked:fabrizio=# with relations as ( select schemaname, relname, relid from pg_stat_user_tables where relname = 'test_size'),sizes as ( select schemaname, r.relname, pg_total_relation_size(relid) AS total_bytes, pg_relation_size(relid, 'main') + pg_relation_size(relid, 'init') + pg_relation_size(relid, 'fsm') + pg_relation_size(relid, 'vm') AS heap_bytes, pg_indexes_size(relid) AS index_bytes, pg_table_size(reltoastrelid) + pg_indexes_size(reltoastrelid) AS toast_bytes from relations r join pg_class on pg_class.oid = r.relid)select total_bytes, heap_bytes, index_bytes, toast_bytes, (total_bytes = (heap_bytes+index_bytes+toast_bytes)) as \\\"Equal?\\\", (total_bytes - (heap_bytes+index_bytes+toast_bytes)) as \\\"Diff\\\"from sizes; total_bytes | heap_bytes | index_bytes | toast_bytes | Equal? | Diff -------------+------------+-------------+-------------+--------+------ 14622720 | 65536 | 40960 | 14516224 | t | 0(1 row)Regards,--Fabrízio de Royes Mello\",\n \"msg_date\": \"Fri, 25 Feb 2022 12:16:48 -0300\",\n \"msg_from\": \"=?UTF-8?Q?Fabr=C3=ADzio_de_Royes_Mello?= <fabriziomello@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Size functions inconsistent results\"\n }\n]","isConversational":false}}},{"rowIdx":53583,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nHere's a tiny patch to do $subject.\n\nRegards,\nBharath Rupireddy.","msg_date":"Fri, 25 Feb 2022 20:39:22 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Fix typo in logicalfuncs.c - :%s/private date/Private data"},{"msg_contents":"On Fri, Feb 25, 2022 at 8:39 PM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> Hi,\n>\n> Here's a tiny patch to do $subject.\n>\n\nLGTM. I'll push this in some time.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Sat, 26 Feb 2022 10:14:45 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Fix typo in logicalfuncs.c - :%s/private date/Private data"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nHere's a tiny patch to do $subject.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Fri, 25 Feb 2022 20:39:22 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Fix typo in logicalfuncs.c - :%s/private date/Private data\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 8:39 PM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> Hi,\\n>\\n> Here's a tiny patch to do $subject.\\n>\\n\\nLGTM. I'll push this in some time.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 10:14:45 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Fix typo in logicalfuncs.c - :%s/private date/Private data\"\n }\n]","isConversational":false}}},{"rowIdx":53584,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"This patch adds a configuration parameter jit_warn_above_fraction that\nwill cause a warning to be logged if the fraction of time spent on\ndoing JIT is bigger than the specified one. For example, this can be\nused to track down those cases where JIT ends up taking 90% of the\nquery runtime because of bad estimates...\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/","msg_date":"Fri, 25 Feb 2022 16:16:01 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Add parameter jit_warn_above_fraction"},{"msg_contents":"Hi,\n\nOn Fri, Feb 25, 2022 at 04:16:01PM +0100, Magnus Hagander wrote:\n> This patch adds a configuration parameter jit_warn_above_fraction that\n> will cause a warning to be logged if the fraction of time spent on\n> doing JIT is bigger than the specified one. For example, this can be\n> used to track down those cases where JIT ends up taking 90% of the\n> query runtime because of bad estimates...\n\nI think that's tremendously useful, huge +1.\n\nJust a few minor nit:\n\n+ A value of 0 (the default)disables the warning.\n\nmissing space\n\n+\t\t\tereport(WARNING,\n+\t\t\t\t\t(errmsg(\"JIT time was %ld ms of %d ms\",\n+\t\t\t\t\t\t\tjit_time, msecs)));\n\n\"JIT time\" may a bit obscure for users, how about \"JIT total processing time\"?\"\n\n+\t\t\tgettext_noop(\"Sets the fraction of query time spent on JIT before writing\"\n+\t\t\t\t\t\t \"a warning to the log.\"),\n+\t\t\tgettext_noop(\"Write a message tot he server log if more than this\"\n+\t\t\t\t\t\t \"fraction of the query runtime is spent on JIT.\"\n+\t\t\t\t\t\t \"Zero turns off the warning.\")\n\nmissing spaces in the concatenated strings.\n\nThe rest of the patch looks good to me.\n\n\n","msg_date":"Sat, 26 Feb 2022 00:01:41 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Hi,\n\nOn 2022-02-25 16:16:01 +0100, Magnus Hagander wrote:\n> This patch adds a configuration parameter jit_warn_above_fraction that\n> will cause a warning to be logged if the fraction of time spent on\n> doing JIT is bigger than the specified one. For example, this can be\n> used to track down those cases where JIT ends up taking 90% of the\n> query runtime because of bad estimates...\n\nHm. Could it make sense to do this as a auto_explain feature?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 25 Feb 2022 08:20:43 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Fri, Feb 25, 2022 at 04:16:01PM +0100, Magnus Hagander wrote:\n> +\t{\n> +\t\t{\"jit_warn_above_fraction\", PGC_SUSET, LOGGING_WHEN,\n> +\t\t\tgettext_noop(\"Sets the fraction of query time spent on JIT before writing\"\n> +\t\t\t\t\t\t \"a warning to the log.\"),\n> +\t\t\tgettext_noop(\"Write a message tot he server log if more than this\"\n> +\t\t\t\t\t\t \"fraction of the query runtime is spent on JIT.\"\n> +\t\t\t\t\t\t \"Zero turns off the warning.\")\n> +\t\t},\n> +\t\t&jit_warn_above_fraction,\n> +\t\t0.0, 0.0, 1.0,\n> +\t\tNULL, NULL, NULL\n> +\t},\n\nShould be PGC_USERSET ?\n\n+ gettext_noop(\"Write a message tot he server log if more than this\" \n\nto the\n\n+ if (jit_enabled && jit_warn_above_fraction > 0) \n+ { \n+ int64 jit_time = \n+ INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.generation_counter) + \n+ INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.inlining_counter) + \n+ INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.optimization_counter) + \n+ INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.emission_counter); \n+ \n+ if (jit_time > msecs * jit_warn_above_fraction) \n+ { \n+ ereport(WARNING, \n+ (errmsg(\"JIT time was %ld ms of %d ms\", \n+ jit_time, msecs))); \n+ } \n+ } \n\n\nI think it should be a NOTICE (or less?)\n\nIs it more useful if this is applied combined with log_min_duration_statement ?\n\nIt's easy to imagine a query for which the planner computes a high cost, but\nactually runs quickly. You might get a bunch of WARNINGs that the query took\n10 MS and JIT was 75% of that, even if you don't care about queries that take\nless than 10 SEC.\n\nI should say that this is already available by processing the output of\nautoexplain.\n\n-- \nJustin\n\n\n","msg_date":"Fri, 25 Feb 2022 10:23:04 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Fri, Feb 25, 2022 at 5:20 PM Andres Freund <andres@anarazel.de> wrote:\n>\n> Hi,\n>\n> On 2022-02-25 16:16:01 +0100, Magnus Hagander wrote:\n> > This patch adds a configuration parameter jit_warn_above_fraction that\n> > will cause a warning to be logged if the fraction of time spent on\n> > doing JIT is bigger than the specified one. For example, this can be\n> > used to track down those cases where JIT ends up taking 90% of the\n> > query runtime because of bad estimates...\n>\n> Hm. Could it make sense to do this as a auto_explain feature?\n\nIt could be. But I was looking for something a lot more \"light weight\"\nthan having to install an extension. But yes, if we wanted to, we\ncould certainly change jit_warn_above_fraction to be\nauto_explain.log_min_jit_fraction or something like that, and do\nbasically the same thing. But then, we could also have\nlog_min_duration_statement be part of auto_explain instead, so it's\nall about where to draw the line :)\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/\n\n\n","msg_date":"Fri, 25 Feb 2022 17:28:41 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Hi,\n\nOn 2022-02-25 17:28:41 +0100, Magnus Hagander wrote:\n> On Fri, Feb 25, 2022 at 5:20 PM Andres Freund <andres@anarazel.de> wrote:\n> > On 2022-02-25 16:16:01 +0100, Magnus Hagander wrote:\n> > > This patch adds a configuration parameter jit_warn_above_fraction that\n> > > will cause a warning to be logged if the fraction of time spent on\n> > > doing JIT is bigger than the specified one. For example, this can be\n> > > used to track down those cases where JIT ends up taking 90% of the\n> > > query runtime because of bad estimates...\n> >\n> > Hm. Could it make sense to do this as a auto_explain feature?\n> \n> It could be. But I was looking for something a lot more \"light weight\"\n> than having to install an extension. But yes, if we wanted to, we\n> could certainly change jit_warn_above_fraction to be\n> auto_explain.log_min_jit_fraction or something like that, and do\n> basically the same thing. But then, we could also have\n> log_min_duration_statement be part of auto_explain instead, so it's\n> all about where to draw the line :)\n\nI guess it feels a tad on the \"too narrow/specific\" side of things for the\ngeneral code. We don't have log_min_duration_{parsing,planning,execution}\neither. But I also get it. So I just wanted to raise it ;)\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 25 Feb 2022 08:47:22 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Fri, Feb 25, 2022 at 5:47 PM Andres Freund <andres@anarazel.de> wrote:\n>\n> Hi,\n>\n> On 2022-02-25 17:28:41 +0100, Magnus Hagander wrote:\n> > On Fri, Feb 25, 2022 at 5:20 PM Andres Freund <andres@anarazel.de> wrote:\n> > > On 2022-02-25 16:16:01 +0100, Magnus Hagander wrote:\n> > > > This patch adds a configuration parameter jit_warn_above_fraction that\n> > > > will cause a warning to be logged if the fraction of time spent on\n> > > > doing JIT is bigger than the specified one. For example, this can be\n> > > > used to track down those cases where JIT ends up taking 90% of the\n> > > > query runtime because of bad estimates...\n> > >\n> > > Hm. Could it make sense to do this as a auto_explain feature?\n> >\n> > It could be. But I was looking for something a lot more \"light weight\"\n> > than having to install an extension. But yes, if we wanted to, we\n> > could certainly change jit_warn_above_fraction to be\n> > auto_explain.log_min_jit_fraction or something like that, and do\n> > basically the same thing. But then, we could also have\n> > log_min_duration_statement be part of auto_explain instead, so it's\n> > all about where to draw the line :)\n>\n> I guess it feels a tad on the \"too narrow/specific\" side of things for the\n> general code. We don't have log_min_duration_{parsing,planning,execution}\n> either. But I also get it. So I just wanted to raise it ;)\n\nOh it's absolutely a valid issue to raise :) In fact, one could\ndefinitely argue that we should have a parameter for auto_explain *as\nwell*.\n\nIt's true we don't have those -- but it's also in my experience a lot\nmore common to be an issue with JIT. And unfortunately the current\n\"solution\" people tend to apply is to globally disable JIT, and\nthere's no similar \"globally disable parsing or \"globally disable\nplanning\" pattern, for obvious reasons.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/\n\n\n","msg_date":"Mon, 7 Mar 2022 12:56:52 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Fri, Feb 25, 2022 at 5:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> On Fri, Feb 25, 2022 at 04:16:01PM +0100, Magnus Hagander wrote:\n> > + {\n> > + {\"jit_warn_above_fraction\", PGC_SUSET, LOGGING_WHEN,\n> > + gettext_noop(\"Sets the fraction of query time spent on JIT before writing\"\n> > + \"a warning to the log.\"),\n> > + gettext_noop(\"Write a message tot he server log if more than this\"\n> > + \"fraction of the query runtime is spent on JIT.\"\n> > + \"Zero turns off the warning.\")\n> > + },\n> > + &jit_warn_above_fraction,\n> > + 0.0, 0.0, 1.0,\n> > + NULL, NULL, NULL\n> > + },\n>\n> Should be PGC_USERSET ?\n\nYes. Definitely. Copy/paste thinko.\n\n\n> + gettext_noop(\"Write a message tot he server log if more than this\"\n>\n> to the\n\nYup.\n\n\n> + if (jit_enabled && jit_warn_above_fraction > 0)\n> + {\n> + int64 jit_time =\n> + INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.generation_counter) +\n> + INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.inlining_counter) +\n> + INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.optimization_counter) +\n> + INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.emission_counter);\n> +\n> + if (jit_time > msecs * jit_warn_above_fraction)\n> + {\n> + ereport(WARNING,\n> + (errmsg(\"JIT time was %ld ms of %d ms\",\n> + jit_time, msecs)));\n> + }\n> + }\n>\n>\n> I think it should be a NOTICE (or less?)\n\nHmm. I'm not so sure.\n\nOther similar parameters often use LOG, but the downside of that is\nthat it won't be sent to the client.\n\nThe problem with using NOTICE is that it won't go to the log by\ndefault. It needs to be at least warning to do that.\n\n\n> Is it more useful if this is applied combined with log_min_duration_statement ?\n>\n> It's easy to imagine a query for which the planner computes a high cost, but\n> actually runs quickly. You might get a bunch of WARNINGs that the query took\n> 10 MS and JIT was 75% of that, even if you don't care about queries that take\n> less than 10 SEC.\n\nYeah, that's definitely a problem. But which way would you want to tie\nit to log_min_duration_statement? That a statement would both have to\ntake longer than log_min_duration_statement *and* have JIT above a\ncertain percentage? In my experience that is instead likely to miss\nmost of the interesting times. Maybe it would need a separate guc for\nthe timing, but I'd like to avoid that, I'm not sure it's a function\nworth *that* much...\n\n\n> I should say that this is already available by processing the output of\n> autoexplain.\n\nTrue. You just can't trigger it based on it. (and it can be somewhat\nof a PITA to parse things out of auto_explain on busy systems, but it\ndoes give a lot of very useful details)\n\n\nMeanwhile here is an updated based on your other comments above, as\nwell as those from Julien.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/","msg_date":"Mon, 7 Mar 2022 13:10:32 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Mon, Mar 07, 2022 at 01:10:32PM +0100, Magnus Hagander wrote:\n> On Fri, Feb 25, 2022 at 5:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> >\n> > I think it should be a NOTICE (or less?)\n> \n> Hmm. I'm not so sure.\n> \n> Other similar parameters often use LOG, but the downside of that is\n> that it won't be sent to the client.\n> \n> The problem with using NOTICE is that it won't go to the log by\n> default. It needs to be at least warning to do that.\n\nAnyone who uses this parameter is aleady going to be changing GUCs, so it\ndoesn't need to work by default. The people who are likely to enable this\nalready monitor their logs and have probably customized their logging\nconfiguration.\n\n> > Is it more useful if this is applied combined with log_min_duration_statement ?\n> >\n> > It's easy to imagine a query for which the planner computes a high cost, but\n> > actually runs quickly. You might get a bunch of WARNINGs that the query took\n> > 10 MS and JIT was 75% of that, even if you don't care about queries that take\n> > less than 10 SEC.\n> \n> Yeah, that's definitely a problem. But which way would you want to tie\n> it to log_min_duration_statement? That a statement would both have to\n> take longer than log_min_duration_statement *and* have JIT above a\n> certain percentage? In my experience that is instead likely to miss\n> most of the interesting times.\n\nI don't understand - why doesn't it combine trivially with\nlog_min_duration_statement? Are you saying that the default / pre-existing min\nduration may not log all of the intended queries ? I think that just means the\nconfiguration needs to be changed. The GUC needs to allow/help finding these\nJIT issues, but going to require an admin's interaction in any case. Avoiding\nfalse positives may be important for it to be useful at all.\n\nHmm .. should it also combine with min_sample_rate ? Maybe that addresses your\nconcern.\n\n-- \nJustin\n\n\n","msg_date":"Mon, 7 Mar 2022 07:09:22 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Mon, Mar 7, 2022 at 2:09 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n>\n> On Mon, Mar 07, 2022 at 01:10:32PM +0100, Magnus Hagander wrote:\n> > On Fri, Feb 25, 2022 at 5:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > >\n> > > I think it should be a NOTICE (or less?)\n> >\n> > Hmm. I'm not so sure.\n> >\n> > Other similar parameters often use LOG, but the downside of that is\n> > that it won't be sent to the client.\n> >\n> > The problem with using NOTICE is that it won't go to the log by\n> > default. It needs to be at least warning to do that.\n>\n> Anyone who uses this parameter is aleady going to be changing GUCs, so it\n> doesn't need to work by default. The people who are likely to enable this\n> already monitor their logs and have probably customized their logging\n> configuration.\n\nSure, but if you set log_min_messagse to NOTICE you are likely going\nto flood your logs beyond usefulness. And the whole idea of this\nparameter is you can keep it on during \"normal times\" to catch\noutliers.\n\n\n> > > Is it more useful if this is applied combined with log_min_duration_statement ?\n> > >\n> > > It's easy to imagine a query for which the planner computes a high cost, but\n> > > actually runs quickly. You might get a bunch of WARNINGs that the query took\n> > > 10 MS and JIT was 75% of that, even if you don't care about queries that take\n> > > less than 10 SEC.\n> >\n> > Yeah, that's definitely a problem. But which way would you want to tie\n> > it to log_min_duration_statement? That a statement would both have to\n> > take longer than log_min_duration_statement *and* have JIT above a\n> > certain percentage? In my experience that is instead likely to miss\n> > most of the interesting times.\n>\n> I don't understand - why doesn't it combine trivially with\n> log_min_duration_statement? Are you saying that the default / pre-existing min\n> duration may not log all of the intended queries ? I think that just means the\n> configuration needs to be changed. The GUC needs to allow/help finding these\n> JIT issues, but going to require an admin's interaction in any case. Avoiding\n> false positives may be important for it to be useful at all.\n>\n> Hmm .. should it also combine with min_sample_rate ? Maybe that addresses your\n> concern.\n\nFor one, what if I don't want to log any queries unless this is the\ncase? I leave log_min_duration_statement at -1...\nOr what if I want to log say only queries taking more than 60 seconds.\nNow I will miss all queries taking 30 seconds where 99.9% of the time\nis spent on JIT which I definitely *would* want.\n\nIf all I wanted was to get the JIT information for all queries over a\ncertain amount of time, then as someone else mentioned as well, I can\njust use auto_explain. The point here is I want to trigger the logging\nfor queries where the runtime is *shorter* than what I have in\nlog_min_duration_statement, but a large portion of the runtime is\nspent on JIT.\n\nI maen. If I have a bunch of queries that take 10ms and 75% is spent\nin JIT (per your example above), that is actually a problem and I'd\nlike to know about it and fix it (by tuning the jit triggers either\nglobally or locally). Yes, at 10ms I *may* ignore it, and in general\nI'd use pg_stat_statements for that. But at 100ms with 75% JIT it\nstarts becoming exactly the situation that I intended this patch to\nhelp with.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/\n Work: https://www.redpill-linpro.com/\n\n\n","msg_date":"Mon, 7 Mar 2022 16:02:16 +0100","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Mon, Mar 07, 2022 at 04:02:16PM +0100, Magnus Hagander wrote:\n> On Mon, Mar 7, 2022 at 2:09 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > On Mon, Mar 07, 2022 at 01:10:32PM +0100, Magnus Hagander wrote:\n> > > On Fri, Feb 25, 2022 at 5:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\n> > > >\n> > > > I think it should be a NOTICE (or less?)\n> > >\n> > > Hmm. I'm not so sure.\n> > >\n> > > Other similar parameters often use LOG, but the downside of that is\n> > > that it won't be sent to the client.\n> > >\n> > > The problem with using NOTICE is that it won't go to the log by\n> > > default. It needs to be at least warning to do that.\n> >\n> > Anyone who uses this parameter is aleady going to be changing GUCs, so it\n> > doesn't need to work by default. The people who are likely to enable this\n> > already monitor their logs and have probably customized their logging\n> > configuration.\n> \n> Sure, but if you set log_min_messagse to NOTICE you are likely going\n> to flood your logs beyond usefulness. And the whole idea of this\n> parameter is you can keep it on during \"normal times\" to catch\n> outliers.\n\nI do set log_min_messages - not to NOTICE, but to INFO ;)\n\nWould NOTICEs really flood most people's logs ? From what ? They're aren't\nthat many, and most seem to be for DDL and utility commands. We do, uh, more\nof that than most people would say is reasonable.\n\nI see a couple hundred of these per day:\n| PostGIS: Unable to compute statistics for...\nAnd during deployment, a few hundred more for IF NOT EXIST commands.\nThat's it.\n\nI still think this GUC should not be WARNING. If it's a warning, then *this*\nwill cause previously-quiet logs to be flooded - the opposite problem. Which\nis not desirable for a new guc.\n\nhttps://www.postgresql.org/docs/current/runtime-config-logging.html\n\nINFO \tProvides information implicitly requested by the user, e.g., output from VACUUM VERBOSE. \tINFO \tINFORMATION\nNOTICE \tProvides information that might be helpful to users, e.g., notice of truncation of long identifiers. \tNOTICE \tINFORMATION\nWARNING Provides warnings of likely problems, e.g., COMMIT outside a transaction block. \tNOTICE \tWARNING\n\n> > I don't understand - why doesn't it combine trivially with\n> > log_min_duration_statement? Are you saying that the default / pre-existing min\n> > duration may not log all of the intended queries ? I think that just means the\n> > configuration needs to be changed. The GUC needs to allow/help finding these\n> > JIT issues, but going to require an admin's interaction in any case. Avoiding\n> > false positives may be important for it to be useful at all.\n> >\n> > Hmm .. should it also combine with min_sample_rate ? Maybe that addresses your\n> > concern.\n> \n> For one, what if I don't want to log any queries unless this is the\n> case? I leave log_min_duration_statement at -1...\n\nThen you will conclude that our logging is inadequate for your case.\nYou need to filter the lines which don't interest you.\n\n> Or what if I want to log say only queries taking more than 60 seconds.\n> Now I will miss all queries taking 30 seconds where 99.9% of the time\n> is spent on JIT which I definitely *would* want.\n\nIf you're unwilling to deal with log entries which are uninteresting, then\nclearly you'll need a separate GUC just for log_min_duration_jit_above_fraction\n(and maybe another one for jit_above_duration_log_level). Or implement generic\nlog filtering based on file/severity/message.\n\nAs I see it: most people won't set this GUC at all. Those who do might enable\nit with a high threshold value of log_min_duration_statement (probably a\npre-existing, configured value) to see what there is to see - the worst issues.\n\nA small fraction of people will enable it with a low threshold value for\nlog_min_duration_statement - maybe only temporarily or per-session. They'll be\nwilling to sift through the logs to look for interesting things, like queries\nthat take 10ms, 75% of which is in JIT, but run hundreds of times per second.\n\nThis feature intends to make it easier to identify queries affected by this\nproblem, but it doesn't need to be possible to do that without logging anything\nelse.\n\n-- \nJustin\n\n\n","msg_date":"Sat, 12 Mar 2022 09:24:11 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Hi,\n\nOn 2022-03-07 13:10:32 +0100, Magnus Hagander wrote:\n> Meanwhile here is an updated based on your other comments above, as\n> well as those from Julien.\n\nThis fails on cfbot, due to compiler warnings:\nhttps://cirrus-ci.com/task/5127667648299008?logs=mingw_cross_warning#L390\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 21 Mar 2022 16:50:48 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Tue, Mar 22, 2022 at 12:50 AM Andres Freund <andres@anarazel.de> wrote:\n\n> Hi,\n>\n> On 2022-03-07 13:10:32 +0100, Magnus Hagander wrote:\n> > Meanwhile here is an updated based on your other comments above, as\n> > well as those from Julien.\n>\n> This fails on cfbot, due to compiler warnings:\n> https://cirrus-ci.com/task/5127667648299008?logs=mingw_cross_warning#L390\n\n\nHuh. That's annoying. I forgot int64 is %d on linux and %lld on Windows :/\n\nPFA a fix for that.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>","msg_date":"Mon, 28 Mar 2022 22:11:16 +0200","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Hi,\n\nOn Mon, Mar 28, 2022 at 10:11:16PM +0200, Magnus Hagander wrote:\n> On Tue, Mar 22, 2022 at 12:50 AM Andres Freund <andres@anarazel.de> wrote:\n> >\n> > This fails on cfbot, due to compiler warnings:\n> > https://cirrus-ci.com/task/5127667648299008?logs=mingw_cross_warning#L390\n>\n>\n> Huh. That's annoying. I forgot int64 is %d on linux and %lld on Windows :/\n>\n> PFA a fix for that.\n\nThe v3 is now stable on all platforms according to the cfbot, and apart from\nthe log level the patch looks good to me.\n\nThe last remaining thing is whether logging at WARNING level is the correct\nchoice. I'm personally fine with it, because the people who are going to use\nit will probably use the same approach as for log_min_duration_statements:\nenable it first with a high value, and check if that just lead to a massive log\nspam. If not, see if there's anything that needs attention and fix it,\notherwise keep lowering it and keep going.\n\nAlso, jitting isn't particularly fast in general, so it's probably somewhat\nunlikely to have a massive amount of queries that both have a cost higher than\nthe jit threshold cost and still run fast enough to spam the log without jit\nslowing everything down.\n\nI'm not sure what status the patch should be. I think it's the committer's\ndecision to chose which log level to use, and that won't affect the patch\nenough to justify another review, so I'm switching the patch to ready for\ncommitter. This doesn't prevent any further discussions for the log level.\n\n\n","msg_date":"Tue, 29 Mar 2022 18:08:46 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Tue, Mar 29, 2022 at 6:09 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> The last remaining thing is whether logging at WARNING level is the correct\n> choice. I'm personally fine with it, because the people who are going to use\n> it will probably use the same approach as for log_min_duration_statements:\n> enable it first with a high value, and check if that just lead to a massive log\n> spam. If not, see if there's anything that needs attention and fix it,\n> otherwise keep lowering it and keep going.\n\nI think WARNING is fine. After all, the parameter is called\n\"jit_warn_above_fraction\". Yeah, we could also rename the parameter,\nbut I think \"jit_notice_above_fraction\" would be harder to understand.\nIt feels very intuitive to just say \"warn me if X thing happens\" and I\ndon't see a reason why we shouldn't just do that.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 29 Mar 2022 09:38:29 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\n> I think WARNING is fine. After all, the parameter is called\n> \"jit_warn_above_fraction\".\n\nI had a think about this patch. I guess it's a little similar to\ncheckpoint_warning. The good thing about the checkpoint_warning is\nthat in the LOG message we give instructions about how the DBA can fix\nthe issue, i.e increase max_wal_size.\n\nWith the proposed patch I see there is no hint about what might be\ndone to remove/reduce the warnings. I imagine that's because it's not\nall that clear which GUC should be changed. In my view, likely\njit_above_cost is the most relevant but there is also\njit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\nand jit_expressions which are relevant too.\n\nIf we go with this patch, the problem I see here is that the amount\nof work the JIT compiler must do for a given query depends mostly on\nthe number of expressions that must be compiled in the query (also to\na lesser extent jit_inline_above_cost, jit_optimize_above_cost,\njit_tuple_deforming and jit_expressions). The DBA does not really have\nmuch control over the number of expressions in the query. All he or\nshe can do to get rid of the warning is something like increase\njit_above_cost. After a few iterations of that, the end result is\nthat jit_above_cost is now high enough that JIT no longer triggers\nfor, say, that query to that table with 1000 partitions where no\nplan-time pruning takes place. Is that really a good thing? It likely\nmeans that we just rarely JIT anything at all!\n\nI really believe that the main problem here is that JIT only enables\nwhen the *total* plan cost reaches a certain threshold. The number of\nexpressions to be compiled is not a factor in the decision at all.\nThat means that even if the total execution time of a plan was a true\nreflection of the total estimated plan cost, then the fraction of time\nspent (as is measured by jit_warn_above_fraction) doing JIT would\nentirely depend on the number of expressions to compile. Of course,\nthe planner's not that good, but does that not indicate that the JIT\ncosting should really account for the number of expressions and not\njust the total plan cost?\n\nAnyway, what I'm trying to indicate here is that JIT is pretty much\nimpossible to tune properly and I don't really see why adding a\nwarning about it not being tuned correctly would help anyone. I think\nit would be better to focus on making improvements to how the JIT\ncosting works.\n\nI did propose a patch to address this in [1]. It does need more work\nand I do plan to come back to it for v16.\n\nI'd much rather see us address the costing problem before adding some\nwarning, especially a warning where it's not clear how to make go\naway.\n\nDavid\n\n[1] https://www.postgresql.org/message-id/CAApHDvpQJqLrNOSi8P1JLM8YE2C+ksKFpSdZg=q6sTbtQ-v=aw@mail.gmail.com\n\n\n","msg_date":"Wed, 30 Mar 2022 09:05:44 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Tue, Mar 29, 2022 at 1:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\n> That means that even if the total execution time of a plan was a true\n> reflection of the total estimated plan cost, then the fraction of time\n> spent (as is measured by jit_warn_above_fraction) doing JIT would\n> entirely depend on the number of expressions to compile. Of course,\n> the planner's not that good, but does that not indicate that the JIT\n> costing should really account for the number of expressions and not\n> just the total plan cost?\n\nThat's a good point. The difference between the actual cost of\nexecuting the query with and without JIT'ing is what we care about,\nfor the most part. Maybe we could do a lot better just by inventing a\nfairly crude model that captures the benefits of JIT'ing -- that's\nwhat \"counting the number of expressions\" sounds like to me. This\nmodel could probably assume that JIT'ing itself was free -- maybe\nsomething this simple would work well.\n\nThe planner has traditionally used the cost units to determine the\ncheapest plan; it compared total plan cost for plans that were taken\nfrom the universe of possible plans for *one specific query*. That's\ncompletely different to any model that expects plan costs to be\nmeaningful in an absolute sense. I'm not completely sure how much that\ndifference matters, but I suspect that the answer is: \"it depends, but\noften it matters a great deal\".\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Tue, 29 Mar 2022 13:35:53 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Peter Geoghegan <pg@bowt.ie> writes:\n> On Tue, Mar 29, 2022 at 1:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\n>> That means that even if the total execution time of a plan was a true\n>> reflection of the total estimated plan cost, then the fraction of time\n>> spent (as is measured by jit_warn_above_fraction) doing JIT would\n>> entirely depend on the number of expressions to compile. Of course,\n>> the planner's not that good, but does that not indicate that the JIT\n>> costing should really account for the number of expressions and not\n>> just the total plan cost?\n\n> That's a good point.\n\nI think David's questions are sufficiently cogent and difficult\nthat we should not add jit_warn_above_fraction at this time.\nMaybe we'll eventually decide it's the best we can do; but I do\nnot think we have a problem there that's so pressing that we need\nto rush out a partially-baked bandaid. Especially not at the\ntail end of the development cycle, with not much time to gain\nexperience with it before we ship.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 29 Mar 2022 18:04:32 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Tue, Mar 29, 2022 at 3:04 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> I think David's questions are sufficiently cogent and difficult\n> that we should not add jit_warn_above_fraction at this time.\n\n+1\n\n-- \nPeter Geoghegan\n\n\n","msg_date":"Tue, 29 Mar 2022 15:19:58 -0700","msg_from":"Peter Geoghegan <pg@bowt.ie>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Hi,\n\nOn 2022-03-30 09:05:44 +1300, David Rowley wrote:\n> I really believe that the main problem here is that JIT only enables\n> when the *total* plan cost reaches a certain threshold.\n\nYes, that is/was a clear design mistake. It wasn't quite as bad back when it\nwas written - partitioning blows the problem up by an order of magnitude or\nthree. . The costing really needs to at least multiply the number of\nto-be-compiled expressions with some cost to decide whether the cost of JITing\nis worth it, rather than making \"flat\" decision.\n\n\n> I did propose a patch to address this in [1]. It does need more work\n> and I do plan to come back to it for v16.\n\nFWIW, that doesn't seem quite right - won't it stop JITing e.g. on the inner\nside of a nested loop, just because it's cheap, even though that's where the\nbulk of the benefits comes from?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 29 Mar 2022 16:16:41 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Wed, 30 Mar 2022 at 12:16, Andres Freund <andres@anarazel.de> wrote:\n> > I did propose a patch to address this in [1]. It does need more work\n> > and I do plan to come back to it for v16.\n>\n> FWIW, that doesn't seem quite right - won't it stop JITing e.g. on the inner\n> side of a nested loop, just because it's cheap, even though that's where the\n> bulk of the benefits comes from?\n\nYeah, I think the total cost would need to be multiplied by the number\nof times we expect to call that part of the plan. I've not yet sat\ndown to figure out if that's easy/cheap or hard/costly information to\nobtain.\n\nDavid\n\n\n","msg_date":"Wed, 30 Mar 2022 12:41:41 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"David Rowley <dgrowleyml@gmail.com> writes:\n> On Wed, 30 Mar 2022 at 12:16, Andres Freund <andres@anarazel.de> wrote:\n>> FWIW, that doesn't seem quite right - won't it stop JITing e.g. on the inner\n>> side of a nested loop, just because it's cheap, even though that's where the\n>> bulk of the benefits comes from?\n\n> Yeah, I think the total cost would need to be multiplied by the number\n> of times we expect to call that part of the plan. I've not yet sat\n> down to figure out if that's easy/cheap or hard/costly information to\n> obtain.\n\nAs long as you don't need the info till the end of planning, it should be\nreasonably simple to determine. I'm not sure if we actually record the\nexpected number of loops in the plan tree today, but the costing\nmechanisms certainly estimate that along the way, so we could store it\nif need be.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Tue, 29 Mar 2022 19:48:21 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Hi,\n\nOn 2022-03-30 12:41:41 +1300, David Rowley wrote:\n> On Wed, 30 Mar 2022 at 12:16, Andres Freund <andres@anarazel.de> wrote:\n> > > I did propose a patch to address this in [1]. It does need more work\n> > > and I do plan to come back to it for v16.\n> >\n> > FWIW, that doesn't seem quite right - won't it stop JITing e.g. on the inner\n> > side of a nested loop, just because it's cheap, even though that's where the\n> > bulk of the benefits comes from?\n>\n> Yeah, I think the total cost would need to be multiplied by the number\n> of times we expect to call that part of the plan. I've not yet sat\n> down to figure out if that's easy/cheap or hard/costly information to\n> obtain.\n\nI wonder whether it'd make sense to combine that with awareness of a few plan\ntypes that can lead to large portions of child nodes never being executed. One\nthe case where the current behaviour is the worst is runtime partition pruning\nin append - we compile expressions for whole subtrees that will never be\nexecuted. We should be much more hesitant to compile there compared to a\ncheap-ish node that we know will be executed as part of a large expensive part\nof the plan tree.\n\n- Andres\n\n\n","msg_date":"Tue, 29 Mar 2022 17:20:30 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Wed, 30 Mar 2022 at 13:20, Andres Freund <andres@anarazel.de> wrote:\n> I wonder whether it'd make sense to combine that with awareness of a few plan\n> types that can lead to large portions of child nodes never being executed. One\n> the case where the current behaviour is the worst is runtime partition pruning\n> in append - we compile expressions for whole subtrees that will never be\n> executed. We should be much more hesitant to compile there compared to a\n> cheap-ish node that we know will be executed as part of a large expensive part\n> of the plan tree.\n\nI think that's also a problem but I think that might be better fixed\nanother way.\n\nThere is a patch [1] around that seems to change things to compile JIT\non-demand. I've not looked at the patch but imagine the overhead\nmight be kept minimal by initially setting the evalfunc to compile and\nrun, then set it to just run the compiled Expr for subsequent\nexecutions. Maybe nodes below an Append/MergeAppend with run-time\npruning could compile on-demand and other nodes up-front. Or maybe\nthere's no problem with making everything on-demand.\n\nDavid\n\n[1] https://commitfest.postgresql.org/37/3071/\n\n\n","msg_date":"Wed, 30 Mar 2022 14:30:32 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Hi,\n\nOn 2022-03-30 14:30:32 +1300, David Rowley wrote:\n> On Wed, 30 Mar 2022 at 13:20, Andres Freund <andres@anarazel.de> wrote:\n> > I wonder whether it'd make sense to combine that with awareness of a few plan\n> > types that can lead to large portions of child nodes never being executed. One\n> > the case where the current behaviour is the worst is runtime partition pruning\n> > in append - we compile expressions for whole subtrees that will never be\n> > executed. We should be much more hesitant to compile there compared to a\n> > cheap-ish node that we know will be executed as part of a large expensive part\n> > of the plan tree.\n> \n> I think that's also a problem but I think that might be better fixed\n> another way.\n> \n> There is a patch [1] around that seems to change things to compile JIT\n> on-demand.\n\nThat's a bad idea idea to do on a per-function basis. Emitting functions one\nby one is considerably slower than doing so in larger chunks. Of course that\ncan be addressed to some degree by something like what you suggest below.\n\n\n> I've not looked at the patch but imagine the overhead might be kept minimal\n> by initially setting the evalfunc to compile and run, then set it to just\n> run the compiled Expr for subsequent executions.\n\nThat part I'm not worried about, there's such an indirection on the first call\neither way IIRC.\n\n\n> Maybe nodes below an Append/MergeAppend with run-time pruning could compile\n> on-demand and other nodes up-front. Or maybe there's no problem with making\n> everything on-demand.\n\nYea, that could work. The expressions for one \"partition query\" would still\nhave to be emitted at once. For each such subtree we should make a separate\ncosting decision. But I think an additional \"will be executed\" sub-node is a\ndifferent story, the threshold shouldn't be done on a per-node basis. That\npartitioning of the plan tree is kind of what I was trying to get at...\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 29 Mar 2022 18:48:30 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Wed, 30 Mar 2022 at 14:48, Andres Freund <andres@anarazel.de> wrote:\n>\n> On 2022-03-30 14:30:32 +1300, David Rowley wrote:\n> > Maybe nodes below an Append/MergeAppend with run-time pruning could compile\n> > on-demand and other nodes up-front. Or maybe there's no problem with making\n> > everything on-demand.\n>\n> Yea, that could work. The expressions for one \"partition query\" would still\n> have to be emitted at once. For each such subtree we should make a separate\n> costing decision. But I think an additional \"will be executed\" sub-node is a\n> different story, the threshold shouldn't be done on a per-node basis. That\n> partitioning of the plan tree is kind of what I was trying to get at...\n\nMaybe this point is moot if we get something like [1]. It might mean\nthat run-time pruning would happen early enough that we could just JIT\ncompile non-pruned subnodes.\n\nDavid\n\n[1] https://commitfest.postgresql.org/37/3478/\n\n\n","msg_date":"Wed, 30 Mar 2022 15:14:12 +1300","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\n\n> On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\n> > I think WARNING is fine. After all, the parameter is called\n> > \"jit_warn_above_fraction\".\n>\n> I had a think about this patch. I guess it's a little similar to\n> checkpoint_warning. The good thing about the checkpoint_warning is\n> that in the LOG message we give instructions about how the DBA can fix\n> the issue, i.e increase max_wal_size.\n>\n> With the proposed patch I see there is no hint about what might be\n> done to remove/reduce the warnings. I imagine that's because it's not\n> all that clear which GUC should be changed. In my view, likely\n> jit_above_cost is the most relevant but there is also\n> jit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\n> and jit_expressions which are relevant too.\n>\n> If we go with this patch, the problem I see here is that the amount\n> of work the JIT compiler must do for a given query depends mostly on\n> the number of expressions that must be compiled in the query (also to\n> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\n> jit_tuple_deforming and jit_expressions). The DBA does not really have\n> much control over the number of expressions in the query. All he or\n> she can do to get rid of the warning is something like increase\n> jit_above_cost. After a few iterations of that, the end result is\n> that jit_above_cost is now high enough that JIT no longer triggers\n> for, say, that query to that table with 1000 partitions where no\n> plan-time pruning takes place. Is that really a good thing? It likely\n> means that we just rarely JIT anything at all!\n>\n\nI don't agree with the conclusion of that.\n\nWhat the parameter would be useful for is to be able to tune those costs\n(or just turn it off) *for that individual query*. That doesn't mean you\n\"rarely JIT anything atll\", it just means you rarely JIT that particular\nquery.\n\nIn fact, my goal is to specifically make people do that and *not* just turn\noff JIT globally.\n\n\nI'd much rather see us address the costing problem before adding some\n> warning, especially a warning where it's not clear how to make go\n> away.\n>\n\nThe easiest way would be to add a HINT that says turn off jit for this\nparticular query or something?\n\nI do agree that if we can make \"spending too much time on JIT vs query\nruntime\" go away completely, then there is no need for a parameter like\nthis.\n\nI still think the warning is useful. And I think it may stay useful even\nafter we have made the JIT costing smarter -- though that's not certain of\ncourse.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\n\nOn Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\n> I think WARNING is fine. After all, the parameter is called\n> \"jit_warn_above_fraction\".\n\nI had a think about this patch. I guess it's a little similar to\ncheckpoint_warning. The good thing about the checkpoint_warning is\nthat in the LOG message we give instructions about how the DBA can fix\nthe issue, i.e increase max_wal_size.\n\nWith the proposed patch I see there is no hint about what might be\ndone to remove/reduce the warnings. I imagine that's because it's not\nall that clear which GUC should be changed. In my view, likely\njit_above_cost is the most relevant but there is also\njit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\nand jit_expressions which are relevant too.\n\nIf we go with this patch, the problem I see here is that the amount\nof work the JIT compiler must do for a given query depends mostly on\nthe number of expressions that must be compiled in the query (also to\na lesser extent jit_inline_above_cost, jit_optimize_above_cost,\njit_tuple_deforming and jit_expressions). The DBA does not really have\nmuch control over the number of expressions in the query. All he or\nshe can do to get rid of the warning is something like increase\njit_above_cost. After a few iterations of that, the end result is\nthat jit_above_cost is now high enough that JIT no longer triggers\nfor, say, that query to that table with 1000 partitions where no\nplan-time pruning takes place. Is that really a good thing? It likely\nmeans that we just rarely JIT anything at all!I don't agree with the conclusion of that.What the parameter would be useful for is to be able to tune those costs (or just turn it off) *for that individual query*. That doesn't mean you \"rarely JIT anything atll\", it just means you rarely JIT that particular query.In fact, my goal is to specifically make people do that and *not* just turn off JIT globally.I'd much rather see us address the costing problem before adding some\nwarning, especially a warning where it's not clear how to make go\naway.The easiest way would be to add a HINT that says turn off jit for this particular query or something?I do agree that if we can make \"spending too much time on JIT vs query runtime\" go away completely, then there is no need for a parameter like this. I still think the warning is useful. And I think it may stay useful even after we have made the JIT costing smarter -- though that's not certain of course. -- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/","msg_date":"Wed, 30 Mar 2022 15:04:30 +0200","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net> wrote:\n\n> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com>\n> wrote:\n>\n>> On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\n>> > I think WARNING is fine. After all, the parameter is called\n>> > \"jit_warn_above_fraction\".\n>>\n>> I had a think about this patch. I guess it's a little similar to\n>> checkpoint_warning. The good thing about the checkpoint_warning is\n>> that in the LOG message we give instructions about how the DBA can fix\n>> the issue, i.e increase max_wal_size.\n>>\n>> With the proposed patch I see there is no hint about what might be\n>> done to remove/reduce the warnings. I imagine that's because it's not\n>> all that clear which GUC should be changed. In my view, likely\n>> jit_above_cost is the most relevant but there is also\n>> jit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\n>> and jit_expressions which are relevant too.\n>>\n>> If we go with this patch, the problem I see here is that the amount\n>> of work the JIT compiler must do for a given query depends mostly on\n>> the number of expressions that must be compiled in the query (also to\n>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\n>> jit_tuple_deforming and jit_expressions). The DBA does not really have\n>> much control over the number of expressions in the query. All he or\n>> she can do to get rid of the warning is something like increase\n>> jit_above_cost. After a few iterations of that, the end result is\n>> that jit_above_cost is now high enough that JIT no longer triggers\n>> for, say, that query to that table with 1000 partitions where no\n>> plan-time pruning takes place. Is that really a good thing? It likely\n>> means that we just rarely JIT anything at all!\n>>\n>\n> I don't agree with the conclusion of that.\n>\n> What the parameter would be useful for is to be able to tune those costs\n> (or just turn it off) *for that individual query*. That doesn't mean you\n> \"rarely JIT anything atll\", it just means you rarely JIT that particular\n> query.\n>\n> In fact, my goal is to specifically make people do that and *not* just\n> turn off JIT globally.\n>\n>\n> I'd much rather see us address the costing problem before adding some\n>> warning, especially a warning where it's not clear how to make go\n>> away.\n>>\n>\n> The easiest way would be to add a HINT that says turn off jit for this\n> particular query or something?\n>\n> I do agree that if we can make \"spending too much time on JIT vs query\n> runtime\" go away completely, then there is no need for a parameter like\n> this.\n>\n> I still think the warning is useful. And I think it may stay useful even\n> after we have made the JIT costing smarter -- though that's not certain of\n> course.\n>\n>\nThis patch is still sitting at \"ready for committer\".\n\nAs an example, I have added such a hint in the attached.\n\nI still stand by that this patch is better than nothing. Sure, I would\nlove for us to adapt the JIT costing model and algorithm to make this not a\nproblem. And once we've done that, we should remove the parameter again.\n\nIt's not on by default, and it's trivial to remove in the future.\n\n\nYes, we're right up at the deadline. I'd still like to get it in, so I'd\nreally appreciate some further voices :)\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>","msg_date":"Fri, 8 Apr 2022 13:27:06 +0200","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Greetings,\n\nOn Fri, Apr 8, 2022 at 07:27 Magnus Hagander <magnus@hagander.net> wrote:\n\n> On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net>\n> wrote:\n>\n>> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com>\n>> wrote:\n>>\n>>> On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\n>>> > I think WARNING is fine. After all, the parameter is called\n>>> > \"jit_warn_above_fraction\".\n>>>\n>>> I had a think about this patch. I guess it's a little similar to\n>>> checkpoint_warning. The good thing about the checkpoint_warning is\n>>> that in the LOG message we give instructions about how the DBA can fix\n>>> the issue, i.e increase max_wal_size.\n>>>\n>>> With the proposed patch I see there is no hint about what might be\n>>> done to remove/reduce the warnings. I imagine that's because it's not\n>>> all that clear which GUC should be changed. In my view, likely\n>>> jit_above_cost is the most relevant but there is also\n>>> jit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\n>>> and jit_expressions which are relevant too.\n>>>\n>>> If we go with this patch, the problem I see here is that the amount\n>>> of work the JIT compiler must do for a given query depends mostly on\n>>> the number of expressions that must be compiled in the query (also to\n>>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\n>>> jit_tuple_deforming and jit_expressions). The DBA does not really have\n>>> much control over the number of expressions in the query. All he or\n>>> she can do to get rid of the warning is something like increase\n>>> jit_above_cost. After a few iterations of that, the end result is\n>>> that jit_above_cost is now high enough that JIT no longer triggers\n>>> for, say, that query to that table with 1000 partitions where no\n>>> plan-time pruning takes place. Is that really a good thing? It likely\n>>> means that we just rarely JIT anything at all!\n>>>\n>>\n>> I don't agree with the conclusion of that.\n>>\n>> What the parameter would be useful for is to be able to tune those costs\n>> (or just turn it off) *for that individual query*. That doesn't mean you\n>> \"rarely JIT anything atll\", it just means you rarely JIT that particular\n>> query.\n>>\n>> In fact, my goal is to specifically make people do that and *not* just\n>> turn off JIT globally.\n>>\n>>\n>> I'd much rather see us address the costing problem before adding some\n>>> warning, especially a warning where it's not clear how to make go\n>>> away.\n>>>\n>>\n>> The easiest way would be to add a HINT that says turn off jit for this\n>> particular query or something?\n>>\n>> I do agree that if we can make \"spending too much time on JIT vs query\n>> runtime\" go away completely, then there is no need for a parameter like\n>> this.\n>>\n>> I still think the warning is useful. And I think it may stay useful even\n>> after we have made the JIT costing smarter -- though that's not certain of\n>> course.\n>>\n>>\n> This patch is still sitting at \"ready for committer\".\n>\n> As an example, I have added such a hint in the attached.\n>\n> I still stand by that this patch is better than nothing. Sure, I would\n> love for us to adapt the JIT costing model and algorithm to make this not a\n> problem. And once we've done that, we should remove the parameter again.\n>\n> It's not on by default, and it's trivial to remove in the future.\n>\n>\n> Yes, we're right up at the deadline. I'd still like to get it in, so I'd\n> really appreciate some further voices :)\n>\n\nLooks reasonable to me, so +1. The default is has it off and so I seriously\ndoubt it’ll cause any issues and it’ll be very handy on large and busy\nsystems with lots of queries finding those that have a serious amount of\ntime being spent in JIT (and hopefully avoid folks just turning jit off\nacross the board, since that’s worse- we need more data on jit and need to\nwork on improving it, not ending up with everyone turning it off).\n\nThanks!\n\nStephen\n\n>\n\nGreetings,On Fri, Apr 8, 2022 at 07:27 Magnus Hagander <magnus@hagander.net> wrote:On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net> wrote:On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\n> I think WARNING is fine. After all, the parameter is called\n> \"jit_warn_above_fraction\".\n\nI had a think about this patch. I guess it's a little similar to\ncheckpoint_warning. The good thing about the checkpoint_warning is\nthat in the LOG message we give instructions about how the DBA can fix\nthe issue, i.e increase max_wal_size.\n\nWith the proposed patch I see there is no hint about what might be\ndone to remove/reduce the warnings. I imagine that's because it's not\nall that clear which GUC should be changed. In my view, likely\njit_above_cost is the most relevant but there is also\njit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\nand jit_expressions which are relevant too.\n\nIf we go with this patch, the problem I see here is that the amount\nof work the JIT compiler must do for a given query depends mostly on\nthe number of expressions that must be compiled in the query (also to\na lesser extent jit_inline_above_cost, jit_optimize_above_cost,\njit_tuple_deforming and jit_expressions). The DBA does not really have\nmuch control over the number of expressions in the query. All he or\nshe can do to get rid of the warning is something like increase\njit_above_cost. After a few iterations of that, the end result is\nthat jit_above_cost is now high enough that JIT no longer triggers\nfor, say, that query to that table with 1000 partitions where no\nplan-time pruning takes place. Is that really a good thing? It likely\nmeans that we just rarely JIT anything at all!I don't agree with the conclusion of that.What the parameter would be useful for is to be able to tune those costs (or just turn it off) *for that individual query*. That doesn't mean you \"rarely JIT anything atll\", it just means you rarely JIT that particular query.In fact, my goal is to specifically make people do that and *not* just turn off JIT globally.I'd much rather see us address the costing problem before adding some\nwarning, especially a warning where it's not clear how to make go\naway.The easiest way would be to add a HINT that says turn off jit for this particular query or something?I do agree that if we can make \"spending too much time on JIT vs query runtime\" go away completely, then there is no need for a parameter like this. I still think the warning is useful. And I think it may stay useful even after we have made the JIT costing smarter -- though that's not certain of course. This patch is still sitting at \"ready for committer\".As an example, I have added such a hint in the attached.I still stand by that this patch is better than nothing. Sure, I would love for us to adapt the JIT costing model and algorithm to make this not a problem. And once we've done that, we should remove the parameter again.It's not on by default, and it's trivial to remove in the future. Yes, we're right up at the deadline. I'd still like to get it in, so I'd really appreciate some further voices :)Looks reasonable to me, so +1. The default is has it off and so I seriously doubt it’ll cause any issues and it’ll be very handy on large and busy systems with lots of queries finding those that have a serious amount of time being spent in JIT (and hopefully avoid folks just turning jit off across the board, since that’s worse- we need more data on jit and need to work on improving it, not ending up with everyone turning it off). Thanks!Stephen","msg_date":"Fri, 8 Apr 2022 07:32:44 -0400","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Fri, 8 Apr 2022 at 23:27, Magnus Hagander <magnus@hagander.net> wrote:\n>\n>\n>\n> On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net> wrote:\n>>\n>> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\n>>>\n>>> If we go with this patch, the problem I see here is that the amount\n>>> of work the JIT compiler must do for a given query depends mostly on\n>>> the number of expressions that must be compiled in the query (also to\n>>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\n>>> jit_tuple_deforming and jit_expressions). The DBA does not really have\n>>> much control over the number of expressions in the query. All he or\n>>> she can do to get rid of the warning is something like increase\n>>> jit_above_cost. After a few iterations of that, the end result is\n>>> that jit_above_cost is now high enough that JIT no longer triggers\n>>> for, say, that query to that table with 1000 partitions where no\n>>> plan-time pruning takes place. Is that really a good thing? It likely\n>>> means that we just rarely JIT anything at all!\n>>\n>>\n>> I don't agree with the conclusion of that.\n>>\n>> What the parameter would be useful for is to be able to tune those costs (or just turn it off) *for that individual query*. That doesn't mean you \"rarely JIT anything atll\", it just means you rarely JIT that particular query.\n\nI just struggle to imagine that anyone is going to spend much effort\ntuning a warning parameter per query. I imagine they're far more\nlikely to just ramp it up to only catch some high percentile problems\nor just (more likely) just not bother with it. It seems more likely\nthat if anyone was to tune anything per query here it would be\njit_above_cost, since that actually might have an affect on the\nperformance of the query, rather than if it spits out some warning\nmessage or not. ISTM that if the user knows what to set it to per\nquery, then there's very little point in having a warning as we'd be\nalerting them to something they already know about.\n\nI looked in the -general list to see if we could get some common\nexplanations to give us an idea of the most common reason for high JIT\ncompilation time. It seems that the plans were never simple. [1] seems\ndue to a complex plan. I'm basing that off the \"Functions: 167\". I\ndidn't catch the full plan. From what I can tell, [2] seems to be due\nto \"lots of empty tables\", so assuming the clamping at 1 page is\ncausing issues there. I think both of those cases could be resolved\nby building the costing the way I mentioned. I admit that 2 cases is\nnot a very large sample size.\n\nDavid\n\n[1] https://www.postgresql.org/message-id/flat/CAPL5KHq8zfWPzueCemXw4c%2BU568PoDfqo3wBDNm3KAyvybdaMQ%40mail.gmail.com#35aca8b42c3862f44b6be5b260c1a109\n[2] https://www.postgresql.org/message-id/flat/CAHOFxGo5xJt02RmwAWrtv2K0jcqqxG-cDiR8FQbvb0WxdKhcgw%40mail.gmail.com#12d91822e869a2e22ca830cb5632f549\n\n\n","msg_date":"Sat, 9 Apr 2022 00:18:44 +1200","msg_from":"David Rowley <dgrowleyml@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Fri, Apr 8, 2022 at 2:19 PM David Rowley <dgrowleyml@gmail.com> wrote:\n\n> On Fri, 8 Apr 2022 at 23:27, Magnus Hagander <magnus@hagander.net> wrote:\n> >\n> >\n> >\n> > On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net>\n> wrote:\n> >>\n> >> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com>\n> wrote:\n> >>>\n> >>> If we go with this patch, the problem I see here is that the amount\n> >>> of work the JIT compiler must do for a given query depends mostly on\n> >>> the number of expressions that must be compiled in the query (also to\n> >>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\n> >>> jit_tuple_deforming and jit_expressions). The DBA does not really have\n> >>> much control over the number of expressions in the query. All he or\n> >>> she can do to get rid of the warning is something like increase\n> >>> jit_above_cost. After a few iterations of that, the end result is\n> >>> that jit_above_cost is now high enough that JIT no longer triggers\n> >>> for, say, that query to that table with 1000 partitions where no\n> >>> plan-time pruning takes place. Is that really a good thing? It likely\n> >>> means that we just rarely JIT anything at all!\n> >>\n> >>\n> >> I don't agree with the conclusion of that.\n> >>\n> >> What the parameter would be useful for is to be able to tune those\n> costs (or just turn it off) *for that individual query*. That doesn't mean\n> you \"rarely JIT anything atll\", it just means you rarely JIT that\n> particular query.\n>\n> I just struggle to imagine that anyone is going to spend much effort\n> tuning a warning parameter per query. I imagine they're far more\n> likely to just ramp it up to only catch some high percentile problems\n> or just (more likely) just not bother with it. It seems more likely\n> that if anyone was to tune anything per query here it would be\n> jit_above_cost, since that actually might have an affect on the\n> performance of the query, rather than if it spits out some warning\n> message or not. ISTM that if the user knows what to set it to per\n> query, then there's very little point in having a warning as we'd be\n> alerting them to something they already know about.\n>\n\nI would not expect people to tune the *warning* at a query level. If\nanything, then ys, they would tune the either jit_above_cost or just\njit=off. But the idea being you can do that on a per query level instead of\nglobally.\n\n\nI looked in the -general list to see if we could get some common\n> explanations to give us an idea of the most common reason for high JIT\n> compilation time. It seems that the plans were never simple. [1] seems\n> due to a complex plan. I'm basing that off the \"Functions: 167\". I\n> didn't catch the full plan. From what I can tell, [2] seems to be due\n> to \"lots of empty tables\", so assuming the clamping at 1 page is\n> causing issues there. I think both of those cases could be resolved\n> by building the costing the way I mentioned. I admit that 2 cases is\n> not a very large sample size.\n>\n\nAgain, I am very much for improvements of the costing model. This is in no\nway intended to be a replacement for that. It's intended to be a stop-gap.\n\nWhat I see much of today are things like\nhttps://dba.stackexchange.com/questions/264955/handling-performance-problems-with-jit-in-postgres-12\nor\nhttps://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\n\nThe bottom line is that people end up with recommendations to turn off JIT\nglobally more or less by default. Because there's no real useful way today\nto figure out when it causes problems vs when it helps.\n\nThe addition to pg_stat_statements I pushed a short while ago would help\nwith that. But I think having a warning like this would also be useful. As\na stop-gap measure, yes, but we really don't know when we will have an\nimproved costing model for it. I hope you're right and that we can have it\nby 16, and then I will definitely advocate for removing the warning again\nif it works.\n\n-- \n Magnus Hagander\n Me: https://www.hagander.net/ <http://www.hagander.net/>\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\n\nOn Fri, Apr 8, 2022 at 2:19 PM David Rowley <dgrowleyml@gmail.com> wrote:On Fri, 8 Apr 2022 at 23:27, Magnus Hagander <magnus@hagander.net> wrote:\n>\n>\n>\n> On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net> wrote:\n>>\n>> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\n>>>\n>>> If we go with this patch, the problem I see here is that the amount\n>>> of work the JIT compiler must do for a given query depends mostly on\n>>> the number of expressions that must be compiled in the query (also to\n>>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\n>>> jit_tuple_deforming and jit_expressions). The DBA does not really have\n>>> much control over the number of expressions in the query. All he or\n>>> she can do to get rid of the warning is something like increase\n>>> jit_above_cost. After a few iterations of that, the end result is\n>>> that jit_above_cost is now high enough that JIT no longer triggers\n>>> for, say, that query to that table with 1000 partitions where no\n>>> plan-time pruning takes place. Is that really a good thing? It likely\n>>> means that we just rarely JIT anything at all!\n>>\n>>\n>> I don't agree with the conclusion of that.\n>>\n>> What the parameter would be useful for is to be able to tune those costs (or just turn it off) *for that individual query*. That doesn't mean you \"rarely JIT anything atll\", it just means you rarely JIT that particular query.\n\nI just struggle to imagine that anyone is going to spend much effort\ntuning a warning parameter per query. I imagine they're far more\nlikely to just ramp it up to only catch some high percentile problems\nor just (more likely) just not bother with it. It seems more likely\nthat if anyone was to tune anything per query here it would be\njit_above_cost, since that actually might have an affect on the\nperformance of the query, rather than if it spits out some warning\nmessage or not. ISTM that if the user knows what to set it to per\nquery, then there's very little point in having a warning as we'd be\nalerting them to something they already know about.I would not expect people to tune the *warning* at a query level. If anything, then ys, they would tune the either jit_above_cost or just jit=off. But the idea being you can do that on a per query level instead of globally.\nI looked in the -general list to see if we could get some common\nexplanations to give us an idea of the most common reason for high JIT\ncompilation time. It seems that the plans were never simple. [1] seems\ndue to a complex plan. I'm basing that off the \"Functions: 167\". I\ndidn't catch the full plan. From what I can tell, [2] seems to be due\nto \"lots of empty tables\", so assuming the clamping at 1 page is\ncausing issues there. I think both of those cases could be resolved\nby building the costing the way I mentioned. I admit that 2 cases is\nnot a very large sample size.Again, I am very much for improvements of the costing model. This is in no way intended to be a replacement for that. It's intended to be a stop-gap. What I see much of today are things like https://dba.stackexchange.com/questions/264955/handling-performance-problems-with-jit-in-postgres-12 or https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2aThe bottom line is that people end up with recommendations to turn off JIT globally more or less by default. Because there's no real useful way today to figure out when it causes problems vs when it helps.The addition to pg_stat_statements I pushed a short while ago would help with that. But I think having a warning like this would also be useful. As a stop-gap measure, yes, but we really don't know when we will have an improved costing model for it. I hope you're right and that we can have it by 16, and then I will definitely advocate for removing the warning again if it works.-- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/","msg_date":"Fri, 8 Apr 2022 14:28:36 +0200","msg_from":"Magnus Hagander <magnus@hagander.net>","msg_from_op":true,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Greetings,\n\n* Magnus Hagander (magnus@hagander.net) wrote:\n> On Fri, Apr 8, 2022 at 2:19 PM David Rowley <dgrowleyml@gmail.com> wrote:\n> > On Fri, 8 Apr 2022 at 23:27, Magnus Hagander <magnus@hagander.net> wrote:\n> > > On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net>\n> > wrote:\n> > >>\n> > >> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com>\n> > wrote:\n> > >>>\n> > >>> If we go with this patch, the problem I see here is that the amount\n> > >>> of work the JIT compiler must do for a given query depends mostly on\n> > >>> the number of expressions that must be compiled in the query (also to\n> > >>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\n> > >>> jit_tuple_deforming and jit_expressions). The DBA does not really have\n> > >>> much control over the number of expressions in the query. All he or\n> > >>> she can do to get rid of the warning is something like increase\n> > >>> jit_above_cost. After a few iterations of that, the end result is\n> > >>> that jit_above_cost is now high enough that JIT no longer triggers\n> > >>> for, say, that query to that table with 1000 partitions where no\n> > >>> plan-time pruning takes place. Is that really a good thing? It likely\n> > >>> means that we just rarely JIT anything at all!\n> > >>\n> > >>\n> > >> I don't agree with the conclusion of that.\n> > >>\n> > >> What the parameter would be useful for is to be able to tune those\n> > costs (or just turn it off) *for that individual query*. That doesn't mean\n> > you \"rarely JIT anything atll\", it just means you rarely JIT that\n> > particular query.\n> >\n> > I just struggle to imagine that anyone is going to spend much effort\n> > tuning a warning parameter per query. I imagine they're far more\n> > likely to just ramp it up to only catch some high percentile problems\n> > or just (more likely) just not bother with it. It seems more likely\n> > that if anyone was to tune anything per query here it would be\n> > jit_above_cost, since that actually might have an affect on the\n> > performance of the query, rather than if it spits out some warning\n> > message or not. ISTM that if the user knows what to set it to per\n> > query, then there's very little point in having a warning as we'd be\n> > alerting them to something they already know about.\n> \n> I would not expect people to tune the *warning* at a query level. If\n> anything, then ys, they would tune the either jit_above_cost or just\n> jit=off. But the idea being you can do that on a per query level instead of\n> globally.\n\nYeah, exactly, this is about having a busy system and wanting to know\nwhich queries are spending a lot of time doing JIT relative to the query\ntime, so that you can go adjust your JIT parameters or possibly disable\nJIT for those queries (or maybe bring those cases to -hackers and try to\nhelp make our costing better).\n\n> > I looked in the -general list to see if we could get some common\n> > explanations to give us an idea of the most common reason for high JIT\n> > compilation time. It seems that the plans were never simple. [1] seems\n> > due to a complex plan. I'm basing that off the \"Functions: 167\". I\n> > didn't catch the full plan. From what I can tell, [2] seems to be due\n> > to \"lots of empty tables\", so assuming the clamping at 1 page is\n> > causing issues there. I think both of those cases could be resolved\n> > by building the costing the way I mentioned. I admit that 2 cases is\n> > not a very large sample size.\n> \n> Again, I am very much for improvements of the costing model. This is in no\n> way intended to be a replacement for that. It's intended to be a stop-gap.\n\nNot sure I'd say it's a 'stop-gap' as it's really very similar, imv\nanyway, to log_min_duration_statement- you want to know what queries are\ntaking a lot of time but you can't log all of them.\n\n> What I see much of today are things like\n> https://dba.stackexchange.com/questions/264955/handling-performance-problems-with-jit-in-postgres-12\n> or\n> https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\n> \n> The bottom line is that people end up with recommendations to turn off JIT\n> globally more or less by default. Because there's no real useful way today\n> to figure out when it causes problems vs when it helps.\n\nYeah, that's frustrating.\n\n> The addition to pg_stat_statements I pushed a short while ago would help\n> with that. But I think having a warning like this would also be useful. As\n> a stop-gap measure, yes, but we really don't know when we will have an\n> improved costing model for it. I hope you're right and that we can have it\n> by 16, and then I will definitely advocate for removing the warning again\n> if it works.\n\nHaving this in pg_stat_statements is certainly helpful but having a\nwarning also is. I don't think we have to address this in only one way.\nA lot faster to flip this guc and then look in the logs on a busy system\nthan to install pg_stat_statements, restart the cluster once you get\npermission to do so, and then query it.\n\nThanks,\n\nStephen","msg_date":"Fri, 8 Apr 2022 09:39:18 -0400","msg_from":"Stephen Frost <sfrost@snowman.net>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Hi,\n\nOn Fri, Apr 08, 2022 at 09:39:18AM -0400, Stephen Frost wrote:\n>\n> * Magnus Hagander (magnus@hagander.net) wrote:\n> > The addition to pg_stat_statements I pushed a short while ago would help\n> > with that. But I think having a warning like this would also be useful. As\n> > a stop-gap measure, yes, but we really don't know when we will have an\n> > improved costing model for it. I hope you're right and that we can have it\n> > by 16, and then I will definitely advocate for removing the warning again\n> > if it works.\n>\n> Having this in pg_stat_statements is certainly helpful but having a\n> warning also is. I don't think we have to address this in only one way.\n> A lot faster to flip this guc and then look in the logs on a busy system\n> than to install pg_stat_statements, restart the cluster once you get\n> permission to do so, and then query it.\n\n+1, especially if you otherwise don't really need or want to have\npg_stat_statements enabled, as it's far from being free. Sure you could enable\nit by default with pg_stat_statements.track = none, but that seems a lot more\ntroublesome than just dynamically enabling a GUC, possibly for a short time\nand/or for a specific database/role.\n\n\n","msg_date":"Sat, 9 Apr 2022 12:20:57 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Julien Rouhaud <rjuju123@gmail.com> writes:\n> On Fri, Apr 08, 2022 at 09:39:18AM -0400, Stephen Frost wrote:\n>> Having this in pg_stat_statements is certainly helpful but having a\n>> warning also is. I don't think we have to address this in only one way.\n>> A lot faster to flip this guc and then look in the logs on a busy system\n>> than to install pg_stat_statements, restart the cluster once you get\n>> permission to do so, and then query it.\n\n> +1, especially if you otherwise don't really need or want to have\n> pg_stat_statements enabled, as it's far from being free. Sure you could enable\n> it by default with pg_stat_statements.track = none, but that seems a lot more\n> troublesome than just dynamically enabling a GUC, possibly for a short time\n> and/or for a specific database/role.\n\nThe arguments against the GUC were not about whether it's convenient.\nThey were about whether the information it gives you is actually going\nto be of any use.\n\nAlso, good luck with \"looking in the logs\", because by default\nWARNING-level messages don't go to the postmaster log. If that's\nthe intended use-case then the message ought to appear at LOG\nlevel (which'd also change the desirable name for the GUC).\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 09 Apr 2022 10:42:12 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\n>\n> Also, good luck with \"looking in the logs\", because by default\n> WARNING-level messages don't go to the postmaster log. If that's\n> the intended use-case then the message ought to appear at LOG\n> level (which'd also change the desirable name for the GUC).\n\nI must be missing something because by default log_min_messages is WARNING, and\nAFAICS I do get WARNING-level messages in some vanilla cluster logs, using\nvanilla .psqlrc.\n\n\n","msg_date":"Sat, 9 Apr 2022 23:21:02 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"Julien Rouhaud <rjuju123@gmail.com> writes:\n> On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\n>> Also, good luck with \"looking in the logs\", because by default\n>> WARNING-level messages don't go to the postmaster log.\n\n> I must be missing something because by default log_min_messages is WARNING, and\n> AFAICS I do get WARNING-level messages in some vanilla cluster logs, using\n> vanilla .psqlrc.\n\nAh, sorry, I was looking at the wrong thing namely\nlog_min_error_statement. The point still holds though: if we emit this\nat level WARNING, what will get logged is just the bare message and not\nthe statement that triggered it. How useful do you think that will be?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 09 Apr 2022 12:31:23 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Sat, Apr 09, 2022 at 12:31:23PM -0400, Tom Lane wrote:\n> Julien Rouhaud <rjuju123@gmail.com> writes:\n> > On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\n> >> Also, good luck with \"looking in the logs\", because by default\n> >> WARNING-level messages don't go to the postmaster log.\n> \n> > I must be missing something because by default log_min_messages is WARNING, and\n> > AFAICS I do get WARNING-level messages in some vanilla cluster logs, using\n> > vanilla .psqlrc.\n> \n> Ah, sorry, I was looking at the wrong thing namely\n> log_min_error_statement. The point still holds though: if we emit this\n> at level WARNING, what will get logged is just the bare message and not\n> the statement that triggered it. How useful do you think that will be?\n\nAh right, I did miss that \"small detail\". And of course I agree, as-is that\nwould be entirely useless and it should be LOG, like the rest of similar\nfeatures.\n\n\n","msg_date":"Sun, 10 Apr 2022 00:43:08 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"},{"msg_contents":"On Sat, Apr 9, 2022 at 8:43 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n\n> On Sat, Apr 09, 2022 at 12:31:23PM -0400, Tom Lane wrote:\n> > Julien Rouhaud <rjuju123@gmail.com> writes:\n> > > On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\n> > >> Also, good luck with \"looking in the logs\", because by default\n> > >> WARNING-level messages don't go to the postmaster log.\n> >\n> > > I must be missing something because by default log_min_messages is\n> WARNING, and\n> > > AFAICS I do get WARNING-level messages in some vanilla cluster logs,\n> using\n> > > vanilla .psqlrc.\n> >\n> > Ah, sorry, I was looking at the wrong thing namely\n> > log_min_error_statement. The point still holds though: if we emit this\n> > at level WARNING, what will get logged is just the bare message and not\n> > the statement that triggered it. How useful do you think that will be?\n>\n> Ah right, I did miss that \"small detail\". And of course I agree, as-is\n> that\n> would be entirely useless and it should be LOG, like the rest of similar\n> features.\n>\n>\n>\nThe patch does not apply successfully; please rebase the patch.\n\npatching file src/backend/utils/misc/guc.c\nHunk #1 FAILED at 642.\nHunk #2 FAILED at 3943.\n2 out of 2 hunks FAILED -- saving rejects to file\nsrc/backend/utils/misc/guc.c.rej\npatching file src/backend/utils/misc/postgresql.conf.sample\npatching file src/include/jit/jit.h\n\n\n-- \nIbrar Ahmed\n\nOn Sat, Apr 9, 2022 at 8:43 PM Julien Rouhaud <rjuju123@gmail.com> wrote:On Sat, Apr 09, 2022 at 12:31:23PM -0400, Tom Lane wrote:\n> Julien Rouhaud <rjuju123@gmail.com> writes:\n> > On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\n> >> Also, good luck with \"looking in the logs\", because by default\n> >> WARNING-level messages don't go to the postmaster log.\n> \n> > I must be missing something because by default log_min_messages is WARNING, and\n> > AFAICS I do get WARNING-level messages in some vanilla cluster logs, using\n> > vanilla .psqlrc.\n> \n> Ah, sorry, I was looking at the wrong thing namely\n> log_min_error_statement. The point still holds though: if we emit this\n> at level WARNING, what will get logged is just the bare message and not\n> the statement that triggered it. How useful do you think that will be?\n\nAh right, I did miss that \"small detail\". And of course I agree, as-is that\nwould be entirely useless and it should be LOG, like the rest of similar\nfeatures.\n\n\nThe patch does not apply successfully; please rebase the patch.patching file src/backend/utils/misc/guc.c\nHunk #1 FAILED at 642.\nHunk #2 FAILED at 3943.\n2 out of 2 hunks FAILED -- saving rejects to file src/backend/utils/misc/guc.c.rej\npatching file src/backend/utils/misc/postgresql.conf.sample\npatching file src/include/jit/jit.h-- Ibrar Ahmed","msg_date":"Thu, 15 Sep 2022 12:02:17 +0400","msg_from":"Ibrar Ahmed <ibrar.ahmad@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add parameter jit_warn_above_fraction"}],"string":"[\n {\n \"msg_contents\": \"This patch adds a configuration parameter jit_warn_above_fraction that\\nwill cause a warning to be logged if the fraction of time spent on\\ndoing JIT is bigger than the specified one. For example, this can be\\nused to track down those cases where JIT ends up taking 90% of the\\nquery runtime because of bad estimates...\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Fri, 25 Feb 2022 16:16:01 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Fri, Feb 25, 2022 at 04:16:01PM +0100, Magnus Hagander wrote:\\n> This patch adds a configuration parameter jit_warn_above_fraction that\\n> will cause a warning to be logged if the fraction of time spent on\\n> doing JIT is bigger than the specified one. For example, this can be\\n> used to track down those cases where JIT ends up taking 90% of the\\n> query runtime because of bad estimates...\\n\\nI think that's tremendously useful, huge +1.\\n\\nJust a few minor nit:\\n\\n+ A value of 0 (the default)disables the warning.\\n\\nmissing space\\n\\n+\\t\\t\\tereport(WARNING,\\n+\\t\\t\\t\\t\\t(errmsg(\\\"JIT time was %ld ms of %d ms\\\",\\n+\\t\\t\\t\\t\\t\\t\\tjit_time, msecs)));\\n\\n\\\"JIT time\\\" may a bit obscure for users, how about \\\"JIT total processing time\\\"?\\\"\\n\\n+\\t\\t\\tgettext_noop(\\\"Sets the fraction of query time spent on JIT before writing\\\"\\n+\\t\\t\\t\\t\\t\\t \\\"a warning to the log.\\\"),\\n+\\t\\t\\tgettext_noop(\\\"Write a message tot he server log if more than this\\\"\\n+\\t\\t\\t\\t\\t\\t \\\"fraction of the query runtime is spent on JIT.\\\"\\n+\\t\\t\\t\\t\\t\\t \\\"Zero turns off the warning.\\\")\\n\\nmissing spaces in the concatenated strings.\\n\\nThe rest of the patch looks good to me.\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 00:01:41 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-25 16:16:01 +0100, Magnus Hagander wrote:\\n> This patch adds a configuration parameter jit_warn_above_fraction that\\n> will cause a warning to be logged if the fraction of time spent on\\n> doing JIT is bigger than the specified one. For example, this can be\\n> used to track down those cases where JIT ends up taking 90% of the\\n> query runtime because of bad estimates...\\n\\nHm. Could it make sense to do this as a auto_explain feature?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 08:20:43 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 04:16:01PM +0100, Magnus Hagander wrote:\\n> +\\t{\\n> +\\t\\t{\\\"jit_warn_above_fraction\\\", PGC_SUSET, LOGGING_WHEN,\\n> +\\t\\t\\tgettext_noop(\\\"Sets the fraction of query time spent on JIT before writing\\\"\\n> +\\t\\t\\t\\t\\t\\t \\\"a warning to the log.\\\"),\\n> +\\t\\t\\tgettext_noop(\\\"Write a message tot he server log if more than this\\\"\\n> +\\t\\t\\t\\t\\t\\t \\\"fraction of the query runtime is spent on JIT.\\\"\\n> +\\t\\t\\t\\t\\t\\t \\\"Zero turns off the warning.\\\")\\n> +\\t\\t},\\n> +\\t\\t&jit_warn_above_fraction,\\n> +\\t\\t0.0, 0.0, 1.0,\\n> +\\t\\tNULL, NULL, NULL\\n> +\\t},\\n\\nShould be PGC_USERSET ?\\n\\n+ gettext_noop(\\\"Write a message tot he server log if more than this\\\" \\n\\nto the\\n\\n+ if (jit_enabled && jit_warn_above_fraction > 0) \\n+ { \\n+ int64 jit_time = \\n+ INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.generation_counter) + \\n+ INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.inlining_counter) + \\n+ INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.optimization_counter) + \\n+ INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.emission_counter); \\n+ \\n+ if (jit_time > msecs * jit_warn_above_fraction) \\n+ { \\n+ ereport(WARNING, \\n+ (errmsg(\\\"JIT time was %ld ms of %d ms\\\", \\n+ jit_time, msecs))); \\n+ } \\n+ } \\n\\n\\nI think it should be a NOTICE (or less?)\\n\\nIs it more useful if this is applied combined with log_min_duration_statement ?\\n\\nIt's easy to imagine a query for which the planner computes a high cost, but\\nactually runs quickly. You might get a bunch of WARNINGs that the query took\\n10 MS and JIT was 75% of that, even if you don't care about queries that take\\nless than 10 SEC.\\n\\nI should say that this is already available by processing the output of\\nautoexplain.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 10:23:04 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 5:20 PM Andres Freund <andres@anarazel.de> wrote:\\n>\\n> Hi,\\n>\\n> On 2022-02-25 16:16:01 +0100, Magnus Hagander wrote:\\n> > This patch adds a configuration parameter jit_warn_above_fraction that\\n> > will cause a warning to be logged if the fraction of time spent on\\n> > doing JIT is bigger than the specified one. For example, this can be\\n> > used to track down those cases where JIT ends up taking 90% of the\\n> > query runtime because of bad estimates...\\n>\\n> Hm. Could it make sense to do this as a auto_explain feature?\\n\\nIt could be. But I was looking for something a lot more \\\"light weight\\\"\\nthan having to install an extension. But yes, if we wanted to, we\\ncould certainly change jit_warn_above_fraction to be\\nauto_explain.log_min_jit_fraction or something like that, and do\\nbasically the same thing. But then, we could also have\\nlog_min_duration_statement be part of auto_explain instead, so it's\\nall about where to draw the line :)\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 17:28:41 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-25 17:28:41 +0100, Magnus Hagander wrote:\\n> On Fri, Feb 25, 2022 at 5:20 PM Andres Freund <andres@anarazel.de> wrote:\\n> > On 2022-02-25 16:16:01 +0100, Magnus Hagander wrote:\\n> > > This patch adds a configuration parameter jit_warn_above_fraction that\\n> > > will cause a warning to be logged if the fraction of time spent on\\n> > > doing JIT is bigger than the specified one. For example, this can be\\n> > > used to track down those cases where JIT ends up taking 90% of the\\n> > > query runtime because of bad estimates...\\n> >\\n> > Hm. Could it make sense to do this as a auto_explain feature?\\n> \\n> It could be. But I was looking for something a lot more \\\"light weight\\\"\\n> than having to install an extension. But yes, if we wanted to, we\\n> could certainly change jit_warn_above_fraction to be\\n> auto_explain.log_min_jit_fraction or something like that, and do\\n> basically the same thing. But then, we could also have\\n> log_min_duration_statement be part of auto_explain instead, so it's\\n> all about where to draw the line :)\\n\\nI guess it feels a tad on the \\\"too narrow/specific\\\" side of things for the\\ngeneral code. We don't have log_min_duration_{parsing,planning,execution}\\neither. But I also get it. So I just wanted to raise it ;)\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 08:47:22 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 5:47 PM Andres Freund <andres@anarazel.de> wrote:\\n>\\n> Hi,\\n>\\n> On 2022-02-25 17:28:41 +0100, Magnus Hagander wrote:\\n> > On Fri, Feb 25, 2022 at 5:20 PM Andres Freund <andres@anarazel.de> wrote:\\n> > > On 2022-02-25 16:16:01 +0100, Magnus Hagander wrote:\\n> > > > This patch adds a configuration parameter jit_warn_above_fraction that\\n> > > > will cause a warning to be logged if the fraction of time spent on\\n> > > > doing JIT is bigger than the specified one. For example, this can be\\n> > > > used to track down those cases where JIT ends up taking 90% of the\\n> > > > query runtime because of bad estimates...\\n> > >\\n> > > Hm. Could it make sense to do this as a auto_explain feature?\\n> >\\n> > It could be. But I was looking for something a lot more \\\"light weight\\\"\\n> > than having to install an extension. But yes, if we wanted to, we\\n> > could certainly change jit_warn_above_fraction to be\\n> > auto_explain.log_min_jit_fraction or something like that, and do\\n> > basically the same thing. But then, we could also have\\n> > log_min_duration_statement be part of auto_explain instead, so it's\\n> > all about where to draw the line :)\\n>\\n> I guess it feels a tad on the \\\"too narrow/specific\\\" side of things for the\\n> general code. We don't have log_min_duration_{parsing,planning,execution}\\n> either. But I also get it. So I just wanted to raise it ;)\\n\\nOh it's absolutely a valid issue to raise :) In fact, one could\\ndefinitely argue that we should have a parameter for auto_explain *as\\nwell*.\\n\\nIt's true we don't have those -- but it's also in my experience a lot\\nmore common to be an issue with JIT. And unfortunately the current\\n\\\"solution\\\" people tend to apply is to globally disable JIT, and\\nthere's no similar \\\"globally disable parsing or \\\"globally disable\\nplanning\\\" pattern, for obvious reasons.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 12:56:52 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Fri, Feb 25, 2022 at 5:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> On Fri, Feb 25, 2022 at 04:16:01PM +0100, Magnus Hagander wrote:\\n> > + {\\n> > + {\\\"jit_warn_above_fraction\\\", PGC_SUSET, LOGGING_WHEN,\\n> > + gettext_noop(\\\"Sets the fraction of query time spent on JIT before writing\\\"\\n> > + \\\"a warning to the log.\\\"),\\n> > + gettext_noop(\\\"Write a message tot he server log if more than this\\\"\\n> > + \\\"fraction of the query runtime is spent on JIT.\\\"\\n> > + \\\"Zero turns off the warning.\\\")\\n> > + },\\n> > + &jit_warn_above_fraction,\\n> > + 0.0, 0.0, 1.0,\\n> > + NULL, NULL, NULL\\n> > + },\\n>\\n> Should be PGC_USERSET ?\\n\\nYes. Definitely. Copy/paste thinko.\\n\\n\\n> + gettext_noop(\\\"Write a message tot he server log if more than this\\\"\\n>\\n> to the\\n\\nYup.\\n\\n\\n> + if (jit_enabled && jit_warn_above_fraction > 0)\\n> + {\\n> + int64 jit_time =\\n> + INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.generation_counter) +\\n> + INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.inlining_counter) +\\n> + INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.optimization_counter) +\\n> + INSTR_TIME_GET_MILLISEC(portal->queryDesc->estate->es_jit->instr.emission_counter);\\n> +\\n> + if (jit_time > msecs * jit_warn_above_fraction)\\n> + {\\n> + ereport(WARNING,\\n> + (errmsg(\\\"JIT time was %ld ms of %d ms\\\",\\n> + jit_time, msecs)));\\n> + }\\n> + }\\n>\\n>\\n> I think it should be a NOTICE (or less?)\\n\\nHmm. I'm not so sure.\\n\\nOther similar parameters often use LOG, but the downside of that is\\nthat it won't be sent to the client.\\n\\nThe problem with using NOTICE is that it won't go to the log by\\ndefault. It needs to be at least warning to do that.\\n\\n\\n> Is it more useful if this is applied combined with log_min_duration_statement ?\\n>\\n> It's easy to imagine a query for which the planner computes a high cost, but\\n> actually runs quickly. You might get a bunch of WARNINGs that the query took\\n> 10 MS and JIT was 75% of that, even if you don't care about queries that take\\n> less than 10 SEC.\\n\\nYeah, that's definitely a problem. But which way would you want to tie\\nit to log_min_duration_statement? That a statement would both have to\\ntake longer than log_min_duration_statement *and* have JIT above a\\ncertain percentage? In my experience that is instead likely to miss\\nmost of the interesting times. Maybe it would need a separate guc for\\nthe timing, but I'd like to avoid that, I'm not sure it's a function\\nworth *that* much...\\n\\n\\n> I should say that this is already available by processing the output of\\n> autoexplain.\\n\\nTrue. You just can't trigger it based on it. (and it can be somewhat\\nof a PITA to parse things out of auto_explain on busy systems, but it\\ndoes give a lot of very useful details)\\n\\n\\nMeanwhile here is an updated based on your other comments above, as\\nwell as those from Julien.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Mon, 7 Mar 2022 13:10:32 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 07, 2022 at 01:10:32PM +0100, Magnus Hagander wrote:\\n> On Fri, Feb 25, 2022 at 5:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> >\\n> > I think it should be a NOTICE (or less?)\\n> \\n> Hmm. I'm not so sure.\\n> \\n> Other similar parameters often use LOG, but the downside of that is\\n> that it won't be sent to the client.\\n> \\n> The problem with using NOTICE is that it won't go to the log by\\n> default. It needs to be at least warning to do that.\\n\\nAnyone who uses this parameter is aleady going to be changing GUCs, so it\\ndoesn't need to work by default. The people who are likely to enable this\\nalready monitor their logs and have probably customized their logging\\nconfiguration.\\n\\n> > Is it more useful if this is applied combined with log_min_duration_statement ?\\n> >\\n> > It's easy to imagine a query for which the planner computes a high cost, but\\n> > actually runs quickly. You might get a bunch of WARNINGs that the query took\\n> > 10 MS and JIT was 75% of that, even if you don't care about queries that take\\n> > less than 10 SEC.\\n> \\n> Yeah, that's definitely a problem. But which way would you want to tie\\n> it to log_min_duration_statement? That a statement would both have to\\n> take longer than log_min_duration_statement *and* have JIT above a\\n> certain percentage? In my experience that is instead likely to miss\\n> most of the interesting times.\\n\\nI don't understand - why doesn't it combine trivially with\\nlog_min_duration_statement? Are you saying that the default / pre-existing min\\nduration may not log all of the intended queries ? I think that just means the\\nconfiguration needs to be changed. The GUC needs to allow/help finding these\\nJIT issues, but going to require an admin's interaction in any case. Avoiding\\nfalse positives may be important for it to be useful at all.\\n\\nHmm .. should it also combine with min_sample_rate ? Maybe that addresses your\\nconcern.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 07:09:22 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 2:09 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n>\\n> On Mon, Mar 07, 2022 at 01:10:32PM +0100, Magnus Hagander wrote:\\n> > On Fri, Feb 25, 2022 at 5:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > >\\n> > > I think it should be a NOTICE (or less?)\\n> >\\n> > Hmm. I'm not so sure.\\n> >\\n> > Other similar parameters often use LOG, but the downside of that is\\n> > that it won't be sent to the client.\\n> >\\n> > The problem with using NOTICE is that it won't go to the log by\\n> > default. It needs to be at least warning to do that.\\n>\\n> Anyone who uses this parameter is aleady going to be changing GUCs, so it\\n> doesn't need to work by default. The people who are likely to enable this\\n> already monitor their logs and have probably customized their logging\\n> configuration.\\n\\nSure, but if you set log_min_messagse to NOTICE you are likely going\\nto flood your logs beyond usefulness. And the whole idea of this\\nparameter is you can keep it on during \\\"normal times\\\" to catch\\noutliers.\\n\\n\\n> > > Is it more useful if this is applied combined with log_min_duration_statement ?\\n> > >\\n> > > It's easy to imagine a query for which the planner computes a high cost, but\\n> > > actually runs quickly. You might get a bunch of WARNINGs that the query took\\n> > > 10 MS and JIT was 75% of that, even if you don't care about queries that take\\n> > > less than 10 SEC.\\n> >\\n> > Yeah, that's definitely a problem. But which way would you want to tie\\n> > it to log_min_duration_statement? That a statement would both have to\\n> > take longer than log_min_duration_statement *and* have JIT above a\\n> > certain percentage? In my experience that is instead likely to miss\\n> > most of the interesting times.\\n>\\n> I don't understand - why doesn't it combine trivially with\\n> log_min_duration_statement? Are you saying that the default / pre-existing min\\n> duration may not log all of the intended queries ? I think that just means the\\n> configuration needs to be changed. The GUC needs to allow/help finding these\\n> JIT issues, but going to require an admin's interaction in any case. Avoiding\\n> false positives may be important for it to be useful at all.\\n>\\n> Hmm .. should it also combine with min_sample_rate ? Maybe that addresses your\\n> concern.\\n\\nFor one, what if I don't want to log any queries unless this is the\\ncase? I leave log_min_duration_statement at -1...\\nOr what if I want to log say only queries taking more than 60 seconds.\\nNow I will miss all queries taking 30 seconds where 99.9% of the time\\nis spent on JIT which I definitely *would* want.\\n\\nIf all I wanted was to get the JIT information for all queries over a\\ncertain amount of time, then as someone else mentioned as well, I can\\njust use auto_explain. The point here is I want to trigger the logging\\nfor queries where the runtime is *shorter* than what I have in\\nlog_min_duration_statement, but a large portion of the runtime is\\nspent on JIT.\\n\\nI maen. If I have a bunch of queries that take 10ms and 75% is spent\\nin JIT (per your example above), that is actually a problem and I'd\\nlike to know about it and fix it (by tuning the jit triggers either\\nglobally or locally). Yes, at 10ms I *may* ignore it, and in general\\nI'd use pg_stat_statements for that. But at 100ms with 75% JIT it\\nstarts becoming exactly the situation that I intended this patch to\\nhelp with.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/\\n Work: https://www.redpill-linpro.com/\\n\\n\\n\",\n \"msg_date\": \"Mon, 7 Mar 2022 16:02:16 +0100\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 07, 2022 at 04:02:16PM +0100, Magnus Hagander wrote:\\n> On Mon, Mar 7, 2022 at 2:09 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > On Mon, Mar 07, 2022 at 01:10:32PM +0100, Magnus Hagander wrote:\\n> > > On Fri, Feb 25, 2022 at 5:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:\\n> > > >\\n> > > > I think it should be a NOTICE (or less?)\\n> > >\\n> > > Hmm. I'm not so sure.\\n> > >\\n> > > Other similar parameters often use LOG, but the downside of that is\\n> > > that it won't be sent to the client.\\n> > >\\n> > > The problem with using NOTICE is that it won't go to the log by\\n> > > default. It needs to be at least warning to do that.\\n> >\\n> > Anyone who uses this parameter is aleady going to be changing GUCs, so it\\n> > doesn't need to work by default. The people who are likely to enable this\\n> > already monitor their logs and have probably customized their logging\\n> > configuration.\\n> \\n> Sure, but if you set log_min_messagse to NOTICE you are likely going\\n> to flood your logs beyond usefulness. And the whole idea of this\\n> parameter is you can keep it on during \\\"normal times\\\" to catch\\n> outliers.\\n\\nI do set log_min_messages - not to NOTICE, but to INFO ;)\\n\\nWould NOTICEs really flood most people's logs ? From what ? They're aren't\\nthat many, and most seem to be for DDL and utility commands. We do, uh, more\\nof that than most people would say is reasonable.\\n\\nI see a couple hundred of these per day:\\n| PostGIS: Unable to compute statistics for...\\nAnd during deployment, a few hundred more for IF NOT EXIST commands.\\nThat's it.\\n\\nI still think this GUC should not be WARNING. If it's a warning, then *this*\\nwill cause previously-quiet logs to be flooded - the opposite problem. Which\\nis not desirable for a new guc.\\n\\nhttps://www.postgresql.org/docs/current/runtime-config-logging.html\\n\\nINFO \\tProvides information implicitly requested by the user, e.g., output from VACUUM VERBOSE. \\tINFO \\tINFORMATION\\nNOTICE \\tProvides information that might be helpful to users, e.g., notice of truncation of long identifiers. \\tNOTICE \\tINFORMATION\\nWARNING Provides warnings of likely problems, e.g., COMMIT outside a transaction block. \\tNOTICE \\tWARNING\\n\\n> > I don't understand - why doesn't it combine trivially with\\n> > log_min_duration_statement? Are you saying that the default / pre-existing min\\n> > duration may not log all of the intended queries ? I think that just means the\\n> > configuration needs to be changed. The GUC needs to allow/help finding these\\n> > JIT issues, but going to require an admin's interaction in any case. Avoiding\\n> > false positives may be important for it to be useful at all.\\n> >\\n> > Hmm .. should it also combine with min_sample_rate ? Maybe that addresses your\\n> > concern.\\n> \\n> For one, what if I don't want to log any queries unless this is the\\n> case? I leave log_min_duration_statement at -1...\\n\\nThen you will conclude that our logging is inadequate for your case.\\nYou need to filter the lines which don't interest you.\\n\\n> Or what if I want to log say only queries taking more than 60 seconds.\\n> Now I will miss all queries taking 30 seconds where 99.9% of the time\\n> is spent on JIT which I definitely *would* want.\\n\\nIf you're unwilling to deal with log entries which are uninteresting, then\\nclearly you'll need a separate GUC just for log_min_duration_jit_above_fraction\\n(and maybe another one for jit_above_duration_log_level). Or implement generic\\nlog filtering based on file/severity/message.\\n\\nAs I see it: most people won't set this GUC at all. Those who do might enable\\nit with a high threshold value of log_min_duration_statement (probably a\\npre-existing, configured value) to see what there is to see - the worst issues.\\n\\nA small fraction of people will enable it with a low threshold value for\\nlog_min_duration_statement - maybe only temporarily or per-session. They'll be\\nwilling to sift through the logs to look for interesting things, like queries\\nthat take 10ms, 75% of which is in JIT, but run hundreds of times per second.\\n\\nThis feature intends to make it easier to identify queries affected by this\\nproblem, but it doesn't need to be possible to do that without logging anything\\nelse.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sat, 12 Mar 2022 09:24:11 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-07 13:10:32 +0100, Magnus Hagander wrote:\\n> Meanwhile here is an updated based on your other comments above, as\\n> well as those from Julien.\\n\\nThis fails on cfbot, due to compiler warnings:\\nhttps://cirrus-ci.com/task/5127667648299008?logs=mingw_cross_warning#L390\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Mar 2022 16:50:48 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 22, 2022 at 12:50 AM Andres Freund <andres@anarazel.de> wrote:\\n\\n> Hi,\\n>\\n> On 2022-03-07 13:10:32 +0100, Magnus Hagander wrote:\\n> > Meanwhile here is an updated based on your other comments above, as\\n> > well as those from Julien.\\n>\\n> This fails on cfbot, due to compiler warnings:\\n> https://cirrus-ci.com/task/5127667648299008?logs=mingw_cross_warning#L390\\n\\n\\nHuh. That's annoying. I forgot int64 is %d on linux and %lld on Windows :/\\n\\nPFA a fix for that.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\",\n \"msg_date\": \"Mon, 28 Mar 2022 22:11:16 +0200\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Mon, Mar 28, 2022 at 10:11:16PM +0200, Magnus Hagander wrote:\\n> On Tue, Mar 22, 2022 at 12:50 AM Andres Freund <andres@anarazel.de> wrote:\\n> >\\n> > This fails on cfbot, due to compiler warnings:\\n> > https://cirrus-ci.com/task/5127667648299008?logs=mingw_cross_warning#L390\\n>\\n>\\n> Huh. That's annoying. I forgot int64 is %d on linux and %lld on Windows :/\\n>\\n> PFA a fix for that.\\n\\nThe v3 is now stable on all platforms according to the cfbot, and apart from\\nthe log level the patch looks good to me.\\n\\nThe last remaining thing is whether logging at WARNING level is the correct\\nchoice. I'm personally fine with it, because the people who are going to use\\nit will probably use the same approach as for log_min_duration_statements:\\nenable it first with a high value, and check if that just lead to a massive log\\nspam. If not, see if there's anything that needs attention and fix it,\\notherwise keep lowering it and keep going.\\n\\nAlso, jitting isn't particularly fast in general, so it's probably somewhat\\nunlikely to have a massive amount of queries that both have a cost higher than\\nthe jit threshold cost and still run fast enough to spam the log without jit\\nslowing everything down.\\n\\nI'm not sure what status the patch should be. I think it's the committer's\\ndecision to chose which log level to use, and that won't affect the patch\\nenough to justify another review, so I'm switching the patch to ready for\\ncommitter. This doesn't prevent any further discussions for the log level.\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 18:08:46 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 29, 2022 at 6:09 AM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> The last remaining thing is whether logging at WARNING level is the correct\\n> choice. I'm personally fine with it, because the people who are going to use\\n> it will probably use the same approach as for log_min_duration_statements:\\n> enable it first with a high value, and check if that just lead to a massive log\\n> spam. If not, see if there's anything that needs attention and fix it,\\n> otherwise keep lowering it and keep going.\\n\\nI think WARNING is fine. After all, the parameter is called\\n\\\"jit_warn_above_fraction\\\". Yeah, we could also rename the parameter,\\nbut I think \\\"jit_notice_above_fraction\\\" would be harder to understand.\\nIt feels very intuitive to just say \\\"warn me if X thing happens\\\" and I\\ndon't see a reason why we shouldn't just do that.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 09:38:29 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\\n> I think WARNING is fine. After all, the parameter is called\\n> \\\"jit_warn_above_fraction\\\".\\n\\nI had a think about this patch. I guess it's a little similar to\\ncheckpoint_warning. The good thing about the checkpoint_warning is\\nthat in the LOG message we give instructions about how the DBA can fix\\nthe issue, i.e increase max_wal_size.\\n\\nWith the proposed patch I see there is no hint about what might be\\ndone to remove/reduce the warnings. I imagine that's because it's not\\nall that clear which GUC should be changed. In my view, likely\\njit_above_cost is the most relevant but there is also\\njit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\\nand jit_expressions which are relevant too.\\n\\nIf we go with this patch, the problem I see here is that the amount\\nof work the JIT compiler must do for a given query depends mostly on\\nthe number of expressions that must be compiled in the query (also to\\na lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\njit_tuple_deforming and jit_expressions). The DBA does not really have\\nmuch control over the number of expressions in the query. All he or\\nshe can do to get rid of the warning is something like increase\\njit_above_cost. After a few iterations of that, the end result is\\nthat jit_above_cost is now high enough that JIT no longer triggers\\nfor, say, that query to that table with 1000 partitions where no\\nplan-time pruning takes place. Is that really a good thing? It likely\\nmeans that we just rarely JIT anything at all!\\n\\nI really believe that the main problem here is that JIT only enables\\nwhen the *total* plan cost reaches a certain threshold. The number of\\nexpressions to be compiled is not a factor in the decision at all.\\nThat means that even if the total execution time of a plan was a true\\nreflection of the total estimated plan cost, then the fraction of time\\nspent (as is measured by jit_warn_above_fraction) doing JIT would\\nentirely depend on the number of expressions to compile. Of course,\\nthe planner's not that good, but does that not indicate that the JIT\\ncosting should really account for the number of expressions and not\\njust the total plan cost?\\n\\nAnyway, what I'm trying to indicate here is that JIT is pretty much\\nimpossible to tune properly and I don't really see why adding a\\nwarning about it not being tuned correctly would help anyone. I think\\nit would be better to focus on making improvements to how the JIT\\ncosting works.\\n\\nI did propose a patch to address this in [1]. It does need more work\\nand I do plan to come back to it for v16.\\n\\nI'd much rather see us address the costing problem before adding some\\nwarning, especially a warning where it's not clear how to make go\\naway.\\n\\nDavid\\n\\n[1] https://www.postgresql.org/message-id/CAApHDvpQJqLrNOSi8P1JLM8YE2C+ksKFpSdZg=q6sTbtQ-v=aw@mail.gmail.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 09:05:44 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 29, 2022 at 1:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n> That means that even if the total execution time of a plan was a true\\n> reflection of the total estimated plan cost, then the fraction of time\\n> spent (as is measured by jit_warn_above_fraction) doing JIT would\\n> entirely depend on the number of expressions to compile. Of course,\\n> the planner's not that good, but does that not indicate that the JIT\\n> costing should really account for the number of expressions and not\\n> just the total plan cost?\\n\\nThat's a good point. The difference between the actual cost of\\nexecuting the query with and without JIT'ing is what we care about,\\nfor the most part. Maybe we could do a lot better just by inventing a\\nfairly crude model that captures the benefits of JIT'ing -- that's\\nwhat \\\"counting the number of expressions\\\" sounds like to me. This\\nmodel could probably assume that JIT'ing itself was free -- maybe\\nsomething this simple would work well.\\n\\nThe planner has traditionally used the cost units to determine the\\ncheapest plan; it compared total plan cost for plans that were taken\\nfrom the universe of possible plans for *one specific query*. That's\\ncompletely different to any model that expects plan costs to be\\nmeaningful in an absolute sense. I'm not completely sure how much that\\ndifference matters, but I suspect that the answer is: \\\"it depends, but\\noften it matters a great deal\\\".\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 13:35:53 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Peter Geoghegan <pg@bowt.ie> writes:\\n> On Tue, Mar 29, 2022 at 1:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n>> That means that even if the total execution time of a plan was a true\\n>> reflection of the total estimated plan cost, then the fraction of time\\n>> spent (as is measured by jit_warn_above_fraction) doing JIT would\\n>> entirely depend on the number of expressions to compile. Of course,\\n>> the planner's not that good, but does that not indicate that the JIT\\n>> costing should really account for the number of expressions and not\\n>> just the total plan cost?\\n\\n> That's a good point.\\n\\nI think David's questions are sufficiently cogent and difficult\\nthat we should not add jit_warn_above_fraction at this time.\\nMaybe we'll eventually decide it's the best we can do; but I do\\nnot think we have a problem there that's so pressing that we need\\nto rush out a partially-baked bandaid. Especially not at the\\ntail end of the development cycle, with not much time to gain\\nexperience with it before we ship.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 18:04:32 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 29, 2022 at 3:04 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> I think David's questions are sufficiently cogent and difficult\\n> that we should not add jit_warn_above_fraction at this time.\\n\\n+1\\n\\n-- \\nPeter Geoghegan\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 15:19:58 -0700\",\n \"msg_from\": \"Peter Geoghegan <pg@bowt.ie>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-30 09:05:44 +1300, David Rowley wrote:\\n> I really believe that the main problem here is that JIT only enables\\n> when the *total* plan cost reaches a certain threshold.\\n\\nYes, that is/was a clear design mistake. It wasn't quite as bad back when it\\nwas written - partitioning blows the problem up by an order of magnitude or\\nthree. . The costing really needs to at least multiply the number of\\nto-be-compiled expressions with some cost to decide whether the cost of JITing\\nis worth it, rather than making \\\"flat\\\" decision.\\n\\n\\n> I did propose a patch to address this in [1]. It does need more work\\n> and I do plan to come back to it for v16.\\n\\nFWIW, that doesn't seem quite right - won't it stop JITing e.g. on the inner\\nside of a nested loop, just because it's cheap, even though that's where the\\nbulk of the benefits comes from?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 16:16:41 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Wed, 30 Mar 2022 at 12:16, Andres Freund <andres@anarazel.de> wrote:\\n> > I did propose a patch to address this in [1]. It does need more work\\n> > and I do plan to come back to it for v16.\\n>\\n> FWIW, that doesn't seem quite right - won't it stop JITing e.g. on the inner\\n> side of a nested loop, just because it's cheap, even though that's where the\\n> bulk of the benefits comes from?\\n\\nYeah, I think the total cost would need to be multiplied by the number\\nof times we expect to call that part of the plan. I've not yet sat\\ndown to figure out if that's easy/cheap or hard/costly information to\\nobtain.\\n\\nDavid\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 12:41:41 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"David Rowley <dgrowleyml@gmail.com> writes:\\n> On Wed, 30 Mar 2022 at 12:16, Andres Freund <andres@anarazel.de> wrote:\\n>> FWIW, that doesn't seem quite right - won't it stop JITing e.g. on the inner\\n>> side of a nested loop, just because it's cheap, even though that's where the\\n>> bulk of the benefits comes from?\\n\\n> Yeah, I think the total cost would need to be multiplied by the number\\n> of times we expect to call that part of the plan. I've not yet sat\\n> down to figure out if that's easy/cheap or hard/costly information to\\n> obtain.\\n\\nAs long as you don't need the info till the end of planning, it should be\\nreasonably simple to determine. I'm not sure if we actually record the\\nexpected number of loops in the plan tree today, but the costing\\nmechanisms certainly estimate that along the way, so we could store it\\nif need be.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 19:48:21 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-30 12:41:41 +1300, David Rowley wrote:\\n> On Wed, 30 Mar 2022 at 12:16, Andres Freund <andres@anarazel.de> wrote:\\n> > > I did propose a patch to address this in [1]. It does need more work\\n> > > and I do plan to come back to it for v16.\\n> >\\n> > FWIW, that doesn't seem quite right - won't it stop JITing e.g. on the inner\\n> > side of a nested loop, just because it's cheap, even though that's where the\\n> > bulk of the benefits comes from?\\n>\\n> Yeah, I think the total cost would need to be multiplied by the number\\n> of times we expect to call that part of the plan. I've not yet sat\\n> down to figure out if that's easy/cheap or hard/costly information to\\n> obtain.\\n\\nI wonder whether it'd make sense to combine that with awareness of a few plan\\ntypes that can lead to large portions of child nodes never being executed. One\\nthe case where the current behaviour is the worst is runtime partition pruning\\nin append - we compile expressions for whole subtrees that will never be\\nexecuted. We should be much more hesitant to compile there compared to a\\ncheap-ish node that we know will be executed as part of a large expensive part\\nof the plan tree.\\n\\n- Andres\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 17:20:30 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Wed, 30 Mar 2022 at 13:20, Andres Freund <andres@anarazel.de> wrote:\\n> I wonder whether it'd make sense to combine that with awareness of a few plan\\n> types that can lead to large portions of child nodes never being executed. One\\n> the case where the current behaviour is the worst is runtime partition pruning\\n> in append - we compile expressions for whole subtrees that will never be\\n> executed. We should be much more hesitant to compile there compared to a\\n> cheap-ish node that we know will be executed as part of a large expensive part\\n> of the plan tree.\\n\\nI think that's also a problem but I think that might be better fixed\\nanother way.\\n\\nThere is a patch [1] around that seems to change things to compile JIT\\non-demand. I've not looked at the patch but imagine the overhead\\nmight be kept minimal by initially setting the evalfunc to compile and\\nrun, then set it to just run the compiled Expr for subsequent\\nexecutions. Maybe nodes below an Append/MergeAppend with run-time\\npruning could compile on-demand and other nodes up-front. Or maybe\\nthere's no problem with making everything on-demand.\\n\\nDavid\\n\\n[1] https://commitfest.postgresql.org/37/3071/\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 14:30:32 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-30 14:30:32 +1300, David Rowley wrote:\\n> On Wed, 30 Mar 2022 at 13:20, Andres Freund <andres@anarazel.de> wrote:\\n> > I wonder whether it'd make sense to combine that with awareness of a few plan\\n> > types that can lead to large portions of child nodes never being executed. One\\n> > the case where the current behaviour is the worst is runtime partition pruning\\n> > in append - we compile expressions for whole subtrees that will never be\\n> > executed. We should be much more hesitant to compile there compared to a\\n> > cheap-ish node that we know will be executed as part of a large expensive part\\n> > of the plan tree.\\n> \\n> I think that's also a problem but I think that might be better fixed\\n> another way.\\n> \\n> There is a patch [1] around that seems to change things to compile JIT\\n> on-demand.\\n\\nThat's a bad idea idea to do on a per-function basis. Emitting functions one\\nby one is considerably slower than doing so in larger chunks. Of course that\\ncan be addressed to some degree by something like what you suggest below.\\n\\n\\n> I've not looked at the patch but imagine the overhead might be kept minimal\\n> by initially setting the evalfunc to compile and run, then set it to just\\n> run the compiled Expr for subsequent executions.\\n\\nThat part I'm not worried about, there's such an indirection on the first call\\neither way IIRC.\\n\\n\\n> Maybe nodes below an Append/MergeAppend with run-time pruning could compile\\n> on-demand and other nodes up-front. Or maybe there's no problem with making\\n> everything on-demand.\\n\\nYea, that could work. The expressions for one \\\"partition query\\\" would still\\nhave to be emitted at once. For each such subtree we should make a separate\\ncosting decision. But I think an additional \\\"will be executed\\\" sub-node is a\\ndifferent story, the threshold shouldn't be done on a per-node basis. That\\npartitioning of the plan tree is kind of what I was trying to get at...\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 29 Mar 2022 18:48:30 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Wed, 30 Mar 2022 at 14:48, Andres Freund <andres@anarazel.de> wrote:\\n>\\n> On 2022-03-30 14:30:32 +1300, David Rowley wrote:\\n> > Maybe nodes below an Append/MergeAppend with run-time pruning could compile\\n> > on-demand and other nodes up-front. Or maybe there's no problem with making\\n> > everything on-demand.\\n>\\n> Yea, that could work. The expressions for one \\\"partition query\\\" would still\\n> have to be emitted at once. For each such subtree we should make a separate\\n> costing decision. But I think an additional \\\"will be executed\\\" sub-node is a\\n> different story, the threshold shouldn't be done on a per-node basis. That\\n> partitioning of the plan tree is kind of what I was trying to get at...\\n\\nMaybe this point is moot if we get something like [1]. It might mean\\nthat run-time pruning would happen early enough that we could just JIT\\ncompile non-pruned subnodes.\\n\\nDavid\\n\\n[1] https://commitfest.postgresql.org/37/3478/\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 15:14:12 +1300\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n\\n> On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\\n> > I think WARNING is fine. After all, the parameter is called\\n> > \\\"jit_warn_above_fraction\\\".\\n>\\n> I had a think about this patch. I guess it's a little similar to\\n> checkpoint_warning. The good thing about the checkpoint_warning is\\n> that in the LOG message we give instructions about how the DBA can fix\\n> the issue, i.e increase max_wal_size.\\n>\\n> With the proposed patch I see there is no hint about what might be\\n> done to remove/reduce the warnings. I imagine that's because it's not\\n> all that clear which GUC should be changed. In my view, likely\\n> jit_above_cost is the most relevant but there is also\\n> jit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\\n> and jit_expressions which are relevant too.\\n>\\n> If we go with this patch, the problem I see here is that the amount\\n> of work the JIT compiler must do for a given query depends mostly on\\n> the number of expressions that must be compiled in the query (also to\\n> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\n> jit_tuple_deforming and jit_expressions). The DBA does not really have\\n> much control over the number of expressions in the query. All he or\\n> she can do to get rid of the warning is something like increase\\n> jit_above_cost. After a few iterations of that, the end result is\\n> that jit_above_cost is now high enough that JIT no longer triggers\\n> for, say, that query to that table with 1000 partitions where no\\n> plan-time pruning takes place. Is that really a good thing? It likely\\n> means that we just rarely JIT anything at all!\\n>\\n\\nI don't agree with the conclusion of that.\\n\\nWhat the parameter would be useful for is to be able to tune those costs\\n(or just turn it off) *for that individual query*. That doesn't mean you\\n\\\"rarely JIT anything atll\\\", it just means you rarely JIT that particular\\nquery.\\n\\nIn fact, my goal is to specifically make people do that and *not* just turn\\noff JIT globally.\\n\\n\\nI'd much rather see us address the costing problem before adding some\\n> warning, especially a warning where it's not clear how to make go\\n> away.\\n>\\n\\nThe easiest way would be to add a HINT that says turn off jit for this\\nparticular query or something?\\n\\nI do agree that if we can make \\\"spending too much time on JIT vs query\\nruntime\\\" go away completely, then there is no need for a parameter like\\nthis.\\n\\nI still think the warning is useful. And I think it may stay useful even\\nafter we have made the JIT costing smarter -- though that's not certain of\\ncourse.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\\n\\nOn Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\\n> I think WARNING is fine. After all, the parameter is called\\n> \\\"jit_warn_above_fraction\\\".\\n\\nI had a think about this patch. I guess it's a little similar to\\ncheckpoint_warning. The good thing about the checkpoint_warning is\\nthat in the LOG message we give instructions about how the DBA can fix\\nthe issue, i.e increase max_wal_size.\\n\\nWith the proposed patch I see there is no hint about what might be\\ndone to remove/reduce the warnings. I imagine that's because it's not\\nall that clear which GUC should be changed. In my view, likely\\njit_above_cost is the most relevant but there is also\\njit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\\nand jit_expressions which are relevant too.\\n\\nIf we go with this patch, the problem I see here is that the amount\\nof work the JIT compiler must do for a given query depends mostly on\\nthe number of expressions that must be compiled in the query (also to\\na lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\njit_tuple_deforming and jit_expressions). The DBA does not really have\\nmuch control over the number of expressions in the query. All he or\\nshe can do to get rid of the warning is something like increase\\njit_above_cost. After a few iterations of that, the end result is\\nthat jit_above_cost is now high enough that JIT no longer triggers\\nfor, say, that query to that table with 1000 partitions where no\\nplan-time pruning takes place. Is that really a good thing? It likely\\nmeans that we just rarely JIT anything at all!I don't agree with the conclusion of that.What the parameter would be useful for is to be able to tune those costs (or just turn it off) *for that individual query*. That doesn't mean you \\\"rarely JIT anything atll\\\", it just means you rarely JIT that particular query.In fact, my goal is to specifically make people do that and *not* just turn off JIT globally.I'd much rather see us address the costing problem before adding some\\nwarning, especially a warning where it's not clear how to make go\\naway.The easiest way would be to add a HINT that says turn off jit for this particular query or something?I do agree that if we can make \\\"spending too much time on JIT vs query runtime\\\" go away completely, then there is no need for a parameter like this. I still think the warning is useful. And I think it may stay useful even after we have made the JIT costing smarter -- though that's not certain of course. -- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Wed, 30 Mar 2022 15:04:30 +0200\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net> wrote:\\n\\n> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com>\\n> wrote:\\n>\\n>> On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\\n>> > I think WARNING is fine. After all, the parameter is called\\n>> > \\\"jit_warn_above_fraction\\\".\\n>>\\n>> I had a think about this patch. I guess it's a little similar to\\n>> checkpoint_warning. The good thing about the checkpoint_warning is\\n>> that in the LOG message we give instructions about how the DBA can fix\\n>> the issue, i.e increase max_wal_size.\\n>>\\n>> With the proposed patch I see there is no hint about what might be\\n>> done to remove/reduce the warnings. I imagine that's because it's not\\n>> all that clear which GUC should be changed. In my view, likely\\n>> jit_above_cost is the most relevant but there is also\\n>> jit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\\n>> and jit_expressions which are relevant too.\\n>>\\n>> If we go with this patch, the problem I see here is that the amount\\n>> of work the JIT compiler must do for a given query depends mostly on\\n>> the number of expressions that must be compiled in the query (also to\\n>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\n>> jit_tuple_deforming and jit_expressions). The DBA does not really have\\n>> much control over the number of expressions in the query. All he or\\n>> she can do to get rid of the warning is something like increase\\n>> jit_above_cost. After a few iterations of that, the end result is\\n>> that jit_above_cost is now high enough that JIT no longer triggers\\n>> for, say, that query to that table with 1000 partitions where no\\n>> plan-time pruning takes place. Is that really a good thing? It likely\\n>> means that we just rarely JIT anything at all!\\n>>\\n>\\n> I don't agree with the conclusion of that.\\n>\\n> What the parameter would be useful for is to be able to tune those costs\\n> (or just turn it off) *for that individual query*. That doesn't mean you\\n> \\\"rarely JIT anything atll\\\", it just means you rarely JIT that particular\\n> query.\\n>\\n> In fact, my goal is to specifically make people do that and *not* just\\n> turn off JIT globally.\\n>\\n>\\n> I'd much rather see us address the costing problem before adding some\\n>> warning, especially a warning where it's not clear how to make go\\n>> away.\\n>>\\n>\\n> The easiest way would be to add a HINT that says turn off jit for this\\n> particular query or something?\\n>\\n> I do agree that if we can make \\\"spending too much time on JIT vs query\\n> runtime\\\" go away completely, then there is no need for a parameter like\\n> this.\\n>\\n> I still think the warning is useful. And I think it may stay useful even\\n> after we have made the JIT costing smarter -- though that's not certain of\\n> course.\\n>\\n>\\nThis patch is still sitting at \\\"ready for committer\\\".\\n\\nAs an example, I have added such a hint in the attached.\\n\\nI still stand by that this patch is better than nothing. Sure, I would\\nlove for us to adapt the JIT costing model and algorithm to make this not a\\nproblem. And once we've done that, we should remove the parameter again.\\n\\nIt's not on by default, and it's trivial to remove in the future.\\n\\n\\nYes, we're right up at the deadline. I'd still like to get it in, so I'd\\nreally appreciate some further voices :)\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\",\n \"msg_date\": \"Fri, 8 Apr 2022 13:27:06 +0200\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\nOn Fri, Apr 8, 2022 at 07:27 Magnus Hagander <magnus@hagander.net> wrote:\\n\\n> On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net>\\n> wrote:\\n>\\n>> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com>\\n>> wrote:\\n>>\\n>>> On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\\n>>> > I think WARNING is fine. After all, the parameter is called\\n>>> > \\\"jit_warn_above_fraction\\\".\\n>>>\\n>>> I had a think about this patch. I guess it's a little similar to\\n>>> checkpoint_warning. The good thing about the checkpoint_warning is\\n>>> that in the LOG message we give instructions about how the DBA can fix\\n>>> the issue, i.e increase max_wal_size.\\n>>>\\n>>> With the proposed patch I see there is no hint about what might be\\n>>> done to remove/reduce the warnings. I imagine that's because it's not\\n>>> all that clear which GUC should be changed. In my view, likely\\n>>> jit_above_cost is the most relevant but there is also\\n>>> jit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\\n>>> and jit_expressions which are relevant too.\\n>>>\\n>>> If we go with this patch, the problem I see here is that the amount\\n>>> of work the JIT compiler must do for a given query depends mostly on\\n>>> the number of expressions that must be compiled in the query (also to\\n>>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\n>>> jit_tuple_deforming and jit_expressions). The DBA does not really have\\n>>> much control over the number of expressions in the query. All he or\\n>>> she can do to get rid of the warning is something like increase\\n>>> jit_above_cost. After a few iterations of that, the end result is\\n>>> that jit_above_cost is now high enough that JIT no longer triggers\\n>>> for, say, that query to that table with 1000 partitions where no\\n>>> plan-time pruning takes place. Is that really a good thing? It likely\\n>>> means that we just rarely JIT anything at all!\\n>>>\\n>>\\n>> I don't agree with the conclusion of that.\\n>>\\n>> What the parameter would be useful for is to be able to tune those costs\\n>> (or just turn it off) *for that individual query*. That doesn't mean you\\n>> \\\"rarely JIT anything atll\\\", it just means you rarely JIT that particular\\n>> query.\\n>>\\n>> In fact, my goal is to specifically make people do that and *not* just\\n>> turn off JIT globally.\\n>>\\n>>\\n>> I'd much rather see us address the costing problem before adding some\\n>>> warning, especially a warning where it's not clear how to make go\\n>>> away.\\n>>>\\n>>\\n>> The easiest way would be to add a HINT that says turn off jit for this\\n>> particular query or something?\\n>>\\n>> I do agree that if we can make \\\"spending too much time on JIT vs query\\n>> runtime\\\" go away completely, then there is no need for a parameter like\\n>> this.\\n>>\\n>> I still think the warning is useful. And I think it may stay useful even\\n>> after we have made the JIT costing smarter -- though that's not certain of\\n>> course.\\n>>\\n>>\\n> This patch is still sitting at \\\"ready for committer\\\".\\n>\\n> As an example, I have added such a hint in the attached.\\n>\\n> I still stand by that this patch is better than nothing. Sure, I would\\n> love for us to adapt the JIT costing model and algorithm to make this not a\\n> problem. And once we've done that, we should remove the parameter again.\\n>\\n> It's not on by default, and it's trivial to remove in the future.\\n>\\n>\\n> Yes, we're right up at the deadline. I'd still like to get it in, so I'd\\n> really appreciate some further voices :)\\n>\\n\\nLooks reasonable to me, so +1. The default is has it off and so I seriously\\ndoubt it’ll cause any issues and it’ll be very handy on large and busy\\nsystems with lots of queries finding those that have a serious amount of\\ntime being spent in JIT (and hopefully avoid folks just turning jit off\\nacross the board, since that’s worse- we need more data on jit and need to\\nwork on improving it, not ending up with everyone turning it off).\\n\\nThanks!\\n\\nStephen\\n\\n>\\n\\nGreetings,On Fri, Apr 8, 2022 at 07:27 Magnus Hagander <magnus@hagander.net> wrote:On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net> wrote:On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:On Wed, 30 Mar 2022 at 02:38, Robert Haas <robertmhaas@gmail.com> wrote:\\n> I think WARNING is fine. After all, the parameter is called\\n> \\\"jit_warn_above_fraction\\\".\\n\\nI had a think about this patch. I guess it's a little similar to\\ncheckpoint_warning. The good thing about the checkpoint_warning is\\nthat in the LOG message we give instructions about how the DBA can fix\\nthe issue, i.e increase max_wal_size.\\n\\nWith the proposed patch I see there is no hint about what might be\\ndone to remove/reduce the warnings. I imagine that's because it's not\\nall that clear which GUC should be changed. In my view, likely\\njit_above_cost is the most relevant but there is also\\njit_inline_above_cost, jit_optimize_above_cost, jit_tuple_deforming\\nand jit_expressions which are relevant too.\\n\\nIf we go with this patch, the problem I see here is that the amount\\nof work the JIT compiler must do for a given query depends mostly on\\nthe number of expressions that must be compiled in the query (also to\\na lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\njit_tuple_deforming and jit_expressions). The DBA does not really have\\nmuch control over the number of expressions in the query. All he or\\nshe can do to get rid of the warning is something like increase\\njit_above_cost. After a few iterations of that, the end result is\\nthat jit_above_cost is now high enough that JIT no longer triggers\\nfor, say, that query to that table with 1000 partitions where no\\nplan-time pruning takes place. Is that really a good thing? It likely\\nmeans that we just rarely JIT anything at all!I don't agree with the conclusion of that.What the parameter would be useful for is to be able to tune those costs (or just turn it off) *for that individual query*. That doesn't mean you \\\"rarely JIT anything atll\\\", it just means you rarely JIT that particular query.In fact, my goal is to specifically make people do that and *not* just turn off JIT globally.I'd much rather see us address the costing problem before adding some\\nwarning, especially a warning where it's not clear how to make go\\naway.The easiest way would be to add a HINT that says turn off jit for this particular query or something?I do agree that if we can make \\\"spending too much time on JIT vs query runtime\\\" go away completely, then there is no need for a parameter like this. I still think the warning is useful. And I think it may stay useful even after we have made the JIT costing smarter -- though that's not certain of course. This patch is still sitting at \\\"ready for committer\\\".As an example, I have added such a hint in the attached.I still stand by that this patch is better than nothing. Sure, I would love for us to adapt the JIT costing model and algorithm to make this not a problem. And once we've done that, we should remove the parameter again.It's not on by default, and it's trivial to remove in the future. Yes, we're right up at the deadline. I'd still like to get it in, so I'd really appreciate some further voices :)Looks reasonable to me, so +1. The default is has it off and so I seriously doubt it’ll cause any issues and it’ll be very handy on large and busy systems with lots of queries finding those that have a serious amount of time being spent in JIT (and hopefully avoid folks just turning jit off across the board, since that’s worse- we need more data on jit and need to work on improving it, not ending up with everyone turning it off). Thanks!Stephen\",\n \"msg_date\": \"Fri, 8 Apr 2022 07:32:44 -0400\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Fri, 8 Apr 2022 at 23:27, Magnus Hagander <magnus@hagander.net> wrote:\\n>\\n>\\n>\\n> On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net> wrote:\\n>>\\n>> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n>>>\\n>>> If we go with this patch, the problem I see here is that the amount\\n>>> of work the JIT compiler must do for a given query depends mostly on\\n>>> the number of expressions that must be compiled in the query (also to\\n>>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\n>>> jit_tuple_deforming and jit_expressions). The DBA does not really have\\n>>> much control over the number of expressions in the query. All he or\\n>>> she can do to get rid of the warning is something like increase\\n>>> jit_above_cost. After a few iterations of that, the end result is\\n>>> that jit_above_cost is now high enough that JIT no longer triggers\\n>>> for, say, that query to that table with 1000 partitions where no\\n>>> plan-time pruning takes place. Is that really a good thing? It likely\\n>>> means that we just rarely JIT anything at all!\\n>>\\n>>\\n>> I don't agree with the conclusion of that.\\n>>\\n>> What the parameter would be useful for is to be able to tune those costs (or just turn it off) *for that individual query*. That doesn't mean you \\\"rarely JIT anything atll\\\", it just means you rarely JIT that particular query.\\n\\nI just struggle to imagine that anyone is going to spend much effort\\ntuning a warning parameter per query. I imagine they're far more\\nlikely to just ramp it up to only catch some high percentile problems\\nor just (more likely) just not bother with it. It seems more likely\\nthat if anyone was to tune anything per query here it would be\\njit_above_cost, since that actually might have an affect on the\\nperformance of the query, rather than if it spits out some warning\\nmessage or not. ISTM that if the user knows what to set it to per\\nquery, then there's very little point in having a warning as we'd be\\nalerting them to something they already know about.\\n\\nI looked in the -general list to see if we could get some common\\nexplanations to give us an idea of the most common reason for high JIT\\ncompilation time. It seems that the plans were never simple. [1] seems\\ndue to a complex plan. I'm basing that off the \\\"Functions: 167\\\". I\\ndidn't catch the full plan. From what I can tell, [2] seems to be due\\nto \\\"lots of empty tables\\\", so assuming the clamping at 1 page is\\ncausing issues there. I think both of those cases could be resolved\\nby building the costing the way I mentioned. I admit that 2 cases is\\nnot a very large sample size.\\n\\nDavid\\n\\n[1] https://www.postgresql.org/message-id/flat/CAPL5KHq8zfWPzueCemXw4c%2BU568PoDfqo3wBDNm3KAyvybdaMQ%40mail.gmail.com#35aca8b42c3862f44b6be5b260c1a109\\n[2] https://www.postgresql.org/message-id/flat/CAHOFxGo5xJt02RmwAWrtv2K0jcqqxG-cDiR8FQbvb0WxdKhcgw%40mail.gmail.com#12d91822e869a2e22ca830cb5632f549\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Apr 2022 00:18:44 +1200\",\n \"msg_from\": \"David Rowley <dgrowleyml@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 8, 2022 at 2:19 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n\\n> On Fri, 8 Apr 2022 at 23:27, Magnus Hagander <magnus@hagander.net> wrote:\\n> >\\n> >\\n> >\\n> > On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net>\\n> wrote:\\n> >>\\n> >> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com>\\n> wrote:\\n> >>>\\n> >>> If we go with this patch, the problem I see here is that the amount\\n> >>> of work the JIT compiler must do for a given query depends mostly on\\n> >>> the number of expressions that must be compiled in the query (also to\\n> >>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\n> >>> jit_tuple_deforming and jit_expressions). The DBA does not really have\\n> >>> much control over the number of expressions in the query. All he or\\n> >>> she can do to get rid of the warning is something like increase\\n> >>> jit_above_cost. After a few iterations of that, the end result is\\n> >>> that jit_above_cost is now high enough that JIT no longer triggers\\n> >>> for, say, that query to that table with 1000 partitions where no\\n> >>> plan-time pruning takes place. Is that really a good thing? It likely\\n> >>> means that we just rarely JIT anything at all!\\n> >>\\n> >>\\n> >> I don't agree with the conclusion of that.\\n> >>\\n> >> What the parameter would be useful for is to be able to tune those\\n> costs (or just turn it off) *for that individual query*. That doesn't mean\\n> you \\\"rarely JIT anything atll\\\", it just means you rarely JIT that\\n> particular query.\\n>\\n> I just struggle to imagine that anyone is going to spend much effort\\n> tuning a warning parameter per query. I imagine they're far more\\n> likely to just ramp it up to only catch some high percentile problems\\n> or just (more likely) just not bother with it. It seems more likely\\n> that if anyone was to tune anything per query here it would be\\n> jit_above_cost, since that actually might have an affect on the\\n> performance of the query, rather than if it spits out some warning\\n> message or not. ISTM that if the user knows what to set it to per\\n> query, then there's very little point in having a warning as we'd be\\n> alerting them to something they already know about.\\n>\\n\\nI would not expect people to tune the *warning* at a query level. If\\nanything, then ys, they would tune the either jit_above_cost or just\\njit=off. But the idea being you can do that on a per query level instead of\\nglobally.\\n\\n\\nI looked in the -general list to see if we could get some common\\n> explanations to give us an idea of the most common reason for high JIT\\n> compilation time. It seems that the plans were never simple. [1] seems\\n> due to a complex plan. I'm basing that off the \\\"Functions: 167\\\". I\\n> didn't catch the full plan. From what I can tell, [2] seems to be due\\n> to \\\"lots of empty tables\\\", so assuming the clamping at 1 page is\\n> causing issues there. I think both of those cases could be resolved\\n> by building the costing the way I mentioned. I admit that 2 cases is\\n> not a very large sample size.\\n>\\n\\nAgain, I am very much for improvements of the costing model. This is in no\\nway intended to be a replacement for that. It's intended to be a stop-gap.\\n\\nWhat I see much of today are things like\\nhttps://dba.stackexchange.com/questions/264955/handling-performance-problems-with-jit-in-postgres-12\\nor\\nhttps://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\\n\\nThe bottom line is that people end up with recommendations to turn off JIT\\nglobally more or less by default. Because there's no real useful way today\\nto figure out when it causes problems vs when it helps.\\n\\nThe addition to pg_stat_statements I pushed a short while ago would help\\nwith that. But I think having a warning like this would also be useful. As\\na stop-gap measure, yes, but we really don't know when we will have an\\nimproved costing model for it. I hope you're right and that we can have it\\nby 16, and then I will definitely advocate for removing the warning again\\nif it works.\\n\\n-- \\n Magnus Hagander\\n Me: https://www.hagander.net/ <http://www.hagander.net/>\\n Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/>\\n\\nOn Fri, Apr 8, 2022 at 2:19 PM David Rowley <dgrowleyml@gmail.com> wrote:On Fri, 8 Apr 2022 at 23:27, Magnus Hagander <magnus@hagander.net> wrote:\\n>\\n>\\n>\\n> On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net> wrote:\\n>>\\n>> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n>>>\\n>>> If we go with this patch, the problem I see here is that the amount\\n>>> of work the JIT compiler must do for a given query depends mostly on\\n>>> the number of expressions that must be compiled in the query (also to\\n>>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\n>>> jit_tuple_deforming and jit_expressions). The DBA does not really have\\n>>> much control over the number of expressions in the query. All he or\\n>>> she can do to get rid of the warning is something like increase\\n>>> jit_above_cost. After a few iterations of that, the end result is\\n>>> that jit_above_cost is now high enough that JIT no longer triggers\\n>>> for, say, that query to that table with 1000 partitions where no\\n>>> plan-time pruning takes place. Is that really a good thing? It likely\\n>>> means that we just rarely JIT anything at all!\\n>>\\n>>\\n>> I don't agree with the conclusion of that.\\n>>\\n>> What the parameter would be useful for is to be able to tune those costs (or just turn it off) *for that individual query*. That doesn't mean you \\\"rarely JIT anything atll\\\", it just means you rarely JIT that particular query.\\n\\nI just struggle to imagine that anyone is going to spend much effort\\ntuning a warning parameter per query. I imagine they're far more\\nlikely to just ramp it up to only catch some high percentile problems\\nor just (more likely) just not bother with it. It seems more likely\\nthat if anyone was to tune anything per query here it would be\\njit_above_cost, since that actually might have an affect on the\\nperformance of the query, rather than if it spits out some warning\\nmessage or not. ISTM that if the user knows what to set it to per\\nquery, then there's very little point in having a warning as we'd be\\nalerting them to something they already know about.I would not expect people to tune the *warning* at a query level. If anything, then ys, they would tune the either jit_above_cost or just jit=off. But the idea being you can do that on a per query level instead of globally.\\nI looked in the -general list to see if we could get some common\\nexplanations to give us an idea of the most common reason for high JIT\\ncompilation time. It seems that the plans were never simple. [1] seems\\ndue to a complex plan. I'm basing that off the \\\"Functions: 167\\\". I\\ndidn't catch the full plan. From what I can tell, [2] seems to be due\\nto \\\"lots of empty tables\\\", so assuming the clamping at 1 page is\\ncausing issues there. I think both of those cases could be resolved\\nby building the costing the way I mentioned. I admit that 2 cases is\\nnot a very large sample size.Again, I am very much for improvements of the costing model. This is in no way intended to be a replacement for that. It's intended to be a stop-gap. What I see much of today are things like https://dba.stackexchange.com/questions/264955/handling-performance-problems-with-jit-in-postgres-12 or https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2aThe bottom line is that people end up with recommendations to turn off JIT globally more or less by default. Because there's no real useful way today to figure out when it causes problems vs when it helps.The addition to pg_stat_statements I pushed a short while ago would help with that. But I think having a warning like this would also be useful. As a stop-gap measure, yes, but we really don't know when we will have an improved costing model for it. I hope you're right and that we can have it by 16, and then I will definitely advocate for removing the warning again if it works.-- Magnus Hagander Me: https://www.hagander.net/ Work: https://www.redpill-linpro.com/\",\n \"msg_date\": \"Fri, 8 Apr 2022 14:28:36 +0200\",\n \"msg_from\": \"Magnus Hagander <magnus@hagander.net>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Greetings,\\n\\n* Magnus Hagander (magnus@hagander.net) wrote:\\n> On Fri, Apr 8, 2022 at 2:19 PM David Rowley <dgrowleyml@gmail.com> wrote:\\n> > On Fri, 8 Apr 2022 at 23:27, Magnus Hagander <magnus@hagander.net> wrote:\\n> > > On Wed, Mar 30, 2022 at 3:04 PM Magnus Hagander <magnus@hagander.net>\\n> > wrote:\\n> > >>\\n> > >> On Tue, Mar 29, 2022 at 10:06 PM David Rowley <dgrowleyml@gmail.com>\\n> > wrote:\\n> > >>>\\n> > >>> If we go with this patch, the problem I see here is that the amount\\n> > >>> of work the JIT compiler must do for a given query depends mostly on\\n> > >>> the number of expressions that must be compiled in the query (also to\\n> > >>> a lesser extent jit_inline_above_cost, jit_optimize_above_cost,\\n> > >>> jit_tuple_deforming and jit_expressions). The DBA does not really have\\n> > >>> much control over the number of expressions in the query. All he or\\n> > >>> she can do to get rid of the warning is something like increase\\n> > >>> jit_above_cost. After a few iterations of that, the end result is\\n> > >>> that jit_above_cost is now high enough that JIT no longer triggers\\n> > >>> for, say, that query to that table with 1000 partitions where no\\n> > >>> plan-time pruning takes place. Is that really a good thing? It likely\\n> > >>> means that we just rarely JIT anything at all!\\n> > >>\\n> > >>\\n> > >> I don't agree with the conclusion of that.\\n> > >>\\n> > >> What the parameter would be useful for is to be able to tune those\\n> > costs (or just turn it off) *for that individual query*. That doesn't mean\\n> > you \\\"rarely JIT anything atll\\\", it just means you rarely JIT that\\n> > particular query.\\n> >\\n> > I just struggle to imagine that anyone is going to spend much effort\\n> > tuning a warning parameter per query. I imagine they're far more\\n> > likely to just ramp it up to only catch some high percentile problems\\n> > or just (more likely) just not bother with it. It seems more likely\\n> > that if anyone was to tune anything per query here it would be\\n> > jit_above_cost, since that actually might have an affect on the\\n> > performance of the query, rather than if it spits out some warning\\n> > message or not. ISTM that if the user knows what to set it to per\\n> > query, then there's very little point in having a warning as we'd be\\n> > alerting them to something they already know about.\\n> \\n> I would not expect people to tune the *warning* at a query level. If\\n> anything, then ys, they would tune the either jit_above_cost or just\\n> jit=off. But the idea being you can do that on a per query level instead of\\n> globally.\\n\\nYeah, exactly, this is about having a busy system and wanting to know\\nwhich queries are spending a lot of time doing JIT relative to the query\\ntime, so that you can go adjust your JIT parameters or possibly disable\\nJIT for those queries (or maybe bring those cases to -hackers and try to\\nhelp make our costing better).\\n\\n> > I looked in the -general list to see if we could get some common\\n> > explanations to give us an idea of the most common reason for high JIT\\n> > compilation time. It seems that the plans were never simple. [1] seems\\n> > due to a complex plan. I'm basing that off the \\\"Functions: 167\\\". I\\n> > didn't catch the full plan. From what I can tell, [2] seems to be due\\n> > to \\\"lots of empty tables\\\", so assuming the clamping at 1 page is\\n> > causing issues there. I think both of those cases could be resolved\\n> > by building the costing the way I mentioned. I admit that 2 cases is\\n> > not a very large sample size.\\n> \\n> Again, I am very much for improvements of the costing model. This is in no\\n> way intended to be a replacement for that. It's intended to be a stop-gap.\\n\\nNot sure I'd say it's a 'stop-gap' as it's really very similar, imv\\nanyway, to log_min_duration_statement- you want to know what queries are\\ntaking a lot of time but you can't log all of them.\\n\\n> What I see much of today are things like\\n> https://dba.stackexchange.com/questions/264955/handling-performance-problems-with-jit-in-postgres-12\\n> or\\n> https://dev.to/xenatisch/cascade-of-doom-jit-and-how-a-postgres-update-led-to-70-failure-on-a-critical-national-service-3f2a\\n> \\n> The bottom line is that people end up with recommendations to turn off JIT\\n> globally more or less by default. Because there's no real useful way today\\n> to figure out when it causes problems vs when it helps.\\n\\nYeah, that's frustrating.\\n\\n> The addition to pg_stat_statements I pushed a short while ago would help\\n> with that. But I think having a warning like this would also be useful. As\\n> a stop-gap measure, yes, but we really don't know when we will have an\\n> improved costing model for it. I hope you're right and that we can have it\\n> by 16, and then I will definitely advocate for removing the warning again\\n> if it works.\\n\\nHaving this in pg_stat_statements is certainly helpful but having a\\nwarning also is. I don't think we have to address this in only one way.\\nA lot faster to flip this guc and then look in the logs on a busy system\\nthan to install pg_stat_statements, restart the cluster once you get\\npermission to do so, and then query it.\\n\\nThanks,\\n\\nStephen\",\n \"msg_date\": \"Fri, 8 Apr 2022 09:39:18 -0400\",\n \"msg_from\": \"Stephen Frost <sfrost@snowman.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Fri, Apr 08, 2022 at 09:39:18AM -0400, Stephen Frost wrote:\\n>\\n> * Magnus Hagander (magnus@hagander.net) wrote:\\n> > The addition to pg_stat_statements I pushed a short while ago would help\\n> > with that. But I think having a warning like this would also be useful. As\\n> > a stop-gap measure, yes, but we really don't know when we will have an\\n> > improved costing model for it. I hope you're right and that we can have it\\n> > by 16, and then I will definitely advocate for removing the warning again\\n> > if it works.\\n>\\n> Having this in pg_stat_statements is certainly helpful but having a\\n> warning also is. I don't think we have to address this in only one way.\\n> A lot faster to flip this guc and then look in the logs on a busy system\\n> than to install pg_stat_statements, restart the cluster once you get\\n> permission to do so, and then query it.\\n\\n+1, especially if you otherwise don't really need or want to have\\npg_stat_statements enabled, as it's far from being free. Sure you could enable\\nit by default with pg_stat_statements.track = none, but that seems a lot more\\ntroublesome than just dynamically enabling a GUC, possibly for a short time\\nand/or for a specific database/role.\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Apr 2022 12:20:57 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Julien Rouhaud <rjuju123@gmail.com> writes:\\n> On Fri, Apr 08, 2022 at 09:39:18AM -0400, Stephen Frost wrote:\\n>> Having this in pg_stat_statements is certainly helpful but having a\\n>> warning also is. I don't think we have to address this in only one way.\\n>> A lot faster to flip this guc and then look in the logs on a busy system\\n>> than to install pg_stat_statements, restart the cluster once you get\\n>> permission to do so, and then query it.\\n\\n> +1, especially if you otherwise don't really need or want to have\\n> pg_stat_statements enabled, as it's far from being free. Sure you could enable\\n> it by default with pg_stat_statements.track = none, but that seems a lot more\\n> troublesome than just dynamically enabling a GUC, possibly for a short time\\n> and/or for a specific database/role.\\n\\nThe arguments against the GUC were not about whether it's convenient.\\nThey were about whether the information it gives you is actually going\\nto be of any use.\\n\\nAlso, good luck with \\\"looking in the logs\\\", because by default\\nWARNING-level messages don't go to the postmaster log. If that's\\nthe intended use-case then the message ought to appear at LOG\\nlevel (which'd also change the desirable name for the GUC).\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 09 Apr 2022 10:42:12 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\\n>\\n> Also, good luck with \\\"looking in the logs\\\", because by default\\n> WARNING-level messages don't go to the postmaster log. If that's\\n> the intended use-case then the message ought to appear at LOG\\n> level (which'd also change the desirable name for the GUC).\\n\\nI must be missing something because by default log_min_messages is WARNING, and\\nAFAICS I do get WARNING-level messages in some vanilla cluster logs, using\\nvanilla .psqlrc.\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Apr 2022 23:21:02 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"Julien Rouhaud <rjuju123@gmail.com> writes:\\n> On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\\n>> Also, good luck with \\\"looking in the logs\\\", because by default\\n>> WARNING-level messages don't go to the postmaster log.\\n\\n> I must be missing something because by default log_min_messages is WARNING, and\\n> AFAICS I do get WARNING-level messages in some vanilla cluster logs, using\\n> vanilla .psqlrc.\\n\\nAh, sorry, I was looking at the wrong thing namely\\nlog_min_error_statement. The point still holds though: if we emit this\\nat level WARNING, what will get logged is just the bare message and not\\nthe statement that triggered it. How useful do you think that will be?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 09 Apr 2022 12:31:23 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Sat, Apr 09, 2022 at 12:31:23PM -0400, Tom Lane wrote:\\n> Julien Rouhaud <rjuju123@gmail.com> writes:\\n> > On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\\n> >> Also, good luck with \\\"looking in the logs\\\", because by default\\n> >> WARNING-level messages don't go to the postmaster log.\\n> \\n> > I must be missing something because by default log_min_messages is WARNING, and\\n> > AFAICS I do get WARNING-level messages in some vanilla cluster logs, using\\n> > vanilla .psqlrc.\\n> \\n> Ah, sorry, I was looking at the wrong thing namely\\n> log_min_error_statement. The point still holds though: if we emit this\\n> at level WARNING, what will get logged is just the bare message and not\\n> the statement that triggered it. How useful do you think that will be?\\n\\nAh right, I did miss that \\\"small detail\\\". And of course I agree, as-is that\\nwould be entirely useless and it should be LOG, like the rest of similar\\nfeatures.\\n\\n\\n\",\n \"msg_date\": \"Sun, 10 Apr 2022 00:43:08 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n },\n {\n \"msg_contents\": \"On Sat, Apr 9, 2022 at 8:43 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n\\n> On Sat, Apr 09, 2022 at 12:31:23PM -0400, Tom Lane wrote:\\n> > Julien Rouhaud <rjuju123@gmail.com> writes:\\n> > > On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\\n> > >> Also, good luck with \\\"looking in the logs\\\", because by default\\n> > >> WARNING-level messages don't go to the postmaster log.\\n> >\\n> > > I must be missing something because by default log_min_messages is\\n> WARNING, and\\n> > > AFAICS I do get WARNING-level messages in some vanilla cluster logs,\\n> using\\n> > > vanilla .psqlrc.\\n> >\\n> > Ah, sorry, I was looking at the wrong thing namely\\n> > log_min_error_statement. The point still holds though: if we emit this\\n> > at level WARNING, what will get logged is just the bare message and not\\n> > the statement that triggered it. How useful do you think that will be?\\n>\\n> Ah right, I did miss that \\\"small detail\\\". And of course I agree, as-is\\n> that\\n> would be entirely useless and it should be LOG, like the rest of similar\\n> features.\\n>\\n>\\n>\\nThe patch does not apply successfully; please rebase the patch.\\n\\npatching file src/backend/utils/misc/guc.c\\nHunk #1 FAILED at 642.\\nHunk #2 FAILED at 3943.\\n2 out of 2 hunks FAILED -- saving rejects to file\\nsrc/backend/utils/misc/guc.c.rej\\npatching file src/backend/utils/misc/postgresql.conf.sample\\npatching file src/include/jit/jit.h\\n\\n\\n-- \\nIbrar Ahmed\\n\\nOn Sat, Apr 9, 2022 at 8:43 PM Julien Rouhaud <rjuju123@gmail.com> wrote:On Sat, Apr 09, 2022 at 12:31:23PM -0400, Tom Lane wrote:\\n> Julien Rouhaud <rjuju123@gmail.com> writes:\\n> > On Sat, Apr 09, 2022 at 10:42:12AM -0400, Tom Lane wrote:\\n> >> Also, good luck with \\\"looking in the logs\\\", because by default\\n> >> WARNING-level messages don't go to the postmaster log.\\n> \\n> > I must be missing something because by default log_min_messages is WARNING, and\\n> > AFAICS I do get WARNING-level messages in some vanilla cluster logs, using\\n> > vanilla .psqlrc.\\n> \\n> Ah, sorry, I was looking at the wrong thing namely\\n> log_min_error_statement. The point still holds though: if we emit this\\n> at level WARNING, what will get logged is just the bare message and not\\n> the statement that triggered it. How useful do you think that will be?\\n\\nAh right, I did miss that \\\"small detail\\\". And of course I agree, as-is that\\nwould be entirely useless and it should be LOG, like the rest of similar\\nfeatures.\\n\\n\\nThe patch does not apply successfully; please rebase the patch.patching file src/backend/utils/misc/guc.c\\nHunk #1 FAILED at 642.\\nHunk #2 FAILED at 3943.\\n2 out of 2 hunks FAILED -- saving rejects to file src/backend/utils/misc/guc.c.rej\\npatching file src/backend/utils/misc/postgresql.conf.sample\\npatching file src/include/jit/jit.h-- Ibrar Ahmed\",\n \"msg_date\": \"Thu, 15 Sep 2022 12:02:17 +0400\",\n \"msg_from\": \"Ibrar Ahmed <ibrar.ahmad@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add parameter jit_warn_above_fraction\"\n }\n]","isConversational":false}}},{"rowIdx":53585,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi All,\n\nI've been working on adding Windows+MinGW environment into cirrus-ci tasks\n(discussion about ci is here [1]).\nIt uses MSYS2 to set the environment. UCRT is chosen as C standard library,\ninstead of MSVCRT.\nThe task configures postgres with features that are available in MSYS2 (see\navailable packages [2]) and tap tests are enabled.\nI already added the necessary docker image, you can find the related PR at\n[3] and a successful cirruc-ci run with these changes at [4].\nAttached patch adds a task runs on Windows with MinGW for cirrus-ci\n\nHowever, I cannot run configure with --with-python, --with-perl or\n--with-tcl.\nThere are two issues I encountered while trying to enable. e.g. for\n--with-python\n\n1- python_ldlibrary is set to \"libpython3.9.dll.a\". So the related line in\nconfigure cannot retrieve \"ldlibrary\"\n2- src/pl/plpython/Makefile looks under \"C:/Windows/system32/\" for\nPYTHONDLL. gendef cannot open that file, give an error like \" failed to\nopen()\" when creating a .def file.\n\nTo overcome those, I added the correct pattern to extract ldlibrary by\nappending \"-e 's/\\.dll.a$//'\" at the end of the related line.\nThen, I tried to use python dll from MSYS instead of windows/system32 by\nchanging PYTHONDLL path to \"/ucrt64/bin/libpython3.9.dll\". I'm not sure if\nthis is the correct dll.\nHere is the diff of these two changes:\n\ndiff --git a/configure b/configure\nindex f3cb5c2b51..42ea580442 100755\n--- a/configure\n+++ b/configure\n@@ -10536,7 +10536,7 @@ python_libdir=`${PYTHON} -c \"import sysconfig;\nprint(' '.join(filter(None,syscon\n python_ldlibrary=`${PYTHON} -c \"import sysconfig; print('\n'.join(filter(None,sysconfig.get_config_vars('LDLIBRARY'))))\"`\n\n # If LDLIBRARY exists and has a shlib extension, use it verbatim.\n-ldlibrary=`echo \"${python_ldlibrary}\" | sed -e 's/\\.so$//' -e 's/\\.dll$//'\n-e 's/\\.dylib$//' -e 's/\\.sl$//'`\n+ldlibrary=`echo \"${python_ldlibrary}\" | sed -e 's/\\.so$//' -e 's/\\.dll$//'\n-e 's/\\.dylib$//' -e 's/\\.sl$//' -e 's/\\.dll.a$//'`\n if test -e \"${python_libdir}/${python_ldlibrary}\" -a\nx\"${python_ldlibrary}\" != x\"${ldlibrary}\"\n then\n ldlibrary=`echo \"${ldlibrary}\" | sed \"s/^lib//\"`\ndiff --git a/src/pl/plpython/Makefile b/src/pl/plpython/Makefile\nindex a83ae8865c..4254ef94d7 100644\n--- a/src/pl/plpython/Makefile\n+++ b/src/pl/plpython/Makefile\n@@ -61,7 +61,8 @@ INCS = plpython.h \\\n ifeq ($(PORTNAME), win32)\n\n pytverstr=$(subst .,,${python_version})\n-PYTHONDLL=$(subst \\,/,$(WINDIR))/system32/python${pytverstr}.dll\n+#PYTHONDLL=$(subst \\,/,$(WINDIR))/system32/python${pytverstr}.dll\n+PYTHONDLL=/ucrt64/bin/libpython3.9.dll\n\n OBJS += libpython${pytverstr}.a\n\n\n\nIn the end, make check-world still fails, even though I was able to run\nconfigure and make without any obvious error.\nI see bunch of errors in tests like:\n+ERROR: language \"plpython3u\" does not exist\n+HINT: Use CREATE EXTENSION to load the language into the database.\n\nHere is the logs from failed ci run:\nhttps://api.cirrus-ci.com/v1/artifact/task/4645682031099904/log/build/src/pl/plpython/regression.diffs\n\n\nAny thoughts on how postgres can be built with --with-python etc. on mingw?\n\nBest,\nMelih\n\n\n[1]\nhttps://www.postgresql.org/message-id/flat/20211001222752.wrz7erzh4cajvgp6%40alap3.anarazel.de\n<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.postgresql.org%2Fmessage-id%2Fflat%2F20211001222752.wrz7erzh4cajvgp6%2540alap3.anarazel.de&data=04%7C01%7CMelih.Mutlu%40microsoft.com%7C7f4780e35ab34e2f38d808d9f7d008e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637813295184245763%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=HkjuBWcQzmiyIyoCuIfKCzAEF9ME6SpGOYjBJMFhQvw%3D&reserved=0>\n[2] https://packages.msys2.org/package/\n<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fpackages.msys2.org%2Fpackage%2F&data=04%7C01%7CMelih.Mutlu%40microsoft.com%7C7f4780e35ab34e2f38d808d9f7d008e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637813295184245763%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=VqfdnrxHv15FUGFXgnNaMli5n2R1tnPPglGq4xWqND0%3D&reserved=0>\n[3] https://github.com/anarazel/pg-vm-images/pull/8\n<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fanarazel%2Fpg-vm-images%2Fpull%2F8&data=04%7C01%7CMelih.Mutlu%40microsoft.com%7C7f4780e35ab34e2f38d808d9f7d008e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637813295184245763%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=IASIgPhLd0%2Fvh1VSn9EC%2BCsH42riE%2Bt6gQPObOb0QmI%3D&reserved=0>\n[4] https://cirrus-ci.com/build/4999469182746624\n<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fcirrus-ci.com%2Fbuild%2F5013577319841792&data=04%7C01%7CMelih.Mutlu%40microsoft.com%7C7f4780e35ab34e2f38d808d9f7d008e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637813295184245763%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=hMGCtgm1uVBWzFHXPEl%2FiaavzWbEqxAuhmKjFcjH%2BF0%3D&reserved=0>","msg_date":"Fri, 25 Feb 2022 19:44:27 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nAndrew, CCIng you both because you might be interested in the CI bit, and\nbecause you might know the answer.\n\nOn 2022-02-25 19:44:27 +0300, Melih Mutlu wrote:\n> I've been working on adding Windows+MinGW environment into cirrus-ci tasks\n> (discussion about ci is here [1]).\n> It uses MSYS2 to set the environment. UCRT is chosen as C standard library,\n> instead of MSVCRT.\n> The task configures postgres with features that are available in MSYS2 (see\n> available packages [2]) and tap tests are enabled.\n> I already added the necessary docker image, you can find the related PR at\n> [3] and a successful cirruc-ci run with these changes at [4].\n> Attached patch adds a task runs on Windows with MinGW for cirrus-ci\n>\n> However, I cannot run configure with --with-python, --with-perl or\n> --with-tcl.\n> There are two issues I encountered while trying to enable. e.g. for\n> --with-python\n>\n> 1- python_ldlibrary is set to \"libpython3.9.dll.a\". So the related line in\n> configure cannot retrieve \"ldlibrary\"\n\nThis presumably is due to using mingw's python rather than python.org's\npython. Seems like a reasonable thing to support for the mingw build.\n\nMelih, you could try to build against the python.org python (i installed in\nthe CI container).\n\n\n> 2- src/pl/plpython/Makefile looks under \"C:/Windows/system32/\" for\n> PYTHONDLL. gendef cannot open that file, give an error like \" failed to\n> open()\" when creating a .def file.\n\nOn my win10 VM in which I installed python.org python I don't see a python dll\nin c:/windows/system32 either. Looks like none of our windows mingw animals\nbuild with python. So it might just be bitrot.\n\n\n> In the end, make check-world still fails, even though I was able to run\n> configure and make without any obvious error.\n> I see bunch of errors in tests like:\n> +ERROR: language \"plpython3u\" does not exist\n> +HINT: Use CREATE EXTENSION to load the language into the database.\n\n> Here is the logs from failed ci run:\n> https://api.cirrus-ci.com/v1/artifact/task/4645682031099904/log/build/src/pl/plpython/regression.diffs\n\nThe URL to the rest of the CI run is https://cirrus-ci.com/task/4645682031099904\n\n\nThe relevant failure is earlier:\n\n+ERROR: could not load library \"C:/cirrus/build/tmp_install/ucrt64/lib/postgresql/plpython3.dll\": The specified module could not be found.\n\nClearly plpython is being built, because we see some warnings:\n\n[22:44:24.456] C:/msys64/ucrt64/include/python3.9/pyconfig.h:1474: warning: \"SIZEOF_OFF_T\" redefined\n[22:44:24.456] 1474 | #define SIZEOF_OFF_T 8\n[22:44:24.456] |\n[22:44:24.456] In file included from c:/cirrus/src/include/c.h:54,\n[22:44:24.456] from c:/cirrus/src/include/postgres.h:46,\n[22:44:24.456] from c:/cirrus/contrib/jsonb_plpython/jsonb_plpython.c:1:\n[22:44:24.456] ../../src/include/pg_config.h:875: note: this is the location of the previous definition\n[22:44:24.456] 875 | #define SIZEOF_OFF_T 4\n\nSeems we're doing something wrong and end up with a 4 byte off_t, whereas\npython ends up with an 8byte one. We probably need to fix that. But it's not\nthe cause of this problem.\n\n\nYou could take out -s from the make flags and see whether plpython3.dll is\nbeing built and installed, and where to.\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 25 Feb 2022 16:27:46 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"\nOn 2/25/22 19:27, Andres Freund wrote:\n> Hi,\n>\n> Andrew, CCIng you both because you might be interested in the CI bit, and\n> because you might know the answer.\n>\n>\n>\n>> 2- src/pl/plpython/Makefile looks under \"C:/Windows/system32/\" for\n>> PYTHONDLL. gendef cannot open that file, give an error like \" failed to\n>> open()\" when creating a .def file.\n> On my win10 VM in which I installed python.org python I don't see a python dll\n> in c:/windows/system32 either. Looks like none of our windows mingw animals\n> build with python. So it might just be bitrot.\n\n\n\nThere certainly was a time when python from python.org used to install\nits DLL in the system32 directory, so I imagine that's why it's there.\nI'm very glad to see that's no longer the case.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Sat, 26 Feb 2022 11:10:45 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi Andres,\n\n\n> This presumably is due to using mingw's python rather than python.org's\n> python. Seems like a reasonable thing to support for the mingw build.\n>\n> Melih, you could try to build against the python.org python (i installed\n> in\n> the CI container).\n>\n\nI tried to use the installed python from python.org in the container.\nThe problem with this is that \"LIBDIR\" and \"LDLIBRARY\" configs of python\nfor windows from python.org are empty. Therefore python_libdir or other\nrelated variables in configure file are not set correctly.\n\n\nSeems we're doing something wrong and end up with a 4 byte off_t, whereas\n> python ends up with an 8byte one. We probably need to fix that. But it's\n> not\n> the cause of this problem.\n>\n>\n> You could take out -s from the make flags and see whether plpython3.dll is\n> being built and installed, and where to.\n>\n\nHere is a run without -s: https://cirrus-ci.com/task/4569363104661504\nI couldn't get what's wrong from these logs to be honest. But I see that\nplpython3.dll exists under src/pl/plpython after build when I run these\nsteps locally.\n\nBest,\nMelih\n\nHi Andres, This presumably is due to using mingw's python rather than python.org's\npython. Seems like a reasonable thing to support for the mingw build.\n\nMelih, you could try to build against the python.org python (i installed in\nthe CI container).I tried to use the installed python from python.org in the container. The problem with this is that \"LIBDIR\" and \"LDLIBRARY\" configs of python for windows from python.org are empty. Therefore python_libdir or other related variables in configure file are not set correctly.\nSeems we're doing something wrong and end up with a 4 byte off_t, whereas\npython ends up with an 8byte one. We probably need to fix that. But it's not\nthe cause of this problem.\n\n\nYou could take out -s from the make flags and see whether plpython3.dll is\nbeing built and installed, and where to.Here is a run without -s: https://cirrus-ci.com/task/4569363104661504 I couldn't get what's wrong from these logs to be honest. But I see that plpython3.dll exists under src/pl/plpython after build when I run these steps locally.Best,Melih","msg_date":"Thu, 3 Mar 2022 13:16:15 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"\nOn 3/3/22 05:16, Melih Mutlu wrote:\n> Hi Andres, \n> \n>\n> This presumably is due to using mingw's python rather than\n> python.org <http://python.org>'s\n> python. Seems like a reasonable thing to support for the mingw build.\n>\n> Melih, you could try to build against the python.org\n> <http://python.org> python (i installed in\n> the CI container).\n>\n>\n> I tried to use the installed python from python.org\n> <http://python.org> in the container. \n> The problem with this is that \"LIBDIR\" and \"LDLIBRARY\" configs of\n> python for windows from python.org <http://python.org> are empty.\n> Therefore python_libdir or other related variables in configure file\n> are not set correctly.\n\n\n\nYeah, here's what it has:\n\n\n# python -c \"import sysconfig; import pprint; pp =\npprint.PrettyPrinter(); pp.pprint(sysconfig.get_config_vars())\"\n{'BINDIR': 'C:\\\\prog\\\\python310',\n 'BINLIBDEST': 'C:\\\\prog\\\\python310\\\\Lib',\n 'EXE': '.exe',\n 'EXT_SUFFIX': '.cp310-win_amd64.pyd',\n 'INCLUDEPY': 'C:\\\\prog\\\\python310\\\\Include',\n 'LIBDEST': 'C:\\\\prog\\\\python310\\\\Lib',\n 'SO': '.cp310-win_amd64.pyd',\n 'TZPATH': '',\n 'VERSION': '310',\n 'abiflags': '',\n 'base': 'C:\\\\prog\\\\python310',\n 'exec_prefix': 'C:\\\\prog\\\\python310',\n 'installed_base': 'C:\\\\prog\\\\python310',\n 'installed_platbase': 'C:\\\\prog\\\\python310',\n 'platbase': 'C:\\\\prog\\\\python310',\n 'platlibdir': 'lib',\n 'prefix': 'C:\\\\prog\\\\python310',\n 'projectbase': 'C:\\\\prog\\\\python310',\n 'py_version': '3.10.2',\n 'py_version_nodot': '310',\n 'py_version_nodot_plat': '310',\n 'py_version_short': '3.10',\n 'srcdir': 'C:\\\\prog\\\\python310',\n 'userbase': 'C:\\\\Users\\\\Administrator\\\\AppData\\\\Roaming\\\\Python'}\n\n\nThe DLL lives in the BINDIR, so maybe I guess we should search there if\nwe can't get the other things.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 3 Mar 2022 07:52:27 -0500","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-02-25 19:44:27 +0300, Melih Mutlu wrote:\n> I've been working on adding Windows+MinGW environment into cirrus-ci tasks\n> (discussion about ci is here [1]).\n\nThis doesn't apply anymore: http://cfbot.cputube.org/patch_37_3575.log\n\nCould you rebase?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 21 Mar 2022 17:15:59 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi Andres,\n\nRebased it.\nI also removed the temp installation task and\nused NoDefaultCurrentDirectoryInExePath env variable instead.\n\nBest,\nMelih","msg_date":"Tue, 22 Mar 2022 19:00:42 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-03-22 19:00:42 +0300, Melih Mutlu wrote:\n> Rebased it.\n> I also removed the temp installation task and\n> used NoDefaultCurrentDirectoryInExePath env variable instead.\n\nHm. But you're still using a separate build directory, from what I can see?\nThe NoDefaultCurrentDirectoryInExePath thing should only have an effect when\nnot using a separate build directory, no?\n\nDoes it work to not use the separate build dir? Without it we don't need the\nthe \"preparing build tree\" step, and that's quite slow on mingw:\nhttps://cirrus-ci.com/task/4713509253545984?logs=configure#L392\n\n[00:23:44.371] preparing build tree... done\n[00:24:25.429] configure: creating ./config.status\n\n\nChatting about this patch with Thomas I started to wonder about other reasons\nfor the slow speed of configure. I briefly experimented locally, and it looks\nlike using 'dash' as the shell makes configure run a good bit quicker.\n\n\n> ---\n> .cirrus.yml | 79 +++++++++++++++++++++++++++++++++++++++++++----------\n> 1 file changed, 65 insertions(+), 14 deletions(-)\n> \n> diff --git a/.cirrus.yml b/.cirrus.yml\n> index e5335fede7..1ed40347cf 100644\n> --- a/.cirrus.yml\n> +++ b/.cirrus.yml\n> @@ -23,7 +23,6 @@ env:\n> CHECKFLAGS: -Otarget\n> PROVE_FLAGS: --timer\n> PGCTLTIMEOUT: 120 # avoids spurious failures during parallel tests\n> - TEMP_CONFIG: ${CIRRUS_WORKING_DIR}/src/tools/ci/pg_ci_base.conf\n> PG_TEST_EXTRA: kerberos ldap ssl\n\nThis removes TEMP_CONFIG from all other tasks. You added it back to the VS\nwindows task, but not the others? I assume that was accidental?\n\n\n> + env:\n> + CCACHE_DIR: C:/msys64/ccache\n> + BUILD_DIR: \"%CIRRUS_WORKING_DIR%/build\"\n\nI think this should use TEMP_CONFIG too. Is the problem that you need to\nchange the path?\n\n\n> + ccache_cache:\n> + folder: ${CCACHE_DIR}\n> +\n> + mingw_info_script:\n> + - C:\\msys64\\usr\\bin\\bash.exe -lc \"where gcc\"\n> + - C:\\msys64\\usr\\bin\\bash.exe -lc \"gcc --version\"\n> + - C:\\msys64\\usr\\bin\\bash.exe -lc \"where perl\"\n> + - C:\\msys64\\usr\\bin\\bash.exe -lc \"perl --version\"\n> +\n> + configure_script:\n> + - C:\\msys64\\usr\\bin\\bash.exe -lc \"mkdir %BUILD_DIR% &&\n> + cd %BUILD_DIR% &&\n> + %CIRRUS_WORKING_DIR%/configure\n\nCould you try using dash to invoke configure here, and whether it makes configure faster?\n\n\n> + --host=x86_64-w64-mingw32\n> + --enable-cassert\n> + --enable-tap-tests\n> + --with-icu\n> + --with-libxml\n> + --with-libxslt\n> + --with-lz4\n> + --enable-debug\n> + CC='ccache gcc'\n> + CXX='ccache g++'\"\n\nI think this task should specify CFLAGS=\"-Og\", CXXFLAGS=\"-Og\" similar to other\ntasks. We end up with -O2 otherwise, which makes the build measurably slower.\n\n\n\n> + tests_script:\n> + - set \"NoDefaultCurrentDirectoryInExePath=0\"\n\nA comment about why NoDefaultCurrentDirectoryInExePath=0 is used would be\ngood.\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 30 Mar 2022 17:26:18 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-03-30 17:26:18 -0700, Andres Freund wrote:\n> Hi,\n>\n> On 2022-03-22 19:00:42 +0300, Melih Mutlu wrote:\n> > Rebased it.\n> > I also removed the temp installation task and\n> > used NoDefaultCurrentDirectoryInExePath env variable instead.\n>\n> Hm. But you're still using a separate build directory, from what I can see?\n> The NoDefaultCurrentDirectoryInExePath thing should only have an effect when\n> not using a separate build directory, no?\n\nMelih chatted with me about making this work. Turns out it doesn't readily -\npg_ctl still fails.\n\n\nThe reason that NoDefaultCurrentDirectoryInExePath doesn't suffice to get a\nworking in-tree build, is that we break it ourselves:\n\nint\nfind_my_exec(const char *argv0, char *retpath)\n...\n#ifdef WIN32\n\t/* Win32 checks the current directory first for names without slashes */\n\tjoin_path_components(retpath, cwd, argv0);\n\tif (validate_exec(retpath) == 0)\n\t\treturn resolve_symlinks(retpath);\n#endif\n\nSo even if windows doesn't actually use the current path, and the current\npg_ctl process isn't the one from the current directory, we *still* return\nthat.\n\nGah.\n\n\nMaybe we should just use GetModuleFileName()?\n\n\nBut even after that the tests don't work. Commands started via IPC::Run do,\nbut when using system() they don't. Looks like perl parses the path the itself\n:(.\n\n- Andres\n\n\n","msg_date":"Mon, 4 Apr 2022 13:41:31 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"\nOn 4/4/22 16:41, Andres Freund wrote:\n> Hi,\n>\n> On 2022-03-30 17:26:18 -0700, Andres Freund wrote:\n>> Hi,\n>>\n>> On 2022-03-22 19:00:42 +0300, Melih Mutlu wrote:\n>>> Rebased it.\n>>> I also removed the temp installation task and\n>>> used NoDefaultCurrentDirectoryInExePath env variable instead.\n>> Hm. But you're still using a separate build directory, from what I can see?\n>> The NoDefaultCurrentDirectoryInExePath thing should only have an effect when\n>> not using a separate build directory, no?\n> Melih chatted with me about making this work. Turns out it doesn't readily -\n> pg_ctl still fails.\n>\n>\n> The reason that NoDefaultCurrentDirectoryInExePath doesn't suffice to get a\n> working in-tree build, is that we break it ourselves:\n>\n> int\n> find_my_exec(const char *argv0, char *retpath)\n> ...\n> #ifdef WIN32\n> \t/* Win32 checks the current directory first for names without slashes */\n> \tjoin_path_components(retpath, cwd, argv0);\n> \tif (validate_exec(retpath) == 0)\n> \t\treturn resolve_symlinks(retpath);\n> #endif\n>\n> So even if windows doesn't actually use the current path, and the current\n> pg_ctl process isn't the one from the current directory, we *still* return\n> that.\n>\n> Gah.\n\n\n\n\nI notice a few things about the latest file in this thread.\n\nYou should set MSYSTEM=UCRT64 in the environment section. Given that,\nthere should be no need to specify a --host= setting for configure.\n\nIf it's not done already, the shebang line in\n/ucrt64/bin/core_perl/prove needs to be modified to use /ucrt64/bin/perl.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Tue, 5 Apr 2022 10:16:36 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"\nOn 3/30/22 20:26, Andres Freund wrote:\n> Could you try using dash to invoke configure here, and whether it makes configure faster?\n>\n>\n\n\nI got weird failures re libxml/parser.h when I tried with dash. See\n<https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\ncould see config.log on failure.)\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Wed, 6 Apr 2022 11:03:37 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\n> On 3/30/22 20:26, Andres Freund wrote:\n> > Could you try using dash to invoke configure here, and whether it makes configure faster?\n> I got weird failures re libxml/parser.h when I tried with dash.\n\nHm. Hadn't enabled that when I tried...\n\n\n> See <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\n> could see config.log on failure.)\n\nAll *.log files are preserved on failure. There's a file directory brower at\nthe top to navigate around:\nhttps://api.cirrus-ci.com/v1/artifact/task/5963254039052288/log/build/config.log\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 6 Apr 2022 09:34:19 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\n> On 3/30/22 20:26, Andres Freund wrote:\n> > Could you try using dash to invoke configure here, and whether it makes configure faster?\n> I got weird failures re libxml/parser.h when I tried with dash. See\n> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\n> could see config.log on failure.)\n\nSince dash won't help us to get the build time down sufficiently, and the\ntests don't pass without a separate build tree, I looked at what makes\nconfig/prep_buildtree so slow.\n\nIt's largely just bad code. The slowest part are spawning one expr and mkdir\n-p for each directory. One 'cmp' for each makefile doesn't help either.\n\nThe expr can be replaced with\n subdir=${item#$sourcetree}\nthat's afaics posix syntax ([1]), not bash.\n\nSpawning one mkdir for each directory can be replaced by a single mkdir\ninvocation with all the directories. On my linux workstation that gets the\ntime for the first loop down from 1005ms to 38ms, really.\n\nThat has the danger of the commandline getting too long. But since we rely on\nthe final link of the backend to be done in a single command, I don't think\nit's making things worse? We could try to use xargs otherwise, iirc that's in\nposix as well.\n\nUsing parameter substitution in the second loop takes it down from 775ms to\n533ms. Not calling cmp when the file doesn't exist cuts it down to 337ms.\n\nI don't know of a way to batch the call to ln. The time with ln replaced with\n: is 151ms, fwiw.\n\nOn windows that makes prep_buildtree go from 42.4s to 5.8s for me.\n\nGreetings,\n\nAndres Freund\n\n[1] https://pubs.opengroup.org/onlinepubs/009604499/utilities/xcu_chap02.html","msg_date":"Thu, 7 Apr 2022 10:10:21 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"\nOn 4/6/22 12:34, Andres Freund wrote:\n> Hi,\n>\n> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\n>> On 3/30/22 20:26, Andres Freund wrote:\n>>> Could you try using dash to invoke configure here, and whether it makes configure faster?\n>> I got weird failures re libxml/parser.h when I tried with dash.\n> Hm. Hadn't enabled that when I tried...\n>\n>\n>> See <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\n>> could see config.log on failure.)\n> All *.log files are preserved on failure. There's a file directory brower at\n> the top to navigate around:\n> https://api.cirrus-ci.com/v1/artifact/task/5963254039052288/log/build/config.log\n\n\n\nThanks.\n\n\nI got it working with this added to the config settings:\n\n\n--with-includes='/ucrt64/include/libxml2 /c/cirrus/src/include/port/win32'\n\n\nI conclude tentatively that while bash translates widows paths to msys\npaths, dash does not.\n\n\nsee https://cirrus-ci.com/task/5968968560148480?logs=configure#L1\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 7 Apr 2022 15:42:13 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"\nOn 4/7/22 13:10, Andres Freund wrote:\n> Hi,\n>\n> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\n>> On 3/30/22 20:26, Andres Freund wrote:\n>>> Could you try using dash to invoke configure here, and whether it makes configure faster?\n>> I got weird failures re libxml/parser.h when I tried with dash. See\n>> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\n>> could see config.log on failure.)\n> Since dash won't help us to get the build time down sufficiently, and the\n> tests don't pass without a separate build tree, I looked at what makes\n> config/prep_buildtree so slow.\n>\n> It's largely just bad code. The slowest part are spawning one expr and mkdir\n> -p for each directory. One 'cmp' for each makefile doesn't help either.\n>\n> The expr can be replaced with\n> subdir=${item#$sourcetree}\n> that's afaics posix syntax ([1]), not bash.\n>\n> Spawning one mkdir for each directory can be replaced by a single mkdir\n> invocation with all the directories. On my linux workstation that gets the\n> time for the first loop down from 1005ms to 38ms, really.\n>\n> That has the danger of the commandline getting too long. But since we rely on\n> the final link of the backend to be done in a single command, I don't think\n> it's making things worse? We could try to use xargs otherwise, iirc that's in\n> posix as well.\n>\n> Using parameter substitution in the second loop takes it down from 775ms to\n> 533ms. Not calling cmp when the file doesn't exist cuts it down to 337ms.\n>\n> I don't know of a way to batch the call to ln. The time with ln replaced with\n> : is 151ms, fwiw.\n\n\nAFAIK Msy2s 'ln -s' by default copies a non-directory rather than\nactually symlinking it. If we want real symlinks, then we need\nMSYS=|winsymlinks:nativestrict set. The is will fail unless the calling\nuser is an Administrator or has the SeCreateSymbolicLink privilege. See\n|\n\n|<https://postgr.es/m/|e05b213c-1257-84d4-f079-5c4d8c79e3ad@dunslane.net>\nfor more details.\n\n\n> On windows that makes prep_buildtree go from 42.4s to 5.8s for me.\n\n\nThat's pretty good.\n\n\nI think we can get rid of the CVS pruning, it's only 15 years or so\nsince we've had that in the tree.\n\n\n+ if test ! -d \"$buildtree/$subdir\"; then\n+ echo \"$buildtree/$subdir\"\n+ fi\n\n\nI would probably just write that as\n\n\ntest -d \"$buildtree/$subdir' || echo \"$buildtree/$subdir\"\n\n\nalthough it's really just a matter of taste.\n\n\ncheers\n\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 7 Apr 2022 16:48:50 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"\nOn 4/7/22 16:48, Andrew Dunstan wrote:\n> On 4/7/22 13:10, Andres Freund wrote:\n>> Hi,\n>>\n>> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\n>>> On 3/30/22 20:26, Andres Freund wrote:\n>>>> Could you try using dash to invoke configure here, and whether it makes configure faster?\n>>> I got weird failures re libxml/parser.h when I tried with dash. See\n>>> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\n>>> could see config.log on failure.)\n>> Since dash won't help us to get the build time down sufficiently, and the\n>> tests don't pass without a separate build tree, I looked at what makes\n>> config/prep_buildtree so slow.\n>>\n>> It's largely just bad code. The slowest part are spawning one expr and mkdir\n>> -p for each directory. One 'cmp' for each makefile doesn't help either.\n>>\n>> The expr can be replaced with\n>> subdir=${item#$sourcetree}\n>> that's afaics posix syntax ([1]), not bash.\n>>\n>> Spawning one mkdir for each directory can be replaced by a single mkdir\n>> invocation with all the directories. On my linux workstation that gets the\n>> time for the first loop down from 1005ms to 38ms, really.\n>>\n>> That has the danger of the commandline getting too long. But since we rely on\n>> the final link of the backend to be done in a single command, I don't think\n>> it's making things worse? We could try to use xargs otherwise, iirc that's in\n>> posix as well.\n>>\n>> Using parameter substitution in the second loop takes it down from 775ms to\n>> 533ms. Not calling cmp when the file doesn't exist cuts it down to 337ms.\n>>\n>> I don't know of a way to batch the call to ln. The time with ln replaced with\n>> : is 151ms, fwiw.\n>\n> AFAIK Msy2s 'ln -s' by default copies a non-directory rather than\n> actually symlinking it. If we want real symlinks, then we need\n> MSYS=|winsymlinks:nativestrict set. The is will fail unless the calling\n> user is an Administrator or has the SeCreateSymbolicLink privilege. See\n> |\n\n\nSometimes I hate Thunderbird. Of course the | is spurious above, we\nwould need\n\nMSYS=winsymlinks:nativestrict\n\nset.\n\n\ncheers\n\nandrew\n\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Thu, 7 Apr 2022 16:52:26 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi Andrew,\n\nYou should set MSYSTEM=UCRT64 in the environment section. Given that,\n> there should be no need to specify a --host= setting for configure.\n>\n\nIt's set to UCRT64 in the docker image side [1]. I didn't know --host isn't\nnecessary on UCRT64 environment. I'll remove it then.\n\n [1]\n https://github.com/anarazel/pg-vm-images/blob/main/docker/windows_ci_mingw64#L11\n<https://github.com/anarazel/pg-vm-images/blob/main/docker/windows_ci_mingw64#L11>\n\nBest,\nMelih\n\nHi Andrew, \nYou should set MSYSTEM=UCRT64 in the environment section. Given that,\nthere should be no need to specify a --host= setting for configure.It's set to UCRT64 in the docker image side [1]. I didn't know --host isn't necessary on UCRT64 environment. I'll remove it then. [1] https://github.com/anarazel/pg-vm-images/blob/main/docker/windows_ci_mingw64#L11Best,Melih","msg_date":"Fri, 8 Apr 2022 17:57:21 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"On 2022-Apr-07, Andres Freund wrote:\n\n> Since dash won't help us to get the build time down sufficiently, and the\n> tests don't pass without a separate build tree, I looked at what makes\n> config/prep_buildtree so slow.\n\nMaybe we can replace prep_buildtree with a Perl script. Surely that\nshould be faster.\n\n-- \nÁlvaro Herrera Breisgau, Deutschland — https://www.EnterpriseDB.com/\n\"Sallah, I said NO camels! That's FIVE camels; can't you count?\"\n(Indiana Jones)\n\n\n","msg_date":"Fri, 8 Apr 2022 17:04:34 +0200","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"> On windows that makes prep_buildtree go from 42.4s to 5.8s for me.\n>\n\nI applied Andres's faster prep build tree changes and triggered some cirrus\nruns\n\nWithout these changes, preparing build tree was taking around 42.3s\n(sometimes even more) [1]\nIt seems like with these changes it drops to around 8s [2]\n\n[1] https://cirrus-ci.com/task/6562493345562624\n[2] https://cirrus-ci.com/task/4836843802853376\n\nBest,\nMelih\n\n\nOn windows that makes prep_buildtree go from 42.4s to 5.8s for me.I applied Andres's faster prep build tree changes and triggered some cirrus runsWithout these changes, preparing build tree was taking around 42.3s (sometimes even more) [1]It seems like with these changes it drops to around 8s [2][1] https://cirrus-ci.com/task/6562493345562624[2] https://cirrus-ci.com/task/4836843802853376Best,Melih","msg_date":"Fri, 8 Apr 2022 18:18:34 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-04-08 17:04:34 +0200, Alvaro Herrera wrote:\n> > Since dash won't help us to get the build time down sufficiently, and the\n> > tests don't pass without a separate build tree, I looked at what makes\n> > config/prep_buildtree so slow.\n>\n> Maybe we can replace prep_buildtree with a Perl script. Surely that\n> should be faster.\n\nCurrently building doesn't depend on perl :(\n\nI think the improvements that I suggested are big enough that they're worth\ndoing on their own, particularly for windows, but also other OSs.\n\n\nI just realized that the second find is pretty expensive compared to the\nfirst.\n\ntime find \"$sourcetree\" -type d \\( \\( -name CVS -prune \\) -o \\( -name .git -prune \\) -o -print \\) | grep -v \"$sourcetree/doc/src/sgml/\\+\" > /dev/null\nreal\t0m0.019s\nuser\t0m0.008s\nsys\t0m0.017s\n\nsecond:\ntime find \"$sourcetree\" -name Makefile -print -o -name GNUmakefile -print | grep -v \"$sourcetree/doc/src/sgml/images/\" > /dev/null\n\nreal\t0m0.118s\nuser\t0m0.071s\nsys\t0m0.053s\n\nIt think we could just obsolete the second find, by checking for the existence\nof Makefile / GNUmakefile in the first loop...\n\n\nThe invocation of ln -s is quite measurable - looks like it's mostly the\nprocess startup overhead (on linux, at least). Doing a ln --version > /dev/null\neach iteration takes about the same time as actually creating the symlinks.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 8 Apr 2022 15:05:57 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"On Thu, Apr 07, 2022 at 10:10:21AM -0700, Andres Freund wrote:\n> Hi,\n> \n> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\n> > On 3/30/22 20:26, Andres Freund wrote:\n> > > Could you try using dash to invoke configure here, and whether it makes configure faster?\n> > I got weird failures re libxml/parser.h when I tried with dash. See\n> > <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\n> > could see config.log on failure.)\n> \n> Since dash won't help us to get the build time down sufficiently, and the\n> tests don't pass without a separate build tree, I looked at what makes\n> config/prep_buildtree so slow.\n> \n> It's largely just bad code. The slowest part are spawning one expr and mkdir\n> -p for each directory. One 'cmp' for each makefile doesn't help either.\n> \n> The expr can be replaced with\n> subdir=${item#$sourcetree}\n> that's afaics posix syntax ([1]), not bash.\n> \n> Spawning one mkdir for each directory can be replaced by a single mkdir\n> invocation with all the directories. On my linux workstation that gets the\n> time for the first loop down from 1005ms to 38ms, really.\n\nEven better?\n\n(cd \"$sourcetree\" && find . -print |grep -E '/Makefile$|/GNUmakefile$' |grep -v \"$sourcetree/doc/src/sgml/images/\" |xargs tar c) |\n (cd \"$buildtree\" && tar x)\n\n\n","msg_date":"Fri, 8 Apr 2022 19:27:58 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-04-08 19:27:58 -0500, Justin Pryzby wrote:\n> On Thu, Apr 07, 2022 at 10:10:21AM -0700, Andres Freund wrote:\n> > On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\n> > > On 3/30/22 20:26, Andres Freund wrote:\n> > > > Could you try using dash to invoke configure here, and whether it makes configure faster?\n> > > I got weird failures re libxml/parser.h when I tried with dash. See\n> > > <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\n> > > could see config.log on failure.)\n> > \n> > Since dash won't help us to get the build time down sufficiently, and the\n> > tests don't pass without a separate build tree, I looked at what makes\n> > config/prep_buildtree so slow.\n> > \n> > It's largely just bad code. The slowest part are spawning one expr and mkdir\n> > -p for each directory. One 'cmp' for each makefile doesn't help either.\n> > \n> > The expr can be replaced with\n> > subdir=${item#$sourcetree}\n> > that's afaics posix syntax ([1]), not bash.\n> > \n> > Spawning one mkdir for each directory can be replaced by a single mkdir\n> > invocation with all the directories. On my linux workstation that gets the\n> > time for the first loop down from 1005ms to 38ms, really.\n> \n> Even better?\n> \n> (cd \"$sourcetree\" && find . -print |grep -E '/Makefile$|/GNUmakefile$' |grep -v \"$sourcetree/doc/src/sgml/images/\" |xargs tar c) |\n> (cd \"$buildtree\" && tar x)\n\nDon't think we depend on tar for building, at the moment. But yes, it'd be\nfaster... Tar is certainly a smaller dependency than perl, not sure if there's\nany relevant platform without it?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Fri, 8 Apr 2022 18:02:41 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"On 2022-Apr-08, Andres Freund wrote:\n\n> I just realized that the second find is pretty expensive compared to the\n> first.\n> \n> time find \"$sourcetree\" -type d \\( \\( -name CVS -prune \\) -o \\( -name .git -prune \\) -o -print \\) | grep -v \"$sourcetree/doc/src/sgml/\\+\" > /dev/null\n> real\t0m0.019s\n> user\t0m0.008s\n> sys\t0m0.017s\n> \n> second:\n> time find \"$sourcetree\" -name Makefile -print -o -name GNUmakefile -print | grep -v \"$sourcetree/doc/src/sgml/images/\" > /dev/null\n> \n> real\t0m0.118s\n> user\t0m0.071s\n> sys\t0m0.053s\n\nHmm, ISTM that time can be reduced a bit with -prune,\n\ntime find \"$sourcetree\" \\( -name .git -prune \\) -o -name Makefile -print -o -name GNUmakefile -print | grep -v \"$sourcetree/doc/src/sgml/images/\" > /dev/null\n\nI thought it might work to do away with the grep and use find's -path\ninstead to prune that subdir, but \"time\" shows almost no difference for\nme:\n\ntime find \"$sourcetree\" \\( -name .git -prune \\) -o \\( -path '*doc/src/sgml/images' -prune \\) -o -name Makefile -print -o -name GNUmakefile -print > /dev/null\n\nMaybe find's -path is equally expensive. Still, that seems a good\nchange anyway. (The times are lower in my system than those you show.)\n\n> It think we could just obsolete the second find, by checking for the existence\n> of Makefile / GNUmakefile in the first loop...\n\nHmm, if that's going to require one more shell command per dir, it\nsounds more expensive. It's worth trying, I guess.\n\n> The invocation of ln -s is quite measurable - looks like it's mostly the\n> process startup overhead (on linux, at least). Doing a ln --version > /dev/null\n> each iteration takes about the same time as actually creating the symlinks.\n\nIs this running with some locale settings enabled? Maybe we can save\nsome time by making sure we're under LC_ALL=C or something like that, to\navoid searching for translation files.\n\n-- \nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\n\"Las navajas y los monos deben estar siempre distantes\" (Germán Poo)\n\n\n","msg_date":"Sat, 9 Apr 2022 12:37:05 +0200","msg_from":"Alvaro Herrera <alvherre@alvh.no-ip.org>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"\nOn 4/8/22 21:02, Andres Freund wrote:\n> Hi,\n>\n> On 2022-04-08 19:27:58 -0500, Justin Pryzby wrote:\n>> On Thu, Apr 07, 2022 at 10:10:21AM -0700, Andres Freund wrote:\n>>> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\n>>>> On 3/30/22 20:26, Andres Freund wrote:\n>>>>> Could you try using dash to invoke configure here, and whether it makes configure faster?\n>>>> I got weird failures re libxml/parser.h when I tried with dash. See\n>>>> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\n>>>> could see config.log on failure.)\n>>> Since dash won't help us to get the build time down sufficiently, and the\n>>> tests don't pass without a separate build tree, I looked at what makes\n>>> config/prep_buildtree so slow.\n>>>\n>>> It's largely just bad code. The slowest part are spawning one expr and mkdir\n>>> -p for each directory. One 'cmp' for each makefile doesn't help either.\n>>>\n>>> The expr can be replaced with\n>>> subdir=${item#$sourcetree}\n>>> that's afaics posix syntax ([1]), not bash.\n>>>\n>>> Spawning one mkdir for each directory can be replaced by a single mkdir\n>>> invocation with all the directories. On my linux workstation that gets the\n>>> time for the first loop down from 1005ms to 38ms, really.\n>> Even better?\n>>\n>> (cd \"$sourcetree\" && find . -print |grep -E '/Makefile$|/GNUmakefile$' |grep -v \"$sourcetree/doc/src/sgml/images/\" |xargs tar c) |\n>> (cd \"$buildtree\" && tar x)\n> Don't think we depend on tar for building, at the moment. But yes, it'd be\n> faster... Tar is certainly a smaller dependency than perl, not sure if there's\n> any relevant platform without it?\n>\n>\n\n\nCouple of things here.\n\n\n1. The second grep should lose \"$sourcetree\" I think.\n\n2. Isn't this going to be a change in behaviour in that it will copy\nrather than symlinking the Makefiles? That is in fact the default\nbehaviour of msys2's 'ln -s', as I pointed out upthread, but I don't\nthink it's what we really want, especially in the general case. If you\nmodify the Makefile and you're using a vpath you want to see the change\nreflected in your vpath.\n\n\ncheers\n\n\nandrew\n\n--\nAndrew Dunstan\nEDB: https://www.enterprisedb.com\n\n\n\n","msg_date":"Sat, 9 Apr 2022 12:34:40 -0400","msg_from":"Andrew Dunstan <andrew@dunslane.net>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi hackers,\n\nI'm sharing the rebased version of this patch, if you're still interested.\n\nI would appreciate any feedback or concerns.\n\nBest,\nMelih\n\n\nAndrew Dunstan <andrew@dunslane.net>, 9 Nis 2022 Cmt, 19:34 tarihinde şunu\nyazdı:\n\n>\n> On 4/8/22 21:02, Andres Freund wrote:\n> > Hi,\n> >\n> > On 2022-04-08 19:27:58 -0500, Justin Pryzby wrote:\n> >> On Thu, Apr 07, 2022 at 10:10:21AM -0700, Andres Freund wrote:\n> >>> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\n> >>>> On 3/30/22 20:26, Andres Freund wrote:\n> >>>>> Could you try using dash to invoke configure here, and whether it\n> makes configure faster?\n> >>>> I got weird failures re libxml/parser.h when I tried with dash. See\n> >>>> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\n> >>>> could see config.log on failure.)\n> >>> Since dash won't help us to get the build time down sufficiently, and\n> the\n> >>> tests don't pass without a separate build tree, I looked at what makes\n> >>> config/prep_buildtree so slow.\n> >>>\n> >>> It's largely just bad code. The slowest part are spawning one expr and\n> mkdir\n> >>> -p for each directory. One 'cmp' for each makefile doesn't help either.\n> >>>\n> >>> The expr can be replaced with\n> >>> subdir=${item#$sourcetree}\n> >>> that's afaics posix syntax ([1]), not bash.\n> >>>\n> >>> Spawning one mkdir for each directory can be replaced by a single mkdir\n> >>> invocation with all the directories. On my linux workstation that gets\n> the\n> >>> time for the first loop down from 1005ms to 38ms, really.\n> >> Even better?\n> >>\n> >> (cd \"$sourcetree\" && find . -print |grep -E '/Makefile$|/GNUmakefile$'\n> |grep -v \"$sourcetree/doc/src/sgml/images/\" |xargs tar c) |\n> >> (cd \"$buildtree\" && tar x)\n> > Don't think we depend on tar for building, at the moment. But yes, it'd\n> be\n> > faster... Tar is certainly a smaller dependency than perl, not sure if\n> there's\n> > any relevant platform without it?\n> >\n> >\n>\n>\n> Couple of things here.\n>\n>\n> 1. The second grep should lose \"$sourcetree\" I think.\n>\n> 2. Isn't this going to be a change in behaviour in that it will copy\n> rather than symlinking the Makefiles? That is in fact the default\n> behaviour of msys2's 'ln -s', as I pointed out upthread, but I don't\n> think it's what we really want, especially in the general case. If you\n> modify the Makefile and you're using a vpath you want to see the change\n> reflected in your vpath.\n>\n>\n> cheers\n>\n>\n> andrew\n>\n> --\n> Andrew Dunstan\n> EDB: https://www.enterprisedb.com\n>\n>","msg_date":"Thu, 28 Jul 2022 17:11:25 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"I think the \"only_if\" should allow separately running one but not both of the\nwindows instances, like:\n\n+ only_if: $CIRRUS_CHANGE_MESSAGE !=~ '.*\\nci-os-only:.*' || $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw64'\n\nI'm not sure, but maybe this task should only run \"by request\", and omit the\nfirst condition:\n\n+ only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw64'\n\nI think it should include something like\n\n+ setup_additional_packages_script: |\n+ REM C:\\msys64\\usr\\bin\\pacman.exe -S --noconfirm ...\n\nLet's see what others think about those.\n\nDo you know if this handles logging of crash dumps ?\n\nWith tweaks to prep_buildtree, and with ./configure --cache-file, that step\ngoes down to ~36sec (unless configure needs to be re-run).\n\nI also looked into using busybox to avoid running separate processes for each\n\"ln\", but I think 36sec is good enough.\n\nAt one point, I tried setting \"CIRRUS_SHELL: bash\" to avoid writing \"bash -c\"\nover and over, but never got it working.\n\n-- \nJustin","msg_date":"Thu, 28 Jul 2022 17:44:28 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"I noticed that this says:\n\n[01:01:45.657] sqlda.pgc: In function 'dump_sqlda':\n[01:01:45.657] sqlda.pgc:45:33: warning: format '%d' expects argument\nof type 'int', but argument 3 has type 'long long int' [-Wformat=]\n[01:01:45.657] 45 | \"name sqlda descriptor: '%s' value %I64d\\n\",\n[01:01:45.657] | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n[01:01:45.657] ......\n[01:01:45.657] 49 | sqlda->sqlvar[i].sqlname.data, *(long long int\n*)sqlda->sqlvar[i].sqldata);\n[01:01:45.657] | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n[01:01:45.657] | |\n[01:01:45.657] | long long int\n\n... but fairywren doesn't. Why would they disagree on recognising %I64d?\n\nThe other warning I'm seeing is present on both, so it's not really\nrelevant to this thread, just mentioning it... seems kinda like a\nworlds-colliding-problem without an elegant fix (POSIX says you have\nto declare it yourself exactly like that, but Windows says linkage\nain't going to work the way you want unless you sprinkle the right\nlinkage dust on it...), so maybe we just want to put #if\n!defined(something something mings) around it and to MinGW's header's\ndeclaration...\n\n[00:48:08.925] c:/cirrus/src/backend/postmaster/postmaster.c: In\nfunction 'PostmasterMain':\n[00:48:08.925] c:/cirrus/src/backend/postmaster/postmaster.c:973:31:\nwarning: '__p__environ' redeclared without dllimport attribute:\nprevious dllimport ignored [-Wattributes]\n[00:48:08.925] 973 | extern char **environ;\n[00:48:08.925] | ^~~~~~~\n\n\n","msg_date":"Thu, 4 Aug 2022 14:04:33 +1200","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"On Thu, Aug 4, 2022 at 2:04 PM Thomas Munro <thomas.munro@gmail.com> wrote:\n> I noticed that this says:\n>\n> [01:01:45.657] sqlda.pgc: In function 'dump_sqlda':\n> [01:01:45.657] sqlda.pgc:45:33: warning: format '%d' expects argument\n> of type 'int', but argument 3 has type 'long long int' [-Wformat=]\n> [01:01:45.657] 45 | \"name sqlda descriptor: '%s' value %I64d\\n\",\n> [01:01:45.657] | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n> [01:01:45.657] ......\n> [01:01:45.657] 49 | sqlda->sqlvar[i].sqlname.data, *(long long int\n> *)sqlda->sqlvar[i].sqldata);\n> [01:01:45.657] | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n> [01:01:45.657] | |\n> [01:01:45.657] | long long int\n>\n> ... but fairywren doesn't. Why would they disagree on recognising %I64d?\n\nOops, I was looking in the wrong place. fairywren does also shows the warning:\n\nhttps://buildfarm.postgresql.org/cgi-bin/show_stage_log.pl?nm=fairywren&dt=2022-08-02%2022%3A05%3A30&stg=ecpg-check\n\nSomething to fix, but not directly relevant to this patch. Sorry for the noise.\n\n\n","msg_date":"Thu, 4 Aug 2022 20:38:34 +1200","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Inline notes about changes since the last version.\n\nOn Thu, Jul 28, 2022 at 05:44:28PM -0500, Justin Pryzby wrote:\n> I think the \"only_if\" should allow separately running one but not both of the\n> windows instances, like:\n> \n> + only_if: $CIRRUS_CHANGE_MESSAGE !=~ '.*\\nci-os-only:.*' || $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw64'\n> \n> I'm not sure, but maybe this task should only run \"by request\", and omit the\n> first condition:\n> \n> + only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw64'\n\nThe patch shouldn't say this during development, or else cfbot doesn't run it..\nOops.\n\n> I think it should include something like\n> \n> + setup_additional_packages_script: |\n> + REM C:\\msys64\\usr\\bin\\pacman.exe -S --noconfirm ...\n> \n> Let's see what others think about those.\n> \n> Do you know if this handles logging of crash dumps ?\n\nIt does now, although I hardcoded \"postgres.exe\" ...\n\n> + setup_additional_packages_script: |\n> + REM C:\\msys64\\usr\\bin\\pacman.exe -S --noconfirm busybox\n\nThis should include choco, too.\n\n> - CXXFLAGS='-Og -ggdb'\"\n> + CXXFLAGS='-Og -ggdb' && break;\n> + rm -v ${CCACHE_DIR}/configure.cache;\n> + done\n\nI noticed that this doesn't seem to do the right thing with the exit status -\nconfigure can fail without cirrusci noticing, and then the build fails at the\nnext step.\n\n> for item in `find \"$sourcetree\" -name Makefile -print -o -name GNUmakefile -print | grep -v \"$sourcetree/doc/src/sgml/images/\"`; do\n> - filename=`expr \"$item\" : \"$sourcetree\\(.*\\)\"`\n> - if test ! -f \"${item}.in\"; then\n> - if cmp \"$item\" \"$buildtree/$filename\" >/dev/null 2>&1; then : ; else\n> - ln -fs \"$item\" \"$buildtree/$filename\" || exit 1\n> - fi\n> - fi\n> + filename=${item#$sourcetree}\n> + [ -e \"$buildtree/$filename\" ] && continue\n\nI fixed this to check for \".in\" files as intended.\n\nIt'd be a lot better if the image didn't take so long to start. :(\n\n-- \nJustin","msg_date":"Thu, 18 Aug 2022 21:33:56 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Justin Pryzby <pryzby@telsasoft.com>, 19 Ağu 2022 Cum, 05:34 tarihinde şunu\nyazdı:\n\n> Inline notes about changes since the last version.\n>\n> On Thu, Jul 28, 2022 at 05:44:28PM -0500, Justin Pryzby wrote:\n> > I think the \"only_if\" should allow separately running one but not both\n> of the\n> > windows instances, like:\n> >\n> > + only_if: $CIRRUS_CHANGE_MESSAGE !=~ '.*\\nci-os-only:.*' ||\n> $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw64'\n> >\n> > I'm not sure, but maybe this task should only run \"by request\", and omit\n> the\n> > first condition:\n> >\n> > + only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw64'\n>\n> The patch shouldn't say this during development, or else cfbot doesn't run\n> it..\n> Oops.\n>\nActually, making MinGW task optional for now might make sense. Due to\nwindows resource limits on Cirrus CI and slow builds on Windows, adding\nthis task as non-optional may not be an efficient decision\nI think that continuing with this patch by changing MinGW to optional for\nnow, instead of waiting for more resource on Cirrus or faster builds on\nWindows, could be better. I don't see any harm.\n\n+ only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-include:[^\\n]*mingw.* ||\n> $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw.*'\n\nAdded this line to allow run only mingw task or run all tasks including\nmingw.\n\nWhat do you all think about this change? Does it make sense?\n\n\nThanks for your contributions/reviews Justin!\n\n> I think it should include something like\n> >\n> > + setup_additional_packages_script: |\n> > + REM C:\\msys64\\usr\\bin\\pacman.exe -S --noconfirm ...\n> >\n> > Let's see what others think about those.\n> >\n> > Do you know if this handles logging of crash dumps ?\n>\n> It does now, although I hardcoded \"postgres.exe\" ...\n>\n\nmerged this with my patch\n\n> + setup_additional_packages_script: |\n> > + REM C:\\msys64\\usr\\bin\\pacman.exe -S --noconfirm busybox\n>\n> This should include choco, too.\n\nAdded pacman.exe line. Do we really need choco here? I don't think mingw\nwould require any package via choco.\nAlso is ending pacman.exe line with busybox intentional? I just added that\nline with \"...\" at the end instead of any package name.\n\n\n>\n>\n> - CXXFLAGS='-Og -ggdb'\"\n> > + CXXFLAGS='-Og -ggdb' && break;\n> > + rm -v ${CCACHE_DIR}/configure.cache;\n> > + done\n>\n> I noticed that this doesn't seem to do the right thing with the exit\n> status -\n> configure can fail without cirrusci noticing, and then the build fails at\n> the\n> next step.\n\n\nmerged.\n\n\n>\n>\n> for item in `find \"$sourcetree\" -name Makefile -print -o -name\n> GNUmakefile -print | grep -v \"$sourcetree/doc/src/sgml/images/\"`; do\n> > - filename=`expr \"$item\" : \"$sourcetree\\(.*\\)\"`\n> > - if test ! -f \"${item}.in\"; then\n> > - if cmp \"$item\" \"$buildtree/$filename\" >/dev/null 2>&1; then : ;\n> else\n> > - ln -fs \"$item\" \"$buildtree/$filename\" || exit 1\n> > - fi\n> > - fi\n> > + filename=${item#$sourcetree}\n> > + [ -e \"$buildtree/$filename\" ] && continue\n>\n> I fixed this to check for \".in\" files as intended.\n>\n> It'd be a lot better if the image didn't take so long to start. :(\n>\n\nOne question would be that should this patch include \"prep_buildtree\"? It\ndoesn't seem to me like it's directly related to adding MinGW into CI but\nmore like an improvement for builds on Windows.\nMaybe we can make it a seperate patch if it's necessary.\n\nWhat do you think?\n\n TAR: \"c:/windows/system32/tar.exe\"\n>\n\nSharing a new version of the patch. It also moves the above line so that it\nwill apply to mingw task too. Otherwise mingw task was failing.\n\nThanks,\nMelih","msg_date":"Sat, 3 Sep 2022 00:52:54 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"On Sat, Sep 03, 2022 at 12:52:54AM +0300, Melih Mutlu wrote:\n> Justin Pryzby <pryzby@telsasoft.com>, 19 Ağu 2022 Cum, 05:34 tarihinde şunu yazdı:\n> \n> > Inline notes about changes since the last version.\n> >\n> > On Thu, Jul 28, 2022 at 05:44:28PM -0500, Justin Pryzby wrote:\n> > > I think the \"only_if\" should allow separately running one but not both\n> > of the\n> > > windows instances, like:\n> > >\n> > > + only_if: $CIRRUS_CHANGE_MESSAGE !=~ '.*\\nci-os-only:.*' || $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw64'\n> > >\n> > > I'm not sure, but maybe this task should only run \"by request\", and omit the\n> > > first condition:\n> > >\n> > > + only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw64'\n> >\n> > The patch shouldn't say this during development, or else cfbot doesn't run it..\n> > Oops.\n>\n> Actually, making MinGW task optional for now might make sense. Due to\n> windows resource limits on Cirrus CI and slow builds on Windows, adding\n> this task as non-optional may not be an efficient decision\n> I think that continuing with this patch by changing MinGW to optional for\n> now, instead of waiting for more resource on Cirrus or faster builds on\n> Windows, could be better. I don't see any harm.\n\nI agree that maybe it should be optional if merged to postgres.\n\nBut cfbot should run the Mingw task for this patch's own commitfest\nentry. But right now (because cfbot doesn't include the original commit\nmessage/s), it doesn't get run :(\n\n> + only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-include:[^\\n]*mingw.* ||\n> $CIRRUS_CHANGE_MESSAGE =~ '.*\\nci-os-only:[^\\n]*mingw.*'\n> \n> Added this line to allow run only mingw task or run all tasks including\n> mingw.\n> \n> What do you all think about this change? Does it make sense?\n\nYou're allowing to write \"ci-os-include: mingw\" to \"opt-in\" to the task,\nwithout opting-out of all the other tasks (and without enumerating all\nthe tasks by writing \"ci-os-only: mingw,windows,macos,freebsd,linux\".\nThat makes sense, and the logic looks right. But that still has to be\ncommented during development to be run by cfbot.\n\nAlso, the first half was missing a closing quote.\nhttps://cirrus-ci.com/build/5874178241855488\n\n> > + setup_additional_packages_script: |\n> > > + REM C:\\msys64\\usr\\bin\\pacman.exe -S --noconfirm busybox\n> >\n> > This should include choco, too.\n> \n> Added pacman.exe line. Do we really need choco here? I don't think mingw\n> would require any package via choco.\n\nI guess choco isn't needed.\n\n> Also is ending pacman.exe line with busybox intentional? I just added that\n> line with \"...\" at the end instead of any package name.\n\nYeah, the busybox part was unintentional.\n\n> > > for item in `find \"$sourcetree\" -name Makefile -print -o -name GNUmakefile -print | grep -v \"$sourcetree/doc/src/sgml/images/\"`; do\n> > > - filename=`expr \"$item\" : \"$sourcetree\\(.*\\)\"`\n> > > - if test ! -f \"${item}.in\"; then\n> > > - if cmp \"$item\" \"$buildtree/$filename\" >/dev/null 2>&1; then : ; else\n> > > - ln -fs \"$item\" \"$buildtree/$filename\" || exit 1\n> > > - fi\n> > > - fi\n> > > + filename=${item#$sourcetree}\n> > > + [ -e \"$buildtree/$filename\" ] && continue\n> >\n> > I fixed this to check for \".in\" files as intended.\n> >\n> > It'd be a lot better if the image didn't take so long to start. :(\n> \n> One question would be that should this patch include \"prep_buildtree\"? It\n> doesn't seem to me like it's directly related to adding MinGW into CI but\n> more like an improvement for builds on Windows.\n> Maybe we can make it a seperate patch if it's necessary.\n\nI don't know what direction that idea is going, but it makes working\nwith this patch a bit easier when configure is less slow. Fine with me\nto split it into a separate patch :)\n\n> Sharing a new version of the patch. It also moves the above line so that it\n> will apply to mingw task too. Otherwise mingw task was failing.\n\nI saw that but hadn't tracked it down yet. Do you know if the tar\nfailures were from a TAP test added since you first posted the mingw\npatch, or ??\n\nAlso: your original patch said --host=x86_64-w64-mingw32, but the task\nis called MinGW64. The most recent patches don't use --host at all, and\nwere building for a 32 bit environment, even though the OS image says\nMSYSTEM=UCRT64.\n\nAlso: right now src/test and src/interfaces/*/test aren't being built\nduring the build phase, which means that they're 1) not compiled in\nparallel; and 2) not cached. This isn't specific to MinGW. Other than\ncompiling those dirs specifically, one option is to put\n\"always: upload_caches: ccache\" after \"test_world_script\" (in that case,\nif the CI instance is rescheduled during tests, the compilation won't be\npushed to cache). Actually, it seems better to compile stuff during\n\"build\" or else any compilation warnings should up in the middle of\n\"check-world..\"\n\nAlso: I'm having second thoughts about the loop around ./configure. It\ncould happen that a cached configure would succeed, but then the build\nwould later fail, and it wouldn't fix itself. I think a slow configure\nis okay for an \"opt-in\" task.\n\nAlso: my backtrace call was using a path to cygwin rather than msys.\nThis seemed to work before, but doesn't seem to be working now...\n\n-- \nJustin","msg_date":"Mon, 5 Sep 2022 06:50:55 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-09-05 06:50:55 -0500, Justin Pryzby wrote:\n> I saw that but hadn't tracked it down yet. Do you know if the tar\n> failures were from a TAP test added since you first posted the mingw\n> patch, or ??\n\nI think it's that msys now includes tar by default, but not sure.\n\n\n> Also: your original patch said --host=x86_64-w64-mingw32, but the task\n> is called MinGW64. The most recent patches don't use --host at all, and\n> were building for a 32 bit environment, even though the OS image says\n> MSYSTEM=UCRT64.\n\nI don't think you should need to use --host, that indicates cross compiling,\nwhich we shouldn't do. I don't think the patch without --host targets 32bit -\nthe last time CI ran the patch, it built a 64bit PG:\n\nhttps://cirrus-ci.com/task/5489471041830912?logs=configure#L231\n[13:47:54.539] checking size of void *... (cached) 8\n[13:47:54.557] checking size of size_t... (cached) 8\n\nand the underlying commit didn't specify --host.\n\n\n> Also: right now src/test and src/interfaces/*/test aren't being built\n> during the build phase, which means that they're 1) not compiled in\n> parallel; and 2) not cached. This isn't specific to MinGW. Other than\n> compiling those dirs specifically, one option is to put\n> \"always: upload_caches: ccache\" after \"test_world_script\" (in that case,\n> if the CI instance is rescheduled during tests, the compilation won't be\n> pushed to cache). Actually, it seems better to compile stuff during\n> \"build\" or else any compilation warnings should up in the middle of\n> \"check-world..\"\n\nI'd tackle that independently of this commit.\n\n\n> Also: I'm having second thoughts about the loop around ./configure. It\n> could happen that a cached configure would succeed, but then the build\n> would later fail, and it wouldn't fix itself. I think a slow configure\n> is okay for an \"opt-in\" task.\n\nAgreed.\n\nI think we can convert this to meson soon, and that seems a *lot* faster at\nconfigure than autoconf on mingw. Not even close to as fast as on a modern-ish\nlinux, but not that painful.\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Mon, 5 Sep 2022 16:52:17 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"On Mon, Sep 05, 2022 at 04:52:17PM -0700, Andres Freund wrote:\n> I don't think you should need to use --host, that indicates cross compiling,\n\nThis made me consider the idea of cross-compiling for windows under a\nnew linux task, and then running tests under Windows with a dependent\ntask. I suppose that use -Og, in addition to CompilerWarnings, which\nuses -O2.\n\ncirrusci caches are (or can be) shared between tasks. I got this mostly\nworking, with a few kludges to compile and install stuff under src/test.\nThis may be yet another idea that's obsoleted by meson. WDYT?\n\n-- \nJustin\n\n\n","msg_date":"Sat, 10 Sep 2022 15:19:44 -0500","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi hackers,\n\nJustin Pryzby <pryzby@telsasoft.com>, 5 Eyl 2022 Pzt, 14:50 tarihinde şunu\nyazdı:\n\n> But cfbot should run the Mingw task for this patch's own commitfest\n> entry. But right now (because cfbot doesn't include the original commit\n> message/s), it doesn't get run :(\n>\n\nI've been thinking about how to make the mingw task run only for this patch\non cfbot and not for others. TBH, I couldn't come up with a nice way to\nachieve this.\n\nDoes anyone have any suggestions on this?\n\nAndres Freund <andres@anarazel.de>, 6 Eyl 2022 Sal, 02:52 tarihinde şunu\nyazdı:\n\n> Hi,\n>\n> On 2022-09-05 06:50:55 -0500, Justin Pryzby wrote:\n> > I saw that but hadn't tracked it down yet. Do you know if the tar\n> > failures were from a TAP test added since you first posted the mingw\n> > patch, or ??\n>\n> I think it's that msys now includes tar by default, but not sure.\n\n\nSeems like msys2 comes with GNU tar now. Previously it was bsd tar under\n\"/c/Windows/System32/tar.exe \" which now we force it to use.\n\n\nThanks,\nMelih\n\nHi hackers,Justin Pryzby <pryzby@telsasoft.com>, 5 Eyl 2022 Pzt, 14:50 tarihinde şunu yazdı:But cfbot should run the Mingw task for this patch's own commitfest\nentry. But right now (because cfbot doesn't include the original commit\nmessage/s), it doesn't get run :( I've been thinking about how to make the mingw task run only for this patch on cfbot and not for others. TBH, I couldn't come up with a nice way to achieve this.Does anyone have any suggestions on this?Andres Freund <andres@anarazel.de>, 6 Eyl 2022 Sal, 02:52 tarihinde şunu yazdı:Hi,On 2022-09-05 06:50:55 -0500, Justin Pryzby wrote:> I saw that but hadn't tracked it down yet. Do you know if the tar> failures were from a TAP test added since you first posted the mingw> patch, or ??I think it's that msys now includes tar by default, but not sure.Seems like msys2 comes with GNU tar now. Previously it was bsd tar under \"/c/Windows/System32/tar.exe \" which now we force it to use.Thanks,Melih","msg_date":"Wed, 21 Sep 2022 16:18:43 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-09-21 16:18:43 +0300, Melih Mutlu wrote:\n> I've been thinking about how to make the mingw task run only for this patch\n> on cfbot and not for others. TBH, I couldn't come up with a nice way to\n> achieve this.\n> \n> Does anyone have any suggestions on this?\n\nAdd a commented-out trigger-type: manual and note in the commit message that\nthe committer needs to adjust that piece.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 21 Sep 2022 07:04:45 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi Melih,\n\nOn 2022-09-05 16:52:17 -0700, Andres Freund wrote:\n> I think we can convert this to meson soon, and that seems a *lot* faster at\n> configure than autoconf on mingw. Not even close to as fast as on a modern-ish\n> linux, but not that painful.\n\nNow that meson has been merged, could you try to adjust this patch so it\nbuilds with meson?\n\nRegards,\n\nAndres\n\n\n","msg_date":"Thu, 22 Sep 2022 10:03:05 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi hackers.\n\nAndres Freund <andres@anarazel.de>, 22 Eyl 2022 Per, 20:03 tarihinde şunu\nyazdı:\n\n> Now that meson has been merged, could you try to adjust this patch so it\n> builds with meson?\n>\n\nAttached patch is the adjusted version to build with meson.\n\nAdd a commented-out trigger-type: manual and note in the commit message that\n> the committer needs to adjust that piece.\n>\n\nDone.\n\n> Also: your original patch said --host=x86_64-w64-mingw32, but the task\n> > is called MinGW64. The most recent patches don't use --host at all, and\n> > were building for a 32 bit environment, even though the OS image says\n> > MSYSTEM=UCRT64.\n\n\n> I don't think you should need to use --host, that indicates cross\n> compiling,\n> which we shouldn't do. I don't think the patch without --host targets\n> 32bit -\n> the last time CI ran the patch, it built a 64bit PG:\n\n\nAlso removed this with meson changes. Can confirm that it builds 64bit.\n\nJustin Pryzby <pryzby@telsasoft.com>, 5 Eyl 2022 Pzt, 14:50 tarihinde şunu\nyazdı:\n\n> I don't know what direction that idea is going, but it makes working\n> with this patch a bit easier when configure is less slow. Fine with me\n> to split it into a separate patch :)\n\n\nIf it's okay, then I would suggest not including that piece into this patch\nsince I don't know what direction this is going either.\n\n\nThe patch has been changed quite a bit due to meson integration.\nI would appreciate any feedback.\n\nBest,\nMelih","msg_date":"Fri, 7 Oct 2022 20:00:56 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-10-07 20:00:56 +0300, Melih Mutlu wrote:\n> Attached patch is the adjusted version to build with meson.\n\nThanks!\n\n\n> diff --git a/.cirrus.yml b/.cirrus.yml\n> index 9f2282471a..7e6ebdf7b7 100644\n> --- a/.cirrus.yml\n> +++ b/.cirrus.yml\n> @@ -35,6 +35,7 @@ on_failure_ac: &on_failure_ac\n> - \"**/*.log\"\n> - \"**/*.diffs\"\n> - \"**/regress_log_*\"\n> + - \"**/*.stackdump\"\n> type: text/plain\n\n\nI think it might be easier to just set MSYS=winjitdebug\nhttps://www.msys2.org/wiki/JIT-Debugging/#native-windows-processes-started-from-msys2\nand then rely on the existing windows crash reporting stuff.\n\n> setup_additional_packages_script: |\n> REM choco install -y --no-progress ...\n> -\n> # Use /DEBUG:FASTLINK to avoid high memory usage during linking\n> configure_script: |\n\nYou removed a bunch of newlines here and nearby - I assume that wasn't\nintentional?\n\n\n> + mingw_info_script:\n> + - C:\\msys64\\usr\\bin\\dash.exe -lc \"where gcc\"\n> + - C:\\msys64\\usr\\bin\\dash.exe -lc \"gcc --version\"\n> + - C:\\msys64\\usr\\bin\\dash.exe -lc \"where perl\"\n> + - C:\\msys64\\usr\\bin\\dash.exe -lc \"perl --version\"\n> +\n> + configure_script:\n> + - C:\\msys64\\usr\\bin\\dash.exe -lc \"cd %CIRRUS_WORKING_DIR% &&\n> + meson setup --buildtype debug -Dcassert=true -Db_pch=true -DTAR=%TAR% build\"\n\nThere's no need to use dash anymore, the amount of shell script run is\nminimal.\n\n\n> + build_script:\n> + - C:\\msys64\\usr\\bin\\dash.exe -lc \"cd %CIRRUS_WORKING_DIR% && ninja -C %BUILD_DIR%\"\n> + upload_caches: ccache\n> +\n> + test_world_script:\n> + - C:\\msys64\\usr\\bin\\dash.exe -lc \"cd %CIRRUS_WORKING_DIR% && meson test --print-errorlogs --num-processes %TEST_JOBS% -C %BUILD_DIR%\"\n> +\n\nI think the \"cd\"s here are superfluous given the ninja -C %BUILD_DIR% etc.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 11 Oct 2022 11:23:36 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-10-11 11:23:36 -0700, Andres Freund wrote:\n> > + build_script:\n> > + - C:\\msys64\\usr\\bin\\dash.exe -lc \"cd %CIRRUS_WORKING_DIR% && ninja -C %BUILD_DIR%\"\n> > + upload_caches: ccache\n\nOnly remembered that just after sending my email: When using b_pch=true (which\nsaves a lot of time on mingw) ccache won't cache much (IIRC just the pch files\nthemselves) unless you do something like\nexport CCACHE_SLOPPINESS=pch_defines,time_macros\n\n$ ccache -z -C\nClearing... 100.0% [==============================================================================]\n\n$ ninja clean && ccache -z && CCACHE_SLOPPINESS= time ninja src/bin/pg_test_fsync/pg_test_fsync.exe && ccache -s\n[2/2] Cleaning\nCleaning... 200 files.\nStatistics zeroed\n[124/124] Linking target src/bin/pg_test_fsync/pg_test_fsync.exe\n0.00user 0.01system 0:07.60elapsed 0%CPU (0avgtext+0avgdata 4936maxresident)k\n0inputs+0outputs (1314major+0minor)pagefaults 0swaps\nSummary:\n Hits: 0 / 1 (0.00 %)\n Direct: 0 / 1 (0.00 %)\n Preprocessed: 0 / 1 (0.00 %)\n Misses: 1\n Direct: 1\n Preprocessed: 1\n Uncacheable: 110\n\nAs you can see, most of the files are determined to be unachable. The build\ntook 7.6s.\n\n\n$ ninja clean && ccache -z && CCACHE_SLOPPINESS=pch_defines,time_macros time ninja src/bin/pg_test_fsync/pg_test_fsync.exe && ccache -s\n[2/2] Cleaning\nCleaning... 200 files.\nStatistics zeroed\n[124/124] Linking target src/bin/pg_test_fsync/pg_test_fsync.exe\n0.00user 0.01system 0:09.91elapsed 0%CPU (0avgtext+0avgdata 4936maxresident)k\n0inputs+0outputs (1313major+0minor)pagefaults 0swaps\nSummary:\n Hits: 0 / 111 (0.00 %)\n Direct: 0 / 111 (0.00 %)\n Preprocessed: 0 / 111 (0.00 %)\n Misses: 111\n Direct: 111\n Preprocessed: 111\nPrimary storage:\n Hits: 0 / 216 (0.00 %)\n Misses: 216\n Cache size (GB): 0.05 / 5.00 (0.98 %)\n\nFiles are cachable, but are cache misses (because we cleaned the cache\nabove). The build took 9.91s.\n\n\n$ ninja clean && ccache -z && CCACHE_SLOPPINESS=pch_defines,time_macros time ninja src/bin/pg_test_fsync/pg_test_fsync.exe && ccache -s\n[2/2] Cleaning\nCleaning... 200 files.\nStatistics zeroed\n[124/124] Linking target src/bin/pg_test_fsync/pg_test_fsync.exe\n0.00user 0.01system 0:02.40elapsed 0%CPU (0avgtext+0avgdata 4936maxresident)k\n0inputs+0outputs (1314major+0minor)pagefaults 0swaps\nSummary:\n Hits: 111 / 111 (100.0 %)\n Direct: 111 / 111 (100.0 %)\n Preprocessed: 0 / 0\n Misses: 0\n Direct: 0\n Preprocessed: 0\nPrimary storage:\n Hits: 222 / 222 (100.0 %)\n Misses: 0\n Cache size (GB): 0.05 / 5.00 (0.98 %)\n\nFiles are cachable, and hit the cache. The build takes 2.4s.\n\n\nUsing ccache's depend mode reduce the cache-miss case to 7.51s and the cache\nhit case to 1.75s. So we should likely export CCACHE_DEPEND=1 as well.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 11 Oct 2022 11:46:12 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi Andres,\n\nAndres Freund <andres@anarazel.de>, 11 Eki 2022 Sal, 21:23 tarihinde şunu\nyazdı:\n\n> I think it might be easier to just set MSYS=winjitdebug\n>\n> https://www.msys2.org/wiki/JIT-Debugging/#native-windows-processes-started-from-msys2\n> and then rely on the existing windows crash reporting stuff.\n>\n\nDone.\n\n\n> You removed a bunch of newlines here and nearby - I assume that wasn't\n> intentional?\n>\n\nYes, that was my mistake. Thanks for pointing it out. Fixed.\n\n\n> There's no need to use dash anymore, the amount of shell script run is\n> minimal.\n>\n\nSwitched back to bash.\n\n\n> I think the \"cd\"s here are superfluous given the ninja -C %BUILD_DIR% etc.\n>\n\nYou're right. Those were needed when it was building with autoconf. Not\nanymore. Removed.\n\nOnly remembered that just after sending my email: When using b_pch=true\n> (which\n> saves a lot of time on mingw) ccache won't cache much (IIRC just the pch\n> files\n> themselves) unless you do something like\n> export CCACHE_SLOPPINESS=pch_defines,time_macros\n>\n\nAdded CCACHE_SLOPPINESS and CCACHE_SLOPPINESS variables.\n\nYou can find the updated patch attached.\n\nBest,\nMelih","msg_date":"Tue, 18 Oct 2022 12:49:02 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nThanks for the patch!\n\nOn 10/18/22 12:49, Melih Mutlu wrote:\n>\n>\n> You can find the updated patch attached.\n>\n\nDoes it makes sense to set these at 'Windows - Server 2019, MinGW64 - \nMeson' task:\n\nenv:\n ...\n CHERE_INVOKING: 1\n BASH_EXE: C:\\msys64\\usr\\bin\\bash.exe\n\n'CHERE_INVOKING: 1' will cause bash.exe to start from current working \ndirectory(%CIRRUS_WORKING_DIR%).\n\nIn this way, there is no need for\n'cd %CIRRUS_WORKING_DIR%' when using bash.exe,\nalso no need for 'BUILD_DIR' environment variable.\n\ni.e:\n\nconfigure_script: |\n %BASH_EXE% -lc \"meson setup --buildtype debug -Dcassert=true \n-Db_pch=true -DTAR=%TAR% build\"\n\nwill work.\n\nRegards,\nNazir Bilal Yavuz\n\n\n\n\n\nHi,\n\n Thanks for the patch!\n\n\nOn 10/18/22 12:49, Melih Mutlu wrote:\n\n\n\n\n\n\n\n\nYou can find the\n updated patch attached.\n\n\n\n\nDoes it makes sense to set\n these at 'Windows - Server 2019, MinGW64 - Meson' task:\n\nenv:\n ...\n CHERE_INVOKING: 1\n BASH_EXE:\n C:\\msys64\\usr\\bin\\bash.exe\n\n'CHERE_INVOKING: 1' will\n cause bash.exe to start from current working\n directory(%CIRRUS_WORKING_DIR%).\n\nIn this way, there is no\n need for\n'cd %CIRRUS_WORKING_DIR%'\n when using bash.exe,\nalso no need for 'BUILD_DIR' environment variable.\n\ni.e:\n\nconfigure_script: |\n %BASH_EXE% -lc \"meson\n setup --buildtype debug -Dcassert=true -Db_pch=true -DTAR=%TAR%\n build\"\n\nwill work.\n\n Regards,\n Nazir Bilal Yavuz","msg_date":"Tue, 18 Oct 2022 15:37:32 +0300","msg_from":"Nazir Bilal Yavuz <byavuz81@gmail.com>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi Bilal,\n\nNazir Bilal Yavuz <byavuz81@gmail.com>, 18 Eki 2022 Sal, 15:37 tarihinde\nşunu yazdı:\n\n> env:\n> ...\n> CHERE_INVOKING: 1\n> BASH_EXE: C:\\msys64\\usr\\bin\\bash.exe\n>\n> 'CHERE_INVOKING: 1' will cause bash.exe to start from current working\n> directory(%CIRRUS_WORKING_DIR%).\n>\n> In this way, there is no need for\n> 'cd %CIRRUS_WORKING_DIR%' when using bash.exe,\n> also no need for 'BUILD_DIR' environment variable.\n>\n\n Right, setting CHERE_INVOKING and removing all cd's work and look better.\nThanks for the suggestion.\n\nHere's the updated patch.\n\nBest,\nMelih","msg_date":"Wed, 19 Oct 2022 00:23:46 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-10-19 00:23:46 +0300, Melih Mutlu wrote:\n> Right, setting CHERE_INVOKING and removing all cd's work and look better.\n> Thanks for the suggestion.\n\nAgreed, good idea.\n\n\n> +\n> + env:\n> + CCACHE_DIR: C:/msys64/ccache\n\nIt's a bit odd to separate the CCACHE_* variables from each other\n(e.g. BUILD_DIR is inbetween them)...\n\n\n> + BUILD_DIR: \"%CIRRUS_WORKING_DIR%/build\"\n> + PYTHONHOME: C:/msys64/ucrt64\n\nPerhaps add a comment explaining that otherwise plpython tests fail?\n\n\n> + MSYS: winjitdebug\n\nWith this src/tools/ci/cores_backtrace.sh shouldn't need to be modified\nanymore afaict?\n\n\n> + CCACHE_SLOPPINESS: pch_defines,time_macros\n> + CCACHE_DEPEND: 1\n\nI experimented a bit and it looks like ccache doesn't yet quite work in CI,\nbut only because the ccache needs to be larger. Looks like we need about\n~400MB.\n\nA fully cached build is ~2min\n\n\n> + configure_script: |\n> + %BASH_EXE% -lc \"meson setup --buildtype debug -Dcassert=true -Db_pch=true -DTAR=%TAR% build\"\n\nWith these buildflags the tests take about 23min37s. Using -Og I saw 18min26s,\nwith -O1 18m57s and with -O2 18m38s. There's obviously a fair bit of variance,\nbut it looks like we should use -Og. I think we considered that making compile\ntimes too bad before, but it seems kinda ok now, with 11min. -O2 is 13min,\nwithout providing further benefits.\n\nI'd replace --buildtype debug with -Ddebug=true -Doptimization=g.\n\n\n> + build_script: |\n> + %BASH_EXE% -lc \"cd %CIRRUS_WORKING_DIR% && ninja -C build\"\n>\nWhy do we still need this cd?\n\n\n> + upload_caches: ccache\n> +\n> + test_world_script: |\n> + %BASH_EXE% -lc \"meson test --print-errorlogs --num-processes %TEST_JOBS% -C build\"\n\nSeems like this could use %MTEST_ARGS%?\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 18 Oct 2022 20:19:28 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nAndres Freund <andres@anarazel.de>, 19 Eki 2022 Çar, 06:19 tarihinde şunu\nyazdı:\n\n> It's a bit odd to separate the CCACHE_* variables from each other\n> (e.g. BUILD_DIR is inbetween them)...\n>\n> Perhaps add a comment explaining that otherwise plpython tests fail?\n\n\n>\nWith this src/tools/ci/cores_backtrace.sh shouldn't need to be modified\n> anymore afaict?\n>\n> I'd replace --buildtype debug with -Ddebug=true -Doptimization=g.\n>\n\n\nSeems like this could use %MTEST_ARGS%?\n\n\nAll the things you mentioned above are done.\n\n\n\n> > + CCACHE_SLOPPINESS: pch_defines,time_macros\n> > + CCACHE_DEPEND: 1\n>\n> I experimented a bit and it looks like ccache doesn't yet quite work in CI,\n> but only because the ccache needs to be larger. Looks like we need about\n> ~400MB.\n>\n> A fully cached build is ~2min\n>\n\nThen I should increase CCACHE_MAXSIZE, right? Made it 500MB for MinGW.\n\nSharing the updated version of the patch.\n\nThanks,\nMelih","msg_date":"Wed, 19 Oct 2022 17:59:22 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-10-19 17:59:22 +0300, Melih Mutlu wrote:\n> All the things you mentioned above are done.\n\n> Then I should increase CCACHE_MAXSIZE, right? Made it 500MB for MinGW.\n\nYes.\n\n> +\n> + on_failure:\n> + <<: *on_failure_meson\n> + cores_script: |\n> + %BASH_EXE% -lc \"cd build src/tools/ci/cores_backtrace.sh msys build/tmp_install\"\n> +\n\nThis is wrong - it should just archive the same files that the current windows\ntask does.\n\nOther than that, I think this is basically ready?\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Wed, 19 Oct 2022 09:23:28 -0700","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\n\n> > +\n> > + on_failure:\n> > + <<: *on_failure_meson\n> > + cores_script: |\n> > + %BASH_EXE% -lc \"cd build src/tools/ci/cores_backtrace.sh msys\n> build/tmp_install\"\n> > +\n>\n> This is wrong - it should just archive the same files that the current\n> windows\n> task does.\n>\n\nChanged it with the on_failure from the other windows task.\n\n\n> Other than that, I think this is basically ready?\n>\n\nIf you say so, then I think it's ready.\n\nBest,\nMelih","msg_date":"Mon, 24 Oct 2022 14:13:06 +0300","msg_from":"Melih Mutlu <m.melihmutlu@gmail.com>","msg_from_op":true,"msg_subject":"Re: Mingw task for Cirrus CI"},{"msg_contents":"Hi,\n\nOn 2022-10-24 14:13:06 +0300, Melih Mutlu wrote:\n> If you say so, then I think it's ready.\n\nI pushed it with a few changes:\n\n- I added an only_if, otherwise the task shows up (without running) even if\n you use ci-os-only: linux\n- I added -Dnls=disabled - It seemed to add about 1.5min to the cached build\n- Inlined the -l into BASH_EXE and renamed it to BASH\n- The indentation of WINDOWS_ENVIRONMENT_BASE was irregular, leading to a\n larger diff. Some other similar cleanups.\n- I added 'mingw' as a ci-os-only choice. Perhaps not the\n be-all-end-all. Perhaps we should use windows-{msvc,mingw,cygwin}? Or add\n ci-task-only?\n\n\nI suspect we can make at least -Dssl=openssl work. Perhaps one of the uuid\nimplementation is available in msys as well?\n\nCurrently we end up with:\n External libraries\n bonjour : NO\n bsd_auth : NO\n gss : NO\n icu : YES 72.1\n ldap : YES\n libxml : YES 2.10.3\n libxslt : YES 1.1.37\n llvm : NO\n lz4 : YES 1.9.4\n nls : NO\n pam : NO\n plperl : YES\n plpython : YES 3.10\n pltcl : YES\n readline : YES\n selinux : NO\n ssl : NO\n systemd : NO\n uuid : NO\n zlib : YES 1.2.13\n zstd : YES 1.5.2\n\ngss should also work if the is available in the msys repo.\n\nBonjour is apple only, bsd_auth bsd specific, selinux and systemd linux\nspecific and pam unixoid specific.\n\nIt could also be interesting to see if llvm works, but that might be more\nwork.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Thu, 17 Nov 2022 16:23:56 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Mingw task for Cirrus CI"}],"string":"[\n {\n \"msg_contents\": \"Hi All,\\n\\nI've been working on adding Windows+MinGW environment into cirrus-ci tasks\\n(discussion about ci is here [1]).\\nIt uses MSYS2 to set the environment. UCRT is chosen as C standard library,\\ninstead of MSVCRT.\\nThe task configures postgres with features that are available in MSYS2 (see\\navailable packages [2]) and tap tests are enabled.\\nI already added the necessary docker image, you can find the related PR at\\n[3] and a successful cirruc-ci run with these changes at [4].\\nAttached patch adds a task runs on Windows with MinGW for cirrus-ci\\n\\nHowever, I cannot run configure with --with-python, --with-perl or\\n--with-tcl.\\nThere are two issues I encountered while trying to enable. e.g. for\\n--with-python\\n\\n1- python_ldlibrary is set to \\\"libpython3.9.dll.a\\\". So the related line in\\nconfigure cannot retrieve \\\"ldlibrary\\\"\\n2- src/pl/plpython/Makefile looks under \\\"C:/Windows/system32/\\\" for\\nPYTHONDLL. gendef cannot open that file, give an error like \\\" failed to\\nopen()\\\" when creating a .def file.\\n\\nTo overcome those, I added the correct pattern to extract ldlibrary by\\nappending \\\"-e 's/\\\\.dll.a$//'\\\" at the end of the related line.\\nThen, I tried to use python dll from MSYS instead of windows/system32 by\\nchanging PYTHONDLL path to \\\"/ucrt64/bin/libpython3.9.dll\\\". I'm not sure if\\nthis is the correct dll.\\nHere is the diff of these two changes:\\n\\ndiff --git a/configure b/configure\\nindex f3cb5c2b51..42ea580442 100755\\n--- a/configure\\n+++ b/configure\\n@@ -10536,7 +10536,7 @@ python_libdir=`${PYTHON} -c \\\"import sysconfig;\\nprint(' '.join(filter(None,syscon\\n python_ldlibrary=`${PYTHON} -c \\\"import sysconfig; print('\\n'.join(filter(None,sysconfig.get_config_vars('LDLIBRARY'))))\\\"`\\n\\n # If LDLIBRARY exists and has a shlib extension, use it verbatim.\\n-ldlibrary=`echo \\\"${python_ldlibrary}\\\" | sed -e 's/\\\\.so$//' -e 's/\\\\.dll$//'\\n-e 's/\\\\.dylib$//' -e 's/\\\\.sl$//'`\\n+ldlibrary=`echo \\\"${python_ldlibrary}\\\" | sed -e 's/\\\\.so$//' -e 's/\\\\.dll$//'\\n-e 's/\\\\.dylib$//' -e 's/\\\\.sl$//' -e 's/\\\\.dll.a$//'`\\n if test -e \\\"${python_libdir}/${python_ldlibrary}\\\" -a\\nx\\\"${python_ldlibrary}\\\" != x\\\"${ldlibrary}\\\"\\n then\\n ldlibrary=`echo \\\"${ldlibrary}\\\" | sed \\\"s/^lib//\\\"`\\ndiff --git a/src/pl/plpython/Makefile b/src/pl/plpython/Makefile\\nindex a83ae8865c..4254ef94d7 100644\\n--- a/src/pl/plpython/Makefile\\n+++ b/src/pl/plpython/Makefile\\n@@ -61,7 +61,8 @@ INCS = plpython.h \\\\\\n ifeq ($(PORTNAME), win32)\\n\\n pytverstr=$(subst .,,${python_version})\\n-PYTHONDLL=$(subst \\\\,/,$(WINDIR))/system32/python${pytverstr}.dll\\n+#PYTHONDLL=$(subst \\\\,/,$(WINDIR))/system32/python${pytverstr}.dll\\n+PYTHONDLL=/ucrt64/bin/libpython3.9.dll\\n\\n OBJS += libpython${pytverstr}.a\\n\\n\\n\\nIn the end, make check-world still fails, even though I was able to run\\nconfigure and make without any obvious error.\\nI see bunch of errors in tests like:\\n+ERROR: language \\\"plpython3u\\\" does not exist\\n+HINT: Use CREATE EXTENSION to load the language into the database.\\n\\nHere is the logs from failed ci run:\\nhttps://api.cirrus-ci.com/v1/artifact/task/4645682031099904/log/build/src/pl/plpython/regression.diffs\\n\\n\\nAny thoughts on how postgres can be built with --with-python etc. on mingw?\\n\\nBest,\\nMelih\\n\\n\\n[1]\\nhttps://www.postgresql.org/message-id/flat/20211001222752.wrz7erzh4cajvgp6%40alap3.anarazel.de\\n<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.postgresql.org%2Fmessage-id%2Fflat%2F20211001222752.wrz7erzh4cajvgp6%2540alap3.anarazel.de&data=04%7C01%7CMelih.Mutlu%40microsoft.com%7C7f4780e35ab34e2f38d808d9f7d008e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637813295184245763%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=HkjuBWcQzmiyIyoCuIfKCzAEF9ME6SpGOYjBJMFhQvw%3D&reserved=0>\\n[2] https://packages.msys2.org/package/\\n<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fpackages.msys2.org%2Fpackage%2F&data=04%7C01%7CMelih.Mutlu%40microsoft.com%7C7f4780e35ab34e2f38d808d9f7d008e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637813295184245763%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=VqfdnrxHv15FUGFXgnNaMli5n2R1tnPPglGq4xWqND0%3D&reserved=0>\\n[3] https://github.com/anarazel/pg-vm-images/pull/8\\n<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fanarazel%2Fpg-vm-images%2Fpull%2F8&data=04%7C01%7CMelih.Mutlu%40microsoft.com%7C7f4780e35ab34e2f38d808d9f7d008e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637813295184245763%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=IASIgPhLd0%2Fvh1VSn9EC%2BCsH42riE%2Bt6gQPObOb0QmI%3D&reserved=0>\\n[4] https://cirrus-ci.com/build/4999469182746624\\n<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fcirrus-ci.com%2Fbuild%2F5013577319841792&data=04%7C01%7CMelih.Mutlu%40microsoft.com%7C7f4780e35ab34e2f38d808d9f7d008e0%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637813295184245763%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=hMGCtgm1uVBWzFHXPEl%2FiaavzWbEqxAuhmKjFcjH%2BF0%3D&reserved=0>\",\n \"msg_date\": \"Fri, 25 Feb 2022 19:44:27 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nAndrew, CCIng you both because you might be interested in the CI bit, and\\nbecause you might know the answer.\\n\\nOn 2022-02-25 19:44:27 +0300, Melih Mutlu wrote:\\n> I've been working on adding Windows+MinGW environment into cirrus-ci tasks\\n> (discussion about ci is here [1]).\\n> It uses MSYS2 to set the environment. UCRT is chosen as C standard library,\\n> instead of MSVCRT.\\n> The task configures postgres with features that are available in MSYS2 (see\\n> available packages [2]) and tap tests are enabled.\\n> I already added the necessary docker image, you can find the related PR at\\n> [3] and a successful cirruc-ci run with these changes at [4].\\n> Attached patch adds a task runs on Windows with MinGW for cirrus-ci\\n>\\n> However, I cannot run configure with --with-python, --with-perl or\\n> --with-tcl.\\n> There are two issues I encountered while trying to enable. e.g. for\\n> --with-python\\n>\\n> 1- python_ldlibrary is set to \\\"libpython3.9.dll.a\\\". So the related line in\\n> configure cannot retrieve \\\"ldlibrary\\\"\\n\\nThis presumably is due to using mingw's python rather than python.org's\\npython. Seems like a reasonable thing to support for the mingw build.\\n\\nMelih, you could try to build against the python.org python (i installed in\\nthe CI container).\\n\\n\\n> 2- src/pl/plpython/Makefile looks under \\\"C:/Windows/system32/\\\" for\\n> PYTHONDLL. gendef cannot open that file, give an error like \\\" failed to\\n> open()\\\" when creating a .def file.\\n\\nOn my win10 VM in which I installed python.org python I don't see a python dll\\nin c:/windows/system32 either. Looks like none of our windows mingw animals\\nbuild with python. So it might just be bitrot.\\n\\n\\n> In the end, make check-world still fails, even though I was able to run\\n> configure and make without any obvious error.\\n> I see bunch of errors in tests like:\\n> +ERROR: language \\\"plpython3u\\\" does not exist\\n> +HINT: Use CREATE EXTENSION to load the language into the database.\\n\\n> Here is the logs from failed ci run:\\n> https://api.cirrus-ci.com/v1/artifact/task/4645682031099904/log/build/src/pl/plpython/regression.diffs\\n\\nThe URL to the rest of the CI run is https://cirrus-ci.com/task/4645682031099904\\n\\n\\nThe relevant failure is earlier:\\n\\n+ERROR: could not load library \\\"C:/cirrus/build/tmp_install/ucrt64/lib/postgresql/plpython3.dll\\\": The specified module could not be found.\\n\\nClearly plpython is being built, because we see some warnings:\\n\\n[22:44:24.456] C:/msys64/ucrt64/include/python3.9/pyconfig.h:1474: warning: \\\"SIZEOF_OFF_T\\\" redefined\\n[22:44:24.456] 1474 | #define SIZEOF_OFF_T 8\\n[22:44:24.456] |\\n[22:44:24.456] In file included from c:/cirrus/src/include/c.h:54,\\n[22:44:24.456] from c:/cirrus/src/include/postgres.h:46,\\n[22:44:24.456] from c:/cirrus/contrib/jsonb_plpython/jsonb_plpython.c:1:\\n[22:44:24.456] ../../src/include/pg_config.h:875: note: this is the location of the previous definition\\n[22:44:24.456] 875 | #define SIZEOF_OFF_T 4\\n\\nSeems we're doing something wrong and end up with a 4 byte off_t, whereas\\npython ends up with an 8byte one. We probably need to fix that. But it's not\\nthe cause of this problem.\\n\\n\\nYou could take out -s from the make flags and see whether plpython3.dll is\\nbeing built and installed, and where to.\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 25 Feb 2022 16:27:46 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"\\nOn 2/25/22 19:27, Andres Freund wrote:\\n> Hi,\\n>\\n> Andrew, CCIng you both because you might be interested in the CI bit, and\\n> because you might know the answer.\\n>\\n>\\n>\\n>> 2- src/pl/plpython/Makefile looks under \\\"C:/Windows/system32/\\\" for\\n>> PYTHONDLL. gendef cannot open that file, give an error like \\\" failed to\\n>> open()\\\" when creating a .def file.\\n> On my win10 VM in which I installed python.org python I don't see a python dll\\n> in c:/windows/system32 either. Looks like none of our windows mingw animals\\n> build with python. So it might just be bitrot.\\n\\n\\n\\nThere certainly was a time when python from python.org used to install\\nits DLL in the system32 directory, so I imagine that's why it's there.\\nI'm very glad to see that's no longer the case.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 11:10:45 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi Andres,\\n\\n\\n> This presumably is due to using mingw's python rather than python.org's\\n> python. Seems like a reasonable thing to support for the mingw build.\\n>\\n> Melih, you could try to build against the python.org python (i installed\\n> in\\n> the CI container).\\n>\\n\\nI tried to use the installed python from python.org in the container.\\nThe problem with this is that \\\"LIBDIR\\\" and \\\"LDLIBRARY\\\" configs of python\\nfor windows from python.org are empty. Therefore python_libdir or other\\nrelated variables in configure file are not set correctly.\\n\\n\\nSeems we're doing something wrong and end up with a 4 byte off_t, whereas\\n> python ends up with an 8byte one. We probably need to fix that. But it's\\n> not\\n> the cause of this problem.\\n>\\n>\\n> You could take out -s from the make flags and see whether plpython3.dll is\\n> being built and installed, and where to.\\n>\\n\\nHere is a run without -s: https://cirrus-ci.com/task/4569363104661504\\nI couldn't get what's wrong from these logs to be honest. But I see that\\nplpython3.dll exists under src/pl/plpython after build when I run these\\nsteps locally.\\n\\nBest,\\nMelih\\n\\nHi Andres, This presumably is due to using mingw's python rather than python.org's\\npython. Seems like a reasonable thing to support for the mingw build.\\n\\nMelih, you could try to build against the python.org python (i installed in\\nthe CI container).I tried to use the installed python from python.org in the container. The problem with this is that \\\"LIBDIR\\\" and \\\"LDLIBRARY\\\" configs of python for windows from python.org are empty. Therefore python_libdir or other related variables in configure file are not set correctly.\\nSeems we're doing something wrong and end up with a 4 byte off_t, whereas\\npython ends up with an 8byte one. We probably need to fix that. But it's not\\nthe cause of this problem.\\n\\n\\nYou could take out -s from the make flags and see whether plpython3.dll is\\nbeing built and installed, and where to.Here is a run without -s: https://cirrus-ci.com/task/4569363104661504 I couldn't get what's wrong from these logs to be honest. But I see that plpython3.dll exists under src/pl/plpython after build when I run these steps locally.Best,Melih\",\n \"msg_date\": \"Thu, 3 Mar 2022 13:16:15 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"\\nOn 3/3/22 05:16, Melih Mutlu wrote:\\n> Hi Andres, \\n> \\n>\\n> This presumably is due to using mingw's python rather than\\n> python.org <http://python.org>'s\\n> python. Seems like a reasonable thing to support for the mingw build.\\n>\\n> Melih, you could try to build against the python.org\\n> <http://python.org> python (i installed in\\n> the CI container).\\n>\\n>\\n> I tried to use the installed python from python.org\\n> <http://python.org> in the container. \\n> The problem with this is that \\\"LIBDIR\\\" and \\\"LDLIBRARY\\\" configs of\\n> python for windows from python.org <http://python.org> are empty.\\n> Therefore python_libdir or other related variables in configure file\\n> are not set correctly.\\n\\n\\n\\nYeah, here's what it has:\\n\\n\\n# python -c \\\"import sysconfig; import pprint; pp =\\npprint.PrettyPrinter(); pp.pprint(sysconfig.get_config_vars())\\\"\\n{'BINDIR': 'C:\\\\\\\\prog\\\\\\\\python310',\\n 'BINLIBDEST': 'C:\\\\\\\\prog\\\\\\\\python310\\\\\\\\Lib',\\n 'EXE': '.exe',\\n 'EXT_SUFFIX': '.cp310-win_amd64.pyd',\\n 'INCLUDEPY': 'C:\\\\\\\\prog\\\\\\\\python310\\\\\\\\Include',\\n 'LIBDEST': 'C:\\\\\\\\prog\\\\\\\\python310\\\\\\\\Lib',\\n 'SO': '.cp310-win_amd64.pyd',\\n 'TZPATH': '',\\n 'VERSION': '310',\\n 'abiflags': '',\\n 'base': 'C:\\\\\\\\prog\\\\\\\\python310',\\n 'exec_prefix': 'C:\\\\\\\\prog\\\\\\\\python310',\\n 'installed_base': 'C:\\\\\\\\prog\\\\\\\\python310',\\n 'installed_platbase': 'C:\\\\\\\\prog\\\\\\\\python310',\\n 'platbase': 'C:\\\\\\\\prog\\\\\\\\python310',\\n 'platlibdir': 'lib',\\n 'prefix': 'C:\\\\\\\\prog\\\\\\\\python310',\\n 'projectbase': 'C:\\\\\\\\prog\\\\\\\\python310',\\n 'py_version': '3.10.2',\\n 'py_version_nodot': '310',\\n 'py_version_nodot_plat': '310',\\n 'py_version_short': '3.10',\\n 'srcdir': 'C:\\\\\\\\prog\\\\\\\\python310',\\n 'userbase': 'C:\\\\\\\\Users\\\\\\\\Administrator\\\\\\\\AppData\\\\\\\\Roaming\\\\\\\\Python'}\\n\\n\\nThe DLL lives in the BINDIR, so maybe I guess we should search there if\\nwe can't get the other things.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Mar 2022 07:52:27 -0500\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-25 19:44:27 +0300, Melih Mutlu wrote:\\n> I've been working on adding Windows+MinGW environment into cirrus-ci tasks\\n> (discussion about ci is here [1]).\\n\\nThis doesn't apply anymore: http://cfbot.cputube.org/patch_37_3575.log\\n\\nCould you rebase?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 21 Mar 2022 17:15:59 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi Andres,\\n\\nRebased it.\\nI also removed the temp installation task and\\nused NoDefaultCurrentDirectoryInExePath env variable instead.\\n\\nBest,\\nMelih\",\n \"msg_date\": \"Tue, 22 Mar 2022 19:00:42 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-22 19:00:42 +0300, Melih Mutlu wrote:\\n> Rebased it.\\n> I also removed the temp installation task and\\n> used NoDefaultCurrentDirectoryInExePath env variable instead.\\n\\nHm. But you're still using a separate build directory, from what I can see?\\nThe NoDefaultCurrentDirectoryInExePath thing should only have an effect when\\nnot using a separate build directory, no?\\n\\nDoes it work to not use the separate build dir? Without it we don't need the\\nthe \\\"preparing build tree\\\" step, and that's quite slow on mingw:\\nhttps://cirrus-ci.com/task/4713509253545984?logs=configure#L392\\n\\n[00:23:44.371] preparing build tree... done\\n[00:24:25.429] configure: creating ./config.status\\n\\n\\nChatting about this patch with Thomas I started to wonder about other reasons\\nfor the slow speed of configure. I briefly experimented locally, and it looks\\nlike using 'dash' as the shell makes configure run a good bit quicker.\\n\\n\\n> ---\\n> .cirrus.yml | 79 +++++++++++++++++++++++++++++++++++++++++++----------\\n> 1 file changed, 65 insertions(+), 14 deletions(-)\\n> \\n> diff --git a/.cirrus.yml b/.cirrus.yml\\n> index e5335fede7..1ed40347cf 100644\\n> --- a/.cirrus.yml\\n> +++ b/.cirrus.yml\\n> @@ -23,7 +23,6 @@ env:\\n> CHECKFLAGS: -Otarget\\n> PROVE_FLAGS: --timer\\n> PGCTLTIMEOUT: 120 # avoids spurious failures during parallel tests\\n> - TEMP_CONFIG: ${CIRRUS_WORKING_DIR}/src/tools/ci/pg_ci_base.conf\\n> PG_TEST_EXTRA: kerberos ldap ssl\\n\\nThis removes TEMP_CONFIG from all other tasks. You added it back to the VS\\nwindows task, but not the others? I assume that was accidental?\\n\\n\\n> + env:\\n> + CCACHE_DIR: C:/msys64/ccache\\n> + BUILD_DIR: \\\"%CIRRUS_WORKING_DIR%/build\\\"\\n\\nI think this should use TEMP_CONFIG too. Is the problem that you need to\\nchange the path?\\n\\n\\n> + ccache_cache:\\n> + folder: ${CCACHE_DIR}\\n> +\\n> + mingw_info_script:\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\bash.exe -lc \\\"where gcc\\\"\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\bash.exe -lc \\\"gcc --version\\\"\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\bash.exe -lc \\\"where perl\\\"\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\bash.exe -lc \\\"perl --version\\\"\\n> +\\n> + configure_script:\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\bash.exe -lc \\\"mkdir %BUILD_DIR% &&\\n> + cd %BUILD_DIR% &&\\n> + %CIRRUS_WORKING_DIR%/configure\\n\\nCould you try using dash to invoke configure here, and whether it makes configure faster?\\n\\n\\n> + --host=x86_64-w64-mingw32\\n> + --enable-cassert\\n> + --enable-tap-tests\\n> + --with-icu\\n> + --with-libxml\\n> + --with-libxslt\\n> + --with-lz4\\n> + --enable-debug\\n> + CC='ccache gcc'\\n> + CXX='ccache g++'\\\"\\n\\nI think this task should specify CFLAGS=\\\"-Og\\\", CXXFLAGS=\\\"-Og\\\" similar to other\\ntasks. We end up with -O2 otherwise, which makes the build measurably slower.\\n\\n\\n\\n> + tests_script:\\n> + - set \\\"NoDefaultCurrentDirectoryInExePath=0\\\"\\n\\nA comment about why NoDefaultCurrentDirectoryInExePath=0 is used would be\\ngood.\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 30 Mar 2022 17:26:18 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-30 17:26:18 -0700, Andres Freund wrote:\\n> Hi,\\n>\\n> On 2022-03-22 19:00:42 +0300, Melih Mutlu wrote:\\n> > Rebased it.\\n> > I also removed the temp installation task and\\n> > used NoDefaultCurrentDirectoryInExePath env variable instead.\\n>\\n> Hm. But you're still using a separate build directory, from what I can see?\\n> The NoDefaultCurrentDirectoryInExePath thing should only have an effect when\\n> not using a separate build directory, no?\\n\\nMelih chatted with me about making this work. Turns out it doesn't readily -\\npg_ctl still fails.\\n\\n\\nThe reason that NoDefaultCurrentDirectoryInExePath doesn't suffice to get a\\nworking in-tree build, is that we break it ourselves:\\n\\nint\\nfind_my_exec(const char *argv0, char *retpath)\\n...\\n#ifdef WIN32\\n\\t/* Win32 checks the current directory first for names without slashes */\\n\\tjoin_path_components(retpath, cwd, argv0);\\n\\tif (validate_exec(retpath) == 0)\\n\\t\\treturn resolve_symlinks(retpath);\\n#endif\\n\\nSo even if windows doesn't actually use the current path, and the current\\npg_ctl process isn't the one from the current directory, we *still* return\\nthat.\\n\\nGah.\\n\\n\\nMaybe we should just use GetModuleFileName()?\\n\\n\\nBut even after that the tests don't work. Commands started via IPC::Run do,\\nbut when using system() they don't. Looks like perl parses the path the itself\\n:(.\\n\\n- Andres\\n\\n\\n\",\n \"msg_date\": \"Mon, 4 Apr 2022 13:41:31 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"\\nOn 4/4/22 16:41, Andres Freund wrote:\\n> Hi,\\n>\\n> On 2022-03-30 17:26:18 -0700, Andres Freund wrote:\\n>> Hi,\\n>>\\n>> On 2022-03-22 19:00:42 +0300, Melih Mutlu wrote:\\n>>> Rebased it.\\n>>> I also removed the temp installation task and\\n>>> used NoDefaultCurrentDirectoryInExePath env variable instead.\\n>> Hm. But you're still using a separate build directory, from what I can see?\\n>> The NoDefaultCurrentDirectoryInExePath thing should only have an effect when\\n>> not using a separate build directory, no?\\n> Melih chatted with me about making this work. Turns out it doesn't readily -\\n> pg_ctl still fails.\\n>\\n>\\n> The reason that NoDefaultCurrentDirectoryInExePath doesn't suffice to get a\\n> working in-tree build, is that we break it ourselves:\\n>\\n> int\\n> find_my_exec(const char *argv0, char *retpath)\\n> ...\\n> #ifdef WIN32\\n> \\t/* Win32 checks the current directory first for names without slashes */\\n> \\tjoin_path_components(retpath, cwd, argv0);\\n> \\tif (validate_exec(retpath) == 0)\\n> \\t\\treturn resolve_symlinks(retpath);\\n> #endif\\n>\\n> So even if windows doesn't actually use the current path, and the current\\n> pg_ctl process isn't the one from the current directory, we *still* return\\n> that.\\n>\\n> Gah.\\n\\n\\n\\n\\nI notice a few things about the latest file in this thread.\\n\\nYou should set MSYSTEM=UCRT64 in the environment section. Given that,\\nthere should be no need to specify a --host= setting for configure.\\n\\nIf it's not done already, the shebang line in\\n/ucrt64/bin/core_perl/prove needs to be modified to use /ucrt64/bin/perl.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Tue, 5 Apr 2022 10:16:36 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"\\nOn 3/30/22 20:26, Andres Freund wrote:\\n> Could you try using dash to invoke configure here, and whether it makes configure faster?\\n>\\n>\\n\\n\\nI got weird failures re libxml/parser.h when I tried with dash. See\\n<https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\ncould see config.log on failure.)\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Apr 2022 11:03:37 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\\n> On 3/30/22 20:26, Andres Freund wrote:\\n> > Could you try using dash to invoke configure here, and whether it makes configure faster?\\n> I got weird failures re libxml/parser.h when I tried with dash.\\n\\nHm. Hadn't enabled that when I tried...\\n\\n\\n> See <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\n> could see config.log on failure.)\\n\\nAll *.log files are preserved on failure. There's a file directory brower at\\nthe top to navigate around:\\nhttps://api.cirrus-ci.com/v1/artifact/task/5963254039052288/log/build/config.log\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 6 Apr 2022 09:34:19 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\\n> On 3/30/22 20:26, Andres Freund wrote:\\n> > Could you try using dash to invoke configure here, and whether it makes configure faster?\\n> I got weird failures re libxml/parser.h when I tried with dash. See\\n> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\n> could see config.log on failure.)\\n\\nSince dash won't help us to get the build time down sufficiently, and the\\ntests don't pass without a separate build tree, I looked at what makes\\nconfig/prep_buildtree so slow.\\n\\nIt's largely just bad code. The slowest part are spawning one expr and mkdir\\n-p for each directory. One 'cmp' for each makefile doesn't help either.\\n\\nThe expr can be replaced with\\n subdir=${item#$sourcetree}\\nthat's afaics posix syntax ([1]), not bash.\\n\\nSpawning one mkdir for each directory can be replaced by a single mkdir\\ninvocation with all the directories. On my linux workstation that gets the\\ntime for the first loop down from 1005ms to 38ms, really.\\n\\nThat has the danger of the commandline getting too long. But since we rely on\\nthe final link of the backend to be done in a single command, I don't think\\nit's making things worse? We could try to use xargs otherwise, iirc that's in\\nposix as well.\\n\\nUsing parameter substitution in the second loop takes it down from 775ms to\\n533ms. Not calling cmp when the file doesn't exist cuts it down to 337ms.\\n\\nI don't know of a way to batch the call to ln. The time with ln replaced with\\n: is 151ms, fwiw.\\n\\nOn windows that makes prep_buildtree go from 42.4s to 5.8s for me.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n[1] https://pubs.opengroup.org/onlinepubs/009604499/utilities/xcu_chap02.html\",\n \"msg_date\": \"Thu, 7 Apr 2022 10:10:21 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"\\nOn 4/6/22 12:34, Andres Freund wrote:\\n> Hi,\\n>\\n> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\\n>> On 3/30/22 20:26, Andres Freund wrote:\\n>>> Could you try using dash to invoke configure here, and whether it makes configure faster?\\n>> I got weird failures re libxml/parser.h when I tried with dash.\\n> Hm. Hadn't enabled that when I tried...\\n>\\n>\\n>> See <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\n>> could see config.log on failure.)\\n> All *.log files are preserved on failure. There's a file directory brower at\\n> the top to navigate around:\\n> https://api.cirrus-ci.com/v1/artifact/task/5963254039052288/log/build/config.log\\n\\n\\n\\nThanks.\\n\\n\\nI got it working with this added to the config settings:\\n\\n\\n--with-includes='/ucrt64/include/libxml2 /c/cirrus/src/include/port/win32'\\n\\n\\nI conclude tentatively that while bash translates widows paths to msys\\npaths, dash does not.\\n\\n\\nsee https://cirrus-ci.com/task/5968968560148480?logs=configure#L1\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 15:42:13 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"\\nOn 4/7/22 13:10, Andres Freund wrote:\\n> Hi,\\n>\\n> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\\n>> On 3/30/22 20:26, Andres Freund wrote:\\n>>> Could you try using dash to invoke configure here, and whether it makes configure faster?\\n>> I got weird failures re libxml/parser.h when I tried with dash. See\\n>> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\n>> could see config.log on failure.)\\n> Since dash won't help us to get the build time down sufficiently, and the\\n> tests don't pass without a separate build tree, I looked at what makes\\n> config/prep_buildtree so slow.\\n>\\n> It's largely just bad code. The slowest part are spawning one expr and mkdir\\n> -p for each directory. One 'cmp' for each makefile doesn't help either.\\n>\\n> The expr can be replaced with\\n> subdir=${item#$sourcetree}\\n> that's afaics posix syntax ([1]), not bash.\\n>\\n> Spawning one mkdir for each directory can be replaced by a single mkdir\\n> invocation with all the directories. On my linux workstation that gets the\\n> time for the first loop down from 1005ms to 38ms, really.\\n>\\n> That has the danger of the commandline getting too long. But since we rely on\\n> the final link of the backend to be done in a single command, I don't think\\n> it's making things worse? We could try to use xargs otherwise, iirc that's in\\n> posix as well.\\n>\\n> Using parameter substitution in the second loop takes it down from 775ms to\\n> 533ms. Not calling cmp when the file doesn't exist cuts it down to 337ms.\\n>\\n> I don't know of a way to batch the call to ln. The time with ln replaced with\\n> : is 151ms, fwiw.\\n\\n\\nAFAIK Msy2s 'ln -s' by default copies a non-directory rather than\\nactually symlinking it. If we want real symlinks, then we need\\nMSYS=|winsymlinks:nativestrict set. The is will fail unless the calling\\nuser is an Administrator or has the SeCreateSymbolicLink privilege. See\\n|\\n\\n|<https://postgr.es/m/|e05b213c-1257-84d4-f079-5c4d8c79e3ad@dunslane.net>\\nfor more details.\\n\\n\\n> On windows that makes prep_buildtree go from 42.4s to 5.8s for me.\\n\\n\\nThat's pretty good.\\n\\n\\nI think we can get rid of the CVS pruning, it's only 15 years or so\\nsince we've had that in the tree.\\n\\n\\n+ if test ! -d \\\"$buildtree/$subdir\\\"; then\\n+ echo \\\"$buildtree/$subdir\\\"\\n+ fi\\n\\n\\nI would probably just write that as\\n\\n\\ntest -d \\\"$buildtree/$subdir' || echo \\\"$buildtree/$subdir\\\"\\n\\n\\nalthough it's really just a matter of taste.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 16:48:50 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"\\nOn 4/7/22 16:48, Andrew Dunstan wrote:\\n> On 4/7/22 13:10, Andres Freund wrote:\\n>> Hi,\\n>>\\n>> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\\n>>> On 3/30/22 20:26, Andres Freund wrote:\\n>>>> Could you try using dash to invoke configure here, and whether it makes configure faster?\\n>>> I got weird failures re libxml/parser.h when I tried with dash. See\\n>>> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\n>>> could see config.log on failure.)\\n>> Since dash won't help us to get the build time down sufficiently, and the\\n>> tests don't pass without a separate build tree, I looked at what makes\\n>> config/prep_buildtree so slow.\\n>>\\n>> It's largely just bad code. The slowest part are spawning one expr and mkdir\\n>> -p for each directory. One 'cmp' for each makefile doesn't help either.\\n>>\\n>> The expr can be replaced with\\n>> subdir=${item#$sourcetree}\\n>> that's afaics posix syntax ([1]), not bash.\\n>>\\n>> Spawning one mkdir for each directory can be replaced by a single mkdir\\n>> invocation with all the directories. On my linux workstation that gets the\\n>> time for the first loop down from 1005ms to 38ms, really.\\n>>\\n>> That has the danger of the commandline getting too long. But since we rely on\\n>> the final link of the backend to be done in a single command, I don't think\\n>> it's making things worse? We could try to use xargs otherwise, iirc that's in\\n>> posix as well.\\n>>\\n>> Using parameter substitution in the second loop takes it down from 775ms to\\n>> 533ms. Not calling cmp when the file doesn't exist cuts it down to 337ms.\\n>>\\n>> I don't know of a way to batch the call to ln. The time with ln replaced with\\n>> : is 151ms, fwiw.\\n>\\n> AFAIK Msy2s 'ln -s' by default copies a non-directory rather than\\n> actually symlinking it. If we want real symlinks, then we need\\n> MSYS=|winsymlinks:nativestrict set. The is will fail unless the calling\\n> user is an Administrator or has the SeCreateSymbolicLink privilege. See\\n> |\\n\\n\\nSometimes I hate Thunderbird. Of course the | is spurious above, we\\nwould need\\n\\nMSYS=winsymlinks:nativestrict\\n\\nset.\\n\\n\\ncheers\\n\\nandrew\\n\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Thu, 7 Apr 2022 16:52:26 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi Andrew,\\n\\nYou should set MSYSTEM=UCRT64 in the environment section. Given that,\\n> there should be no need to specify a --host= setting for configure.\\n>\\n\\nIt's set to UCRT64 in the docker image side [1]. I didn't know --host isn't\\nnecessary on UCRT64 environment. I'll remove it then.\\n\\n [1]\\n https://github.com/anarazel/pg-vm-images/blob/main/docker/windows_ci_mingw64#L11\\n<https://github.com/anarazel/pg-vm-images/blob/main/docker/windows_ci_mingw64#L11>\\n\\nBest,\\nMelih\\n\\nHi Andrew, \\nYou should set MSYSTEM=UCRT64 in the environment section. Given that,\\nthere should be no need to specify a --host= setting for configure.It's set to UCRT64 in the docker image side [1]. I didn't know --host isn't necessary on UCRT64 environment. I'll remove it then. [1] https://github.com/anarazel/pg-vm-images/blob/main/docker/windows_ci_mingw64#L11Best,Melih\",\n \"msg_date\": \"Fri, 8 Apr 2022 17:57:21 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"On 2022-Apr-07, Andres Freund wrote:\\n\\n> Since dash won't help us to get the build time down sufficiently, and the\\n> tests don't pass without a separate build tree, I looked at what makes\\n> config/prep_buildtree so slow.\\n\\nMaybe we can replace prep_buildtree with a Perl script. Surely that\\nshould be faster.\\n\\n-- \\nÁlvaro Herrera Breisgau, Deutschland — https://www.EnterpriseDB.com/\\n\\\"Sallah, I said NO camels! That's FIVE camels; can't you count?\\\"\\n(Indiana Jones)\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Apr 2022 17:04:34 +0200\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"> On windows that makes prep_buildtree go from 42.4s to 5.8s for me.\\n>\\n\\nI applied Andres's faster prep build tree changes and triggered some cirrus\\nruns\\n\\nWithout these changes, preparing build tree was taking around 42.3s\\n(sometimes even more) [1]\\nIt seems like with these changes it drops to around 8s [2]\\n\\n[1] https://cirrus-ci.com/task/6562493345562624\\n[2] https://cirrus-ci.com/task/4836843802853376\\n\\nBest,\\nMelih\\n\\n\\nOn windows that makes prep_buildtree go from 42.4s to 5.8s for me.I applied Andres's faster prep build tree changes and triggered some cirrus runsWithout these changes, preparing build tree was taking around 42.3s (sometimes even more) [1]It seems like with these changes it drops to around 8s [2][1] https://cirrus-ci.com/task/6562493345562624[2] https://cirrus-ci.com/task/4836843802853376Best,Melih\",\n \"msg_date\": \"Fri, 8 Apr 2022 18:18:34 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-08 17:04:34 +0200, Alvaro Herrera wrote:\\n> > Since dash won't help us to get the build time down sufficiently, and the\\n> > tests don't pass without a separate build tree, I looked at what makes\\n> > config/prep_buildtree so slow.\\n>\\n> Maybe we can replace prep_buildtree with a Perl script. Surely that\\n> should be faster.\\n\\nCurrently building doesn't depend on perl :(\\n\\nI think the improvements that I suggested are big enough that they're worth\\ndoing on their own, particularly for windows, but also other OSs.\\n\\n\\nI just realized that the second find is pretty expensive compared to the\\nfirst.\\n\\ntime find \\\"$sourcetree\\\" -type d \\\\( \\\\( -name CVS -prune \\\\) -o \\\\( -name .git -prune \\\\) -o -print \\\\) | grep -v \\\"$sourcetree/doc/src/sgml/\\\\+\\\" > /dev/null\\nreal\\t0m0.019s\\nuser\\t0m0.008s\\nsys\\t0m0.017s\\n\\nsecond:\\ntime find \\\"$sourcetree\\\" -name Makefile -print -o -name GNUmakefile -print | grep -v \\\"$sourcetree/doc/src/sgml/images/\\\" > /dev/null\\n\\nreal\\t0m0.118s\\nuser\\t0m0.071s\\nsys\\t0m0.053s\\n\\nIt think we could just obsolete the second find, by checking for the existence\\nof Makefile / GNUmakefile in the first loop...\\n\\n\\nThe invocation of ln -s is quite measurable - looks like it's mostly the\\nprocess startup overhead (on linux, at least). Doing a ln --version > /dev/null\\neach iteration takes about the same time as actually creating the symlinks.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Apr 2022 15:05:57 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"On Thu, Apr 07, 2022 at 10:10:21AM -0700, Andres Freund wrote:\\n> Hi,\\n> \\n> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\\n> > On 3/30/22 20:26, Andres Freund wrote:\\n> > > Could you try using dash to invoke configure here, and whether it makes configure faster?\\n> > I got weird failures re libxml/parser.h when I tried with dash. See\\n> > <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\n> > could see config.log on failure.)\\n> \\n> Since dash won't help us to get the build time down sufficiently, and the\\n> tests don't pass without a separate build tree, I looked at what makes\\n> config/prep_buildtree so slow.\\n> \\n> It's largely just bad code. The slowest part are spawning one expr and mkdir\\n> -p for each directory. One 'cmp' for each makefile doesn't help either.\\n> \\n> The expr can be replaced with\\n> subdir=${item#$sourcetree}\\n> that's afaics posix syntax ([1]), not bash.\\n> \\n> Spawning one mkdir for each directory can be replaced by a single mkdir\\n> invocation with all the directories. On my linux workstation that gets the\\n> time for the first loop down from 1005ms to 38ms, really.\\n\\nEven better?\\n\\n(cd \\\"$sourcetree\\\" && find . -print |grep -E '/Makefile$|/GNUmakefile$' |grep -v \\\"$sourcetree/doc/src/sgml/images/\\\" |xargs tar c) |\\n (cd \\\"$buildtree\\\" && tar x)\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Apr 2022 19:27:58 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-04-08 19:27:58 -0500, Justin Pryzby wrote:\\n> On Thu, Apr 07, 2022 at 10:10:21AM -0700, Andres Freund wrote:\\n> > On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\\n> > > On 3/30/22 20:26, Andres Freund wrote:\\n> > > > Could you try using dash to invoke configure here, and whether it makes configure faster?\\n> > > I got weird failures re libxml/parser.h when I tried with dash. See\\n> > > <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\n> > > could see config.log on failure.)\\n> > \\n> > Since dash won't help us to get the build time down sufficiently, and the\\n> > tests don't pass without a separate build tree, I looked at what makes\\n> > config/prep_buildtree so slow.\\n> > \\n> > It's largely just bad code. The slowest part are spawning one expr and mkdir\\n> > -p for each directory. One 'cmp' for each makefile doesn't help either.\\n> > \\n> > The expr can be replaced with\\n> > subdir=${item#$sourcetree}\\n> > that's afaics posix syntax ([1]), not bash.\\n> > \\n> > Spawning one mkdir for each directory can be replaced by a single mkdir\\n> > invocation with all the directories. On my linux workstation that gets the\\n> > time for the first loop down from 1005ms to 38ms, really.\\n> \\n> Even better?\\n> \\n> (cd \\\"$sourcetree\\\" && find . -print |grep -E '/Makefile$|/GNUmakefile$' |grep -v \\\"$sourcetree/doc/src/sgml/images/\\\" |xargs tar c) |\\n> (cd \\\"$buildtree\\\" && tar x)\\n\\nDon't think we depend on tar for building, at the moment. But yes, it'd be\\nfaster... Tar is certainly a smaller dependency than perl, not sure if there's\\nany relevant platform without it?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Apr 2022 18:02:41 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"On 2022-Apr-08, Andres Freund wrote:\\n\\n> I just realized that the second find is pretty expensive compared to the\\n> first.\\n> \\n> time find \\\"$sourcetree\\\" -type d \\\\( \\\\( -name CVS -prune \\\\) -o \\\\( -name .git -prune \\\\) -o -print \\\\) | grep -v \\\"$sourcetree/doc/src/sgml/\\\\+\\\" > /dev/null\\n> real\\t0m0.019s\\n> user\\t0m0.008s\\n> sys\\t0m0.017s\\n> \\n> second:\\n> time find \\\"$sourcetree\\\" -name Makefile -print -o -name GNUmakefile -print | grep -v \\\"$sourcetree/doc/src/sgml/images/\\\" > /dev/null\\n> \\n> real\\t0m0.118s\\n> user\\t0m0.071s\\n> sys\\t0m0.053s\\n\\nHmm, ISTM that time can be reduced a bit with -prune,\\n\\ntime find \\\"$sourcetree\\\" \\\\( -name .git -prune \\\\) -o -name Makefile -print -o -name GNUmakefile -print | grep -v \\\"$sourcetree/doc/src/sgml/images/\\\" > /dev/null\\n\\nI thought it might work to do away with the grep and use find's -path\\ninstead to prune that subdir, but \\\"time\\\" shows almost no difference for\\nme:\\n\\ntime find \\\"$sourcetree\\\" \\\\( -name .git -prune \\\\) -o \\\\( -path '*doc/src/sgml/images' -prune \\\\) -o -name Makefile -print -o -name GNUmakefile -print > /dev/null\\n\\nMaybe find's -path is equally expensive. Still, that seems a good\\nchange anyway. (The times are lower in my system than those you show.)\\n\\n> It think we could just obsolete the second find, by checking for the existence\\n> of Makefile / GNUmakefile in the first loop...\\n\\nHmm, if that's going to require one more shell command per dir, it\\nsounds more expensive. It's worth trying, I guess.\\n\\n> The invocation of ln -s is quite measurable - looks like it's mostly the\\n> process startup overhead (on linux, at least). Doing a ln --version > /dev/null\\n> each iteration takes about the same time as actually creating the symlinks.\\n\\nIs this running with some locale settings enabled? Maybe we can save\\nsome time by making sure we're under LC_ALL=C or something like that, to\\navoid searching for translation files.\\n\\n-- \\nÁlvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/\\n\\\"Las navajas y los monos deben estar siempre distantes\\\" (Germán Poo)\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Apr 2022 12:37:05 +0200\",\n \"msg_from\": \"Alvaro Herrera <alvherre@alvh.no-ip.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"\\nOn 4/8/22 21:02, Andres Freund wrote:\\n> Hi,\\n>\\n> On 2022-04-08 19:27:58 -0500, Justin Pryzby wrote:\\n>> On Thu, Apr 07, 2022 at 10:10:21AM -0700, Andres Freund wrote:\\n>>> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\\n>>>> On 3/30/22 20:26, Andres Freund wrote:\\n>>>>> Could you try using dash to invoke configure here, and whether it makes configure faster?\\n>>>> I got weird failures re libxml/parser.h when I tried with dash. See\\n>>>> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\n>>>> could see config.log on failure.)\\n>>> Since dash won't help us to get the build time down sufficiently, and the\\n>>> tests don't pass without a separate build tree, I looked at what makes\\n>>> config/prep_buildtree so slow.\\n>>>\\n>>> It's largely just bad code. The slowest part are spawning one expr and mkdir\\n>>> -p for each directory. One 'cmp' for each makefile doesn't help either.\\n>>>\\n>>> The expr can be replaced with\\n>>> subdir=${item#$sourcetree}\\n>>> that's afaics posix syntax ([1]), not bash.\\n>>>\\n>>> Spawning one mkdir for each directory can be replaced by a single mkdir\\n>>> invocation with all the directories. On my linux workstation that gets the\\n>>> time for the first loop down from 1005ms to 38ms, really.\\n>> Even better?\\n>>\\n>> (cd \\\"$sourcetree\\\" && find . -print |grep -E '/Makefile$|/GNUmakefile$' |grep -v \\\"$sourcetree/doc/src/sgml/images/\\\" |xargs tar c) |\\n>> (cd \\\"$buildtree\\\" && tar x)\\n> Don't think we depend on tar for building, at the moment. But yes, it'd be\\n> faster... Tar is certainly a smaller dependency than perl, not sure if there's\\n> any relevant platform without it?\\n>\\n>\\n\\n\\nCouple of things here.\\n\\n\\n1. The second grep should lose \\\"$sourcetree\\\" I think.\\n\\n2. Isn't this going to be a change in behaviour in that it will copy\\nrather than symlinking the Makefiles? That is in fact the default\\nbehaviour of msys2's 'ln -s', as I pointed out upthread, but I don't\\nthink it's what we really want, especially in the general case. If you\\nmodify the Makefile and you're using a vpath you want to see the change\\nreflected in your vpath.\\n\\n\\ncheers\\n\\n\\nandrew\\n\\n--\\nAndrew Dunstan\\nEDB: https://www.enterprisedb.com\\n\\n\\n\\n\",\n \"msg_date\": \"Sat, 9 Apr 2022 12:34:40 -0400\",\n \"msg_from\": \"Andrew Dunstan <andrew@dunslane.net>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi hackers,\\n\\nI'm sharing the rebased version of this patch, if you're still interested.\\n\\nI would appreciate any feedback or concerns.\\n\\nBest,\\nMelih\\n\\n\\nAndrew Dunstan <andrew@dunslane.net>, 9 Nis 2022 Cmt, 19:34 tarihinde şunu\\nyazdı:\\n\\n>\\n> On 4/8/22 21:02, Andres Freund wrote:\\n> > Hi,\\n> >\\n> > On 2022-04-08 19:27:58 -0500, Justin Pryzby wrote:\\n> >> On Thu, Apr 07, 2022 at 10:10:21AM -0700, Andres Freund wrote:\\n> >>> On 2022-04-06 11:03:37 -0400, Andrew Dunstan wrote:\\n> >>>> On 3/30/22 20:26, Andres Freund wrote:\\n> >>>>> Could you try using dash to invoke configure here, and whether it\\n> makes configure faster?\\n> >>>> I got weird failures re libxml/parser.h when I tried with dash. See\\n> >>>> <https://cirrus-ci.com/task/5963254039052288> (It would be nice if we\\n> >>>> could see config.log on failure.)\\n> >>> Since dash won't help us to get the build time down sufficiently, and\\n> the\\n> >>> tests don't pass without a separate build tree, I looked at what makes\\n> >>> config/prep_buildtree so slow.\\n> >>>\\n> >>> It's largely just bad code. The slowest part are spawning one expr and\\n> mkdir\\n> >>> -p for each directory. One 'cmp' for each makefile doesn't help either.\\n> >>>\\n> >>> The expr can be replaced with\\n> >>> subdir=${item#$sourcetree}\\n> >>> that's afaics posix syntax ([1]), not bash.\\n> >>>\\n> >>> Spawning one mkdir for each directory can be replaced by a single mkdir\\n> >>> invocation with all the directories. On my linux workstation that gets\\n> the\\n> >>> time for the first loop down from 1005ms to 38ms, really.\\n> >> Even better?\\n> >>\\n> >> (cd \\\"$sourcetree\\\" && find . -print |grep -E '/Makefile$|/GNUmakefile$'\\n> |grep -v \\\"$sourcetree/doc/src/sgml/images/\\\" |xargs tar c) |\\n> >> (cd \\\"$buildtree\\\" && tar x)\\n> > Don't think we depend on tar for building, at the moment. But yes, it'd\\n> be\\n> > faster... Tar is certainly a smaller dependency than perl, not sure if\\n> there's\\n> > any relevant platform without it?\\n> >\\n> >\\n>\\n>\\n> Couple of things here.\\n>\\n>\\n> 1. The second grep should lose \\\"$sourcetree\\\" I think.\\n>\\n> 2. Isn't this going to be a change in behaviour in that it will copy\\n> rather than symlinking the Makefiles? That is in fact the default\\n> behaviour of msys2's 'ln -s', as I pointed out upthread, but I don't\\n> think it's what we really want, especially in the general case. If you\\n> modify the Makefile and you're using a vpath you want to see the change\\n> reflected in your vpath.\\n>\\n>\\n> cheers\\n>\\n>\\n> andrew\\n>\\n> --\\n> Andrew Dunstan\\n> EDB: https://www.enterprisedb.com\\n>\\n>\",\n \"msg_date\": \"Thu, 28 Jul 2022 17:11:25 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"I think the \\\"only_if\\\" should allow separately running one but not both of the\\nwindows instances, like:\\n\\n+ only_if: $CIRRUS_CHANGE_MESSAGE !=~ '.*\\\\nci-os-only:.*' || $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw64'\\n\\nI'm not sure, but maybe this task should only run \\\"by request\\\", and omit the\\nfirst condition:\\n\\n+ only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw64'\\n\\nI think it should include something like\\n\\n+ setup_additional_packages_script: |\\n+ REM C:\\\\msys64\\\\usr\\\\bin\\\\pacman.exe -S --noconfirm ...\\n\\nLet's see what others think about those.\\n\\nDo you know if this handles logging of crash dumps ?\\n\\nWith tweaks to prep_buildtree, and with ./configure --cache-file, that step\\ngoes down to ~36sec (unless configure needs to be re-run).\\n\\nI also looked into using busybox to avoid running separate processes for each\\n\\\"ln\\\", but I think 36sec is good enough.\\n\\nAt one point, I tried setting \\\"CIRRUS_SHELL: bash\\\" to avoid writing \\\"bash -c\\\"\\nover and over, but never got it working.\\n\\n-- \\nJustin\",\n \"msg_date\": \"Thu, 28 Jul 2022 17:44:28 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"I noticed that this says:\\n\\n[01:01:45.657] sqlda.pgc: In function 'dump_sqlda':\\n[01:01:45.657] sqlda.pgc:45:33: warning: format '%d' expects argument\\nof type 'int', but argument 3 has type 'long long int' [-Wformat=]\\n[01:01:45.657] 45 | \\\"name sqlda descriptor: '%s' value %I64d\\\\n\\\",\\n[01:01:45.657] | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\n[01:01:45.657] ......\\n[01:01:45.657] 49 | sqlda->sqlvar[i].sqlname.data, *(long long int\\n*)sqlda->sqlvar[i].sqldata);\\n[01:01:45.657] | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\n[01:01:45.657] | |\\n[01:01:45.657] | long long int\\n\\n... but fairywren doesn't. Why would they disagree on recognising %I64d?\\n\\nThe other warning I'm seeing is present on both, so it's not really\\nrelevant to this thread, just mentioning it... seems kinda like a\\nworlds-colliding-problem without an elegant fix (POSIX says you have\\nto declare it yourself exactly like that, but Windows says linkage\\nain't going to work the way you want unless you sprinkle the right\\nlinkage dust on it...), so maybe we just want to put #if\\n!defined(something something mings) around it and to MinGW's header's\\ndeclaration...\\n\\n[00:48:08.925] c:/cirrus/src/backend/postmaster/postmaster.c: In\\nfunction 'PostmasterMain':\\n[00:48:08.925] c:/cirrus/src/backend/postmaster/postmaster.c:973:31:\\nwarning: '__p__environ' redeclared without dllimport attribute:\\nprevious dllimport ignored [-Wattributes]\\n[00:48:08.925] 973 | extern char **environ;\\n[00:48:08.925] | ^~~~~~~\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Aug 2022 14:04:33 +1200\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"On Thu, Aug 4, 2022 at 2:04 PM Thomas Munro <thomas.munro@gmail.com> wrote:\\n> I noticed that this says:\\n>\\n> [01:01:45.657] sqlda.pgc: In function 'dump_sqlda':\\n> [01:01:45.657] sqlda.pgc:45:33: warning: format '%d' expects argument\\n> of type 'int', but argument 3 has type 'long long int' [-Wformat=]\\n> [01:01:45.657] 45 | \\\"name sqlda descriptor: '%s' value %I64d\\\\n\\\",\\n> [01:01:45.657] | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\n> [01:01:45.657] ......\\n> [01:01:45.657] 49 | sqlda->sqlvar[i].sqlname.data, *(long long int\\n> *)sqlda->sqlvar[i].sqldata);\\n> [01:01:45.657] | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\n> [01:01:45.657] | |\\n> [01:01:45.657] | long long int\\n>\\n> ... but fairywren doesn't. Why would they disagree on recognising %I64d?\\n\\nOops, I was looking in the wrong place. fairywren does also shows the warning:\\n\\nhttps://buildfarm.postgresql.org/cgi-bin/show_stage_log.pl?nm=fairywren&dt=2022-08-02%2022%3A05%3A30&stg=ecpg-check\\n\\nSomething to fix, but not directly relevant to this patch. Sorry for the noise.\\n\\n\\n\",\n \"msg_date\": \"Thu, 4 Aug 2022 20:38:34 +1200\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Inline notes about changes since the last version.\\n\\nOn Thu, Jul 28, 2022 at 05:44:28PM -0500, Justin Pryzby wrote:\\n> I think the \\\"only_if\\\" should allow separately running one but not both of the\\n> windows instances, like:\\n> \\n> + only_if: $CIRRUS_CHANGE_MESSAGE !=~ '.*\\\\nci-os-only:.*' || $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw64'\\n> \\n> I'm not sure, but maybe this task should only run \\\"by request\\\", and omit the\\n> first condition:\\n> \\n> + only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw64'\\n\\nThe patch shouldn't say this during development, or else cfbot doesn't run it..\\nOops.\\n\\n> I think it should include something like\\n> \\n> + setup_additional_packages_script: |\\n> + REM C:\\\\msys64\\\\usr\\\\bin\\\\pacman.exe -S --noconfirm ...\\n> \\n> Let's see what others think about those.\\n> \\n> Do you know if this handles logging of crash dumps ?\\n\\nIt does now, although I hardcoded \\\"postgres.exe\\\" ...\\n\\n> + setup_additional_packages_script: |\\n> + REM C:\\\\msys64\\\\usr\\\\bin\\\\pacman.exe -S --noconfirm busybox\\n\\nThis should include choco, too.\\n\\n> - CXXFLAGS='-Og -ggdb'\\\"\\n> + CXXFLAGS='-Og -ggdb' && break;\\n> + rm -v ${CCACHE_DIR}/configure.cache;\\n> + done\\n\\nI noticed that this doesn't seem to do the right thing with the exit status -\\nconfigure can fail without cirrusci noticing, and then the build fails at the\\nnext step.\\n\\n> for item in `find \\\"$sourcetree\\\" -name Makefile -print -o -name GNUmakefile -print | grep -v \\\"$sourcetree/doc/src/sgml/images/\\\"`; do\\n> - filename=`expr \\\"$item\\\" : \\\"$sourcetree\\\\(.*\\\\)\\\"`\\n> - if test ! -f \\\"${item}.in\\\"; then\\n> - if cmp \\\"$item\\\" \\\"$buildtree/$filename\\\" >/dev/null 2>&1; then : ; else\\n> - ln -fs \\\"$item\\\" \\\"$buildtree/$filename\\\" || exit 1\\n> - fi\\n> - fi\\n> + filename=${item#$sourcetree}\\n> + [ -e \\\"$buildtree/$filename\\\" ] && continue\\n\\nI fixed this to check for \\\".in\\\" files as intended.\\n\\nIt'd be a lot better if the image didn't take so long to start. :(\\n\\n-- \\nJustin\",\n \"msg_date\": \"Thu, 18 Aug 2022 21:33:56 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Justin Pryzby <pryzby@telsasoft.com>, 19 Ağu 2022 Cum, 05:34 tarihinde şunu\\nyazdı:\\n\\n> Inline notes about changes since the last version.\\n>\\n> On Thu, Jul 28, 2022 at 05:44:28PM -0500, Justin Pryzby wrote:\\n> > I think the \\\"only_if\\\" should allow separately running one but not both\\n> of the\\n> > windows instances, like:\\n> >\\n> > + only_if: $CIRRUS_CHANGE_MESSAGE !=~ '.*\\\\nci-os-only:.*' ||\\n> $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw64'\\n> >\\n> > I'm not sure, but maybe this task should only run \\\"by request\\\", and omit\\n> the\\n> > first condition:\\n> >\\n> > + only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw64'\\n>\\n> The patch shouldn't say this during development, or else cfbot doesn't run\\n> it..\\n> Oops.\\n>\\nActually, making MinGW task optional for now might make sense. Due to\\nwindows resource limits on Cirrus CI and slow builds on Windows, adding\\nthis task as non-optional may not be an efficient decision\\nI think that continuing with this patch by changing MinGW to optional for\\nnow, instead of waiting for more resource on Cirrus or faster builds on\\nWindows, could be better. I don't see any harm.\\n\\n+ only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-include:[^\\\\n]*mingw.* ||\\n> $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw.*'\\n\\nAdded this line to allow run only mingw task or run all tasks including\\nmingw.\\n\\nWhat do you all think about this change? Does it make sense?\\n\\n\\nThanks for your contributions/reviews Justin!\\n\\n> I think it should include something like\\n> >\\n> > + setup_additional_packages_script: |\\n> > + REM C:\\\\msys64\\\\usr\\\\bin\\\\pacman.exe -S --noconfirm ...\\n> >\\n> > Let's see what others think about those.\\n> >\\n> > Do you know if this handles logging of crash dumps ?\\n>\\n> It does now, although I hardcoded \\\"postgres.exe\\\" ...\\n>\\n\\nmerged this with my patch\\n\\n> + setup_additional_packages_script: |\\n> > + REM C:\\\\msys64\\\\usr\\\\bin\\\\pacman.exe -S --noconfirm busybox\\n>\\n> This should include choco, too.\\n\\nAdded pacman.exe line. Do we really need choco here? I don't think mingw\\nwould require any package via choco.\\nAlso is ending pacman.exe line with busybox intentional? I just added that\\nline with \\\"...\\\" at the end instead of any package name.\\n\\n\\n>\\n>\\n> - CXXFLAGS='-Og -ggdb'\\\"\\n> > + CXXFLAGS='-Og -ggdb' && break;\\n> > + rm -v ${CCACHE_DIR}/configure.cache;\\n> > + done\\n>\\n> I noticed that this doesn't seem to do the right thing with the exit\\n> status -\\n> configure can fail without cirrusci noticing, and then the build fails at\\n> the\\n> next step.\\n\\n\\nmerged.\\n\\n\\n>\\n>\\n> for item in `find \\\"$sourcetree\\\" -name Makefile -print -o -name\\n> GNUmakefile -print | grep -v \\\"$sourcetree/doc/src/sgml/images/\\\"`; do\\n> > - filename=`expr \\\"$item\\\" : \\\"$sourcetree\\\\(.*\\\\)\\\"`\\n> > - if test ! -f \\\"${item}.in\\\"; then\\n> > - if cmp \\\"$item\\\" \\\"$buildtree/$filename\\\" >/dev/null 2>&1; then : ;\\n> else\\n> > - ln -fs \\\"$item\\\" \\\"$buildtree/$filename\\\" || exit 1\\n> > - fi\\n> > - fi\\n> > + filename=${item#$sourcetree}\\n> > + [ -e \\\"$buildtree/$filename\\\" ] && continue\\n>\\n> I fixed this to check for \\\".in\\\" files as intended.\\n>\\n> It'd be a lot better if the image didn't take so long to start. :(\\n>\\n\\nOne question would be that should this patch include \\\"prep_buildtree\\\"? It\\ndoesn't seem to me like it's directly related to adding MinGW into CI but\\nmore like an improvement for builds on Windows.\\nMaybe we can make it a seperate patch if it's necessary.\\n\\nWhat do you think?\\n\\n TAR: \\\"c:/windows/system32/tar.exe\\\"\\n>\\n\\nSharing a new version of the patch. It also moves the above line so that it\\nwill apply to mingw task too. Otherwise mingw task was failing.\\n\\nThanks,\\nMelih\",\n \"msg_date\": \"Sat, 3 Sep 2022 00:52:54 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"On Sat, Sep 03, 2022 at 12:52:54AM +0300, Melih Mutlu wrote:\\n> Justin Pryzby <pryzby@telsasoft.com>, 19 Ağu 2022 Cum, 05:34 tarihinde şunu yazdı:\\n> \\n> > Inline notes about changes since the last version.\\n> >\\n> > On Thu, Jul 28, 2022 at 05:44:28PM -0500, Justin Pryzby wrote:\\n> > > I think the \\\"only_if\\\" should allow separately running one but not both\\n> > of the\\n> > > windows instances, like:\\n> > >\\n> > > + only_if: $CIRRUS_CHANGE_MESSAGE !=~ '.*\\\\nci-os-only:.*' || $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw64'\\n> > >\\n> > > I'm not sure, but maybe this task should only run \\\"by request\\\", and omit the\\n> > > first condition:\\n> > >\\n> > > + only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw64'\\n> >\\n> > The patch shouldn't say this during development, or else cfbot doesn't run it..\\n> > Oops.\\n>\\n> Actually, making MinGW task optional for now might make sense. Due to\\n> windows resource limits on Cirrus CI and slow builds on Windows, adding\\n> this task as non-optional may not be an efficient decision\\n> I think that continuing with this patch by changing MinGW to optional for\\n> now, instead of waiting for more resource on Cirrus or faster builds on\\n> Windows, could be better. I don't see any harm.\\n\\nI agree that maybe it should be optional if merged to postgres.\\n\\nBut cfbot should run the Mingw task for this patch's own commitfest\\nentry. But right now (because cfbot doesn't include the original commit\\nmessage/s), it doesn't get run :(\\n\\n> + only_if: $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-include:[^\\\\n]*mingw.* ||\\n> $CIRRUS_CHANGE_MESSAGE =~ '.*\\\\nci-os-only:[^\\\\n]*mingw.*'\\n> \\n> Added this line to allow run only mingw task or run all tasks including\\n> mingw.\\n> \\n> What do you all think about this change? Does it make sense?\\n\\nYou're allowing to write \\\"ci-os-include: mingw\\\" to \\\"opt-in\\\" to the task,\\nwithout opting-out of all the other tasks (and without enumerating all\\nthe tasks by writing \\\"ci-os-only: mingw,windows,macos,freebsd,linux\\\".\\nThat makes sense, and the logic looks right. But that still has to be\\ncommented during development to be run by cfbot.\\n\\nAlso, the first half was missing a closing quote.\\nhttps://cirrus-ci.com/build/5874178241855488\\n\\n> > + setup_additional_packages_script: |\\n> > > + REM C:\\\\msys64\\\\usr\\\\bin\\\\pacman.exe -S --noconfirm busybox\\n> >\\n> > This should include choco, too.\\n> \\n> Added pacman.exe line. Do we really need choco here? I don't think mingw\\n> would require any package via choco.\\n\\nI guess choco isn't needed.\\n\\n> Also is ending pacman.exe line with busybox intentional? I just added that\\n> line with \\\"...\\\" at the end instead of any package name.\\n\\nYeah, the busybox part was unintentional.\\n\\n> > > for item in `find \\\"$sourcetree\\\" -name Makefile -print -o -name GNUmakefile -print | grep -v \\\"$sourcetree/doc/src/sgml/images/\\\"`; do\\n> > > - filename=`expr \\\"$item\\\" : \\\"$sourcetree\\\\(.*\\\\)\\\"`\\n> > > - if test ! -f \\\"${item}.in\\\"; then\\n> > > - if cmp \\\"$item\\\" \\\"$buildtree/$filename\\\" >/dev/null 2>&1; then : ; else\\n> > > - ln -fs \\\"$item\\\" \\\"$buildtree/$filename\\\" || exit 1\\n> > > - fi\\n> > > - fi\\n> > > + filename=${item#$sourcetree}\\n> > > + [ -e \\\"$buildtree/$filename\\\" ] && continue\\n> >\\n> > I fixed this to check for \\\".in\\\" files as intended.\\n> >\\n> > It'd be a lot better if the image didn't take so long to start. :(\\n> \\n> One question would be that should this patch include \\\"prep_buildtree\\\"? It\\n> doesn't seem to me like it's directly related to adding MinGW into CI but\\n> more like an improvement for builds on Windows.\\n> Maybe we can make it a seperate patch if it's necessary.\\n\\nI don't know what direction that idea is going, but it makes working\\nwith this patch a bit easier when configure is less slow. Fine with me\\nto split it into a separate patch :)\\n\\n> Sharing a new version of the patch. It also moves the above line so that it\\n> will apply to mingw task too. Otherwise mingw task was failing.\\n\\nI saw that but hadn't tracked it down yet. Do you know if the tar\\nfailures were from a TAP test added since you first posted the mingw\\npatch, or ??\\n\\nAlso: your original patch said --host=x86_64-w64-mingw32, but the task\\nis called MinGW64. The most recent patches don't use --host at all, and\\nwere building for a 32 bit environment, even though the OS image says\\nMSYSTEM=UCRT64.\\n\\nAlso: right now src/test and src/interfaces/*/test aren't being built\\nduring the build phase, which means that they're 1) not compiled in\\nparallel; and 2) not cached. This isn't specific to MinGW. Other than\\ncompiling those dirs specifically, one option is to put\\n\\\"always: upload_caches: ccache\\\" after \\\"test_world_script\\\" (in that case,\\nif the CI instance is rescheduled during tests, the compilation won't be\\npushed to cache). Actually, it seems better to compile stuff during\\n\\\"build\\\" or else any compilation warnings should up in the middle of\\n\\\"check-world..\\\"\\n\\nAlso: I'm having second thoughts about the loop around ./configure. It\\ncould happen that a cached configure would succeed, but then the build\\nwould later fail, and it wouldn't fix itself. I think a slow configure\\nis okay for an \\\"opt-in\\\" task.\\n\\nAlso: my backtrace call was using a path to cygwin rather than msys.\\nThis seemed to work before, but doesn't seem to be working now...\\n\\n-- \\nJustin\",\n \"msg_date\": \"Mon, 5 Sep 2022 06:50:55 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-09-05 06:50:55 -0500, Justin Pryzby wrote:\\n> I saw that but hadn't tracked it down yet. Do you know if the tar\\n> failures were from a TAP test added since you first posted the mingw\\n> patch, or ??\\n\\nI think it's that msys now includes tar by default, but not sure.\\n\\n\\n> Also: your original patch said --host=x86_64-w64-mingw32, but the task\\n> is called MinGW64. The most recent patches don't use --host at all, and\\n> were building for a 32 bit environment, even though the OS image says\\n> MSYSTEM=UCRT64.\\n\\nI don't think you should need to use --host, that indicates cross compiling,\\nwhich we shouldn't do. I don't think the patch without --host targets 32bit -\\nthe last time CI ran the patch, it built a 64bit PG:\\n\\nhttps://cirrus-ci.com/task/5489471041830912?logs=configure#L231\\n[13:47:54.539] checking size of void *... (cached) 8\\n[13:47:54.557] checking size of size_t... (cached) 8\\n\\nand the underlying commit didn't specify --host.\\n\\n\\n> Also: right now src/test and src/interfaces/*/test aren't being built\\n> during the build phase, which means that they're 1) not compiled in\\n> parallel; and 2) not cached. This isn't specific to MinGW. Other than\\n> compiling those dirs specifically, one option is to put\\n> \\\"always: upload_caches: ccache\\\" after \\\"test_world_script\\\" (in that case,\\n> if the CI instance is rescheduled during tests, the compilation won't be\\n> pushed to cache). Actually, it seems better to compile stuff during\\n> \\\"build\\\" or else any compilation warnings should up in the middle of\\n> \\\"check-world..\\\"\\n\\nI'd tackle that independently of this commit.\\n\\n\\n> Also: I'm having second thoughts about the loop around ./configure. It\\n> could happen that a cached configure would succeed, but then the build\\n> would later fail, and it wouldn't fix itself. I think a slow configure\\n> is okay for an \\\"opt-in\\\" task.\\n\\nAgreed.\\n\\nI think we can convert this to meson soon, and that seems a *lot* faster at\\nconfigure than autoconf on mingw. Not even close to as fast as on a modern-ish\\nlinux, but not that painful.\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Mon, 5 Sep 2022 16:52:17 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"On Mon, Sep 05, 2022 at 04:52:17PM -0700, Andres Freund wrote:\\n> I don't think you should need to use --host, that indicates cross compiling,\\n\\nThis made me consider the idea of cross-compiling for windows under a\\nnew linux task, and then running tests under Windows with a dependent\\ntask. I suppose that use -Og, in addition to CompilerWarnings, which\\nuses -O2.\\n\\ncirrusci caches are (or can be) shared between tasks. I got this mostly\\nworking, with a few kludges to compile and install stuff under src/test.\\nThis may be yet another idea that's obsoleted by meson. WDYT?\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sat, 10 Sep 2022 15:19:44 -0500\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi hackers,\\n\\nJustin Pryzby <pryzby@telsasoft.com>, 5 Eyl 2022 Pzt, 14:50 tarihinde şunu\\nyazdı:\\n\\n> But cfbot should run the Mingw task for this patch's own commitfest\\n> entry. But right now (because cfbot doesn't include the original commit\\n> message/s), it doesn't get run :(\\n>\\n\\nI've been thinking about how to make the mingw task run only for this patch\\non cfbot and not for others. TBH, I couldn't come up with a nice way to\\nachieve this.\\n\\nDoes anyone have any suggestions on this?\\n\\nAndres Freund <andres@anarazel.de>, 6 Eyl 2022 Sal, 02:52 tarihinde şunu\\nyazdı:\\n\\n> Hi,\\n>\\n> On 2022-09-05 06:50:55 -0500, Justin Pryzby wrote:\\n> > I saw that but hadn't tracked it down yet. Do you know if the tar\\n> > failures were from a TAP test added since you first posted the mingw\\n> > patch, or ??\\n>\\n> I think it's that msys now includes tar by default, but not sure.\\n\\n\\nSeems like msys2 comes with GNU tar now. Previously it was bsd tar under\\n\\\"/c/Windows/System32/tar.exe \\\" which now we force it to use.\\n\\n\\nThanks,\\nMelih\\n\\nHi hackers,Justin Pryzby <pryzby@telsasoft.com>, 5 Eyl 2022 Pzt, 14:50 tarihinde şunu yazdı:But cfbot should run the Mingw task for this patch's own commitfest\\nentry. But right now (because cfbot doesn't include the original commit\\nmessage/s), it doesn't get run :( I've been thinking about how to make the mingw task run only for this patch on cfbot and not for others. TBH, I couldn't come up with a nice way to achieve this.Does anyone have any suggestions on this?Andres Freund <andres@anarazel.de>, 6 Eyl 2022 Sal, 02:52 tarihinde şunu yazdı:Hi,On 2022-09-05 06:50:55 -0500, Justin Pryzby wrote:> I saw that but hadn't tracked it down yet. Do you know if the tar> failures were from a TAP test added since you first posted the mingw> patch, or ??I think it's that msys now includes tar by default, but not sure.Seems like msys2 comes with GNU tar now. Previously it was bsd tar under \\\"/c/Windows/System32/tar.exe \\\" which now we force it to use.Thanks,Melih\",\n \"msg_date\": \"Wed, 21 Sep 2022 16:18:43 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-09-21 16:18:43 +0300, Melih Mutlu wrote:\\n> I've been thinking about how to make the mingw task run only for this patch\\n> on cfbot and not for others. TBH, I couldn't come up with a nice way to\\n> achieve this.\\n> \\n> Does anyone have any suggestions on this?\\n\\nAdd a commented-out trigger-type: manual and note in the commit message that\\nthe committer needs to adjust that piece.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 21 Sep 2022 07:04:45 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi Melih,\\n\\nOn 2022-09-05 16:52:17 -0700, Andres Freund wrote:\\n> I think we can convert this to meson soon, and that seems a *lot* faster at\\n> configure than autoconf on mingw. Not even close to as fast as on a modern-ish\\n> linux, but not that painful.\\n\\nNow that meson has been merged, could you try to adjust this patch so it\\nbuilds with meson?\\n\\nRegards,\\n\\nAndres\\n\\n\\n\",\n \"msg_date\": \"Thu, 22 Sep 2022 10:03:05 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi hackers.\\n\\nAndres Freund <andres@anarazel.de>, 22 Eyl 2022 Per, 20:03 tarihinde şunu\\nyazdı:\\n\\n> Now that meson has been merged, could you try to adjust this patch so it\\n> builds with meson?\\n>\\n\\nAttached patch is the adjusted version to build with meson.\\n\\nAdd a commented-out trigger-type: manual and note in the commit message that\\n> the committer needs to adjust that piece.\\n>\\n\\nDone.\\n\\n> Also: your original patch said --host=x86_64-w64-mingw32, but the task\\n> > is called MinGW64. The most recent patches don't use --host at all, and\\n> > were building for a 32 bit environment, even though the OS image says\\n> > MSYSTEM=UCRT64.\\n\\n\\n> I don't think you should need to use --host, that indicates cross\\n> compiling,\\n> which we shouldn't do. I don't think the patch without --host targets\\n> 32bit -\\n> the last time CI ran the patch, it built a 64bit PG:\\n\\n\\nAlso removed this with meson changes. Can confirm that it builds 64bit.\\n\\nJustin Pryzby <pryzby@telsasoft.com>, 5 Eyl 2022 Pzt, 14:50 tarihinde şunu\\nyazdı:\\n\\n> I don't know what direction that idea is going, but it makes working\\n> with this patch a bit easier when configure is less slow. Fine with me\\n> to split it into a separate patch :)\\n\\n\\nIf it's okay, then I would suggest not including that piece into this patch\\nsince I don't know what direction this is going either.\\n\\n\\nThe patch has been changed quite a bit due to meson integration.\\nI would appreciate any feedback.\\n\\nBest,\\nMelih\",\n \"msg_date\": \"Fri, 7 Oct 2022 20:00:56 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-10-07 20:00:56 +0300, Melih Mutlu wrote:\\n> Attached patch is the adjusted version to build with meson.\\n\\nThanks!\\n\\n\\n> diff --git a/.cirrus.yml b/.cirrus.yml\\n> index 9f2282471a..7e6ebdf7b7 100644\\n> --- a/.cirrus.yml\\n> +++ b/.cirrus.yml\\n> @@ -35,6 +35,7 @@ on_failure_ac: &on_failure_ac\\n> - \\\"**/*.log\\\"\\n> - \\\"**/*.diffs\\\"\\n> - \\\"**/regress_log_*\\\"\\n> + - \\\"**/*.stackdump\\\"\\n> type: text/plain\\n\\n\\nI think it might be easier to just set MSYS=winjitdebug\\nhttps://www.msys2.org/wiki/JIT-Debugging/#native-windows-processes-started-from-msys2\\nand then rely on the existing windows crash reporting stuff.\\n\\n> setup_additional_packages_script: |\\n> REM choco install -y --no-progress ...\\n> -\\n> # Use /DEBUG:FASTLINK to avoid high memory usage during linking\\n> configure_script: |\\n\\nYou removed a bunch of newlines here and nearby - I assume that wasn't\\nintentional?\\n\\n\\n> + mingw_info_script:\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\dash.exe -lc \\\"where gcc\\\"\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\dash.exe -lc \\\"gcc --version\\\"\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\dash.exe -lc \\\"where perl\\\"\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\dash.exe -lc \\\"perl --version\\\"\\n> +\\n> + configure_script:\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\dash.exe -lc \\\"cd %CIRRUS_WORKING_DIR% &&\\n> + meson setup --buildtype debug -Dcassert=true -Db_pch=true -DTAR=%TAR% build\\\"\\n\\nThere's no need to use dash anymore, the amount of shell script run is\\nminimal.\\n\\n\\n> + build_script:\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\dash.exe -lc \\\"cd %CIRRUS_WORKING_DIR% && ninja -C %BUILD_DIR%\\\"\\n> + upload_caches: ccache\\n> +\\n> + test_world_script:\\n> + - C:\\\\msys64\\\\usr\\\\bin\\\\dash.exe -lc \\\"cd %CIRRUS_WORKING_DIR% && meson test --print-errorlogs --num-processes %TEST_JOBS% -C %BUILD_DIR%\\\"\\n> +\\n\\nI think the \\\"cd\\\"s here are superfluous given the ninja -C %BUILD_DIR% etc.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 11 Oct 2022 11:23:36 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-10-11 11:23:36 -0700, Andres Freund wrote:\\n> > + build_script:\\n> > + - C:\\\\msys64\\\\usr\\\\bin\\\\dash.exe -lc \\\"cd %CIRRUS_WORKING_DIR% && ninja -C %BUILD_DIR%\\\"\\n> > + upload_caches: ccache\\n\\nOnly remembered that just after sending my email: When using b_pch=true (which\\nsaves a lot of time on mingw) ccache won't cache much (IIRC just the pch files\\nthemselves) unless you do something like\\nexport CCACHE_SLOPPINESS=pch_defines,time_macros\\n\\n$ ccache -z -C\\nClearing... 100.0% [==============================================================================]\\n\\n$ ninja clean && ccache -z && CCACHE_SLOPPINESS= time ninja src/bin/pg_test_fsync/pg_test_fsync.exe && ccache -s\\n[2/2] Cleaning\\nCleaning... 200 files.\\nStatistics zeroed\\n[124/124] Linking target src/bin/pg_test_fsync/pg_test_fsync.exe\\n0.00user 0.01system 0:07.60elapsed 0%CPU (0avgtext+0avgdata 4936maxresident)k\\n0inputs+0outputs (1314major+0minor)pagefaults 0swaps\\nSummary:\\n Hits: 0 / 1 (0.00 %)\\n Direct: 0 / 1 (0.00 %)\\n Preprocessed: 0 / 1 (0.00 %)\\n Misses: 1\\n Direct: 1\\n Preprocessed: 1\\n Uncacheable: 110\\n\\nAs you can see, most of the files are determined to be unachable. The build\\ntook 7.6s.\\n\\n\\n$ ninja clean && ccache -z && CCACHE_SLOPPINESS=pch_defines,time_macros time ninja src/bin/pg_test_fsync/pg_test_fsync.exe && ccache -s\\n[2/2] Cleaning\\nCleaning... 200 files.\\nStatistics zeroed\\n[124/124] Linking target src/bin/pg_test_fsync/pg_test_fsync.exe\\n0.00user 0.01system 0:09.91elapsed 0%CPU (0avgtext+0avgdata 4936maxresident)k\\n0inputs+0outputs (1313major+0minor)pagefaults 0swaps\\nSummary:\\n Hits: 0 / 111 (0.00 %)\\n Direct: 0 / 111 (0.00 %)\\n Preprocessed: 0 / 111 (0.00 %)\\n Misses: 111\\n Direct: 111\\n Preprocessed: 111\\nPrimary storage:\\n Hits: 0 / 216 (0.00 %)\\n Misses: 216\\n Cache size (GB): 0.05 / 5.00 (0.98 %)\\n\\nFiles are cachable, but are cache misses (because we cleaned the cache\\nabove). The build took 9.91s.\\n\\n\\n$ ninja clean && ccache -z && CCACHE_SLOPPINESS=pch_defines,time_macros time ninja src/bin/pg_test_fsync/pg_test_fsync.exe && ccache -s\\n[2/2] Cleaning\\nCleaning... 200 files.\\nStatistics zeroed\\n[124/124] Linking target src/bin/pg_test_fsync/pg_test_fsync.exe\\n0.00user 0.01system 0:02.40elapsed 0%CPU (0avgtext+0avgdata 4936maxresident)k\\n0inputs+0outputs (1314major+0minor)pagefaults 0swaps\\nSummary:\\n Hits: 111 / 111 (100.0 %)\\n Direct: 111 / 111 (100.0 %)\\n Preprocessed: 0 / 0\\n Misses: 0\\n Direct: 0\\n Preprocessed: 0\\nPrimary storage:\\n Hits: 222 / 222 (100.0 %)\\n Misses: 0\\n Cache size (GB): 0.05 / 5.00 (0.98 %)\\n\\nFiles are cachable, and hit the cache. The build takes 2.4s.\\n\\n\\nUsing ccache's depend mode reduce the cache-miss case to 7.51s and the cache\\nhit case to 1.75s. So we should likely export CCACHE_DEPEND=1 as well.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 11 Oct 2022 11:46:12 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi Andres,\\n\\nAndres Freund <andres@anarazel.de>, 11 Eki 2022 Sal, 21:23 tarihinde şunu\\nyazdı:\\n\\n> I think it might be easier to just set MSYS=winjitdebug\\n>\\n> https://www.msys2.org/wiki/JIT-Debugging/#native-windows-processes-started-from-msys2\\n> and then rely on the existing windows crash reporting stuff.\\n>\\n\\nDone.\\n\\n\\n> You removed a bunch of newlines here and nearby - I assume that wasn't\\n> intentional?\\n>\\n\\nYes, that was my mistake. Thanks for pointing it out. Fixed.\\n\\n\\n> There's no need to use dash anymore, the amount of shell script run is\\n> minimal.\\n>\\n\\nSwitched back to bash.\\n\\n\\n> I think the \\\"cd\\\"s here are superfluous given the ninja -C %BUILD_DIR% etc.\\n>\\n\\nYou're right. Those were needed when it was building with autoconf. Not\\nanymore. Removed.\\n\\nOnly remembered that just after sending my email: When using b_pch=true\\n> (which\\n> saves a lot of time on mingw) ccache won't cache much (IIRC just the pch\\n> files\\n> themselves) unless you do something like\\n> export CCACHE_SLOPPINESS=pch_defines,time_macros\\n>\\n\\nAdded CCACHE_SLOPPINESS and CCACHE_SLOPPINESS variables.\\n\\nYou can find the updated patch attached.\\n\\nBest,\\nMelih\",\n \"msg_date\": \"Tue, 18 Oct 2022 12:49:02 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nThanks for the patch!\\n\\nOn 10/18/22 12:49, Melih Mutlu wrote:\\n>\\n>\\n> You can find the updated patch attached.\\n>\\n\\nDoes it makes sense to set these at 'Windows - Server 2019, MinGW64 - \\nMeson' task:\\n\\nenv:\\n ...\\n CHERE_INVOKING: 1\\n BASH_EXE: C:\\\\msys64\\\\usr\\\\bin\\\\bash.exe\\n\\n'CHERE_INVOKING: 1' will cause bash.exe to start from current working \\ndirectory(%CIRRUS_WORKING_DIR%).\\n\\nIn this way, there is no need for\\n'cd %CIRRUS_WORKING_DIR%' when using bash.exe,\\nalso no need for 'BUILD_DIR' environment variable.\\n\\ni.e:\\n\\nconfigure_script: |\\n %BASH_EXE% -lc \\\"meson setup --buildtype debug -Dcassert=true \\n-Db_pch=true -DTAR=%TAR% build\\\"\\n\\nwill work.\\n\\nRegards,\\nNazir Bilal Yavuz\\n\\n\\n\\n\\n\\nHi,\\n\\n Thanks for the patch!\\n\\n\\nOn 10/18/22 12:49, Melih Mutlu wrote:\\n\\n\\n\\n\\n\\n\\n\\n\\nYou can find the\\n updated patch attached.\\n\\n\\n\\n\\nDoes it makes sense to set\\n these at 'Windows - Server 2019, MinGW64 - Meson' task:\\n\\nenv:\\n ...\\n CHERE_INVOKING: 1\\n BASH_EXE:\\n C:\\\\msys64\\\\usr\\\\bin\\\\bash.exe\\n\\n'CHERE_INVOKING: 1' will\\n cause bash.exe to start from current working\\n directory(%CIRRUS_WORKING_DIR%).\\n\\nIn this way, there is no\\n need for\\n'cd %CIRRUS_WORKING_DIR%'\\n when using bash.exe,\\nalso no need for 'BUILD_DIR' environment variable.\\n\\ni.e:\\n\\nconfigure_script: |\\n %BASH_EXE% -lc \\\"meson\\n setup --buildtype debug -Dcassert=true -Db_pch=true -DTAR=%TAR%\\n build\\\"\\n\\nwill work.\\n\\n Regards,\\n Nazir Bilal Yavuz\",\n \"msg_date\": \"Tue, 18 Oct 2022 15:37:32 +0300\",\n \"msg_from\": \"Nazir Bilal Yavuz <byavuz81@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi Bilal,\\n\\nNazir Bilal Yavuz <byavuz81@gmail.com>, 18 Eki 2022 Sal, 15:37 tarihinde\\nşunu yazdı:\\n\\n> env:\\n> ...\\n> CHERE_INVOKING: 1\\n> BASH_EXE: C:\\\\msys64\\\\usr\\\\bin\\\\bash.exe\\n>\\n> 'CHERE_INVOKING: 1' will cause bash.exe to start from current working\\n> directory(%CIRRUS_WORKING_DIR%).\\n>\\n> In this way, there is no need for\\n> 'cd %CIRRUS_WORKING_DIR%' when using bash.exe,\\n> also no need for 'BUILD_DIR' environment variable.\\n>\\n\\n Right, setting CHERE_INVOKING and removing all cd's work and look better.\\nThanks for the suggestion.\\n\\nHere's the updated patch.\\n\\nBest,\\nMelih\",\n \"msg_date\": \"Wed, 19 Oct 2022 00:23:46 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-10-19 00:23:46 +0300, Melih Mutlu wrote:\\n> Right, setting CHERE_INVOKING and removing all cd's work and look better.\\n> Thanks for the suggestion.\\n\\nAgreed, good idea.\\n\\n\\n> +\\n> + env:\\n> + CCACHE_DIR: C:/msys64/ccache\\n\\nIt's a bit odd to separate the CCACHE_* variables from each other\\n(e.g. BUILD_DIR is inbetween them)...\\n\\n\\n> + BUILD_DIR: \\\"%CIRRUS_WORKING_DIR%/build\\\"\\n> + PYTHONHOME: C:/msys64/ucrt64\\n\\nPerhaps add a comment explaining that otherwise plpython tests fail?\\n\\n\\n> + MSYS: winjitdebug\\n\\nWith this src/tools/ci/cores_backtrace.sh shouldn't need to be modified\\nanymore afaict?\\n\\n\\n> + CCACHE_SLOPPINESS: pch_defines,time_macros\\n> + CCACHE_DEPEND: 1\\n\\nI experimented a bit and it looks like ccache doesn't yet quite work in CI,\\nbut only because the ccache needs to be larger. Looks like we need about\\n~400MB.\\n\\nA fully cached build is ~2min\\n\\n\\n> + configure_script: |\\n> + %BASH_EXE% -lc \\\"meson setup --buildtype debug -Dcassert=true -Db_pch=true -DTAR=%TAR% build\\\"\\n\\nWith these buildflags the tests take about 23min37s. Using -Og I saw 18min26s,\\nwith -O1 18m57s and with -O2 18m38s. There's obviously a fair bit of variance,\\nbut it looks like we should use -Og. I think we considered that making compile\\ntimes too bad before, but it seems kinda ok now, with 11min. -O2 is 13min,\\nwithout providing further benefits.\\n\\nI'd replace --buildtype debug with -Ddebug=true -Doptimization=g.\\n\\n\\n> + build_script: |\\n> + %BASH_EXE% -lc \\\"cd %CIRRUS_WORKING_DIR% && ninja -C build\\\"\\n>\\nWhy do we still need this cd?\\n\\n\\n> + upload_caches: ccache\\n> +\\n> + test_world_script: |\\n> + %BASH_EXE% -lc \\\"meson test --print-errorlogs --num-processes %TEST_JOBS% -C build\\\"\\n\\nSeems like this could use %MTEST_ARGS%?\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 18 Oct 2022 20:19:28 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nAndres Freund <andres@anarazel.de>, 19 Eki 2022 Çar, 06:19 tarihinde şunu\\nyazdı:\\n\\n> It's a bit odd to separate the CCACHE_* variables from each other\\n> (e.g. BUILD_DIR is inbetween them)...\\n>\\n> Perhaps add a comment explaining that otherwise plpython tests fail?\\n\\n\\n>\\nWith this src/tools/ci/cores_backtrace.sh shouldn't need to be modified\\n> anymore afaict?\\n>\\n> I'd replace --buildtype debug with -Ddebug=true -Doptimization=g.\\n>\\n\\n\\nSeems like this could use %MTEST_ARGS%?\\n\\n\\nAll the things you mentioned above are done.\\n\\n\\n\\n> > + CCACHE_SLOPPINESS: pch_defines,time_macros\\n> > + CCACHE_DEPEND: 1\\n>\\n> I experimented a bit and it looks like ccache doesn't yet quite work in CI,\\n> but only because the ccache needs to be larger. Looks like we need about\\n> ~400MB.\\n>\\n> A fully cached build is ~2min\\n>\\n\\nThen I should increase CCACHE_MAXSIZE, right? Made it 500MB for MinGW.\\n\\nSharing the updated version of the patch.\\n\\nThanks,\\nMelih\",\n \"msg_date\": \"Wed, 19 Oct 2022 17:59:22 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-10-19 17:59:22 +0300, Melih Mutlu wrote:\\n> All the things you mentioned above are done.\\n\\n> Then I should increase CCACHE_MAXSIZE, right? Made it 500MB for MinGW.\\n\\nYes.\\n\\n> +\\n> + on_failure:\\n> + <<: *on_failure_meson\\n> + cores_script: |\\n> + %BASH_EXE% -lc \\\"cd build src/tools/ci/cores_backtrace.sh msys build/tmp_install\\\"\\n> +\\n\\nThis is wrong - it should just archive the same files that the current windows\\ntask does.\\n\\nOther than that, I think this is basically ready?\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Wed, 19 Oct 2022 09:23:28 -0700\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\n\\n> > +\\n> > + on_failure:\\n> > + <<: *on_failure_meson\\n> > + cores_script: |\\n> > + %BASH_EXE% -lc \\\"cd build src/tools/ci/cores_backtrace.sh msys\\n> build/tmp_install\\\"\\n> > +\\n>\\n> This is wrong - it should just archive the same files that the current\\n> windows\\n> task does.\\n>\\n\\nChanged it with the on_failure from the other windows task.\\n\\n\\n> Other than that, I think this is basically ready?\\n>\\n\\nIf you say so, then I think it's ready.\\n\\nBest,\\nMelih\",\n \"msg_date\": \"Mon, 24 Oct 2022 14:13:06 +0300\",\n \"msg_from\": \"Melih Mutlu <m.melihmutlu@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-10-24 14:13:06 +0300, Melih Mutlu wrote:\\n> If you say so, then I think it's ready.\\n\\nI pushed it with a few changes:\\n\\n- I added an only_if, otherwise the task shows up (without running) even if\\n you use ci-os-only: linux\\n- I added -Dnls=disabled - It seemed to add about 1.5min to the cached build\\n- Inlined the -l into BASH_EXE and renamed it to BASH\\n- The indentation of WINDOWS_ENVIRONMENT_BASE was irregular, leading to a\\n larger diff. Some other similar cleanups.\\n- I added 'mingw' as a ci-os-only choice. Perhaps not the\\n be-all-end-all. Perhaps we should use windows-{msvc,mingw,cygwin}? Or add\\n ci-task-only?\\n\\n\\nI suspect we can make at least -Dssl=openssl work. Perhaps one of the uuid\\nimplementation is available in msys as well?\\n\\nCurrently we end up with:\\n External libraries\\n bonjour : NO\\n bsd_auth : NO\\n gss : NO\\n icu : YES 72.1\\n ldap : YES\\n libxml : YES 2.10.3\\n libxslt : YES 1.1.37\\n llvm : NO\\n lz4 : YES 1.9.4\\n nls : NO\\n pam : NO\\n plperl : YES\\n plpython : YES 3.10\\n pltcl : YES\\n readline : YES\\n selinux : NO\\n ssl : NO\\n systemd : NO\\n uuid : NO\\n zlib : YES 1.2.13\\n zstd : YES 1.5.2\\n\\ngss should also work if the is available in the msys repo.\\n\\nBonjour is apple only, bsd_auth bsd specific, selinux and systemd linux\\nspecific and pam unixoid specific.\\n\\nIt could also be interesting to see if llvm works, but that might be more\\nwork.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Nov 2022 16:23:56 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Mingw task for Cirrus CI\"\n }\n]","isConversational":false}}},{"rowIdx":53586,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi all,\n\nI've seen various discussions around whether PG makes any guarantees on the\nordering of rows returned by the RETURNING clause (e.g. [1]). In a\nnutshell, when executing a statement such as the following:\n\nCREATE TABLE foo (id INT PRIMARY KEY GENERATED ALWAYS AS IDENTITY, data\nINT);\nINSERT INTO foo (data) VALUES (8), (9), (10) RETURNING id, data;\n\n... us the INSERT guaranteed to return the rows (1,8), (2,9) and (3,10)\n(and in that order)? This point is important when inserting multiple rows\nand wanting to e.g. match a database-generated ID back to memory structures\non the client.\n\nFWIW I've received feedback from a SQL Server engineer that one definitely\nshould *not* depend on such ordering there, and that future optimizations\n(e.g. parallel insertion of many rows) could result in row ordering which\ndiffers from the lexical ordering of the VALUES clause. That seems very\nreasonable; if the situation is similar on PostgreSQL, then I'd suggest\nmaking that very clear in the INSERT[2] and UPDATE[3] docs. I'd also\npossibly point to the workaround of wrapping the INSERT/UPDATE in a CTE\nwhich then defines the ordering.\n\nThanks,\n\nShay\n\n[1]\nhttps://stackoverflow.com/questions/5439293/is-insert-returning-guaranteed-to-return-things-in-the-right-order\n[2] https://www.postgresql.org/docs/current/sql-insert.html\n[3] https://www.postgresql.org/docs/current/sql-update.html\n\nHi all,I've seen various discussions around whether PG makes any guarantees on the ordering of rows returned by the RETURNING clause (e.g. [1]). In a nutshell, when executing a statement such as the following:CREATE TABLE foo (id INT PRIMARY KEY GENERATED ALWAYS AS IDENTITY, data INT);INSERT INTO foo (data) VALUES (8), (9), (10) RETURNING id, data;... us the INSERT guaranteed to return the rows (1,8), (2,9) and (3,10) (and in that order)? This point is important when inserting multiple rows and wanting to e.g. match a database-generated ID back to memory structures on the client.FWIW I've received feedback from a SQL Server engineer that one definitely should *not* depend on such ordering there, and that future optimizations (e.g. parallel insertion of many rows) could result in row ordering which differs from the lexical ordering of the VALUES clause. That seems very reasonable; if the situation is similar on PostgreSQL, then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs. I'd also possibly point to the workaround of wrapping the INSERT/UPDATE in a CTE which then defines the ordering.Thanks,Shay[1] https://stackoverflow.com/questions/5439293/is-insert-returning-guaranteed-to-return-things-in-the-right-order[2] https://www.postgresql.org/docs/current/sql-insert.html[3] https://www.postgresql.org/docs/current/sql-update.html","msg_date":"Sat, 26 Feb 2022 13:41:47 +0100","msg_from":"Shay Rojansky <roji@roji.org>","msg_from_op":true,"msg_subject":"Document ordering guarantees on INSERT/UPDATE RETURNING clause"},{"msg_contents":"On Sat, Feb 26, 2022 at 5:42 AM Shay Rojansky <roji@roji.org> wrote:\n\n> FWIW I've received feedback from a SQL Server engineer that one definitely\n> should *not* depend on such ordering there, and that future optimizations\n> (e.g. parallel insertion of many rows) could result in row ordering which\n> differs from the lexical ordering of the VALUES clause.\n>\n\n\n> That seems very reasonable; if the situation is similar on PostgreSQL,\n> then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs.\n>\n\nThere is clearly no mention of such a guarantee in our documentation.\n\nDavid J.\n\nOn Sat, Feb 26, 2022 at 5:42 AM Shay Rojansky <roji@roji.org> wrote:FWIW I've received feedback from a SQL Server engineer that one definitely should *not* depend on such ordering there, and that future optimizations (e.g. parallel insertion of many rows) could result in row ordering which differs from the lexical ordering of the VALUES clause. That seems very reasonable; if the situation is similar on PostgreSQL, then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs.There is clearly no mention of such a guarantee in our documentation.David J.","msg_date":"Sat, 26 Feb 2022 06:25:22 -0700","msg_from":"\"David G. Johnston\" <david.g.johnston@gmail.com>","msg_from_op":false,"msg_subject":"Re: Document ordering guarantees on INSERT/UPDATE RETURNING clause"},{"msg_contents":"Hi,\n\nOn Sat, Feb 26, 2022 at 06:25:22AM -0700, David G. Johnston wrote:\n> On Sat, Feb 26, 2022 at 5:42 AM Shay Rojansky <roji@roji.org> wrote:\n> \n> > FWIW I've received feedback from a SQL Server engineer that one definitely\n> > should *not* depend on such ordering there, and that future optimizations\n> > (e.g. parallel insertion of many rows) could result in row ordering which\n> > differs from the lexical ordering of the VALUES clause.\n> >\n> \n> \n> > That seems very reasonable; if the situation is similar on PostgreSQL,\n> > then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs.\n> >\n> \n> There is clearly no mention of such a guarantee in our documentation.\n\nYes, which is just how SQL works: a set doesn't have any ordering unless an\nexplicit one has been defined, RETURNING is no exception to that.\n\n\n","msg_date":"Sat, 26 Feb 2022 22:59:01 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Document ordering guarantees on INSERT/UPDATE RETURNING clause"},{"msg_contents":"> > > That seems very reasonable; if the situation is similar on PostgreSQL,\n> > > then I'd suggest making that very clear in the INSERT[2] and\nUPDATE[3] docs.\n> >\n> > There is clearly no mention of such a guarantee in our documentation.\n>\n> Yes, which is just how SQL works: a set doesn't have any ordering unless\nan\n> explicit one has been defined, RETURNING is no exception to that.\n\nThanks for confirming that such a guarantee doesn't exist. I would still\nsuggest explicitly calling that out in the docs around RETURNING, since\nthat seems like an understand pitfall; personally-speaking, this certainly\nwasn't clear to me when first looking at it (even if it is now).\n\n> > > That seems very reasonable; if the situation is similar on PostgreSQL,> > > then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs.> >> > There is clearly no mention of such a guarantee in our documentation.>> Yes, which is just how SQL works: a set doesn't have any ordering unless an> explicit one has been defined, RETURNING is no exception to that.Thanks for confirming that such a guarantee doesn't exist. I would still suggest explicitly calling that out in the docs around RETURNING, since that seems like an understand pitfall; personally-speaking, this certainly wasn't clear to me when first looking at it (even if it is now).","msg_date":"Sat, 26 Feb 2022 17:51:25 +0100","msg_from":"Shay Rojansky <roji@roji.org>","msg_from_op":true,"msg_subject":"Re: Document ordering guarantees on INSERT/UPDATE RETURNING clause"}],"string":"[\n {\n \"msg_contents\": \"Hi all,\\n\\nI've seen various discussions around whether PG makes any guarantees on the\\nordering of rows returned by the RETURNING clause (e.g. [1]). In a\\nnutshell, when executing a statement such as the following:\\n\\nCREATE TABLE foo (id INT PRIMARY KEY GENERATED ALWAYS AS IDENTITY, data\\nINT);\\nINSERT INTO foo (data) VALUES (8), (9), (10) RETURNING id, data;\\n\\n... us the INSERT guaranteed to return the rows (1,8), (2,9) and (3,10)\\n(and in that order)? This point is important when inserting multiple rows\\nand wanting to e.g. match a database-generated ID back to memory structures\\non the client.\\n\\nFWIW I've received feedback from a SQL Server engineer that one definitely\\nshould *not* depend on such ordering there, and that future optimizations\\n(e.g. parallel insertion of many rows) could result in row ordering which\\ndiffers from the lexical ordering of the VALUES clause. That seems very\\nreasonable; if the situation is similar on PostgreSQL, then I'd suggest\\nmaking that very clear in the INSERT[2] and UPDATE[3] docs. I'd also\\npossibly point to the workaround of wrapping the INSERT/UPDATE in a CTE\\nwhich then defines the ordering.\\n\\nThanks,\\n\\nShay\\n\\n[1]\\nhttps://stackoverflow.com/questions/5439293/is-insert-returning-guaranteed-to-return-things-in-the-right-order\\n[2] https://www.postgresql.org/docs/current/sql-insert.html\\n[3] https://www.postgresql.org/docs/current/sql-update.html\\n\\nHi all,I've seen various discussions around whether PG makes any guarantees on the ordering of rows returned by the RETURNING clause (e.g. [1]). In a nutshell, when executing a statement such as the following:CREATE TABLE foo (id INT PRIMARY KEY GENERATED ALWAYS AS IDENTITY, data INT);INSERT INTO foo (data) VALUES (8), (9), (10) RETURNING id, data;... us the INSERT guaranteed to return the rows (1,8), (2,9) and (3,10) (and in that order)? This point is important when inserting multiple rows and wanting to e.g. match a database-generated ID back to memory structures on the client.FWIW I've received feedback from a SQL Server engineer that one definitely should *not* depend on such ordering there, and that future optimizations (e.g. parallel insertion of many rows) could result in row ordering which differs from the lexical ordering of the VALUES clause. That seems very reasonable; if the situation is similar on PostgreSQL, then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs. I'd also possibly point to the workaround of wrapping the INSERT/UPDATE in a CTE which then defines the ordering.Thanks,Shay[1] https://stackoverflow.com/questions/5439293/is-insert-returning-guaranteed-to-return-things-in-the-right-order[2] https://www.postgresql.org/docs/current/sql-insert.html[3] https://www.postgresql.org/docs/current/sql-update.html\",\n \"msg_date\": \"Sat, 26 Feb 2022 13:41:47 +0100\",\n \"msg_from\": \"Shay Rojansky <roji@roji.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Document ordering guarantees on INSERT/UPDATE RETURNING clause\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 5:42 AM Shay Rojansky <roji@roji.org> wrote:\\n\\n> FWIW I've received feedback from a SQL Server engineer that one definitely\\n> should *not* depend on such ordering there, and that future optimizations\\n> (e.g. parallel insertion of many rows) could result in row ordering which\\n> differs from the lexical ordering of the VALUES clause.\\n>\\n\\n\\n> That seems very reasonable; if the situation is similar on PostgreSQL,\\n> then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs.\\n>\\n\\nThere is clearly no mention of such a guarantee in our documentation.\\n\\nDavid J.\\n\\nOn Sat, Feb 26, 2022 at 5:42 AM Shay Rojansky <roji@roji.org> wrote:FWIW I've received feedback from a SQL Server engineer that one definitely should *not* depend on such ordering there, and that future optimizations (e.g. parallel insertion of many rows) could result in row ordering which differs from the lexical ordering of the VALUES clause. That seems very reasonable; if the situation is similar on PostgreSQL, then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs.There is clearly no mention of such a guarantee in our documentation.David J.\",\n \"msg_date\": \"Sat, 26 Feb 2022 06:25:22 -0700\",\n \"msg_from\": \"\\\"David G. Johnston\\\" <david.g.johnston@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Document ordering guarantees on INSERT/UPDATE RETURNING clause\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Sat, Feb 26, 2022 at 06:25:22AM -0700, David G. Johnston wrote:\\n> On Sat, Feb 26, 2022 at 5:42 AM Shay Rojansky <roji@roji.org> wrote:\\n> \\n> > FWIW I've received feedback from a SQL Server engineer that one definitely\\n> > should *not* depend on such ordering there, and that future optimizations\\n> > (e.g. parallel insertion of many rows) could result in row ordering which\\n> > differs from the lexical ordering of the VALUES clause.\\n> >\\n> \\n> \\n> > That seems very reasonable; if the situation is similar on PostgreSQL,\\n> > then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs.\\n> >\\n> \\n> There is clearly no mention of such a guarantee in our documentation.\\n\\nYes, which is just how SQL works: a set doesn't have any ordering unless an\\nexplicit one has been defined, RETURNING is no exception to that.\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 22:59:01 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Document ordering guarantees on INSERT/UPDATE RETURNING clause\"\n },\n {\n \"msg_contents\": \"> > > That seems very reasonable; if the situation is similar on PostgreSQL,\\n> > > then I'd suggest making that very clear in the INSERT[2] and\\nUPDATE[3] docs.\\n> >\\n> > There is clearly no mention of such a guarantee in our documentation.\\n>\\n> Yes, which is just how SQL works: a set doesn't have any ordering unless\\nan\\n> explicit one has been defined, RETURNING is no exception to that.\\n\\nThanks for confirming that such a guarantee doesn't exist. I would still\\nsuggest explicitly calling that out in the docs around RETURNING, since\\nthat seems like an understand pitfall; personally-speaking, this certainly\\nwasn't clear to me when first looking at it (even if it is now).\\n\\n> > > That seems very reasonable; if the situation is similar on PostgreSQL,> > > then I'd suggest making that very clear in the INSERT[2] and UPDATE[3] docs.> >> > There is clearly no mention of such a guarantee in our documentation.>> Yes, which is just how SQL works: a set doesn't have any ordering unless an> explicit one has been defined, RETURNING is no exception to that.Thanks for confirming that such a guarantee doesn't exist. I would still suggest explicitly calling that out in the docs around RETURNING, since that seems like an understand pitfall; personally-speaking, this certainly wasn't clear to me when first looking at it (even if it is now).\",\n \"msg_date\": \"Sat, 26 Feb 2022 17:51:25 +0100\",\n \"msg_from\": \"Shay Rojansky <roji@roji.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Document ordering guarantees on INSERT/UPDATE RETURNING clause\"\n }\n]","isConversational":false}}},{"rowIdx":53587,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nIt looks like there's a typo in pg_receivewal.c's\nget_destination_dir(), that is, use the function parameter dest_folder\ninstead of global variable basedir in the error message. Although\nthere's no harm in it as basedir is a global variable in\npg_receivewal.c, let's use the function parameter to be in sync with\nclose_destination_dir.\n\nAttaching a tiny patch to fix it.\n\nRegards,\nBharath Rupireddy.","msg_date":"Sat, 26 Feb 2022 22:04:13 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Fix a typo in pg_receivewal.c's get_destination_dir()"},{"msg_contents":"Hi,\n\nOn 2022-02-26 22:04:13 +0530, Bharath Rupireddy wrote:\n> It looks like there's a typo in pg_receivewal.c's\n> get_destination_dir(), that is, use the function parameter dest_folder\n> instead of global variable basedir in the error message. Although\n> there's no harm in it as basedir is a global variable in\n> pg_receivewal.c, let's use the function parameter to be in sync with\n> close_destination_dir.\n> \n> Attaching a tiny patch to fix it.\n\nGood catch, pushed.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sat, 26 Feb 2022 16:46:37 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Fix a typo in pg_receivewal.c's get_destination_dir()"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nIt looks like there's a typo in pg_receivewal.c's\\nget_destination_dir(), that is, use the function parameter dest_folder\\ninstead of global variable basedir in the error message. Although\\nthere's no harm in it as basedir is a global variable in\\npg_receivewal.c, let's use the function parameter to be in sync with\\nclose_destination_dir.\\n\\nAttaching a tiny patch to fix it.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sat, 26 Feb 2022 22:04:13 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Fix a typo in pg_receivewal.c's get_destination_dir()\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-26 22:04:13 +0530, Bharath Rupireddy wrote:\\n> It looks like there's a typo in pg_receivewal.c's\\n> get_destination_dir(), that is, use the function parameter dest_folder\\n> instead of global variable basedir in the error message. Although\\n> there's no harm in it as basedir is a global variable in\\n> pg_receivewal.c, let's use the function parameter to be in sync with\\n> close_destination_dir.\\n> \\n> Attaching a tiny patch to fix it.\\n\\nGood catch, pushed.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 16:46:37 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Fix a typo in pg_receivewal.c's get_destination_dir()\"\n }\n]","isConversational":false}}},{"rowIdx":53588,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"About once a month over the last six months, my buildfarm animal\nflorican has gotten stuck while running the core regression tests.\nThe symptoms have looked very much the same each time: there is\na backend with two parallel worker processes that are just sitting\nand not consuming any CPU time. Each time I've attached to these\nprocesses with gdb to check their stack traces, and seen pretty\nmuch the same story every time (traces below). What is really\ninteresting is that after I detach from the second worker, the\nprocesses resume running and finish out the test successfully!\nI don't know much about how gdb interacts with kernel calls on\nFreeBSD, but I speculate that the poll(2) call returns with EINTR\nafter gdb releases the process, and then things resume fine,\nsuggesting that we lost an interrupt somewhere.\n\nI have observed this three times in the REL_11 branch, once\nin REL_12, and a couple of times last summer before it occurred\nto me to start keeping notes. Over that time the machine has\nbeen running various patchlevels of FreeBSD 13.0.\n\nHere's the stack trace from the leader process in the most\nrecent event (REL_11 as of yesterday). It's not always the\nsame query that gets stuck, but it's always a parallel hash join:\n\n(gdb) p debug_query_string \n$1 = 0x21873090 \"select count(*) from simple r join simple s using (id);\"\n(gdb) bt\n#0 _poll () at _poll.S:4\n#1 0x21701361 in __thr_poll (fds=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\n#2 0x215eaf3f in poll (pfd=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\n#3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\n#4 WaitEventSetWait (set=0x219dc138, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217738) at latch.c:1000\n#5 0x0099842b in ConditionVariableSleep (cv=0x2bf0b230, wait_event_info=134217738) at condition_variable.c:157\n#6 0x00977e12 in BarrierArriveAndWait (barrier=0x2bf0b218, wait_event_info=134217738) at barrier.c:191\n#7 0x00873441 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:328\n#8 ExecParallelHashJoin (pstate=0x2a887740) at nodeHashjoin.c:600\n#9 0x0086a509 in ExecProcNode (node=0x2a887740) at ../../../src/include/executor/executor.h:248\n#10 fetch_input_tuple (aggstate=aggstate@entry=0x2a887500) at nodeAgg.c:407\n#11 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\n#12 ExecAgg (pstate=0x2a887500) at nodeAgg.c:1561\n#13 0x0086e181 in ExecProcNode (node=0x2a887500) at ../../../src/include/executor/executor.h:248\n#14 gather_getnext (gatherstate=0x2a887414) at nodeGather.c:276\n#15 ExecGather (pstate=0x2a887414) at nodeGather.c:207\n#16 0x0086a509 in ExecProcNode (node=0x2a887414) at ../../../src/include/executor/executor.h:248\n#17 fetch_input_tuple (aggstate=aggstate@entry=0x2a8871b8) at nodeAgg.c:407\n#18 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\n#19 ExecAgg (pstate=0x2a8871b8) at nodeAgg.c:1561\n#20 0x0085956b in ExecProcNode (node=0x2a8871b8) at ../../../src/include/executor/executor.h:248\n#21 ExecutePlan (estate=0x2a887090, planstate=0x2a8871b8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>, \n numberTuples=<optimized out>, direction=<optimized out>, dest=0x2ab629b4, execute_once=<optimized out>) at execMain.c:1712\n#22 standard_ExecutorRun (queryDesc=0x218a1490, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\n#23 0x00859445 in ExecutorRun (queryDesc=0x218a1490, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\n#24 0x009a3d57 in PortalRunSelect (portal=portal@entry=0x2197f090, forward=<optimized out>, count=0, dest=0x2ab629b4) at pquery.c:941\n#25 0x009a39d6 in PortalRun (portal=0x2197f090, count=2147483647, isTopLevel=<optimized out>, run_once=<optimized out>, dest=0x2ab629b4, altdest=0x2ab629b4, \n completionTag=0xffbfdc70 \"\") at pquery.c:782\n#26 0x009a2b53 in exec_simple_query (query_string=query_string@entry=0x21873090 \"select count(*) from simple r join simple s using (id);\") at postgres.c:1145\n#27 0x009a04ec in PostgresMain (argc=1, argv=0x21954ed4, dbname=<optimized out>, username=0x21954d38 \"buildfarm\") at postgres.c:4052\n\nWorker 1 looks like this:\n\n(gdb) bt\n#0 _poll () at _poll.S:4\n#1 0x21701361 in __thr_poll (fds=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\n#2 0x215eaf3f in poll (pfd=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\n#3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\n#4 WaitEventSetWait (set=0x2187cb10, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217738) at latch.c:1000\n#5 0x0099842b in ConditionVariableSleep (cv=0x2196b230, wait_event_info=134217738) at condition_variable.c:157\n#6 0x00977e12 in BarrierArriveAndWait (barrier=0x2196b218, wait_event_info=134217738) at barrier.c:191\n#7 0x00873441 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:328\n#8 ExecParallelHashJoin (pstate=0x2a8378d8) at nodeHashjoin.c:600\n#9 0x0086a509 in ExecProcNode (node=0x2a8378d8) at ../../../src/include/executor/executor.h:248\n#10 fetch_input_tuple (aggstate=aggstate@entry=0x2a8376c8) at nodeAgg.c:407\n#11 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\n#12 ExecAgg (pstate=0x2a8376c8) at nodeAgg.c:1561\n#13 0x0085956b in ExecProcNode (node=0x2a8376c8) at ../../../src/include/executor/executor.h:248\n#14 ExecutePlan (estate=0x2a837090, planstate=0x2a8376c8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>, \n numberTuples=<optimized out>, direction=<optimized out>, dest=0x219ac780, execute_once=<optimized out>) at execMain.c:1712\n#15 standard_ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\n#16 0x00859445 in ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\n#17 0x0085d6de in ParallelQueryMain (seg=0x2187c89c, toc=0x2196b000) at execParallel.c:1402\n#18 0x00750ee0 in ParallelWorkerMain (main_arg=1903254317) at parallel.c:1433\n\nWorker 2 looks like this:\n\n(gdb) bt\n#0 _poll () at _poll.S:4\n#1 0x21701361 in __thr_poll (fds=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\n#2 0x215eaf3f in poll (pfd=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\n#3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\n#4 WaitEventSetWait (set=0x2187cb10, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217737) at latch.c:1000\n#5 0x0099842b in ConditionVariableSleep (cv=0x2196b230, wait_event_info=134217737) at condition_variable.c:157\n#6 0x00977e12 in BarrierArriveAndWait (barrier=0x2196b218, wait_event_info=134217737) at barrier.c:191\n#7 0x0086ea44 in MultiExecParallelHash (node=<optimized out>) at nodeHash.c:312\n#8 MultiExecHash (node=0x2a837e88) at nodeHash.c:112\n#9 0x0085fa03 in MultiExecProcNode (node=0x2a837e88) at execProcnode.c:501\n#10 0x00872f80 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:290\n#11 ExecParallelHashJoin (pstate=0x2a8378d8) at nodeHashjoin.c:600\n#12 0x0086a509 in ExecProcNode (node=0x2a8378d8) at ../../../src/include/executor/executor.h:248\n#13 fetch_input_tuple (aggstate=aggstate@entry=0x2a8376c8) at nodeAgg.c:407\n#14 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\n#15 ExecAgg (pstate=0x2a8376c8) at nodeAgg.c:1561\n#16 0x0085956b in ExecProcNode (node=0x2a8376c8) at ../../../src/include/executor/executor.h:248\n#17 ExecutePlan (estate=0x2a837090, planstate=0x2a8376c8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>, \n numberTuples=<optimized out>, direction=<optimized out>, dest=0x219ac780, execute_once=<optimized out>) at execMain.c:1712\n#18 standard_ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\n#19 0x00859445 in ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\n#20 0x0085d6de in ParallelQueryMain (seg=0x2187c89c, toc=0x2196b000) at execParallel.c:1402\n#21 0x00750ee0 in ParallelWorkerMain (main_arg=1903254317) at parallel.c:1433\n#22 0x009193d3 in StartBackgroundWorker () at bgworker.c:911\n\nIn each case I attached to the process, ran the gdb commands shown\nabove, and detached. As I've grown used to seeing, the processes\nresumed running after I detached from the second worker. I've not\ntried attaching in something other than PID order, but I probably\nwill next time.\n\nThoughts? Ideas on debugging this?\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 26 Feb 2022 14:07:05 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Missed condition-variable wakeups on FreeBSD"},{"msg_contents":"Hi,\n\nOn 2022-02-26 14:07:05 -0500, Tom Lane wrote:\n> About once a month over the last six months, my buildfarm animal\n> florican has gotten stuck while running the core regression tests.\n> The symptoms have looked very much the same each time: there is\n> a backend with two parallel worker processes that are just sitting\n> and not consuming any CPU time. Each time I've attached to these\n> processes with gdb to check their stack traces, and seen pretty\n> much the same story every time (traces below). What is really\n> interesting is that after I detach from the second worker, the\n> processes resume running and finish out the test successfully!\n> I don't know much about how gdb interacts with kernel calls on\n> FreeBSD, but I speculate that the poll(2) call returns with EINTR\n> after gdb releases the process, and then things resume fine,\n> suggesting that we lost an interrupt somewhere.\n>\n> I have observed this three times in the REL_11 branch, once\n> in REL_12, and a couple of times last summer before it occurred\n> to me to start keeping notes. Over that time the machine has\n> been running various patchlevels of FreeBSD 13.0.\n\nIt's certainly interesting that it appears to happen only in the branches\nusing poll rather than kqueue to implement latches. That changed between 12\nand 13.\n\nHave you tried running the core regression tests with force_parallel_mode =\non, or with the parallel costs lowered, to see if that makes the problem\nappear more often?\n\n\n> Here's the stack trace from the leader process in the most\n> recent event (REL_11 as of yesterday). It's not always the\n> same query that gets stuck, but it's always a parallel hash join:\n\nWhich does make me wonder if it's a latch issue or a logic issue in hashjoin /\nbarriers. HJ is the only user of barrier.c...\n\n\nThere's this commit, which subsequently was reverted due to some issue, and\nnot re-applied since...\n\ncommit 4e0f0995e923948631c4114ab353b256b51b58ad\nAuthor: Thomas Munro <tmunro@postgresql.org>\nDate: 2021-03-17 17:46:39 +1300\n \n Fix race in Parallel Hash Join batch cleanup.\n\nIt doesn't seem super likely to be related, but...\n\n\n> (gdb) p debug_query_string\n> $1 = 0x21873090 \"select count(*) from simple r join simple s using (id);\"\n> (gdb) bt\n> #0 _poll () at _poll.S:4\n> #1 0x21701361 in __thr_poll (fds=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\n> #2 0x215eaf3f in poll (pfd=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\n> #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\n\nThe next time this happens / if you still have this open, perhaps it could be\nworth checking if there's a byte in the self pipe?\n\n\n> Thoughts? Ideas on debugging this?\n\nBesides trying to make the issue more likely as suggested above, it might be\nworth checking if signalling the stuck processes with SIGUSR1 gets them\nunstuck.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sat, 26 Feb 2022 12:44:31 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":false,"msg_subject":"Re: Missed condition-variable wakeups on FreeBSD"},{"msg_contents":"On Sat, Feb 26, 2022 at 02:07:05PM -0500, Tom Lane wrote:\n> I don't know much about how gdb interacts with kernel calls on\n> FreeBSD, but I speculate that the poll(2) call returns with EINTR\n> after gdb releases the process, and then things resume fine,\n> suggesting that we lost an interrupt somewhere.\n\nI've seen some similar interactions with strace under linux causing a process\nto be woken up or otherwise incur a different behavior (not necessarily\npostgres). \n\n> Thoughts? Ideas on debugging this?\n\nBefore attaching a debugger, figure out what syscall each process is in.\nIn linux, that's ps O wchan PID.\n\n>> Besides trying to make the issue more likely as suggested above, it might be\n>> worth checking if signalling the stuck processes with SIGUSR1 gets them\n>> unstuck.\n\nAnd SIGCONT.\n\nMaybe already did this, but you can dump a corefile of the running processes to\nallow future inspection.\n\n-- \nJustin\n\n\n","msg_date":"Sat, 26 Feb 2022 15:06:25 -0600","msg_from":"Justin Pryzby <pryzby@telsasoft.com>","msg_from_op":false,"msg_subject":"Re: Missed condition-variable wakeups on FreeBSD"},{"msg_contents":"On Sat, Feb 26, 2022 at 2:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n>\n> About once a month over the last six months, my buildfarm animal\n> florican has gotten stuck while running the core regression tests.\n> The symptoms have looked very much the same each time: there is\n> a backend with two parallel worker processes that are just sitting\n> and not consuming any CPU time. Each time I've attached to these\n> processes with gdb to check their stack traces, and seen pretty\n> much the same story every time (traces below). What is really\n> interesting is that after I detach from the second worker, the\n> processes resume running and finish out the test successfully!\n> I don't know much about how gdb interacts with kernel calls on\n> FreeBSD, but I speculate that the poll(2) call returns with EINTR\n> after gdb releases the process, and then things resume fine,\n> suggesting that we lost an interrupt somewhere.\n>\n> I have observed this three times in the REL_11 branch, once\n> in REL_12, and a couple of times last summer before it occurred\n> to me to start keeping notes. Over that time the machine has\n> been running various patchlevels of FreeBSD 13.0.\n>\n> Here's the stack trace from the leader process in the most\n> recent event (REL_11 as of yesterday). It's not always the\n> same query that gets stuck, but it's always a parallel hash join:\n>\n> (gdb) p debug_query_string\n> $1 = 0x21873090 \"select count(*) from simple r join simple s using (id);\"\n> (gdb) bt\n> #0 _poll () at _poll.S:4\n> #1 0x21701361 in __thr_poll (fds=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\n> #2 0x215eaf3f in poll (pfd=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\n> #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\n> #4 WaitEventSetWait (set=0x219dc138, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217738) at latch.c:1000\n> #5 0x0099842b in ConditionVariableSleep (cv=0x2bf0b230, wait_event_info=134217738) at condition_variable.c:157\n> #6 0x00977e12 in BarrierArriveAndWait (barrier=0x2bf0b218, wait_event_info=134217738) at barrier.c:191\n> #7 0x00873441 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:328\n> #8 ExecParallelHashJoin (pstate=0x2a887740) at nodeHashjoin.c:600\n> #9 0x0086a509 in ExecProcNode (node=0x2a887740) at ../../../src/include/executor/executor.h:248\n> #10 fetch_input_tuple (aggstate=aggstate@entry=0x2a887500) at nodeAgg.c:407\n> #11 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\n> #12 ExecAgg (pstate=0x2a887500) at nodeAgg.c:1561\n> #13 0x0086e181 in ExecProcNode (node=0x2a887500) at ../../../src/include/executor/executor.h:248\n> #14 gather_getnext (gatherstate=0x2a887414) at nodeGather.c:276\n> #15 ExecGather (pstate=0x2a887414) at nodeGather.c:207\n> #16 0x0086a509 in ExecProcNode (node=0x2a887414) at ../../../src/include/executor/executor.h:248\n> #17 fetch_input_tuple (aggstate=aggstate@entry=0x2a8871b8) at nodeAgg.c:407\n> #18 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\n> #19 ExecAgg (pstate=0x2a8871b8) at nodeAgg.c:1561\n> #20 0x0085956b in ExecProcNode (node=0x2a8871b8) at ../../../src/include/executor/executor.h:248\n> #21 ExecutePlan (estate=0x2a887090, planstate=0x2a8871b8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>,\n> numberTuples=<optimized out>, direction=<optimized out>, dest=0x2ab629b4, execute_once=<optimized out>) at execMain.c:1712\n> #22 standard_ExecutorRun (queryDesc=0x218a1490, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\n> #23 0x00859445 in ExecutorRun (queryDesc=0x218a1490, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\n> #24 0x009a3d57 in PortalRunSelect (portal=portal@entry=0x2197f090, forward=<optimized out>, count=0, dest=0x2ab629b4) at pquery.c:941\n> #25 0x009a39d6 in PortalRun (portal=0x2197f090, count=2147483647, isTopLevel=<optimized out>, run_once=<optimized out>, dest=0x2ab629b4, altdest=0x2ab629b4,\n> completionTag=0xffbfdc70 \"\") at pquery.c:782\n> #26 0x009a2b53 in exec_simple_query (query_string=query_string@entry=0x21873090 \"select count(*) from simple r join simple s using (id);\") at postgres.c:1145\n> #27 0x009a04ec in PostgresMain (argc=1, argv=0x21954ed4, dbname=<optimized out>, username=0x21954d38 \"buildfarm\") at postgres.c:4052\n>\n> Worker 1 looks like this:\n>\n> (gdb) bt\n> #0 _poll () at _poll.S:4\n> #1 0x21701361 in __thr_poll (fds=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\n> #2 0x215eaf3f in poll (pfd=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\n> #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\n> #4 WaitEventSetWait (set=0x2187cb10, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217738) at latch.c:1000\n> #5 0x0099842b in ConditionVariableSleep (cv=0x2196b230, wait_event_info=134217738) at condition_variable.c:157\n> #6 0x00977e12 in BarrierArriveAndWait (barrier=0x2196b218, wait_event_info=134217738) at barrier.c:191\n> #7 0x00873441 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:328\n> #8 ExecParallelHashJoin (pstate=0x2a8378d8) at nodeHashjoin.c:600\n> #9 0x0086a509 in ExecProcNode (node=0x2a8378d8) at ../../../src/include/executor/executor.h:248\n> #10 fetch_input_tuple (aggstate=aggstate@entry=0x2a8376c8) at nodeAgg.c:407\n> #11 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\n> #12 ExecAgg (pstate=0x2a8376c8) at nodeAgg.c:1561\n> #13 0x0085956b in ExecProcNode (node=0x2a8376c8) at ../../../src/include/executor/executor.h:248\n> #14 ExecutePlan (estate=0x2a837090, planstate=0x2a8376c8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>,\n> numberTuples=<optimized out>, direction=<optimized out>, dest=0x219ac780, execute_once=<optimized out>) at execMain.c:1712\n> #15 standard_ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\n> #16 0x00859445 in ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\n> #17 0x0085d6de in ParallelQueryMain (seg=0x2187c89c, toc=0x2196b000) at execParallel.c:1402\n> #18 0x00750ee0 in ParallelWorkerMain (main_arg=1903254317) at parallel.c:1433\n>\n> Worker 2 looks like this:\n>\n> (gdb) bt\n> #0 _poll () at _poll.S:4\n> #1 0x21701361 in __thr_poll (fds=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\n> #2 0x215eaf3f in poll (pfd=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\n> #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\n> #4 WaitEventSetWait (set=0x2187cb10, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217737) at latch.c:1000\n> #5 0x0099842b in ConditionVariableSleep (cv=0x2196b230, wait_event_info=134217737) at condition_variable.c:157\n> #6 0x00977e12 in BarrierArriveAndWait (barrier=0x2196b218, wait_event_info=134217737) at barrier.c:191\n> #7 0x0086ea44 in MultiExecParallelHash (node=<optimized out>) at nodeHash.c:312\n> #8 MultiExecHash (node=0x2a837e88) at nodeHash.c:112\n> #9 0x0085fa03 in MultiExecProcNode (node=0x2a837e88) at execProcnode.c:501\n> #10 0x00872f80 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:290\n> #11 ExecParallelHashJoin (pstate=0x2a8378d8) at nodeHashjoin.c:600\n> #12 0x0086a509 in ExecProcNode (node=0x2a8378d8) at ../../../src/include/executor/executor.h:248\n> #13 fetch_input_tuple (aggstate=aggstate@entry=0x2a8376c8) at nodeAgg.c:407\n> #14 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\n> #15 ExecAgg (pstate=0x2a8376c8) at nodeAgg.c:1561\n> #16 0x0085956b in ExecProcNode (node=0x2a8376c8) at ../../../src/include/executor/executor.h:248\n> #17 ExecutePlan (estate=0x2a837090, planstate=0x2a8376c8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>,\n> numberTuples=<optimized out>, direction=<optimized out>, dest=0x219ac780, execute_once=<optimized out>) at execMain.c:1712\n> #18 standard_ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\n> #19 0x00859445 in ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\n> #20 0x0085d6de in ParallelQueryMain (seg=0x2187c89c, toc=0x2196b000) at execParallel.c:1402\n> #21 0x00750ee0 in ParallelWorkerMain (main_arg=1903254317) at parallel.c:1433\n> #22 0x009193d3 in StartBackgroundWorker () at bgworker.c:911\n>\n> In each case I attached to the process, ran the gdb commands shown\n> above, and detached. As I've grown used to seeing, the processes\n> resumed running after I detached from the second worker. I've not\n> tried attaching in something other than PID order, but I probably\n> will next time.\n>\n> Thoughts? Ideas on debugging this?\n\nHow could it be that worker 2 is waiting on the build barrier in\nPHJ_BUILD_HASHING_INNER and worker 1 and the leader are waiting on it\nwith it supposedly in PHJ_BUILD_HASHING_OUTER?\n\n\n","msg_date":"Sat, 26 Feb 2022 17:18:18 -0500","msg_from":"Melanie Plageman <melanieplageman@gmail.com>","msg_from_op":false,"msg_subject":"Re: Missed condition-variable wakeups on FreeBSD"},{"msg_contents":"Andres Freund <andres@anarazel.de> writes:\n> On 2022-02-26 14:07:05 -0500, Tom Lane wrote:\n>> I have observed this three times in the REL_11 branch, once\n>> in REL_12, and a couple of times last summer before it occurred\n>> to me to start keeping notes. Over that time the machine has\n>> been running various patchlevels of FreeBSD 13.0.\n\n> It's certainly interesting that it appears to happen only in the branches\n> using poll rather than kqueue to implement latches. That changed between 12\n> and 13.\n\nYeah, and there was no PHJ in v10, so that's a pretty good theory as\nto why I've only seen it in those two branches.\n\n> Have you tried running the core regression tests with force_parallel_mode =\n> on, or with the parallel costs lowered, to see if that makes the problem\n> appear more often?\n> The next time this happens / if you still have this open, perhaps it could be\n> worth checking if there's a byte in the self pipe?\n> Besides trying to make the issue more likely as suggested above, it might be\n> worth checking if signalling the stuck processes with SIGUSR1 gets them\n> unstuck.\n\nI've now wasted a bunch of kilowatt-hours fruitlessly trying to\nreproduce this outside the confines of the buildfarm script.\nI'm at a loss to figure out what the buildfarm is doing differently,\nbut apparently there's something. I'm going to re-enable the\nmachine's buildfarm job and just wait for it to hang up again.\nMore info eventually ...\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 27 Feb 2022 17:24:48 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":true,"msg_subject":"Re: Missed condition-variable wakeups on FreeBSD"},{"msg_contents":"On Sun, Feb 27, 2022 at 8:07 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> I don't know much about how gdb interacts with kernel calls on\n> FreeBSD, but I speculate that the poll(2) call returns with EINTR\n> after gdb releases the process, and then things resume fine,\n\nYeah, at least FreeBSD and macOS interrupt system calls when you send\nSIGSTOP + SIGCONT directly, or when gdb and lldb do something similar\nvia ptrace. The same applies on Linux, except that Linux restarts the\nsystem call automatically (even though the man page has this in the\nlist of system calls that never restart) so you don't usually notice\n(though you can see it with strace). It's not really clear to me what\nshould happen given the language around restarts is all about signal\nhandlers, and stop + cont are system magic, not signal handlers.\nAnyway...\n\n> suggesting that we lost an interrupt somewhere.\n\nSo it's happening on an i386 kernel, with WAIT_USE_POLL (not\nWAIT_USE_KQUEUE), but only under the build farm script (not when you\nran it manually in loop).... hmmm.\n\n\n","msg_date":"Mon, 28 Feb 2022 13:00:26 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Missed condition-variable wakeups on FreeBSD"},{"msg_contents":"On Sun, Feb 27, 2022 at 11:18 AM Melanie Plageman\n<melanieplageman@gmail.com> wrote:\n> How could it be that worker 2 is waiting on the build barrier in\n> PHJ_BUILD_HASHING_INNER and worker 1 and the leader are waiting on it\n> with it supposedly in PHJ_BUILD_HASHING_OUTER?\n\nThat'd be consistent with a wakeup going missing, somehow. W2 was\nsupposed to be released but somehow didn't get the message (until a\nrandom debugger-induced EINTR releases it and it sees all the state it\nneeds to proceed in shared memory), meanwhile the other two got to the\nend of the next phase and are waiting for it.\n\n\n","msg_date":"Tue, 1 Mar 2022 00:02:08 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Missed condition-variable wakeups on FreeBSD"},{"msg_contents":"On Sun, Feb 27, 2022 at 9:44 AM Andres Freund <andres@anarazel.de> wrote:\n> > (gdb) p debug_query_string\n> > $1 = 0x21873090 \"select count(*) from simple r join simple s using (id);\"\n> > (gdb) bt\n> > #0 _poll () at _poll.S:4\n> > #1 0x21701361 in __thr_poll (fds=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\n> > #2 0x215eaf3f in poll (pfd=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\n> > #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\n>\n> The next time this happens / if you still have this open, perhaps it could be\n> worth checking if there's a byte in the self pipe?\n\nAs a basic sanity check I'd like to see pfd[0], pfd[1] and the output\nof fstat -p PID to see unconsumed bytes in the pipe. For example\n\"echo hello | sleep 60\" shows \"0* pipe ... 6 rw\" for sleep, but \"echo\nhello | more\" shows \"0* pipe ... 0 rw\".\n\n\n","msg_date":"Tue, 1 Mar 2022 10:17:26 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Missed condition-variable wakeups on FreeBSD"},{"msg_contents":"On Sun, Feb 27, 2022 at 8:07 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> I have observed this three times in the REL_11 branch, once\n> in REL_12, and a couple of times last summer before it occurred\n> to me to start keeping notes. Over that time the machine has\n> been running various patchlevels of FreeBSD 13.0.\n\nFTR I noticed that my animal elver hadn't reported for a while and\nlogged and found it locked up, a bit like that, but in test_shm_mq\nrather than PHJ. It's FreeBSD 13.1 on a fairly fast 16 way amd64 (in\na jail), and it was testing REL_12_STABLE, so using poll() for\nWaitEventSetWaitBlock(). When I connected a debugger then\ndisconnected, it was unblocked and continued. Unfortunately I didn't\nget any new information :-(\n\n\n","msg_date":"Fri, 13 Jan 2023 15:00:58 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Missed condition-variable wakeups on FreeBSD"}],"string":"[\n {\n \"msg_contents\": \"About once a month over the last six months, my buildfarm animal\\nflorican has gotten stuck while running the core regression tests.\\nThe symptoms have looked very much the same each time: there is\\na backend with two parallel worker processes that are just sitting\\nand not consuming any CPU time. Each time I've attached to these\\nprocesses with gdb to check their stack traces, and seen pretty\\nmuch the same story every time (traces below). What is really\\ninteresting is that after I detach from the second worker, the\\nprocesses resume running and finish out the test successfully!\\nI don't know much about how gdb interacts with kernel calls on\\nFreeBSD, but I speculate that the poll(2) call returns with EINTR\\nafter gdb releases the process, and then things resume fine,\\nsuggesting that we lost an interrupt somewhere.\\n\\nI have observed this three times in the REL_11 branch, once\\nin REL_12, and a couple of times last summer before it occurred\\nto me to start keeping notes. Over that time the machine has\\nbeen running various patchlevels of FreeBSD 13.0.\\n\\nHere's the stack trace from the leader process in the most\\nrecent event (REL_11 as of yesterday). It's not always the\\nsame query that gets stuck, but it's always a parallel hash join:\\n\\n(gdb) p debug_query_string \\n$1 = 0x21873090 \\\"select count(*) from simple r join simple s using (id);\\\"\\n(gdb) bt\\n#0 _poll () at _poll.S:4\\n#1 0x21701361 in __thr_poll (fds=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\\n#2 0x215eaf3f in poll (pfd=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\\n#3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\\n#4 WaitEventSetWait (set=0x219dc138, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217738) at latch.c:1000\\n#5 0x0099842b in ConditionVariableSleep (cv=0x2bf0b230, wait_event_info=134217738) at condition_variable.c:157\\n#6 0x00977e12 in BarrierArriveAndWait (barrier=0x2bf0b218, wait_event_info=134217738) at barrier.c:191\\n#7 0x00873441 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:328\\n#8 ExecParallelHashJoin (pstate=0x2a887740) at nodeHashjoin.c:600\\n#9 0x0086a509 in ExecProcNode (node=0x2a887740) at ../../../src/include/executor/executor.h:248\\n#10 fetch_input_tuple (aggstate=aggstate@entry=0x2a887500) at nodeAgg.c:407\\n#11 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\\n#12 ExecAgg (pstate=0x2a887500) at nodeAgg.c:1561\\n#13 0x0086e181 in ExecProcNode (node=0x2a887500) at ../../../src/include/executor/executor.h:248\\n#14 gather_getnext (gatherstate=0x2a887414) at nodeGather.c:276\\n#15 ExecGather (pstate=0x2a887414) at nodeGather.c:207\\n#16 0x0086a509 in ExecProcNode (node=0x2a887414) at ../../../src/include/executor/executor.h:248\\n#17 fetch_input_tuple (aggstate=aggstate@entry=0x2a8871b8) at nodeAgg.c:407\\n#18 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\\n#19 ExecAgg (pstate=0x2a8871b8) at nodeAgg.c:1561\\n#20 0x0085956b in ExecProcNode (node=0x2a8871b8) at ../../../src/include/executor/executor.h:248\\n#21 ExecutePlan (estate=0x2a887090, planstate=0x2a8871b8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>, \\n numberTuples=<optimized out>, direction=<optimized out>, dest=0x2ab629b4, execute_once=<optimized out>) at execMain.c:1712\\n#22 standard_ExecutorRun (queryDesc=0x218a1490, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\\n#23 0x00859445 in ExecutorRun (queryDesc=0x218a1490, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\\n#24 0x009a3d57 in PortalRunSelect (portal=portal@entry=0x2197f090, forward=<optimized out>, count=0, dest=0x2ab629b4) at pquery.c:941\\n#25 0x009a39d6 in PortalRun (portal=0x2197f090, count=2147483647, isTopLevel=<optimized out>, run_once=<optimized out>, dest=0x2ab629b4, altdest=0x2ab629b4, \\n completionTag=0xffbfdc70 \\\"\\\") at pquery.c:782\\n#26 0x009a2b53 in exec_simple_query (query_string=query_string@entry=0x21873090 \\\"select count(*) from simple r join simple s using (id);\\\") at postgres.c:1145\\n#27 0x009a04ec in PostgresMain (argc=1, argv=0x21954ed4, dbname=<optimized out>, username=0x21954d38 \\\"buildfarm\\\") at postgres.c:4052\\n\\nWorker 1 looks like this:\\n\\n(gdb) bt\\n#0 _poll () at _poll.S:4\\n#1 0x21701361 in __thr_poll (fds=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\\n#2 0x215eaf3f in poll (pfd=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\\n#3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\\n#4 WaitEventSetWait (set=0x2187cb10, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217738) at latch.c:1000\\n#5 0x0099842b in ConditionVariableSleep (cv=0x2196b230, wait_event_info=134217738) at condition_variable.c:157\\n#6 0x00977e12 in BarrierArriveAndWait (barrier=0x2196b218, wait_event_info=134217738) at barrier.c:191\\n#7 0x00873441 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:328\\n#8 ExecParallelHashJoin (pstate=0x2a8378d8) at nodeHashjoin.c:600\\n#9 0x0086a509 in ExecProcNode (node=0x2a8378d8) at ../../../src/include/executor/executor.h:248\\n#10 fetch_input_tuple (aggstate=aggstate@entry=0x2a8376c8) at nodeAgg.c:407\\n#11 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\\n#12 ExecAgg (pstate=0x2a8376c8) at nodeAgg.c:1561\\n#13 0x0085956b in ExecProcNode (node=0x2a8376c8) at ../../../src/include/executor/executor.h:248\\n#14 ExecutePlan (estate=0x2a837090, planstate=0x2a8376c8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>, \\n numberTuples=<optimized out>, direction=<optimized out>, dest=0x219ac780, execute_once=<optimized out>) at execMain.c:1712\\n#15 standard_ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\\n#16 0x00859445 in ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\\n#17 0x0085d6de in ParallelQueryMain (seg=0x2187c89c, toc=0x2196b000) at execParallel.c:1402\\n#18 0x00750ee0 in ParallelWorkerMain (main_arg=1903254317) at parallel.c:1433\\n\\nWorker 2 looks like this:\\n\\n(gdb) bt\\n#0 _poll () at _poll.S:4\\n#1 0x21701361 in __thr_poll (fds=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\\n#2 0x215eaf3f in poll (pfd=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\\n#3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\\n#4 WaitEventSetWait (set=0x2187cb10, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217737) at latch.c:1000\\n#5 0x0099842b in ConditionVariableSleep (cv=0x2196b230, wait_event_info=134217737) at condition_variable.c:157\\n#6 0x00977e12 in BarrierArriveAndWait (barrier=0x2196b218, wait_event_info=134217737) at barrier.c:191\\n#7 0x0086ea44 in MultiExecParallelHash (node=<optimized out>) at nodeHash.c:312\\n#8 MultiExecHash (node=0x2a837e88) at nodeHash.c:112\\n#9 0x0085fa03 in MultiExecProcNode (node=0x2a837e88) at execProcnode.c:501\\n#10 0x00872f80 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:290\\n#11 ExecParallelHashJoin (pstate=0x2a8378d8) at nodeHashjoin.c:600\\n#12 0x0086a509 in ExecProcNode (node=0x2a8378d8) at ../../../src/include/executor/executor.h:248\\n#13 fetch_input_tuple (aggstate=aggstate@entry=0x2a8376c8) at nodeAgg.c:407\\n#14 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\\n#15 ExecAgg (pstate=0x2a8376c8) at nodeAgg.c:1561\\n#16 0x0085956b in ExecProcNode (node=0x2a8376c8) at ../../../src/include/executor/executor.h:248\\n#17 ExecutePlan (estate=0x2a837090, planstate=0x2a8376c8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>, \\n numberTuples=<optimized out>, direction=<optimized out>, dest=0x219ac780, execute_once=<optimized out>) at execMain.c:1712\\n#18 standard_ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\\n#19 0x00859445 in ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\\n#20 0x0085d6de in ParallelQueryMain (seg=0x2187c89c, toc=0x2196b000) at execParallel.c:1402\\n#21 0x00750ee0 in ParallelWorkerMain (main_arg=1903254317) at parallel.c:1433\\n#22 0x009193d3 in StartBackgroundWorker () at bgworker.c:911\\n\\nIn each case I attached to the process, ran the gdb commands shown\\nabove, and detached. As I've grown used to seeing, the processes\\nresumed running after I detached from the second worker. I've not\\ntried attaching in something other than PID order, but I probably\\nwill next time.\\n\\nThoughts? Ideas on debugging this?\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 14:07:05 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Missed condition-variable wakeups on FreeBSD\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-02-26 14:07:05 -0500, Tom Lane wrote:\\n> About once a month over the last six months, my buildfarm animal\\n> florican has gotten stuck while running the core regression tests.\\n> The symptoms have looked very much the same each time: there is\\n> a backend with two parallel worker processes that are just sitting\\n> and not consuming any CPU time. Each time I've attached to these\\n> processes with gdb to check their stack traces, and seen pretty\\n> much the same story every time (traces below). What is really\\n> interesting is that after I detach from the second worker, the\\n> processes resume running and finish out the test successfully!\\n> I don't know much about how gdb interacts with kernel calls on\\n> FreeBSD, but I speculate that the poll(2) call returns with EINTR\\n> after gdb releases the process, and then things resume fine,\\n> suggesting that we lost an interrupt somewhere.\\n>\\n> I have observed this three times in the REL_11 branch, once\\n> in REL_12, and a couple of times last summer before it occurred\\n> to me to start keeping notes. Over that time the machine has\\n> been running various patchlevels of FreeBSD 13.0.\\n\\nIt's certainly interesting that it appears to happen only in the branches\\nusing poll rather than kqueue to implement latches. That changed between 12\\nand 13.\\n\\nHave you tried running the core regression tests with force_parallel_mode =\\non, or with the parallel costs lowered, to see if that makes the problem\\nappear more often?\\n\\n\\n> Here's the stack trace from the leader process in the most\\n> recent event (REL_11 as of yesterday). It's not always the\\n> same query that gets stuck, but it's always a parallel hash join:\\n\\nWhich does make me wonder if it's a latch issue or a logic issue in hashjoin /\\nbarriers. HJ is the only user of barrier.c...\\n\\n\\nThere's this commit, which subsequently was reverted due to some issue, and\\nnot re-applied since...\\n\\ncommit 4e0f0995e923948631c4114ab353b256b51b58ad\\nAuthor: Thomas Munro <tmunro@postgresql.org>\\nDate: 2021-03-17 17:46:39 +1300\\n \\n Fix race in Parallel Hash Join batch cleanup.\\n\\nIt doesn't seem super likely to be related, but...\\n\\n\\n> (gdb) p debug_query_string\\n> $1 = 0x21873090 \\\"select count(*) from simple r join simple s using (id);\\\"\\n> (gdb) bt\\n> #0 _poll () at _poll.S:4\\n> #1 0x21701361 in __thr_poll (fds=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\\n> #2 0x215eaf3f in poll (pfd=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\\n> #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\\n\\nThe next time this happens / if you still have this open, perhaps it could be\\nworth checking if there's a byte in the self pipe?\\n\\n\\n> Thoughts? Ideas on debugging this?\\n\\nBesides trying to make the issue more likely as suggested above, it might be\\nworth checking if signalling the stuck processes with SIGUSR1 gets them\\nunstuck.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 12:44:31 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Missed condition-variable wakeups on FreeBSD\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 02:07:05PM -0500, Tom Lane wrote:\\n> I don't know much about how gdb interacts with kernel calls on\\n> FreeBSD, but I speculate that the poll(2) call returns with EINTR\\n> after gdb releases the process, and then things resume fine,\\n> suggesting that we lost an interrupt somewhere.\\n\\nI've seen some similar interactions with strace under linux causing a process\\nto be woken up or otherwise incur a different behavior (not necessarily\\npostgres). \\n\\n> Thoughts? Ideas on debugging this?\\n\\nBefore attaching a debugger, figure out what syscall each process is in.\\nIn linux, that's ps O wchan PID.\\n\\n>> Besides trying to make the issue more likely as suggested above, it might be\\n>> worth checking if signalling the stuck processes with SIGUSR1 gets them\\n>> unstuck.\\n\\nAnd SIGCONT.\\n\\nMaybe already did this, but you can dump a corefile of the running processes to\\nallow future inspection.\\n\\n-- \\nJustin\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 15:06:25 -0600\",\n \"msg_from\": \"Justin Pryzby <pryzby@telsasoft.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Missed condition-variable wakeups on FreeBSD\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 2:07 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n>\\n> About once a month over the last six months, my buildfarm animal\\n> florican has gotten stuck while running the core regression tests.\\n> The symptoms have looked very much the same each time: there is\\n> a backend with two parallel worker processes that are just sitting\\n> and not consuming any CPU time. Each time I've attached to these\\n> processes with gdb to check their stack traces, and seen pretty\\n> much the same story every time (traces below). What is really\\n> interesting is that after I detach from the second worker, the\\n> processes resume running and finish out the test successfully!\\n> I don't know much about how gdb interacts with kernel calls on\\n> FreeBSD, but I speculate that the poll(2) call returns with EINTR\\n> after gdb releases the process, and then things resume fine,\\n> suggesting that we lost an interrupt somewhere.\\n>\\n> I have observed this three times in the REL_11 branch, once\\n> in REL_12, and a couple of times last summer before it occurred\\n> to me to start keeping notes. Over that time the machine has\\n> been running various patchlevels of FreeBSD 13.0.\\n>\\n> Here's the stack trace from the leader process in the most\\n> recent event (REL_11 as of yesterday). It's not always the\\n> same query that gets stuck, but it's always a parallel hash join:\\n>\\n> (gdb) p debug_query_string\\n> $1 = 0x21873090 \\\"select count(*) from simple r join simple s using (id);\\\"\\n> (gdb) bt\\n> #0 _poll () at _poll.S:4\\n> #1 0x21701361 in __thr_poll (fds=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\\n> #2 0x215eaf3f in poll (pfd=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\\n> #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\\n> #4 WaitEventSetWait (set=0x219dc138, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217738) at latch.c:1000\\n> #5 0x0099842b in ConditionVariableSleep (cv=0x2bf0b230, wait_event_info=134217738) at condition_variable.c:157\\n> #6 0x00977e12 in BarrierArriveAndWait (barrier=0x2bf0b218, wait_event_info=134217738) at barrier.c:191\\n> #7 0x00873441 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:328\\n> #8 ExecParallelHashJoin (pstate=0x2a887740) at nodeHashjoin.c:600\\n> #9 0x0086a509 in ExecProcNode (node=0x2a887740) at ../../../src/include/executor/executor.h:248\\n> #10 fetch_input_tuple (aggstate=aggstate@entry=0x2a887500) at nodeAgg.c:407\\n> #11 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\\n> #12 ExecAgg (pstate=0x2a887500) at nodeAgg.c:1561\\n> #13 0x0086e181 in ExecProcNode (node=0x2a887500) at ../../../src/include/executor/executor.h:248\\n> #14 gather_getnext (gatherstate=0x2a887414) at nodeGather.c:276\\n> #15 ExecGather (pstate=0x2a887414) at nodeGather.c:207\\n> #16 0x0086a509 in ExecProcNode (node=0x2a887414) at ../../../src/include/executor/executor.h:248\\n> #17 fetch_input_tuple (aggstate=aggstate@entry=0x2a8871b8) at nodeAgg.c:407\\n> #18 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\\n> #19 ExecAgg (pstate=0x2a8871b8) at nodeAgg.c:1561\\n> #20 0x0085956b in ExecProcNode (node=0x2a8871b8) at ../../../src/include/executor/executor.h:248\\n> #21 ExecutePlan (estate=0x2a887090, planstate=0x2a8871b8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>,\\n> numberTuples=<optimized out>, direction=<optimized out>, dest=0x2ab629b4, execute_once=<optimized out>) at execMain.c:1712\\n> #22 standard_ExecutorRun (queryDesc=0x218a1490, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\\n> #23 0x00859445 in ExecutorRun (queryDesc=0x218a1490, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\\n> #24 0x009a3d57 in PortalRunSelect (portal=portal@entry=0x2197f090, forward=<optimized out>, count=0, dest=0x2ab629b4) at pquery.c:941\\n> #25 0x009a39d6 in PortalRun (portal=0x2197f090, count=2147483647, isTopLevel=<optimized out>, run_once=<optimized out>, dest=0x2ab629b4, altdest=0x2ab629b4,\\n> completionTag=0xffbfdc70 \\\"\\\") at pquery.c:782\\n> #26 0x009a2b53 in exec_simple_query (query_string=query_string@entry=0x21873090 \\\"select count(*) from simple r join simple s using (id);\\\") at postgres.c:1145\\n> #27 0x009a04ec in PostgresMain (argc=1, argv=0x21954ed4, dbname=<optimized out>, username=0x21954d38 \\\"buildfarm\\\") at postgres.c:4052\\n>\\n> Worker 1 looks like this:\\n>\\n> (gdb) bt\\n> #0 _poll () at _poll.S:4\\n> #1 0x21701361 in __thr_poll (fds=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\\n> #2 0x215eaf3f in poll (pfd=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\\n> #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\\n> #4 WaitEventSetWait (set=0x2187cb10, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217738) at latch.c:1000\\n> #5 0x0099842b in ConditionVariableSleep (cv=0x2196b230, wait_event_info=134217738) at condition_variable.c:157\\n> #6 0x00977e12 in BarrierArriveAndWait (barrier=0x2196b218, wait_event_info=134217738) at barrier.c:191\\n> #7 0x00873441 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:328\\n> #8 ExecParallelHashJoin (pstate=0x2a8378d8) at nodeHashjoin.c:600\\n> #9 0x0086a509 in ExecProcNode (node=0x2a8378d8) at ../../../src/include/executor/executor.h:248\\n> #10 fetch_input_tuple (aggstate=aggstate@entry=0x2a8376c8) at nodeAgg.c:407\\n> #11 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\\n> #12 ExecAgg (pstate=0x2a8376c8) at nodeAgg.c:1561\\n> #13 0x0085956b in ExecProcNode (node=0x2a8376c8) at ../../../src/include/executor/executor.h:248\\n> #14 ExecutePlan (estate=0x2a837090, planstate=0x2a8376c8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>,\\n> numberTuples=<optimized out>, direction=<optimized out>, dest=0x219ac780, execute_once=<optimized out>) at execMain.c:1712\\n> #15 standard_ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\\n> #16 0x00859445 in ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\\n> #17 0x0085d6de in ParallelQueryMain (seg=0x2187c89c, toc=0x2196b000) at execParallel.c:1402\\n> #18 0x00750ee0 in ParallelWorkerMain (main_arg=1903254317) at parallel.c:1433\\n>\\n> Worker 2 looks like this:\\n>\\n> (gdb) bt\\n> #0 _poll () at _poll.S:4\\n> #1 0x21701361 in __thr_poll (fds=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\\n> #2 0x215eaf3f in poll (pfd=0x2187cb48, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\\n> #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\\n> #4 WaitEventSetWait (set=0x2187cb10, timeout=-1, occurred_events=<optimized out>, nevents=1, wait_event_info=134217737) at latch.c:1000\\n> #5 0x0099842b in ConditionVariableSleep (cv=0x2196b230, wait_event_info=134217737) at condition_variable.c:157\\n> #6 0x00977e12 in BarrierArriveAndWait (barrier=0x2196b218, wait_event_info=134217737) at barrier.c:191\\n> #7 0x0086ea44 in MultiExecParallelHash (node=<optimized out>) at nodeHash.c:312\\n> #8 MultiExecHash (node=0x2a837e88) at nodeHash.c:112\\n> #9 0x0085fa03 in MultiExecProcNode (node=0x2a837e88) at execProcnode.c:501\\n> #10 0x00872f80 in ExecHashJoinImpl (pstate=<optimized out>, parallel=true) at nodeHashjoin.c:290\\n> #11 ExecParallelHashJoin (pstate=0x2a8378d8) at nodeHashjoin.c:600\\n> #12 0x0086a509 in ExecProcNode (node=0x2a8378d8) at ../../../src/include/executor/executor.h:248\\n> #13 fetch_input_tuple (aggstate=aggstate@entry=0x2a8376c8) at nodeAgg.c:407\\n> #14 0x00869646 in agg_retrieve_direct (aggstate=<optimized out>) at nodeAgg.c:1746\\n> #15 ExecAgg (pstate=0x2a8376c8) at nodeAgg.c:1561\\n> #16 0x0085956b in ExecProcNode (node=0x2a8376c8) at ../../../src/include/executor/executor.h:248\\n> #17 ExecutePlan (estate=0x2a837090, planstate=0x2a8376c8, use_parallel_mode=<optimized out>, operation=CMD_SELECT, sendTuples=<optimized out>,\\n> numberTuples=<optimized out>, direction=<optimized out>, dest=0x219ac780, execute_once=<optimized out>) at execMain.c:1712\\n> #18 standard_ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:353\\n> #19 0x00859445 in ExecutorRun (queryDesc=0x219ad3e0, direction=ForwardScanDirection, count=0, execute_once=<optimized out>) at execMain.c:296\\n> #20 0x0085d6de in ParallelQueryMain (seg=0x2187c89c, toc=0x2196b000) at execParallel.c:1402\\n> #21 0x00750ee0 in ParallelWorkerMain (main_arg=1903254317) at parallel.c:1433\\n> #22 0x009193d3 in StartBackgroundWorker () at bgworker.c:911\\n>\\n> In each case I attached to the process, ran the gdb commands shown\\n> above, and detached. As I've grown used to seeing, the processes\\n> resumed running after I detached from the second worker. I've not\\n> tried attaching in something other than PID order, but I probably\\n> will next time.\\n>\\n> Thoughts? Ideas on debugging this?\\n\\nHow could it be that worker 2 is waiting on the build barrier in\\nPHJ_BUILD_HASHING_INNER and worker 1 and the leader are waiting on it\\nwith it supposedly in PHJ_BUILD_HASHING_OUTER?\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 17:18:18 -0500\",\n \"msg_from\": \"Melanie Plageman <melanieplageman@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Missed condition-variable wakeups on FreeBSD\"\n },\n {\n \"msg_contents\": \"Andres Freund <andres@anarazel.de> writes:\\n> On 2022-02-26 14:07:05 -0500, Tom Lane wrote:\\n>> I have observed this three times in the REL_11 branch, once\\n>> in REL_12, and a couple of times last summer before it occurred\\n>> to me to start keeping notes. Over that time the machine has\\n>> been running various patchlevels of FreeBSD 13.0.\\n\\n> It's certainly interesting that it appears to happen only in the branches\\n> using poll rather than kqueue to implement latches. That changed between 12\\n> and 13.\\n\\nYeah, and there was no PHJ in v10, so that's a pretty good theory as\\nto why I've only seen it in those two branches.\\n\\n> Have you tried running the core regression tests with force_parallel_mode =\\n> on, or with the parallel costs lowered, to see if that makes the problem\\n> appear more often?\\n> The next time this happens / if you still have this open, perhaps it could be\\n> worth checking if there's a byte in the self pipe?\\n> Besides trying to make the issue more likely as suggested above, it might be\\n> worth checking if signalling the stuck processes with SIGUSR1 gets them\\n> unstuck.\\n\\nI've now wasted a bunch of kilowatt-hours fruitlessly trying to\\nreproduce this outside the confines of the buildfarm script.\\nI'm at a loss to figure out what the buildfarm is doing differently,\\nbut apparently there's something. I'm going to re-enable the\\nmachine's buildfarm job and just wait for it to hang up again.\\nMore info eventually ...\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Feb 2022 17:24:48 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Missed condition-variable wakeups on FreeBSD\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 8:07 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> I don't know much about how gdb interacts with kernel calls on\\n> FreeBSD, but I speculate that the poll(2) call returns with EINTR\\n> after gdb releases the process, and then things resume fine,\\n\\nYeah, at least FreeBSD and macOS interrupt system calls when you send\\nSIGSTOP + SIGCONT directly, or when gdb and lldb do something similar\\nvia ptrace. The same applies on Linux, except that Linux restarts the\\nsystem call automatically (even though the man page has this in the\\nlist of system calls that never restart) so you don't usually notice\\n(though you can see it with strace). It's not really clear to me what\\nshould happen given the language around restarts is all about signal\\nhandlers, and stop + cont are system magic, not signal handlers.\\nAnyway...\\n\\n> suggesting that we lost an interrupt somewhere.\\n\\nSo it's happening on an i386 kernel, with WAIT_USE_POLL (not\\nWAIT_USE_KQUEUE), but only under the build farm script (not when you\\nran it manually in loop).... hmmm.\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 13:00:26 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Missed condition-variable wakeups on FreeBSD\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 11:18 AM Melanie Plageman\\n<melanieplageman@gmail.com> wrote:\\n> How could it be that worker 2 is waiting on the build barrier in\\n> PHJ_BUILD_HASHING_INNER and worker 1 and the leader are waiting on it\\n> with it supposedly in PHJ_BUILD_HASHING_OUTER?\\n\\nThat'd be consistent with a wakeup going missing, somehow. W2 was\\nsupposed to be released but somehow didn't get the message (until a\\nrandom debugger-induced EINTR releases it and it sees all the state it\\nneeds to proceed in shared memory), meanwhile the other two got to the\\nend of the next phase and are waiting for it.\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 00:02:08 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Missed condition-variable wakeups on FreeBSD\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 9:44 AM Andres Freund <andres@anarazel.de> wrote:\\n> > (gdb) p debug_query_string\\n> > $1 = 0x21873090 \\\"select count(*) from simple r join simple s using (id);\\\"\\n> > (gdb) bt\\n> > #0 _poll () at _poll.S:4\\n> > #1 0x21701361 in __thr_poll (fds=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libthr/thread/thr_syscalls.c:338\\n> > #2 0x215eaf3f in poll (pfd=0x219dc170, nfds=2, timeout=-1) at /usr/src/lib/libc/sys/poll.c:47\\n> > #3 0x0097b0fd in WaitEventSetWaitBlock (set=<optimized out>, cur_timeout=-1, occurred_events=<optimized out>, nevents=<optimized out>) at latch.c:1171\\n>\\n> The next time this happens / if you still have this open, perhaps it could be\\n> worth checking if there's a byte in the self pipe?\\n\\nAs a basic sanity check I'd like to see pfd[0], pfd[1] and the output\\nof fstat -p PID to see unconsumed bytes in the pipe. For example\\n\\\"echo hello | sleep 60\\\" shows \\\"0* pipe ... 6 rw\\\" for sleep, but \\\"echo\\nhello | more\\\" shows \\\"0* pipe ... 0 rw\\\".\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 10:17:26 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Missed condition-variable wakeups on FreeBSD\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 8:07 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> I have observed this three times in the REL_11 branch, once\\n> in REL_12, and a couple of times last summer before it occurred\\n> to me to start keeping notes. Over that time the machine has\\n> been running various patchlevels of FreeBSD 13.0.\\n\\nFTR I noticed that my animal elver hadn't reported for a while and\\nlogged and found it locked up, a bit like that, but in test_shm_mq\\nrather than PHJ. It's FreeBSD 13.1 on a fairly fast 16 way amd64 (in\\na jail), and it was testing REL_12_STABLE, so using poll() for\\nWaitEventSetWaitBlock(). When I connected a debugger then\\ndisconnected, it was unblocked and continued. Unfortunately I didn't\\nget any new information :-(\\n\\n\\n\",\n \"msg_date\": \"Fri, 13 Jan 2023 15:00:58 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Missed condition-variable wakeups on FreeBSD\"\n }\n]","isConversational":false}}},{"rowIdx":53589,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nIn two recent investigations in occasional test failures\n(019_replslot_limit.pl failures, AIO rebase) the problems are somehow tied to\ncheckpointer.\n\nI don't yet know if actually causally related to precisely those failures, but\nwhen running e.g. 027_stream_regress.pl, I see phases in which many backends\nare looping in RegisterSyncRequest() repeatedly, each time sleeping with\npg_usleep(10000L).\n\n\nWithout adding instrumentation this is completely invisible at any log\nlevel. There's no log messages, there's no wait events, nothing.\n\nISTM, we should not have any loops around pg_usleep(). And shorter term, we\nshouldn't have any loops around pg_usleep() that don't emit log messages / set\nwait events. Therefore I propose that we \"prohibit\" such loops without at\nleast a DEBUG2 elog() or so. It's just too hard to debug.\n\n\nThe reason for the sync queue filling up in 027_stream_regress.pl is actually\nfairly simple:\n\n1) The test runs with shared_buffers = 1MB, leading to a small sync queue of\n 128 entries.\n2) CheckpointWriteDelay() does pg_usleep(100000L)\n\nForwardSyncRequest() wakes up the checkpointer using SetLatch() if the sync\nqueue is more than half full.\n\nBut at least on linux and freebsd that doesn't actually interrupt pg_usleep()\nanymore (due to using signalfd / kqueue rather than a signal handler). And on\nall platforms the signal might arrive just before the pg_usleep() rather than\nduring, also not causing usleep to be interrupted.\n\nIf I shorten the sleep in CheckpointWriteDelay() the problem goes away. This\nactually reduces the time for a single run of 027_stream_regress.pl on my\nworkstation noticably. With default sleep time it's ~32s, with shortened time\nit's ~27s.\n\nI suspect we need to do something about this concrete problem for 14 and\nmaster, because it's certainly worse than before on linux / freebsd.\n\nI suspect the easiest is to just convert that usleep to a WaitLatch(). That'd\nrequire adding a new enum value to WaitEventTimeout in 14. Which probably is\nfine?\n\n\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Sat, 26 Feb 2022 13:39:42 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Checkpointer sync queue fills up / loops around pg_usleep() are bad"},{"msg_contents":"On Sat, Feb 26, 2022 at 01:39:42PM -0800, Andres Freund wrote:\n> I suspect the easiest is to just convert that usleep to a WaitLatch(). That'd\n> require adding a new enum value to WaitEventTimeout in 14. Which probably is\n> fine?\n\nWe've added wait events in back-branches in the past, so this does not\nstrike me as a problem as long as you add the new entry at the end of\nthe enum, while keeping things ordered on HEAD. In recent memory, I\nthink that only some of the extensions published by PostgresPro rely\non the enums in this area. \n--\nMichael","msg_date":"Sun, 27 Feb 2022 18:10:45 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\n bad"},{"msg_contents":"On Sun, Feb 27, 2022 at 06:10:45PM +0900, Michael Paquier wrote:\n> On Sat, Feb 26, 2022 at 01:39:42PM -0800, Andres Freund wrote:\n> > I suspect the easiest is to just convert that usleep to a WaitLatch(). That'd\n> > require adding a new enum value to WaitEventTimeout in 14. Which probably is\n> > fine?\n> \n> We've added wait events in back-branches in the past, so this does not\n> strike me as a problem as long as you add the new entry at the end of\n> the enum, while keeping things ordered on HEAD.\n\n+1\n\n> In recent memory, I\n> think that only some of the extensions published by PostgresPro rely\n> on the enums in this area. \n\nIndeed, I only know of pg_wait_sampling which uses it. Note that it relies on\npgstat_get_wait_event* functions, so it should only returns \"???\" / \"unknown\nwait event\" until you recompile it for a newer minor version and not report\nerrors or crash. All other extensions I know of simply use whatever\npg_stat_activity returns, so no impact.\n\n\n","msg_date":"Sun, 27 Feb 2022 17:29:19 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\n bad"},{"msg_contents":"On Sun, Feb 27, 2022 at 10:29 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\n> On Sun, Feb 27, 2022 at 06:10:45PM +0900, Michael Paquier wrote:\n> > On Sat, Feb 26, 2022 at 01:39:42PM -0800, Andres Freund wrote:\n> > > I suspect the easiest is to just convert that usleep to a WaitLatch(). That'd\n> > > require adding a new enum value to WaitEventTimeout in 14. Which probably is\n> > > fine?\n> >\n> > We've added wait events in back-branches in the past, so this does not\n> > strike me as a problem as long as you add the new entry at the end of\n> > the enum, while keeping things ordered on HEAD.\n>\n> +1\n\n+1\n\nSleeps like these are also really bad for ProcSignalBarrier, which I\nwas expecting to be the impetus for fixing any remaining loops like\nthis.\n\nWith the attached, 027_stream_regress.pl drops from ~29.5s to ~19.6s\non my FreeBSD workstation!\n\nIt seems a little strange to introduce a new wait event that will very\noften appear into a stable branch, but ... it is actually telling the\ntruth, so there is that.\n\nThe sleep/poll loop in RegisterSyncRequest() may also have another\nproblem. The comment explains that it was a deliberate choice not to\ndo CHECK_FOR_INTERRUPTS() here, which may be debatable, but I don't\nthink there's an excuse to ignore postmaster death in a loop that\npresumably becomes infinite if the checkpointer exits. I guess we\ncould do:\n\n- pg_usleep(10000L);\n+ WaitLatch(NULL, WL_EXIT_ON_PM_DEATH | WL_TIMEOUT, 10,\nWAIT_EVENT_SYNC_REQUEST);\n\nBut... really, this should be waiting on a condition variable that the\ncheckpointer broadcasts on when the queue goes from full to not full,\nno? Perhaps for master only?","msg_date":"Mon, 28 Feb 2022 13:19:21 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\n bad"},{"msg_contents":"Hi, \n\nOn February 27, 2022 4:19:21 PM PST, Thomas Munro <thomas.munro@gmail.com> wrote:\n>With the attached, 027_stream_regress.pl drops from ~29.5s to ~19.6s\n>on my FreeBSD workstation!\n\nThat's impressive - wouldn't have guessed it to make that much of a difference. I assume that running the tests on freebsd for an older pg with a similar s_b & max_wal_size doesn't benefit as much? I wonder how much windows will improve.\n\n\n>It seems a little strange to introduce a new wait event that will very\n>often appear into a stable branch, but ... it is actually telling the\n>truth, so there is that.\n\nIn the back branches it needs to be at the end of the enum - I assume you intended that just to be for HEAD.\n\nI wonder whether in HEAD we shouldn't make that sleep duration be computed from the calculation in IsOnSchedule...\n\n\n>The sleep/poll loop in RegisterSyncRequest() may also have another\n>problem. The comment explains that it was a deliberate choice not to\n>do CHECK_FOR_INTERRUPTS() here, which may be debatable, but I don't\n>think there's an excuse to ignore postmaster death in a loop that\n>presumably becomes infinite if the checkpointer exits. I guess we\n>could do:\n>\n>- pg_usleep(10000L);\n>+ WaitLatch(NULL, WL_EXIT_ON_PM_DEATH | WL_TIMEOUT, 10,\n>WAIT_EVENT_SYNC_REQUEST);\n>\n>But... really, this should be waiting on a condition variable that the\n>checkpointer broadcasts on when the queue goes from full to not full,\n>no? Perhaps for master only?\n\nLooks worth improving, but yes, I'd not do it in the back branches. \n\nI do think it's worth giving that sleep a proper wait event though, even in the back branches.\n\nAndres\n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.\n\n\n","msg_date":"Sun, 27 Feb 2022 17:36:20 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"=?US-ASCII?Q?Re=3A_Checkpointer_sync_queue_fills_u?=\n =?US-ASCII?Q?p_/_loops_around_pg=5Fusleep=28=29_are_bad?="},{"msg_contents":"On Mon, Feb 28, 2022 at 2:36 PM Andres Freund <andres@anarazel.de> wrote:\n> On February 27, 2022 4:19:21 PM PST, Thomas Munro <thomas.munro@gmail.com> wrote:\n> >It seems a little strange to introduce a new wait event that will very\n> >often appear into a stable branch, but ... it is actually telling the\n> >truth, so there is that.\n>\n> In the back branches it needs to be at the end of the enum - I assume you intended that just to be for HEAD.\n\nYeah.\n\n> I wonder whether in HEAD we shouldn't make that sleep duration be computed from the calculation in IsOnSchedule...\n\nI might look into this.\n\n> >The sleep/poll loop in RegisterSyncRequest() may also have another\n> >problem. The comment explains that it was a deliberate choice not to\n> >do CHECK_FOR_INTERRUPTS() here, which may be debatable, but I don't\n> >think there's an excuse to ignore postmaster death in a loop that\n> >presumably becomes infinite if the checkpointer exits. I guess we\n> >could do:\n> >\n> >- pg_usleep(10000L);\n> >+ WaitLatch(NULL, WL_EXIT_ON_PM_DEATH | WL_TIMEOUT, 10,\n> >WAIT_EVENT_SYNC_REQUEST);\n> >\n> >But... really, this should be waiting on a condition variable that the\n> >checkpointer broadcasts on when the queue goes from full to not full,\n> >no? Perhaps for master only?\n>\n> Looks worth improving, but yes, I'd not do it in the back branches.\n\n0003 is a first attempt at that, for master only (on top of 0002 which\nis the minimal fix). This shaves another second off\n027_stream_regress.pl on my workstation. The main thing I realised is\nthat I needed to hold interrupts while waiting, which seems like it\nshould go away with 'tombstone' files as discussed in other threads.\nThat's not a new problem in this patch, it just looks more offensive\nto the eye when you spell it out, instead of hiding it with an\nunreported sleep/poll loop...\n\n> I do think it's worth giving that sleep a proper wait event though, even in the back branches.\n\nI'm thinking that 0002 should be back-patched all the way, but 0001\ncould be limited to 14.","msg_date":"Wed, 2 Mar 2022 06:46:23 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\n bad"},{"msg_contents":"Hi,\n\nOn 2022-03-02 06:46:23 +1300, Thomas Munro wrote:\n> > I do think it's worth giving that sleep a proper wait event though, even in the back branches.\n> \n> I'm thinking that 0002 should be back-patched all the way, but 0001\n> could be limited to 14.\n\nNo strong opinion on back to where to backpatch. It'd be nice to have a proper\nwait event everywhere, but especially < 12 it looks different enough to be\nsome effort.\n\n\n> From a9344bb2fb2a363bec4be526f87560cb212ca10b Mon Sep 17 00:00:00 2001\n> From: Thomas Munro <thomas.munro@gmail.com>\n> Date: Mon, 28 Feb 2022 11:27:05 +1300\n> Subject: [PATCH v2 1/3] Wake up for latches in CheckpointWriteDelay().\n\nLGTM. Would be nice to have this fixed soon, even if it's just to reduce test\ntimes :)\n\n\n\n> From 1eb0266fed7ccb63a2430e4fbbaef2300f2aa0d0 Mon Sep 17 00:00:00 2001\n> From: Thomas Munro <thomas.munro@gmail.com>\n> Date: Tue, 1 Mar 2022 11:38:27 +1300\n> Subject: [PATCH v2 2/3] Fix waiting in RegisterSyncRequest().\n\nLGTM.\n\n\n> From 50060e5a0ed66762680ddee9e30acbad905c6e98 Mon Sep 17 00:00:00 2001\n> From: Thomas Munro <thomas.munro@gmail.com>\n> Date: Tue, 1 Mar 2022 17:34:43 +1300\n> Subject: [PATCH v2 3/3] Use condition variable to wait when sync request queue\n> is full.\n> \n> Previously, in the (hopefully) rare case that we need to wait for the\n> checkpointer to create space in the sync request queue, we'd enter a\n> sleep/retry loop. Instead, create a condition variable so the\n> checkpointer can wake us up whenever there is a transition from 'full'\n> to 'not full'.\n\n\n> @@ -1076,10 +1078,11 @@ RequestCheckpoint(int flags)\n> * to perform its own fsync, which is far more expensive in practice. It\n> * is theoretically possible a backend fsync might still be necessary, if\n> * the queue is full and contains no duplicate entries. In that case, we\n> - * let the backend know by returning false.\n> + * let the backend know by returning false, or if 'wait' is true, then we\n> + * wait for space to become available.\n> */\n> bool\n> -ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)\n> +ForwardSyncRequest(const FileTag *ftag, SyncRequestType type, bool wait)\n> {\n> \tCheckpointerRequest *request;\n> \tbool\t\ttoo_full;\n> @@ -1101,9 +1104,9 @@ ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)\n> \t * backend will have to perform its own fsync request. But before forcing\n> \t * that to happen, we can try to compact the request queue.\n> \t */\n> -\tif (CheckpointerShmem->checkpointer_pid == 0 ||\n> -\t\t(CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&\n> -\t\t !CompactCheckpointerRequestQueue()))\n> +\tif (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&\n> +\t\t!CompactCheckpointerRequestQueue() &&\n> +\t\t!wait)\n\nBit confused about the addition of the wait parameter / removal of the\nCheckpointerShmem->checkpointer_pid check. What's that about?\n\n\n> +\t/*\n> +\t * If we still don't have enough space and the caller asked us to wait,\n> +\t * wait for the checkpointer to advertise that there is space.\n> +\t */\n> +\tif (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests)\n> +\t{\n> +\t\tConditionVariablePrepareToSleep(&CheckpointerShmem->requests_not_full_cv);\n> +\t\twhile (CheckpointerShmem->num_requests >=\n> +\t\t\t CheckpointerShmem->max_requests)\n> +\t\t{\n> +\t\t\tLWLockRelease(CheckpointerCommLock);\n> +\t\t\tConditionVariableSleep(&CheckpointerShmem->requests_not_full_cv,\n> +\t\t\t\t\t\t\t\t WAIT_EVENT_FORWARD_SYNC_REQUEST);\n> +\t\t\tLWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);\n> +\t\t}\n> +\t\tConditionVariableCancelSleep();\n> +\t}\n\nCould there be a problem with a lot of backends trying to acquire\nCheckpointerCommLock in exclusive mode? I'm inclined to think it's rare enough\nto not worry.\n\nGreetings,\n\nAndres Freund\n\n\n","msg_date":"Tue, 1 Mar 2022 13:58:48 -0800","msg_from":"Andres Freund <andres@anarazel.de>","msg_from_op":true,"msg_subject":"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\n bad"},{"msg_contents":"On Wed, Mar 2, 2022 at 10:58 AM Andres Freund <andres@anarazel.de> wrote:\n> On 2022-03-02 06:46:23 +1300, Thomas Munro wrote:\n> > From a9344bb2fb2a363bec4be526f87560cb212ca10b Mon Sep 17 00:00:00 2001\n> > From: Thomas Munro <thomas.munro@gmail.com>\n> > Date: Mon, 28 Feb 2022 11:27:05 +1300\n> > Subject: [PATCH v2 1/3] Wake up for latches in CheckpointWriteDelay().\n>\n> LGTM. Would be nice to have this fixed soon, even if it's just to reduce test\n> times :)\n\nThanks for the review. Pushed to master and 14, with the wait event\nmoved to the end of the enum for the back-patch.\n\n> > From 1eb0266fed7ccb63a2430e4fbbaef2300f2aa0d0 Mon Sep 17 00:00:00 2001\n> > From: Thomas Munro <thomas.munro@gmail.com>\n> > Date: Tue, 1 Mar 2022 11:38:27 +1300\n> > Subject: [PATCH v2 2/3] Fix waiting in RegisterSyncRequest().\n>\n> LGTM.\n\nPushed as far back as 12. It could be done for 10 & 11, but indeed\nthe code starts getting quite different back there, and since there\nare no field reports, I think that's OK for now.\n\nA simple repro, for the record: run installcheck with\nshared_buffers=256kB, and then partway through, kill -STOP\n$checkpointer to simulate being stalled on IO for a while. Backends\nwill soon start waiting for the checkpointer to drain the queue while\ndropping relations. This state was invisible to pg_stat_activity, and\nhangs forever if you kill the postmaster and CONT the checkpointer.\n\n> > From 50060e5a0ed66762680ddee9e30acbad905c6e98 Mon Sep 17 00:00:00 2001\n> > From: Thomas Munro <thomas.munro@gmail.com>\n> > Date: Tue, 1 Mar 2022 17:34:43 +1300\n> > Subject: [PATCH v2 3/3] Use condition variable to wait when sync request queue\n> > is full.\n\n> [review]\n\nI'll come back to 0003 (condition variable-based improvement) a bit later.\n\n\n","msg_date":"Wed, 16 Mar 2022 16:04:30 +1300","msg_from":"Thomas Munro <thomas.munro@gmail.com>","msg_from_op":false,"msg_subject":"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\n bad"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nIn two recent investigations in occasional test failures\\n(019_replslot_limit.pl failures, AIO rebase) the problems are somehow tied to\\ncheckpointer.\\n\\nI don't yet know if actually causally related to precisely those failures, but\\nwhen running e.g. 027_stream_regress.pl, I see phases in which many backends\\nare looping in RegisterSyncRequest() repeatedly, each time sleeping with\\npg_usleep(10000L).\\n\\n\\nWithout adding instrumentation this is completely invisible at any log\\nlevel. There's no log messages, there's no wait events, nothing.\\n\\nISTM, we should not have any loops around pg_usleep(). And shorter term, we\\nshouldn't have any loops around pg_usleep() that don't emit log messages / set\\nwait events. Therefore I propose that we \\\"prohibit\\\" such loops without at\\nleast a DEBUG2 elog() or so. It's just too hard to debug.\\n\\n\\nThe reason for the sync queue filling up in 027_stream_regress.pl is actually\\nfairly simple:\\n\\n1) The test runs with shared_buffers = 1MB, leading to a small sync queue of\\n 128 entries.\\n2) CheckpointWriteDelay() does pg_usleep(100000L)\\n\\nForwardSyncRequest() wakes up the checkpointer using SetLatch() if the sync\\nqueue is more than half full.\\n\\nBut at least on linux and freebsd that doesn't actually interrupt pg_usleep()\\nanymore (due to using signalfd / kqueue rather than a signal handler). And on\\nall platforms the signal might arrive just before the pg_usleep() rather than\\nduring, also not causing usleep to be interrupted.\\n\\nIf I shorten the sleep in CheckpointWriteDelay() the problem goes away. This\\nactually reduces the time for a single run of 027_stream_regress.pl on my\\nworkstation noticably. With default sleep time it's ~32s, with shortened time\\nit's ~27s.\\n\\nI suspect we need to do something about this concrete problem for 14 and\\nmaster, because it's certainly worse than before on linux / freebsd.\\n\\nI suspect the easiest is to just convert that usleep to a WaitLatch(). That'd\\nrequire adding a new enum value to WaitEventTimeout in 14. Which probably is\\nfine?\\n\\n\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Sat, 26 Feb 2022 13:39:42 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Checkpointer sync queue fills up / loops around pg_usleep() are bad\"\n },\n {\n \"msg_contents\": \"On Sat, Feb 26, 2022 at 01:39:42PM -0800, Andres Freund wrote:\\n> I suspect the easiest is to just convert that usleep to a WaitLatch(). That'd\\n> require adding a new enum value to WaitEventTimeout in 14. Which probably is\\n> fine?\\n\\nWe've added wait events in back-branches in the past, so this does not\\nstrike me as a problem as long as you add the new entry at the end of\\nthe enum, while keeping things ordered on HEAD. In recent memory, I\\nthink that only some of the extensions published by PostgresPro rely\\non the enums in this area. \\n--\\nMichael\",\n \"msg_date\": \"Sun, 27 Feb 2022 18:10:45 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\\n bad\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 06:10:45PM +0900, Michael Paquier wrote:\\n> On Sat, Feb 26, 2022 at 01:39:42PM -0800, Andres Freund wrote:\\n> > I suspect the easiest is to just convert that usleep to a WaitLatch(). That'd\\n> > require adding a new enum value to WaitEventTimeout in 14. Which probably is\\n> > fine?\\n> \\n> We've added wait events in back-branches in the past, so this does not\\n> strike me as a problem as long as you add the new entry at the end of\\n> the enum, while keeping things ordered on HEAD.\\n\\n+1\\n\\n> In recent memory, I\\n> think that only some of the extensions published by PostgresPro rely\\n> on the enums in this area. \\n\\nIndeed, I only know of pg_wait_sampling which uses it. Note that it relies on\\npgstat_get_wait_event* functions, so it should only returns \\\"???\\\" / \\\"unknown\\nwait event\\\" until you recompile it for a newer minor version and not report\\nerrors or crash. All other extensions I know of simply use whatever\\npg_stat_activity returns, so no impact.\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Feb 2022 17:29:19 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\\n bad\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 10:29 PM Julien Rouhaud <rjuju123@gmail.com> wrote:\\n> On Sun, Feb 27, 2022 at 06:10:45PM +0900, Michael Paquier wrote:\\n> > On Sat, Feb 26, 2022 at 01:39:42PM -0800, Andres Freund wrote:\\n> > > I suspect the easiest is to just convert that usleep to a WaitLatch(). That'd\\n> > > require adding a new enum value to WaitEventTimeout in 14. Which probably is\\n> > > fine?\\n> >\\n> > We've added wait events in back-branches in the past, so this does not\\n> > strike me as a problem as long as you add the new entry at the end of\\n> > the enum, while keeping things ordered on HEAD.\\n>\\n> +1\\n\\n+1\\n\\nSleeps like these are also really bad for ProcSignalBarrier, which I\\nwas expecting to be the impetus for fixing any remaining loops like\\nthis.\\n\\nWith the attached, 027_stream_regress.pl drops from ~29.5s to ~19.6s\\non my FreeBSD workstation!\\n\\nIt seems a little strange to introduce a new wait event that will very\\noften appear into a stable branch, but ... it is actually telling the\\ntruth, so there is that.\\n\\nThe sleep/poll loop in RegisterSyncRequest() may also have another\\nproblem. The comment explains that it was a deliberate choice not to\\ndo CHECK_FOR_INTERRUPTS() here, which may be debatable, but I don't\\nthink there's an excuse to ignore postmaster death in a loop that\\npresumably becomes infinite if the checkpointer exits. I guess we\\ncould do:\\n\\n- pg_usleep(10000L);\\n+ WaitLatch(NULL, WL_EXIT_ON_PM_DEATH | WL_TIMEOUT, 10,\\nWAIT_EVENT_SYNC_REQUEST);\\n\\nBut... really, this should be waiting on a condition variable that the\\ncheckpointer broadcasts on when the queue goes from full to not full,\\nno? Perhaps for master only?\",\n \"msg_date\": \"Mon, 28 Feb 2022 13:19:21 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\\n bad\"\n },\n {\n \"msg_contents\": \"Hi, \\n\\nOn February 27, 2022 4:19:21 PM PST, Thomas Munro <thomas.munro@gmail.com> wrote:\\n>With the attached, 027_stream_regress.pl drops from ~29.5s to ~19.6s\\n>on my FreeBSD workstation!\\n\\nThat's impressive - wouldn't have guessed it to make that much of a difference. I assume that running the tests on freebsd for an older pg with a similar s_b & max_wal_size doesn't benefit as much? I wonder how much windows will improve.\\n\\n\\n>It seems a little strange to introduce a new wait event that will very\\n>often appear into a stable branch, but ... it is actually telling the\\n>truth, so there is that.\\n\\nIn the back branches it needs to be at the end of the enum - I assume you intended that just to be for HEAD.\\n\\nI wonder whether in HEAD we shouldn't make that sleep duration be computed from the calculation in IsOnSchedule...\\n\\n\\n>The sleep/poll loop in RegisterSyncRequest() may also have another\\n>problem. The comment explains that it was a deliberate choice not to\\n>do CHECK_FOR_INTERRUPTS() here, which may be debatable, but I don't\\n>think there's an excuse to ignore postmaster death in a loop that\\n>presumably becomes infinite if the checkpointer exits. I guess we\\n>could do:\\n>\\n>- pg_usleep(10000L);\\n>+ WaitLatch(NULL, WL_EXIT_ON_PM_DEATH | WL_TIMEOUT, 10,\\n>WAIT_EVENT_SYNC_REQUEST);\\n>\\n>But... really, this should be waiting on a condition variable that the\\n>checkpointer broadcasts on when the queue goes from full to not full,\\n>no? Perhaps for master only?\\n\\nLooks worth improving, but yes, I'd not do it in the back branches. \\n\\nI do think it's worth giving that sleep a proper wait event though, even in the back branches.\\n\\nAndres\\n-- \\nSent from my Android device with K-9 Mail. Please excuse my brevity.\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Feb 2022 17:36:20 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"=?US-ASCII?Q?Re=3A_Checkpointer_sync_queue_fills_u?=\\n =?US-ASCII?Q?p_/_loops_around_pg=5Fusleep=28=29_are_bad?=\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 2:36 PM Andres Freund <andres@anarazel.de> wrote:\\n> On February 27, 2022 4:19:21 PM PST, Thomas Munro <thomas.munro@gmail.com> wrote:\\n> >It seems a little strange to introduce a new wait event that will very\\n> >often appear into a stable branch, but ... it is actually telling the\\n> >truth, so there is that.\\n>\\n> In the back branches it needs to be at the end of the enum - I assume you intended that just to be for HEAD.\\n\\nYeah.\\n\\n> I wonder whether in HEAD we shouldn't make that sleep duration be computed from the calculation in IsOnSchedule...\\n\\nI might look into this.\\n\\n> >The sleep/poll loop in RegisterSyncRequest() may also have another\\n> >problem. The comment explains that it was a deliberate choice not to\\n> >do CHECK_FOR_INTERRUPTS() here, which may be debatable, but I don't\\n> >think there's an excuse to ignore postmaster death in a loop that\\n> >presumably becomes infinite if the checkpointer exits. I guess we\\n> >could do:\\n> >\\n> >- pg_usleep(10000L);\\n> >+ WaitLatch(NULL, WL_EXIT_ON_PM_DEATH | WL_TIMEOUT, 10,\\n> >WAIT_EVENT_SYNC_REQUEST);\\n> >\\n> >But... really, this should be waiting on a condition variable that the\\n> >checkpointer broadcasts on when the queue goes from full to not full,\\n> >no? Perhaps for master only?\\n>\\n> Looks worth improving, but yes, I'd not do it in the back branches.\\n\\n0003 is a first attempt at that, for master only (on top of 0002 which\\nis the minimal fix). This shaves another second off\\n027_stream_regress.pl on my workstation. The main thing I realised is\\nthat I needed to hold interrupts while waiting, which seems like it\\nshould go away with 'tombstone' files as discussed in other threads.\\nThat's not a new problem in this patch, it just looks more offensive\\nto the eye when you spell it out, instead of hiding it with an\\nunreported sleep/poll loop...\\n\\n> I do think it's worth giving that sleep a proper wait event though, even in the back branches.\\n\\nI'm thinking that 0002 should be back-patched all the way, but 0001\\ncould be limited to 14.\",\n \"msg_date\": \"Wed, 2 Mar 2022 06:46:23 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\\n bad\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn 2022-03-02 06:46:23 +1300, Thomas Munro wrote:\\n> > I do think it's worth giving that sleep a proper wait event though, even in the back branches.\\n> \\n> I'm thinking that 0002 should be back-patched all the way, but 0001\\n> could be limited to 14.\\n\\nNo strong opinion on back to where to backpatch. It'd be nice to have a proper\\nwait event everywhere, but especially < 12 it looks different enough to be\\nsome effort.\\n\\n\\n> From a9344bb2fb2a363bec4be526f87560cb212ca10b Mon Sep 17 00:00:00 2001\\n> From: Thomas Munro <thomas.munro@gmail.com>\\n> Date: Mon, 28 Feb 2022 11:27:05 +1300\\n> Subject: [PATCH v2 1/3] Wake up for latches in CheckpointWriteDelay().\\n\\nLGTM. Would be nice to have this fixed soon, even if it's just to reduce test\\ntimes :)\\n\\n\\n\\n> From 1eb0266fed7ccb63a2430e4fbbaef2300f2aa0d0 Mon Sep 17 00:00:00 2001\\n> From: Thomas Munro <thomas.munro@gmail.com>\\n> Date: Tue, 1 Mar 2022 11:38:27 +1300\\n> Subject: [PATCH v2 2/3] Fix waiting in RegisterSyncRequest().\\n\\nLGTM.\\n\\n\\n> From 50060e5a0ed66762680ddee9e30acbad905c6e98 Mon Sep 17 00:00:00 2001\\n> From: Thomas Munro <thomas.munro@gmail.com>\\n> Date: Tue, 1 Mar 2022 17:34:43 +1300\\n> Subject: [PATCH v2 3/3] Use condition variable to wait when sync request queue\\n> is full.\\n> \\n> Previously, in the (hopefully) rare case that we need to wait for the\\n> checkpointer to create space in the sync request queue, we'd enter a\\n> sleep/retry loop. Instead, create a condition variable so the\\n> checkpointer can wake us up whenever there is a transition from 'full'\\n> to 'not full'.\\n\\n\\n> @@ -1076,10 +1078,11 @@ RequestCheckpoint(int flags)\\n> * to perform its own fsync, which is far more expensive in practice. It\\n> * is theoretically possible a backend fsync might still be necessary, if\\n> * the queue is full and contains no duplicate entries. In that case, we\\n> - * let the backend know by returning false.\\n> + * let the backend know by returning false, or if 'wait' is true, then we\\n> + * wait for space to become available.\\n> */\\n> bool\\n> -ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)\\n> +ForwardSyncRequest(const FileTag *ftag, SyncRequestType type, bool wait)\\n> {\\n> \\tCheckpointerRequest *request;\\n> \\tbool\\t\\ttoo_full;\\n> @@ -1101,9 +1104,9 @@ ForwardSyncRequest(const FileTag *ftag, SyncRequestType type)\\n> \\t * backend will have to perform its own fsync request. But before forcing\\n> \\t * that to happen, we can try to compact the request queue.\\n> \\t */\\n> -\\tif (CheckpointerShmem->checkpointer_pid == 0 ||\\n> -\\t\\t(CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&\\n> -\\t\\t !CompactCheckpointerRequestQueue()))\\n> +\\tif (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests &&\\n> +\\t\\t!CompactCheckpointerRequestQueue() &&\\n> +\\t\\t!wait)\\n\\nBit confused about the addition of the wait parameter / removal of the\\nCheckpointerShmem->checkpointer_pid check. What's that about?\\n\\n\\n> +\\t/*\\n> +\\t * If we still don't have enough space and the caller asked us to wait,\\n> +\\t * wait for the checkpointer to advertise that there is space.\\n> +\\t */\\n> +\\tif (CheckpointerShmem->num_requests >= CheckpointerShmem->max_requests)\\n> +\\t{\\n> +\\t\\tConditionVariablePrepareToSleep(&CheckpointerShmem->requests_not_full_cv);\\n> +\\t\\twhile (CheckpointerShmem->num_requests >=\\n> +\\t\\t\\t CheckpointerShmem->max_requests)\\n> +\\t\\t{\\n> +\\t\\t\\tLWLockRelease(CheckpointerCommLock);\\n> +\\t\\t\\tConditionVariableSleep(&CheckpointerShmem->requests_not_full_cv,\\n> +\\t\\t\\t\\t\\t\\t\\t\\t WAIT_EVENT_FORWARD_SYNC_REQUEST);\\n> +\\t\\t\\tLWLockAcquire(CheckpointerCommLock, LW_EXCLUSIVE);\\n> +\\t\\t}\\n> +\\t\\tConditionVariableCancelSleep();\\n> +\\t}\\n\\nCould there be a problem with a lot of backends trying to acquire\\nCheckpointerCommLock in exclusive mode? I'm inclined to think it's rare enough\\nto not worry.\\n\\nGreetings,\\n\\nAndres Freund\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 13:58:48 -0800\",\n \"msg_from\": \"Andres Freund <andres@anarazel.de>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\\n bad\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 10:58 AM Andres Freund <andres@anarazel.de> wrote:\\n> On 2022-03-02 06:46:23 +1300, Thomas Munro wrote:\\n> > From a9344bb2fb2a363bec4be526f87560cb212ca10b Mon Sep 17 00:00:00 2001\\n> > From: Thomas Munro <thomas.munro@gmail.com>\\n> > Date: Mon, 28 Feb 2022 11:27:05 +1300\\n> > Subject: [PATCH v2 1/3] Wake up for latches in CheckpointWriteDelay().\\n>\\n> LGTM. Would be nice to have this fixed soon, even if it's just to reduce test\\n> times :)\\n\\nThanks for the review. Pushed to master and 14, with the wait event\\nmoved to the end of the enum for the back-patch.\\n\\n> > From 1eb0266fed7ccb63a2430e4fbbaef2300f2aa0d0 Mon Sep 17 00:00:00 2001\\n> > From: Thomas Munro <thomas.munro@gmail.com>\\n> > Date: Tue, 1 Mar 2022 11:38:27 +1300\\n> > Subject: [PATCH v2 2/3] Fix waiting in RegisterSyncRequest().\\n>\\n> LGTM.\\n\\nPushed as far back as 12. It could be done for 10 & 11, but indeed\\nthe code starts getting quite different back there, and since there\\nare no field reports, I think that's OK for now.\\n\\nA simple repro, for the record: run installcheck with\\nshared_buffers=256kB, and then partway through, kill -STOP\\n$checkpointer to simulate being stalled on IO for a while. Backends\\nwill soon start waiting for the checkpointer to drain the queue while\\ndropping relations. This state was invisible to pg_stat_activity, and\\nhangs forever if you kill the postmaster and CONT the checkpointer.\\n\\n> > From 50060e5a0ed66762680ddee9e30acbad905c6e98 Mon Sep 17 00:00:00 2001\\n> > From: Thomas Munro <thomas.munro@gmail.com>\\n> > Date: Tue, 1 Mar 2022 17:34:43 +1300\\n> > Subject: [PATCH v2 3/3] Use condition variable to wait when sync request queue\\n> > is full.\\n\\n> [review]\\n\\nI'll come back to 0003 (condition variable-based improvement) a bit later.\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 16:04:30 +1300\",\n \"msg_from\": \"Thomas Munro <thomas.munro@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Checkpointer sync queue fills up / loops around pg_usleep() are\\n bad\"\n }\n]","isConversational":false}}},{"rowIdx":53590,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nFirst time PostgreSQL contributor here :)\n\nI wanted to be able to allocate a bunch of numbers from a sequence at \nonce. Multiple people seem to be struggling with this \n(https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \nhttps://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\n\nI propose to add an extra argument to nextval() that specifies how many \nnumbers you want to allocate (default 1).\n\nThe attached patch (based on master) passes `./configure \n--enable-cassert --enable-debug && make && make check`, including the \nnewly added regression tests.\n\nIt does change the signature of nextval_internal(), not sure if that's \nconsidered backwards compatibility breaking (for extensions?).\n\n-- Jille","msg_date":"Sun, 27 Feb 2022 10:42:25 +0100","msg_from":"Jille Timmermans <jille@quis.cx>","msg_from_op":true,"msg_subject":"Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"Hi,\n\nOn Sun, Feb 27, 2022 at 10:42:25AM +0100, Jille Timmermans wrote:\n>\n> First time PostgreSQL contributor here :)\n\nWelcome!\n\n> I wanted to be able to allocate a bunch of numbers from a sequence at once.\n> Multiple people seem to be struggling with this (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence,\n> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\n>\n> I propose to add an extra argument to nextval() that specifies how many\n> numbers you want to allocate (default 1).\n>\n> The attached patch (based on master) passes `./configure --enable-cassert\n> --enable-debug && make && make check`, including the newly added regression\n> tests.\n>\n> It does change the signature of nextval_internal(), not sure if that's\n> considered backwards compatibility breaking (for extensions?).\n\nPlease register this patch to the next commit fest (and last for pg15\ninclusion) at https://commitfest.postgresql.org/37/ if not done already.\n\n\n","msg_date":"Sun, 27 Feb 2022 21:22:50 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On 2022-02-27 14:22, Julien Rouhaud wrote:\n> Hi,\n> \n> On Sun, Feb 27, 2022 at 10:42:25AM +0100, Jille Timmermans wrote:\n>> \n>> First time PostgreSQL contributor here :)\n> \n> Welcome!\nThanks!\n\n> \n>> I wanted to be able to allocate a bunch of numbers from a sequence at \n>> once.\n>> Multiple people seem to be struggling with this \n>> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence,\n>> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\n>> \n>> I propose to add an extra argument to nextval() that specifies how \n>> many\n>> numbers you want to allocate (default 1).\n>> \n>> The attached patch (based on master) passes `./configure \n>> --enable-cassert\n>> --enable-debug && make && make check`, including the newly added \n>> regression\n>> tests.\n>> \n>> It does change the signature of nextval_internal(), not sure if that's\n>> considered backwards compatibility breaking (for extensions?).\n> \n> Please register this patch to the next commit fest (and last for pg15\n> inclusion) at https://commitfest.postgresql.org/37/ if not done \n> already.\nDone: https://commitfest.postgresql.org/37/3577/ (I was waiting for \nmailman approval before I got the thread id.)\n\n\n","msg_date":"Sun, 27 Feb 2022 15:27:38 +0100","msg_from":"Jille Timmermans <jille@quis.cx>","msg_from_op":true,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On 27.02.22 10:42, Jille Timmermans wrote:\n> I wanted to be able to allocate a bunch of numbers from a sequence at \n> once. Multiple people seem to be struggling with this \n> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \n> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\n> \n> I propose to add an extra argument to nextval() that specifies how many \n> numbers you want to allocate (default 1).\n\nWhat is the use of this?\n\nI note that the stackoverflow question wanted to return multiple \nsequence values as a result set, whereas your implementation just skips \na number of values and returns the last one. At least we should be \nclear about what we are trying to achieve.\n\n\n\n","msg_date":"Mon, 28 Feb 2022 11:13:23 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On 2022-02-28 11:13, Peter Eisentraut wrote:\n> On 27.02.22 10:42, Jille Timmermans wrote:\n>> I wanted to be able to allocate a bunch of numbers from a sequence at \n>> once. Multiple people seem to be struggling with this \n>> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \n>> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\n>> \n>> I propose to add an extra argument to nextval() that specifies how \n>> many numbers you want to allocate (default 1).\n> \n> What is the use of this?\n> \n> I note that the stackoverflow question wanted to return multiple\n> sequence values as a result set, whereas your implementation just\n> skips a number of values and returns the last one. At least we should\n> be clear about what we are trying to achieve.\nBoth would work for me actually. I'm using COPY FROM to insert many rows \nand need to know their ids and COPY FROM doesn't support RETURNING.\n\nI implemented this approach because:\n- smaller diff\n- maybe someone benefits from them being consecutive\n- less data to send between client/server\n\nThe obvious downside is that people can make mistakes in whether the \nreturned number is the first or last number of the series.\n\n\n","msg_date":"Wed, 02 Mar 2022 20:12:20 +0100","msg_from":"Jille Timmermans <jille@quis.cx>","msg_from_op":true,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On 02.03.22 20:12, Jille Timmermans wrote:\n> On 2022-02-28 11:13, Peter Eisentraut wrote:\n>> On 27.02.22 10:42, Jille Timmermans wrote:\n>>> I wanted to be able to allocate a bunch of numbers from a sequence at \n>>> once. Multiple people seem to be struggling with this \n>>> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \n>>> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/). \n>>>\n>>>\n>>> I propose to add an extra argument to nextval() that specifies how \n>>> many numbers you want to allocate (default 1).\n>>\n>> What is the use of this?\n>>\n>> I note that the stackoverflow question wanted to return multiple\n>> sequence values as a result set, whereas your implementation just\n>> skips a number of values and returns the last one. At least we should\n>> be clear about what we are trying to achieve.\n> Both would work for me actually. I'm using COPY FROM to insert many rows \n> and need to know their ids and COPY FROM doesn't support RETURNING.\n\nI don't understand this use case. COPY FROM copies from a file. So you \nwant to preallocate the sequence numbers before you copy the new data \nin? How do you know how many rows are in the file?\n\n\n","msg_date":"Thu, 3 Mar 2022 10:10:34 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On 2022-03-03 10:10, Peter Eisentraut wrote:\n> On 02.03.22 20:12, Jille Timmermans wrote:\n>> On 2022-02-28 11:13, Peter Eisentraut wrote:\n>>> On 27.02.22 10:42, Jille Timmermans wrote:\n>>>> I wanted to be able to allocate a bunch of numbers from a sequence \n>>>> at once. Multiple people seem to be struggling with this \n>>>> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \n>>>> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/). \n>>>> I propose to add an extra argument to nextval() that specifies how \n>>>> many numbers you want to allocate (default 1).\n>>> \n>>> What is the use of this?\n>>> \n>>> I note that the stackoverflow question wanted to return multiple\n>>> sequence values as a result set, whereas your implementation just\n>>> skips a number of values and returns the last one. At least we \n>>> should\n>>> be clear about what we are trying to achieve.\n>> Both would work for me actually. I'm using COPY FROM to insert many \n>> rows and need to know their ids and COPY FROM doesn't support \n>> RETURNING.\n> \n> I don't understand this use case. COPY FROM copies from a file. So\n> you want to preallocate the sequence numbers before you copy the new\n> data in?\nYes\n\n> How do you know how many rows are in the file?\nI'm using https://pkg.go.dev/github.com/jackc/pgx/v4#Conn.CopyFrom, \nwhich uses the COPY FROM protocol but doesn't actually have to originate \nfrom a file.\n\n\n","msg_date":"Thu, 03 Mar 2022 10:21:05 +0100","msg_from":"Jille Timmermans <jille@quis.cx>","msg_from_op":true,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On Sun, 27 Feb 2022 at 07:09, Jille Timmermans <jille@quis.cx> wrote:\n>\n> Hi,\n>\n> First time PostgreSQL contributor here :)\n\nI wish I had noticed this patch during the CF. It seems like a nice\nself-contained feature that could have been easily reviewed and\ncommitted and it's always good to see first-time contributions.\nHopefully it'll get committed early in the next cycle.\n\n\n-- \ngreg\n\n\n","msg_date":"Fri, 8 Apr 2022 09:33:14 -0400","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On 2022-04-08 15:33, Greg Stark wrote:\n> On Sun, 27 Feb 2022 at 07:09, Jille Timmermans <jille@quis.cx> wrote:\n>> First time PostgreSQL contributor here :)\n> \n> I wish I had noticed this patch during the CF. It seems like a nice\n> self-contained feature that could have been easily reviewed and\n> committed and it's always good to see first-time contributions.\n> Hopefully it'll get committed early in the next cycle.\n\nIf anyone is looking for a small patch to review, here's one for you :)\n\n(https://commitfest.postgresql.org/38/3577/)\n\n\n","msg_date":"Fri, 01 Jul 2022 10:21:26 +0200","msg_from":"Jille Timmermans <jille@quis.cx>","msg_from_op":true,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"Hello,\n\nReading the thread, I think the feature has value: it would basically transfer \ncontrol of the sequence cache to the client application.\n\nHowever, I don't think that returning only the last value is a sensible thing \nto do. The client will need to know the details of the sequence to do anything \nuseful about this, especially it's increment, minvalue, maxvalue and cycle \noptions. \n\nAs suggested upthread, returning a resultset would probably be better. If the \nclient application is concerned about the volume of data exchanged with the \nserver, and is willing to deal with handling the knowledge of the sequence \ndetails themselves, they can always wrap it in an aggregate:\n\nSELECT min(newvals), max(newvals) FROM nextvals(<num_nextvals>) as newvals\n\nRegards,\n\n--\nRonan Dunklau\n\n\n\n\n","msg_date":"Fri, 08 Jul 2022 10:14:06 +0200","msg_from":"Ronan Dunklau <ronan.dunklau@aiven.io>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On Thu, Mar 03, 2022 at 10:21:05AM +0100, Jille Timmermans wrote:\n> I'm using https://pkg.go.dev/github.com/jackc/pgx/v4#Conn.CopyFrom, which\n> uses the COPY FROM protocol but doesn't actually have to originate from a\n> file.\n\nIt is Friday here, so I would easily miss something.. It is possible\nto use COPY FROM with a list of columns, so assuming that you could\nuse a default expression with nextval() or just a SERIAL column not\nlisted in the COPY FROM query to do the job, what do we gain with this\nfeature? In which aspect does the preallocation of a range handled\non the client side after being allocated in the backend make things\nbetter?\n--\nMichael","msg_date":"Fri, 8 Jul 2022 19:27:27 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"> It is Friday here, so I would easily miss something.. It is possible\n> to use COPY FROM with a list of columns, so assuming that you could\n> use a default expression with nextval() or just a SERIAL column not\n> listed in the COPY FROM query to do the job, what do we gain with this\n> feature? In which aspect does the preallocation of a range handled\n> on the client side after being allocated in the backend make things\n> better?\n\nThe problem the author wants to solve is the fact they don't have a way of \nreturning the ids when using COPY FROM. Pre-allocating them and assigning them \nto the individual records before sending them via COPY FROM would solve that \nfor them.\n\n-- \nRonan Dunklau\n\n\n\n\n","msg_date":"Wed, 13 Jul 2022 15:47:13 +0200","msg_from":"Ronan Dunklau <ronan.dunklau@aiven.io>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"Ronan Dunklau <ronan.dunklau@aiven.io> writes:\n> The problem the author wants to solve is the fact they don't have a way of \n> returning the ids when using COPY FROM. Pre-allocating them and assigning them \n> to the individual records before sending them via COPY FROM would solve that \n> for them.\n\nTrue.\n\nI took a quick look at this patch and am not pleased at all with the\nimplementation. That loop in nextval_internal is okay performance-wise\ntoday, since there's a small upper bound on the number of times it will\niterate. But with this patch, a user can trivially lock up a backend for\nup to 2^63 iterations; let's just say that's longer than you want to wait.\nThere's not even a CHECK_FOR_INTERRUPTS() in the loop :-(. Even without\nmistakes or deliberate DoS attempts, looping means holding the sequence\nlock for longer than we really want to.\n\nI think to seriously consider a feature like this, nextval_internal\nwould have to be rewritten so that it can advance the counter the\ncorrect number of steps without using a loop. That would be quite a\nheadache, once you've dealt with integer overflow, cyclic sequences,\nand suchlike complications, but it's probably do-able.\n\nI don't think I agree with Ronan's upthread comment that\n\n>> However, I don't think that returning only the last value is a sensible thing \n>> to do. The client will need to know the details of the sequence to do anything \n>> useful about this, especially it's increment, minvalue, maxvalue and cycle \n>> options. \n\nMost applications are probably quite happy to assume that they know the\nsequence's static parameters, and those that aren't can easily fetch them\nusing existing facilities. So I don't think that returning them in this\nfunction's result is really necessary.\n\nI've got no strong opinion about this bit:\n\n> As suggested upthread, returning a resultset would probably be better.\n\nThere are use-cases for that, sure, but there are also use-cases for\nreturning just the first or just the last value --- I'd think \"just the\nfirst\" is the more common need of those two. Aggregating over a resultset\nis a remarkably inefficient answer when that's what you want.\n\nIn any case, \"nextval()\" is an increasingly poor name for these different\ndefinitions, so I counsel picking some other name instead of overloading\nnextval(). \"nextvals()\" would be a pretty good choice for the resultset\ncase, I think.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 28 Jul 2022 12:20:56 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"I wrote:\n> I've got no strong opinion about this bit:\n>> As suggested upthread, returning a resultset would probably be better.\n\nActually, on further thought, I do like the resultset idea, because\nit'd remove the need for a complex rewrite of nextval_internal.\nAssuming the SRF is written in ValuePerCall style, each iteration\ncan just call nextval_internal with no modifications needed in that\nfunction. There'll be a CHECK_FOR_INTERRUPTS somewhere in the\nquery-level loop, or at least it's not nextval's fault if there's not.\nThe situation is then no different from generate_series with a large\nloop count, or any other query that can generate lots of data.\n\nOf course, this does imply a lot more cycles expended per generated value\n--- but most of that is inherent in the larger amount of data being\nhanded back.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Thu, 28 Jul 2022 12:47:10 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On Thu, Jul 28, 2022 at 12:47:10PM -0400, Tom Lane wrote:\n> Actually, on further thought, I do like the resultset idea, because\n> it'd remove the need for a complex rewrite of nextval_internal.\n> Assuming the SRF is written in ValuePerCall style, each iteration\n> can just call nextval_internal with no modifications needed in that\n> function. There'll be a CHECK_FOR_INTERRUPTS somewhere in the\n> query-level loop, or at least it's not nextval's fault if there's not.\n> The situation is then no different from generate_series with a large\n> loop count, or any other query that can generate lots of data.\n> \n> Of course, this does imply a lot more cycles expended per generated value\n> --- but most of that is inherent in the larger amount of data being\n> handed back.\n\nFWIW, I find the result set approach more intuitive and robust,\nparticularly in the case of a sequence has non-default values\nINCREMENT and min/max values. This reduces the dependency of what an\napplication needs to know about the details of a given sequence. With\nonly the last value reported, the application would need to compile\nthings by itself.\n--\nMichael","msg_date":"Sat, 30 Jul 2022 16:21:07 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"},{"msg_contents":"On Sat, Jul 30, 2022 at 04:21:07PM +0900, Michael Paquier wrote:\n> FWIW, I find the result set approach more intuitive and robust,\n> particularly in the case of a sequence has non-default values\n> INCREMENT and min/max values. This reduces the dependency of what an\n> application needs to know about the details of a given sequence. With\n> only the last value reported, the application would need to compile\n> things by itself.\n\nIt seems like there is a consensus here, but the thread has no\nactivity for the past two months, so I have marked the patch as\nreturned with feedback for now.\n--\nMichael","msg_date":"Wed, 12 Oct 2022 14:38:12 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Support for grabbing multiple consecutive values with nextval()"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nFirst time PostgreSQL contributor here :)\\n\\nI wanted to be able to allocate a bunch of numbers from a sequence at \\nonce. Multiple people seem to be struggling with this \\n(https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \\nhttps://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\\n\\nI propose to add an extra argument to nextval() that specifies how many \\nnumbers you want to allocate (default 1).\\n\\nThe attached patch (based on master) passes `./configure \\n--enable-cassert --enable-debug && make && make check`, including the \\nnewly added regression tests.\\n\\nIt does change the signature of nextval_internal(), not sure if that's \\nconsidered backwards compatibility breaking (for extensions?).\\n\\n-- Jille\",\n \"msg_date\": \"Sun, 27 Feb 2022 10:42:25 +0100\",\n \"msg_from\": \"Jille Timmermans <jille@quis.cx>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Sun, Feb 27, 2022 at 10:42:25AM +0100, Jille Timmermans wrote:\\n>\\n> First time PostgreSQL contributor here :)\\n\\nWelcome!\\n\\n> I wanted to be able to allocate a bunch of numbers from a sequence at once.\\n> Multiple people seem to be struggling with this (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence,\\n> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\\n>\\n> I propose to add an extra argument to nextval() that specifies how many\\n> numbers you want to allocate (default 1).\\n>\\n> The attached patch (based on master) passes `./configure --enable-cassert\\n> --enable-debug && make && make check`, including the newly added regression\\n> tests.\\n>\\n> It does change the signature of nextval_internal(), not sure if that's\\n> considered backwards compatibility breaking (for extensions?).\\n\\nPlease register this patch to the next commit fest (and last for pg15\\ninclusion) at https://commitfest.postgresql.org/37/ if not done already.\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Feb 2022 21:22:50 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On 2022-02-27 14:22, Julien Rouhaud wrote:\\n> Hi,\\n> \\n> On Sun, Feb 27, 2022 at 10:42:25AM +0100, Jille Timmermans wrote:\\n>> \\n>> First time PostgreSQL contributor here :)\\n> \\n> Welcome!\\nThanks!\\n\\n> \\n>> I wanted to be able to allocate a bunch of numbers from a sequence at \\n>> once.\\n>> Multiple people seem to be struggling with this \\n>> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence,\\n>> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\\n>> \\n>> I propose to add an extra argument to nextval() that specifies how \\n>> many\\n>> numbers you want to allocate (default 1).\\n>> \\n>> The attached patch (based on master) passes `./configure \\n>> --enable-cassert\\n>> --enable-debug && make && make check`, including the newly added \\n>> regression\\n>> tests.\\n>> \\n>> It does change the signature of nextval_internal(), not sure if that's\\n>> considered backwards compatibility breaking (for extensions?).\\n> \\n> Please register this patch to the next commit fest (and last for pg15\\n> inclusion) at https://commitfest.postgresql.org/37/ if not done \\n> already.\\nDone: https://commitfest.postgresql.org/37/3577/ (I was waiting for \\nmailman approval before I got the thread id.)\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Feb 2022 15:27:38 +0100\",\n \"msg_from\": \"Jille Timmermans <jille@quis.cx>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On 27.02.22 10:42, Jille Timmermans wrote:\\n> I wanted to be able to allocate a bunch of numbers from a sequence at \\n> once. Multiple people seem to be struggling with this \\n> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \\n> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\\n> \\n> I propose to add an extra argument to nextval() that specifies how many \\n> numbers you want to allocate (default 1).\\n\\nWhat is the use of this?\\n\\nI note that the stackoverflow question wanted to return multiple \\nsequence values as a result set, whereas your implementation just skips \\na number of values and returns the last one. At least we should be \\nclear about what we are trying to achieve.\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 11:13:23 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On 2022-02-28 11:13, Peter Eisentraut wrote:\\n> On 27.02.22 10:42, Jille Timmermans wrote:\\n>> I wanted to be able to allocate a bunch of numbers from a sequence at \\n>> once. Multiple people seem to be struggling with this \\n>> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \\n>> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/).\\n>> \\n>> I propose to add an extra argument to nextval() that specifies how \\n>> many numbers you want to allocate (default 1).\\n> \\n> What is the use of this?\\n> \\n> I note that the stackoverflow question wanted to return multiple\\n> sequence values as a result set, whereas your implementation just\\n> skips a number of values and returns the last one. At least we should\\n> be clear about what we are trying to achieve.\\nBoth would work for me actually. I'm using COPY FROM to insert many rows \\nand need to know their ids and COPY FROM doesn't support RETURNING.\\n\\nI implemented this approach because:\\n- smaller diff\\n- maybe someone benefits from them being consecutive\\n- less data to send between client/server\\n\\nThe obvious downside is that people can make mistakes in whether the \\nreturned number is the first or last number of the series.\\n\\n\\n\",\n \"msg_date\": \"Wed, 02 Mar 2022 20:12:20 +0100\",\n \"msg_from\": \"Jille Timmermans <jille@quis.cx>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On 02.03.22 20:12, Jille Timmermans wrote:\\n> On 2022-02-28 11:13, Peter Eisentraut wrote:\\n>> On 27.02.22 10:42, Jille Timmermans wrote:\\n>>> I wanted to be able to allocate a bunch of numbers from a sequence at \\n>>> once. Multiple people seem to be struggling with this \\n>>> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \\n>>> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/). \\n>>>\\n>>>\\n>>> I propose to add an extra argument to nextval() that specifies how \\n>>> many numbers you want to allocate (default 1).\\n>>\\n>> What is the use of this?\\n>>\\n>> I note that the stackoverflow question wanted to return multiple\\n>> sequence values as a result set, whereas your implementation just\\n>> skips a number of values and returns the last one. At least we should\\n>> be clear about what we are trying to achieve.\\n> Both would work for me actually. I'm using COPY FROM to insert many rows \\n> and need to know their ids and COPY FROM doesn't support RETURNING.\\n\\nI don't understand this use case. COPY FROM copies from a file. So you \\nwant to preallocate the sequence numbers before you copy the new data \\nin? How do you know how many rows are in the file?\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Mar 2022 10:10:34 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On 2022-03-03 10:10, Peter Eisentraut wrote:\\n> On 02.03.22 20:12, Jille Timmermans wrote:\\n>> On 2022-02-28 11:13, Peter Eisentraut wrote:\\n>>> On 27.02.22 10:42, Jille Timmermans wrote:\\n>>>> I wanted to be able to allocate a bunch of numbers from a sequence \\n>>>> at once. Multiple people seem to be struggling with this \\n>>>> (https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \\n>>>> https://www.depesz.com/2008/03/20/getting-multiple-values-from-sequences/). \\n>>>> I propose to add an extra argument to nextval() that specifies how \\n>>>> many numbers you want to allocate (default 1).\\n>>> \\n>>> What is the use of this?\\n>>> \\n>>> I note that the stackoverflow question wanted to return multiple\\n>>> sequence values as a result set, whereas your implementation just\\n>>> skips a number of values and returns the last one. At least we \\n>>> should\\n>>> be clear about what we are trying to achieve.\\n>> Both would work for me actually. I'm using COPY FROM to insert many \\n>> rows and need to know their ids and COPY FROM doesn't support \\n>> RETURNING.\\n> \\n> I don't understand this use case. COPY FROM copies from a file. So\\n> you want to preallocate the sequence numbers before you copy the new\\n> data in?\\nYes\\n\\n> How do you know how many rows are in the file?\\nI'm using https://pkg.go.dev/github.com/jackc/pgx/v4#Conn.CopyFrom, \\nwhich uses the COPY FROM protocol but doesn't actually have to originate \\nfrom a file.\\n\\n\\n\",\n \"msg_date\": \"Thu, 03 Mar 2022 10:21:05 +0100\",\n \"msg_from\": \"Jille Timmermans <jille@quis.cx>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On Sun, 27 Feb 2022 at 07:09, Jille Timmermans <jille@quis.cx> wrote:\\n>\\n> Hi,\\n>\\n> First time PostgreSQL contributor here :)\\n\\nI wish I had noticed this patch during the CF. It seems like a nice\\nself-contained feature that could have been easily reviewed and\\ncommitted and it's always good to see first-time contributions.\\nHopefully it'll get committed early in the next cycle.\\n\\n\\n-- \\ngreg\\n\\n\\n\",\n \"msg_date\": \"Fri, 8 Apr 2022 09:33:14 -0400\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On 2022-04-08 15:33, Greg Stark wrote:\\n> On Sun, 27 Feb 2022 at 07:09, Jille Timmermans <jille@quis.cx> wrote:\\n>> First time PostgreSQL contributor here :)\\n> \\n> I wish I had noticed this patch during the CF. It seems like a nice\\n> self-contained feature that could have been easily reviewed and\\n> committed and it's always good to see first-time contributions.\\n> Hopefully it'll get committed early in the next cycle.\\n\\nIf anyone is looking for a small patch to review, here's one for you :)\\n\\n(https://commitfest.postgresql.org/38/3577/)\\n\\n\\n\",\n \"msg_date\": \"Fri, 01 Jul 2022 10:21:26 +0200\",\n \"msg_from\": \"Jille Timmermans <jille@quis.cx>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"Hello,\\n\\nReading the thread, I think the feature has value: it would basically transfer \\ncontrol of the sequence cache to the client application.\\n\\nHowever, I don't think that returning only the last value is a sensible thing \\nto do. The client will need to know the details of the sequence to do anything \\nuseful about this, especially it's increment, minvalue, maxvalue and cycle \\noptions. \\n\\nAs suggested upthread, returning a resultset would probably be better. If the \\nclient application is concerned about the volume of data exchanged with the \\nserver, and is willing to deal with handling the knowledge of the sequence \\ndetails themselves, they can always wrap it in an aggregate:\\n\\nSELECT min(newvals), max(newvals) FROM nextvals(<num_nextvals>) as newvals\\n\\nRegards,\\n\\n--\\nRonan Dunklau\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Fri, 08 Jul 2022 10:14:06 +0200\",\n \"msg_from\": \"Ronan Dunklau <ronan.dunklau@aiven.io>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 03, 2022 at 10:21:05AM +0100, Jille Timmermans wrote:\\n> I'm using https://pkg.go.dev/github.com/jackc/pgx/v4#Conn.CopyFrom, which\\n> uses the COPY FROM protocol but doesn't actually have to originate from a\\n> file.\\n\\nIt is Friday here, so I would easily miss something.. It is possible\\nto use COPY FROM with a list of columns, so assuming that you could\\nuse a default expression with nextval() or just a SERIAL column not\\nlisted in the COPY FROM query to do the job, what do we gain with this\\nfeature? In which aspect does the preallocation of a range handled\\non the client side after being allocated in the backend make things\\nbetter?\\n--\\nMichael\",\n \"msg_date\": \"Fri, 8 Jul 2022 19:27:27 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"> It is Friday here, so I would easily miss something.. It is possible\\n> to use COPY FROM with a list of columns, so assuming that you could\\n> use a default expression with nextval() or just a SERIAL column not\\n> listed in the COPY FROM query to do the job, what do we gain with this\\n> feature? In which aspect does the preallocation of a range handled\\n> on the client side after being allocated in the backend make things\\n> better?\\n\\nThe problem the author wants to solve is the fact they don't have a way of \\nreturning the ids when using COPY FROM. Pre-allocating them and assigning them \\nto the individual records before sending them via COPY FROM would solve that \\nfor them.\\n\\n-- \\nRonan Dunklau\\n\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 13 Jul 2022 15:47:13 +0200\",\n \"msg_from\": \"Ronan Dunklau <ronan.dunklau@aiven.io>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"Ronan Dunklau <ronan.dunklau@aiven.io> writes:\\n> The problem the author wants to solve is the fact they don't have a way of \\n> returning the ids when using COPY FROM. Pre-allocating them and assigning them \\n> to the individual records before sending them via COPY FROM would solve that \\n> for them.\\n\\nTrue.\\n\\nI took a quick look at this patch and am not pleased at all with the\\nimplementation. That loop in nextval_internal is okay performance-wise\\ntoday, since there's a small upper bound on the number of times it will\\niterate. But with this patch, a user can trivially lock up a backend for\\nup to 2^63 iterations; let's just say that's longer than you want to wait.\\nThere's not even a CHECK_FOR_INTERRUPTS() in the loop :-(. Even without\\nmistakes or deliberate DoS attempts, looping means holding the sequence\\nlock for longer than we really want to.\\n\\nI think to seriously consider a feature like this, nextval_internal\\nwould have to be rewritten so that it can advance the counter the\\ncorrect number of steps without using a loop. That would be quite a\\nheadache, once you've dealt with integer overflow, cyclic sequences,\\nand suchlike complications, but it's probably do-able.\\n\\nI don't think I agree with Ronan's upthread comment that\\n\\n>> However, I don't think that returning only the last value is a sensible thing \\n>> to do. The client will need to know the details of the sequence to do anything \\n>> useful about this, especially it's increment, minvalue, maxvalue and cycle \\n>> options. \\n\\nMost applications are probably quite happy to assume that they know the\\nsequence's static parameters, and those that aren't can easily fetch them\\nusing existing facilities. So I don't think that returning them in this\\nfunction's result is really necessary.\\n\\nI've got no strong opinion about this bit:\\n\\n> As suggested upthread, returning a resultset would probably be better.\\n\\nThere are use-cases for that, sure, but there are also use-cases for\\nreturning just the first or just the last value --- I'd think \\\"just the\\nfirst\\\" is the more common need of those two. Aggregating over a resultset\\nis a remarkably inefficient answer when that's what you want.\\n\\nIn any case, \\\"nextval()\\\" is an increasingly poor name for these different\\ndefinitions, so I counsel picking some other name instead of overloading\\nnextval(). \\\"nextvals()\\\" would be a pretty good choice for the resultset\\ncase, I think.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 28 Jul 2022 12:20:56 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"I wrote:\\n> I've got no strong opinion about this bit:\\n>> As suggested upthread, returning a resultset would probably be better.\\n\\nActually, on further thought, I do like the resultset idea, because\\nit'd remove the need for a complex rewrite of nextval_internal.\\nAssuming the SRF is written in ValuePerCall style, each iteration\\ncan just call nextval_internal with no modifications needed in that\\nfunction. There'll be a CHECK_FOR_INTERRUPTS somewhere in the\\nquery-level loop, or at least it's not nextval's fault if there's not.\\nThe situation is then no different from generate_series with a large\\nloop count, or any other query that can generate lots of data.\\n\\nOf course, this does imply a lot more cycles expended per generated value\\n--- but most of that is inherent in the larger amount of data being\\nhanded back.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Thu, 28 Jul 2022 12:47:10 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On Thu, Jul 28, 2022 at 12:47:10PM -0400, Tom Lane wrote:\\n> Actually, on further thought, I do like the resultset idea, because\\n> it'd remove the need for a complex rewrite of nextval_internal.\\n> Assuming the SRF is written in ValuePerCall style, each iteration\\n> can just call nextval_internal with no modifications needed in that\\n> function. There'll be a CHECK_FOR_INTERRUPTS somewhere in the\\n> query-level loop, or at least it's not nextval's fault if there's not.\\n> The situation is then no different from generate_series with a large\\n> loop count, or any other query that can generate lots of data.\\n> \\n> Of course, this does imply a lot more cycles expended per generated value\\n> --- but most of that is inherent in the larger amount of data being\\n> handed back.\\n\\nFWIW, I find the result set approach more intuitive and robust,\\nparticularly in the case of a sequence has non-default values\\nINCREMENT and min/max values. This reduces the dependency of what an\\napplication needs to know about the details of a given sequence. With\\nonly the last value reported, the application would need to compile\\nthings by itself.\\n--\\nMichael\",\n \"msg_date\": \"Sat, 30 Jul 2022 16:21:07 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n },\n {\n \"msg_contents\": \"On Sat, Jul 30, 2022 at 04:21:07PM +0900, Michael Paquier wrote:\\n> FWIW, I find the result set approach more intuitive and robust,\\n> particularly in the case of a sequence has non-default values\\n> INCREMENT and min/max values. This reduces the dependency of what an\\n> application needs to know about the details of a given sequence. With\\n> only the last value reported, the application would need to compile\\n> things by itself.\\n\\nIt seems like there is a consensus here, but the thread has no\\nactivity for the past two months, so I have marked the patch as\\nreturned with feedback for now.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 12 Oct 2022 14:38:12 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Support for grabbing multiple consecutive values with nextval()\"\n }\n]","isConversational":false}}},{"rowIdx":53591,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nI found some redundant function calls in\npg_stat_statements.c/pg_stat_statements_internal(),\nThere is no need to call GetUserId() again because the value was\npreviously obtained.\nso I propose a patch to fix it.\n\n--- a/contrib/pg_stat_statements/pg_stat_statements.c\n+++ b/contrib/pg_stat_statements/pg_stat_statements.c\n@@ -1508,7 +1508,7 @@ pg_stat_statements_internal(FunctionCallInfo fcinfo,\n pgssEntry *entry;\n\n /* Superusers or members of pg_read_all_stats members are allowed */\n- is_allowed_role = is_member_of_role(GetUserId(),\nROLE_PG_READ_ALL_STATS);\n+ is_allowed_role = is_member_of_role(userid, ROLE_PG_READ_ALL_STATS);\n\n /* hash table must exist already */\n if (!pgss || !pgss_hash)\n\nRegards,\nLee Dong Wook.","msg_date":"Sun, 27 Feb 2022 20:45:13 +0900","msg_from":"Dong Wook Lee <sh95119@gmail.com>","msg_from_op":true,"msg_subject":"pg_stat_statements: remove redundant function call in\n pg_stat_statements_internal"},{"msg_contents":"Hi,\n\nOn Sun, Feb 27, 2022 at 08:45:13PM +0900, Dong Wook Lee wrote:\n> \n> I found some redundant function calls in\n> pg_stat_statements.c/pg_stat_statements_internal(),\n> There is no need to call GetUserId() again because the value was\n> previously obtained.\n\nIndeed. I doubt it will make any real difference but it doesn't hurt to fix\nit.\n\nPatch looks good to me.\n\n\n","msg_date":"Sun, 27 Feb 2022 19:55:26 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: pg_stat_statements: remove redundant function call in\n pg_stat_statements_internal"},{"msg_contents":"On Sun, Feb 27, 2022 at 07:55:26PM +0800, Julien Rouhaud wrote:\n> Indeed. I doubt it will make any real difference but it doesn't hurt to fix\n> it.\n> \n> Patch looks good to me.\n\nYes, let's clean up that on HEAD. No objections from here. I'll do\nthat tomorrow or so.\n--\nMichael","msg_date":"Sun, 27 Feb 2022 21:08:56 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_stat_statements: remove redundant function call in\n pg_stat_statements_internal"},{"msg_contents":"On Sun, Feb 27, 2022 at 09:08:56PM +0900, Michael Paquier wrote:\n> Yes, let's clean up that on HEAD. No objections from here. I'll do\n> that tomorrow or so.\n\nAnd done.\n--\nMichael","msg_date":"Mon, 28 Feb 2022 11:02:09 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: pg_stat_statements: remove redundant function call in\n pg_stat_statements_internal"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nI found some redundant function calls in\\npg_stat_statements.c/pg_stat_statements_internal(),\\nThere is no need to call GetUserId() again because the value was\\npreviously obtained.\\nso I propose a patch to fix it.\\n\\n--- a/contrib/pg_stat_statements/pg_stat_statements.c\\n+++ b/contrib/pg_stat_statements/pg_stat_statements.c\\n@@ -1508,7 +1508,7 @@ pg_stat_statements_internal(FunctionCallInfo fcinfo,\\n pgssEntry *entry;\\n\\n /* Superusers or members of pg_read_all_stats members are allowed */\\n- is_allowed_role = is_member_of_role(GetUserId(),\\nROLE_PG_READ_ALL_STATS);\\n+ is_allowed_role = is_member_of_role(userid, ROLE_PG_READ_ALL_STATS);\\n\\n /* hash table must exist already */\\n if (!pgss || !pgss_hash)\\n\\nRegards,\\nLee Dong Wook.\",\n \"msg_date\": \"Sun, 27 Feb 2022 20:45:13 +0900\",\n \"msg_from\": \"Dong Wook Lee <sh95119@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"pg_stat_statements: remove redundant function call in\\n pg_stat_statements_internal\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Sun, Feb 27, 2022 at 08:45:13PM +0900, Dong Wook Lee wrote:\\n> \\n> I found some redundant function calls in\\n> pg_stat_statements.c/pg_stat_statements_internal(),\\n> There is no need to call GetUserId() again because the value was\\n> previously obtained.\\n\\nIndeed. I doubt it will make any real difference but it doesn't hurt to fix\\nit.\\n\\nPatch looks good to me.\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Feb 2022 19:55:26 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_stat_statements: remove redundant function call in\\n pg_stat_statements_internal\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 07:55:26PM +0800, Julien Rouhaud wrote:\\n> Indeed. I doubt it will make any real difference but it doesn't hurt to fix\\n> it.\\n> \\n> Patch looks good to me.\\n\\nYes, let's clean up that on HEAD. No objections from here. I'll do\\nthat tomorrow or so.\\n--\\nMichael\",\n \"msg_date\": \"Sun, 27 Feb 2022 21:08:56 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_stat_statements: remove redundant function call in\\n pg_stat_statements_internal\"\n },\n {\n \"msg_contents\": \"On Sun, Feb 27, 2022 at 09:08:56PM +0900, Michael Paquier wrote:\\n> Yes, let's clean up that on HEAD. No objections from here. I'll do\\n> that tomorrow or so.\\n\\nAnd done.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 28 Feb 2022 11:02:09 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: pg_stat_statements: remove redundant function call in\\n pg_stat_statements_internal\"\n }\n]","isConversational":false}}},{"rowIdx":53592,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi, hackers\n\nWhen I try to use CREATE DATABASE IF NOT EXISTS in PostgreSQL, it complains\nthis syntax is not supported. We can use the following command to achieve\nthis, however, it's not straightforward.\n\n SELECT 'CREATE DATABASE mydb'\n WHERE NOT EXISTS (SELECT FROM pg_database WHERE datname = 'mydb')\\gexec\n\nWhy don't support CREATE DATABASE IF NOT EXISTS syntax in PostgreSQL?\n\nI create a patch for this, any suggestions?\n\n-- \nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.","msg_date":"Sun, 27 Feb 2022 22:13:13 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":true,"msg_subject":"CREATE DATABASE IF NOT EXISTS in PostgreSQL"},{"msg_contents":"Japin Li <japinli@hotmail.com> writes:\n> Why don't support CREATE DATABASE IF NOT EXISTS syntax in PostgreSQL?\n\nFWIW, I'm generally hostile to CREATE IF NOT EXISTS semantics across\nthe board, because of its exceedingly squishy semantics: it ensures\nthat an object by that name exists, but you have exactly no guarantees\nabout its properties or contents. The more complex the object, the\nbigger that problem becomes ... and a whole database is the most\ncomplex sort of object we have. So IMV, the fact that we don't have\nthis \"feature\" is a good thing.\n\nWe do have DROP DATABASE IF EXISTS, and I think using that followed\nby CREATE is a much better-defined approach.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sun, 27 Feb 2022 12:53:00 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: CREATE DATABASE IF NOT EXISTS in PostgreSQL"},{"msg_contents":"\nOn Mon, 28 Feb 2022 at 01:53, Tom Lane <tgl@sss.pgh.pa.us> wrote:\n> Japin Li <japinli@hotmail.com> writes:\n>> Why don't support CREATE DATABASE IF NOT EXISTS syntax in PostgreSQL?\n>\n> FWIW, I'm generally hostile to CREATE IF NOT EXISTS semantics across\n> the board, because of its exceedingly squishy semantics: it ensures\n> that an object by that name exists, but you have exactly no guarantees\n> about its properties or contents.\n\nThanks for the explanation! I think it is the database user who should\nguarantee the properties and contents of a database.\nCREATE IF NOT EXISTS is just a syntax sugar.\n\n-- \nRegrads,\nJapin Li.\nChengDu WenWu Information Technology Co.,Ltd.\n\n\n","msg_date":"Mon, 28 Feb 2022 09:04:50 +0800","msg_from":"Japin Li <japinli@hotmail.com>","msg_from_op":true,"msg_subject":"Re: CREATE DATABASE IF NOT EXISTS in PostgreSQL"}],"string":"[\n {\n \"msg_contents\": \"Hi, hackers\\n\\nWhen I try to use CREATE DATABASE IF NOT EXISTS in PostgreSQL, it complains\\nthis syntax is not supported. We can use the following command to achieve\\nthis, however, it's not straightforward.\\n\\n SELECT 'CREATE DATABASE mydb'\\n WHERE NOT EXISTS (SELECT FROM pg_database WHERE datname = 'mydb')\\\\gexec\\n\\nWhy don't support CREATE DATABASE IF NOT EXISTS syntax in PostgreSQL?\\n\\nI create a patch for this, any suggestions?\\n\\n-- \\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\",\n \"msg_date\": \"Sun, 27 Feb 2022 22:13:13 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"CREATE DATABASE IF NOT EXISTS in PostgreSQL\"\n },\n {\n \"msg_contents\": \"Japin Li <japinli@hotmail.com> writes:\\n> Why don't support CREATE DATABASE IF NOT EXISTS syntax in PostgreSQL?\\n\\nFWIW, I'm generally hostile to CREATE IF NOT EXISTS semantics across\\nthe board, because of its exceedingly squishy semantics: it ensures\\nthat an object by that name exists, but you have exactly no guarantees\\nabout its properties or contents. The more complex the object, the\\nbigger that problem becomes ... and a whole database is the most\\ncomplex sort of object we have. So IMV, the fact that we don't have\\nthis \\\"feature\\\" is a good thing.\\n\\nWe do have DROP DATABASE IF EXISTS, and I think using that followed\\nby CREATE is a much better-defined approach.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sun, 27 Feb 2022 12:53:00 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: CREATE DATABASE IF NOT EXISTS in PostgreSQL\"\n },\n {\n \"msg_contents\": \"\\nOn Mon, 28 Feb 2022 at 01:53, Tom Lane <tgl@sss.pgh.pa.us> wrote:\\n> Japin Li <japinli@hotmail.com> writes:\\n>> Why don't support CREATE DATABASE IF NOT EXISTS syntax in PostgreSQL?\\n>\\n> FWIW, I'm generally hostile to CREATE IF NOT EXISTS semantics across\\n> the board, because of its exceedingly squishy semantics: it ensures\\n> that an object by that name exists, but you have exactly no guarantees\\n> about its properties or contents.\\n\\nThanks for the explanation! I think it is the database user who should\\nguarantee the properties and contents of a database.\\nCREATE IF NOT EXISTS is just a syntax sugar.\\n\\n-- \\nRegrads,\\nJapin Li.\\nChengDu WenWu Information Technology Co.,Ltd.\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 09:04:50 +0800\",\n \"msg_from\": \"Japin Li <japinli@hotmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: CREATE DATABASE IF NOT EXISTS in PostgreSQL\"\n }\n]","isConversational":false}}},{"rowIdx":53593,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Currently, when you connect with psql over SSL, you get a display like\nthis:\n\npsql (15devel)\nSSL connection (protocol: TLSv1.3, cipher: TLS_AES_256_GCM_SHA384, bits: 256, compression: off)\nType \"help\" for help.\n\nSince support for SSL compression has been removed from PostgreSQL, it\ndoesn't seem sensible to display it anymore. And while we're there, I\nthink the bits information is redundant, since it can be derived from\nthe cipher suite, either because it's part of the name (as in the\nexample) or by looking it up somewhere. So I propose that we make this\ndisplay a bit more compact like this:\n\npsql (15devel)\nSSL connection (protocol: TLSv1.3, cipher: TLS_AES_256_GCM_SHA384)\nType \"help\" for help.\n\nSee attached patch.","msg_date":"Mon, 28 Feb 2022 10:02:51 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":true,"msg_subject":"psql: Make SSL info display more compact"},{"msg_contents":"> On 28 Feb 2022, at 10:02, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n\n> Since support for SSL compression has been removed from PostgreSQL, it\n> doesn't seem sensible to display it anymore.\n\nThis was originally done, but all client side changes reverted as there still\nare server versions in production which allow compression.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Mon, 28 Feb 2022 10:21:58 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: psql: Make SSL info display more compact"},{"msg_contents":"Daniel Gustafsson <daniel@yesql.se> writes:\n\n>> On 28 Feb 2022, at 10:02, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n>\n>> Since support for SSL compression has been removed from PostgreSQL, it\n>> doesn't seem sensible to display it anymore.\n>\n> This was originally done, but all client side changes reverted as there still\n> are server versions in production which allow compression.\n\nHow about making it show \"compression: on\" if compression is on, but\nnothing in the common \"off\" case?\n\n- ilmari\n\n\n","msg_date":"Mon, 28 Feb 2022 10:50:01 +0000","msg_from":"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>","msg_from_op":false,"msg_subject":"Re: psql: Make SSL info display more compact"},{"msg_contents":"On Mon, Feb 28, 2022 at 10:50:01AM +0000, Dagfinn Ilmari Mannsåker wrote:\n> Daniel Gustafsson <daniel@yesql.se> writes:\n>> On 28 Feb 2022, at 10:02, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n>> This was originally done, but all client side changes reverted as there still\n>> are server versions in production which allow compression.\n> \n> How about making it show \"compression: on\" if compression is on, but\n> nothing in the common \"off\" case?\n\nHm, no, as it can be useful to know that compression is disabled when\nconnecting to an old server. What about that making the information\nshown version-aware? I would choose to show the compression part only\nfor server versions where it is settable.\n--\nMichael","msg_date":"Mon, 28 Feb 2022 20:21:32 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: psql: Make SSL info display more compact"},{"msg_contents":"On 28.02.22 11:50, Dagfinn Ilmari Mannsåker wrote:\n> Daniel Gustafsson <daniel@yesql.se> writes:\n> \n>>> On 28 Feb 2022, at 10:02, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n>>\n>>> Since support for SSL compression has been removed from PostgreSQL, it\n>>> doesn't seem sensible to display it anymore.\n>>\n>> This was originally done, but all client side changes reverted as there still\n>> are server versions in production which allow compression.\n> \n> How about making it show \"compression: on\" if compression is on, but\n> nothing in the common \"off\" case?\n\nThat would work for me.\n\n\n","msg_date":"Mon, 28 Feb 2022 12:56:42 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: psql: Make SSL info display more compact"},{"msg_contents":"On 28.02.22 12:21, Michael Paquier wrote:\n> What about that making the information\n> shown version-aware? I would choose to show the compression part only\n> for server versions where it is settable.\n\nThat might lead to confusing results if you are not connecting to \nsomething that is a stock PostgreSQL server.\n\n\n","msg_date":"Mon, 28 Feb 2022 12:57:39 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: psql: Make SSL info display more compact"},{"msg_contents":"> On 28 Feb 2022, at 12:56, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n> On 28.02.22 11:50, Dagfinn Ilmari Mannsåker wrote:\n>> Daniel Gustafsson <daniel@yesql.se> writes:\n\n>>> This was originally done, but all client side changes reverted as there still\n>>> are server versions in production which allow compression.\n>> How about making it show \"compression: on\" if compression is on, but\n>> nothing in the common \"off\" case?\n> \n> That would work for me.\n\nOn POLA grounds I would prefer to never to show it. If we ever get libpq\ncompression and want to show that, I'd prefer that we didn't end up\n\"compression\" meaning one thing except when it means two things.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Mon, 28 Feb 2022 13:48:43 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: psql: Make SSL info display more compact"},{"msg_contents":"Daniel Gustafsson <daniel@yesql.se> writes:\n> On 28 Feb 2022, at 12:56, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n>> On 28.02.22 11:50, Dagfinn Ilmari Mannsåker wrote:\n>>> How about making it show \"compression: on\" if compression is on, but\n>>> nothing in the common \"off\" case?\n\n>> That would work for me.\n\n> On POLA grounds I would prefer to never to show it. If we ever get libpq\n> compression and want to show that, I'd prefer that we didn't end up\n> \"compression\" meaning one thing except when it means two things.\n\nWell, any such output would presumably be on a different line and\nthus distinguishable from the SSL property; plus, it'd be impossible\nfor both forms to show up in the same connection.\n\nHowever, how about writing \"SSL compression: on\" versus writing\nnothing? That avoids doubt about what it means. I don't buy\nMichael's argument that this is ambiguous, either.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Mon, 28 Feb 2022 10:12:42 -0500","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: psql: Make SSL info display more compact"},{"msg_contents":"On 28.02.22 16:12, Tom Lane wrote:\n> Daniel Gustafsson <daniel@yesql.se> writes:\n>> On 28 Feb 2022, at 12:56, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n>>> On 28.02.22 11:50, Dagfinn Ilmari Mannsåker wrote:\n>>>> How about making it show \"compression: on\" if compression is on, but\n>>>> nothing in the common \"off\" case?\n> \n>>> That would work for me.\n> \n>> On POLA grounds I would prefer to never to show it. If we ever get libpq\n>> compression and want to show that, I'd prefer that we didn't end up\n>> \"compression\" meaning one thing except when it means two things.\n> \n> Well, any such output would presumably be on a different line and\n> thus distinguishable from the SSL property; plus, it'd be impossible\n> for both forms to show up in the same connection.\n> \n> However, how about writing \"SSL compression: on\" versus writing\n> nothing? That avoids doubt about what it means. I don't buy\n> Michael's argument that this is ambiguous, either.\n\nI didn't mean to reopen the whole SSL compression can of worms here, I \nwas mistaken about the level of support left after PG14. I was merely \nlightly annoyed that the psql startup display got quite long with not \nvery interesting information.\n\nI propose a reduced patch that just removes the \"bits\" display, since \nthat is redundant with the \"cipher\" display.","msg_date":"Tue, 1 Mar 2022 09:44:29 +0100","msg_from":"Peter Eisentraut <peter.eisentraut@enterprisedb.com>","msg_from_op":true,"msg_subject":"Re: psql: Make SSL info display more compact"},{"msg_contents":"> On 1 Mar 2022, at 09:44, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\n\n> I propose a reduced patch that just removes the \"bits\" display, since that is\n> redundant with the \"cipher\"\n\n\nNo objections from me.\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/\n\n\n\n","msg_date":"Wed, 2 Mar 2022 11:24:05 +0100","msg_from":"Daniel Gustafsson <daniel@yesql.se>","msg_from_op":false,"msg_subject":"Re: psql: Make SSL info display more compact"}],"string":"[\n {\n \"msg_contents\": \"Currently, when you connect with psql over SSL, you get a display like\\nthis:\\n\\npsql (15devel)\\nSSL connection (protocol: TLSv1.3, cipher: TLS_AES_256_GCM_SHA384, bits: 256, compression: off)\\nType \\\"help\\\" for help.\\n\\nSince support for SSL compression has been removed from PostgreSQL, it\\ndoesn't seem sensible to display it anymore. And while we're there, I\\nthink the bits information is redundant, since it can be derived from\\nthe cipher suite, either because it's part of the name (as in the\\nexample) or by looking it up somewhere. So I propose that we make this\\ndisplay a bit more compact like this:\\n\\npsql (15devel)\\nSSL connection (protocol: TLSv1.3, cipher: TLS_AES_256_GCM_SHA384)\\nType \\\"help\\\" for help.\\n\\nSee attached patch.\",\n \"msg_date\": \"Mon, 28 Feb 2022 10:02:51 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"psql: Make SSL info display more compact\"\n },\n {\n \"msg_contents\": \"> On 28 Feb 2022, at 10:02, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n\\n> Since support for SSL compression has been removed from PostgreSQL, it\\n> doesn't seem sensible to display it anymore.\\n\\nThis was originally done, but all client side changes reverted as there still\\nare server versions in production which allow compression.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 10:21:58 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: psql: Make SSL info display more compact\"\n },\n {\n \"msg_contents\": \"Daniel Gustafsson <daniel@yesql.se> writes:\\n\\n>> On 28 Feb 2022, at 10:02, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n>\\n>> Since support for SSL compression has been removed from PostgreSQL, it\\n>> doesn't seem sensible to display it anymore.\\n>\\n> This was originally done, but all client side changes reverted as there still\\n> are server versions in production which allow compression.\\n\\nHow about making it show \\\"compression: on\\\" if compression is on, but\\nnothing in the common \\\"off\\\" case?\\n\\n- ilmari\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 10:50:01 +0000\",\n \"msg_from\": \"=?utf-8?Q?Dagfinn_Ilmari_Manns=C3=A5ker?= <ilmari@ilmari.org>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: psql: Make SSL info display more compact\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 10:50:01AM +0000, Dagfinn Ilmari Mannsåker wrote:\\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n>> On 28 Feb 2022, at 10:02, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n>> This was originally done, but all client side changes reverted as there still\\n>> are server versions in production which allow compression.\\n> \\n> How about making it show \\\"compression: on\\\" if compression is on, but\\n> nothing in the common \\\"off\\\" case?\\n\\nHm, no, as it can be useful to know that compression is disabled when\\nconnecting to an old server. What about that making the information\\nshown version-aware? I would choose to show the compression part only\\nfor server versions where it is settable.\\n--\\nMichael\",\n \"msg_date\": \"Mon, 28 Feb 2022 20:21:32 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: psql: Make SSL info display more compact\"\n },\n {\n \"msg_contents\": \"On 28.02.22 11:50, Dagfinn Ilmari Mannsåker wrote:\\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n> \\n>>> On 28 Feb 2022, at 10:02, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n>>\\n>>> Since support for SSL compression has been removed from PostgreSQL, it\\n>>> doesn't seem sensible to display it anymore.\\n>>\\n>> This was originally done, but all client side changes reverted as there still\\n>> are server versions in production which allow compression.\\n> \\n> How about making it show \\\"compression: on\\\" if compression is on, but\\n> nothing in the common \\\"off\\\" case?\\n\\nThat would work for me.\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 12:56:42 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: psql: Make SSL info display more compact\"\n },\n {\n \"msg_contents\": \"On 28.02.22 12:21, Michael Paquier wrote:\\n> What about that making the information\\n> shown version-aware? I would choose to show the compression part only\\n> for server versions where it is settable.\\n\\nThat might lead to confusing results if you are not connecting to \\nsomething that is a stock PostgreSQL server.\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 12:57:39 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: psql: Make SSL info display more compact\"\n },\n {\n \"msg_contents\": \"> On 28 Feb 2022, at 12:56, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n> On 28.02.22 11:50, Dagfinn Ilmari Mannsåker wrote:\\n>> Daniel Gustafsson <daniel@yesql.se> writes:\\n\\n>>> This was originally done, but all client side changes reverted as there still\\n>>> are server versions in production which allow compression.\\n>> How about making it show \\\"compression: on\\\" if compression is on, but\\n>> nothing in the common \\\"off\\\" case?\\n> \\n> That would work for me.\\n\\nOn POLA grounds I would prefer to never to show it. If we ever get libpq\\ncompression and want to show that, I'd prefer that we didn't end up\\n\\\"compression\\\" meaning one thing except when it means two things.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 13:48:43 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: psql: Make SSL info display more compact\"\n },\n {\n \"msg_contents\": \"Daniel Gustafsson <daniel@yesql.se> writes:\\n> On 28 Feb 2022, at 12:56, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n>> On 28.02.22 11:50, Dagfinn Ilmari Mannsåker wrote:\\n>>> How about making it show \\\"compression: on\\\" if compression is on, but\\n>>> nothing in the common \\\"off\\\" case?\\n\\n>> That would work for me.\\n\\n> On POLA grounds I would prefer to never to show it. If we ever get libpq\\n> compression and want to show that, I'd prefer that we didn't end up\\n> \\\"compression\\\" meaning one thing except when it means two things.\\n\\nWell, any such output would presumably be on a different line and\\nthus distinguishable from the SSL property; plus, it'd be impossible\\nfor both forms to show up in the same connection.\\n\\nHowever, how about writing \\\"SSL compression: on\\\" versus writing\\nnothing? That avoids doubt about what it means. I don't buy\\nMichael's argument that this is ambiguous, either.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 10:12:42 -0500\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: psql: Make SSL info display more compact\"\n },\n {\n \"msg_contents\": \"On 28.02.22 16:12, Tom Lane wrote:\\n> Daniel Gustafsson <daniel@yesql.se> writes:\\n>> On 28 Feb 2022, at 12:56, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n>>> On 28.02.22 11:50, Dagfinn Ilmari Mannsåker wrote:\\n>>>> How about making it show \\\"compression: on\\\" if compression is on, but\\n>>>> nothing in the common \\\"off\\\" case?\\n> \\n>>> That would work for me.\\n> \\n>> On POLA grounds I would prefer to never to show it. If we ever get libpq\\n>> compression and want to show that, I'd prefer that we didn't end up\\n>> \\\"compression\\\" meaning one thing except when it means two things.\\n> \\n> Well, any such output would presumably be on a different line and\\n> thus distinguishable from the SSL property; plus, it'd be impossible\\n> for both forms to show up in the same connection.\\n> \\n> However, how about writing \\\"SSL compression: on\\\" versus writing\\n> nothing? That avoids doubt about what it means. I don't buy\\n> Michael's argument that this is ambiguous, either.\\n\\nI didn't mean to reopen the whole SSL compression can of worms here, I \\nwas mistaken about the level of support left after PG14. I was merely \\nlightly annoyed that the psql startup display got quite long with not \\nvery interesting information.\\n\\nI propose a reduced patch that just removes the \\\"bits\\\" display, since \\nthat is redundant with the \\\"cipher\\\" display.\",\n \"msg_date\": \"Tue, 1 Mar 2022 09:44:29 +0100\",\n \"msg_from\": \"Peter Eisentraut <peter.eisentraut@enterprisedb.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: psql: Make SSL info display more compact\"\n },\n {\n \"msg_contents\": \"> On 1 Mar 2022, at 09:44, Peter Eisentraut <peter.eisentraut@enterprisedb.com> wrote:\\n\\n> I propose a reduced patch that just removes the \\\"bits\\\" display, since that is\\n> redundant with the \\\"cipher\\\"\\n\\n\\nNo objections from me.\\n\\n--\\nDaniel Gustafsson\\t\\thttps://vmware.com/\\n\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 11:24:05 +0100\",\n \"msg_from\": \"Daniel Gustafsson <daniel@yesql.se>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: psql: Make SSL info display more compact\"\n }\n]","isConversational":false}}},{"rowIdx":53594,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nFor Toast storage [1] in PostgreSQL, first, the attribute value is\ncompressed\nand then divided into chunks. The problem with storing compressed value is,\nif we\nare not saving enough space such that it reduces the #chunks then we end up\nadding extra decompression cost on every read.\nBased on the discussion with Robert and Dilip, we tried to optimize this\nprocess. The idea is to TOAST the compressed value only if we are saving at\nleast\n1 chunk(2KB default) of disk storage else use the uncompressed one. In this\nway,\nwe will save decompression costs without losing much on storage.\n\nIn our tests, we have observed improvement in the read performance up to\n28% by\ngiving up at most TOAST_MAX_CHUNK_SIZE (2KB) bytes for storage on disk. The\ngain is more on large attributes (size > 4KB) because\ncompression/decompression\ncost increases with size.\nHowever, We have found, this assumption is not true when the data\ncompression\nratio is more and the size is barely big enough to cause TOASTing. For\nexample,\nin the following test 4. b, we didn't get any performance advantage but the\ntable\nsize grew by 42% by storing uncompressed values.\nTest Setup.\n\nCreate table t1_lz4 ( a text compression lz4, b text compression lz4);\n-- Generate random data\ncreate or replace function generate_att_data(len_info int)\nreturns text\nlanguage plpgsql\nas\n$$\ndeclare\n value text;\nbegin\n select array_agg(md5(g::text))\n into value\n from generate_series(1, round(len_info/33)::int) g;\n return value;\nend;\n$$;\n\n--Test\nSelect b from t1_lz4;\n\nTest 1:\nData: rows 200000\ninsert into t1_lz4(a, b) select generate_att_data(364), repeat\n(generate_att_data(1980), 2);\nSummary:\nAttribute size: original: 7925 bytes, after compression: 7845 bytes\nTime for select: head: 42 sec, patch: 37 sec, *Performance Gain: 11%*\ntable size: Head 1662 MB, Patch: 1662 MB\n\nTest 2:\nData: rows 100000\ninsert into t1_lz4(a, b) select generate_att_data(364),\ngenerate_att_data(16505);\nSummary:\nAttribute size: original: 16505 bytes, after compression: 16050 bytes\nTime for select: head: 35.6 sec, patch: 30 sec, *Performance Gain: 14%*\ntable size: Head 1636 MB, Patch: 1688 MB\n\nTest 3:\nData: rows 50000\ninsert into t1_lz4(a, b) select generate_att_data(364),\ngenerate_att_data(31685);\nSummary:\nAttribute size: original: 31685 bytes, after compression: 30263 bytes\nTime for select: head: 35.4 sec, patch: 25.5 sec, *Performance Gain: 28%*\ntable size: Head 1601 MB, Patch: 1601 MB\n\nTest 4.a:\nData: rows 200000\ninsert into t1_lz4(a, b) select generate_att_data(11), repeat ('b', 250)\n|| generate_att_data(3885);\nSummary:\nAttribute size: original: 3885 bytes, after compression: 3645 bytes\nTime for select: head: 28 sec, patch: 26 sec, *Performance Gain: 7%*\n*table size: Head 872 MB, Patch: 872 MB*\n\nTest 4.b (High compression):\nData: rows 200000\ninsert into t1_lz4(a, b) select generate_att_data(364), repeat\n(generate_att_data(1980), 2);\nSummary:\nAttribute size: original: 3966 bytes, after compression: 2012 bytes\nTime for select: head: 27 sec, patch: 26 sec, *Performance Gain: 0%*\n*table size: Head 612 MB, Patch: 872 MB*\n\nThis is the worst case for this optimization because of the following 2\nreasons.\nFirst, the table size would grow by 50% when compressed size is half of the\noriginal size and\nyet barely large enough for TOASTing. Table size can't grow more than that\nbecause If\ncompression reduces the size even more then it will reduce the #chunks as\nwell and it stores\nthe compressed value in the table.\n\nSecond, not much gain in performance because of the small attribute size,\nmore attributes fit\n\nin page (almost twice), on each page access it can access twice the number\nof rows. And\n\nalso small value means low compression/decompression costs.\n\n\nWe have avoided such cases by applying the optimization when attribute size\n> 4KB.\n\nRegards,\nDavinder\nEnterpriseDB: http://www.enterprisedb.com","msg_date":"Mon, 28 Feb 2022 15:22:11 +0530","msg_from":"davinder singh <davindersingh2692@gmail.com>","msg_from_op":true,"msg_subject":"Optimize external TOAST storage"},{"msg_contents":"On Mon, Feb 28, 2022 at 3:22 PM davinder singh\n<davindersingh2692@gmail.com> wrote:\n>\n> Hi,\n>\n> For Toast storage [1] in PostgreSQL, first, the attribute value is compressed\n> and then divided into chunks. The problem with storing compressed value is, if we\n> are not saving enough space such that it reduces the #chunks then we end up\n> adding extra decompression cost on every read.\n> Based on the discussion with Robert and Dilip, we tried to optimize this\n> process. The idea is to TOAST the compressed value only if we are saving at least\n> 1 chunk(2KB default) of disk storage else use the uncompressed one. In this way,\n> we will save decompression costs without losing much on storage.\n>\n> In our tests, we have observed improvement in the read performance up to 28% by\n> giving up at most TOAST_MAX_CHUNK_SIZE (2KB) bytes for storage on disk. The\n> gain is more on large attributes (size > 4KB) because compression/decompression\n> cost increases with size.\n> However, We have found, this assumption is not true when the data compression\n> ratio is more and the size is barely big enough to cause TOASTing. For example,\n> in the following test 4. b, we didn't get any performance advantage but the table\n> size grew by 42% by storing uncompressed values.\n> Test Setup.\n>\n> Create table t1_lz4 ( a text compression lz4, b text compression lz4);\n> -- Generate random data\n> create or replace function generate_att_data(len_info int)\n> returns text\n> language plpgsql\n> as\n> $$\n> declare\n> value text;\n> begin\n> select array_agg(md5(g::text))\n> into value\n> from generate_series(1, round(len_info/33)::int) g;\n> return value;\n> end;\n> $$;\n>\n> --Test\n> Select b from t1_lz4;\n>\n> Test 1:\n> Data: rows 200000\n> insert into t1_lz4(a, b) select generate_att_data(364), repeat (generate_att_data(1980), 2);\n> Summary:\n> Attribute size: original: 7925 bytes, after compression: 7845 bytes\n> Time for select: head: 42 sec, patch: 37 sec, Performance Gain: 11%\n> table size: Head 1662 MB, Patch: 1662 MB\n>\n> Test 2:\n> Data: rows 100000\n> insert into t1_lz4(a, b) select generate_att_data(364), generate_att_data(16505);\n> Summary:\n> Attribute size: original: 16505 bytes, after compression: 16050 bytes\n> Time for select: head: 35.6 sec, patch: 30 sec, Performance Gain: 14%\n> table size: Head 1636 MB, Patch: 1688 MB\n>\n> Test 3:\n> Data: rows 50000\n> insert into t1_lz4(a, b) select generate_att_data(364), generate_att_data(31685);\n> Summary:\n> Attribute size: original: 31685 bytes, after compression: 30263 bytes\n> Time for select: head: 35.4 sec, patch: 25.5 sec, Performance Gain: 28%\n> table size: Head 1601 MB, Patch: 1601 MB\n>\n> Test 4.a:\n> Data: rows 200000\n> insert into t1_lz4(a, b) select generate_att_data(11), repeat ('b', 250) || generate_att_data(3885);\n> Summary:\n> Attribute size: original: 3885 bytes, after compression: 3645 bytes\n> Time for select: head: 28 sec, patch: 26 sec, Performance Gain: 7%\n> table size: Head 872 MB, Patch: 872 MB\n>\n> Test 4.b (High compression):\n> Data: rows 200000\n> insert into t1_lz4(a, b) select generate_att_data(364), repeat (generate_att_data(1980), 2);\n> Summary:\n> Attribute size: original: 3966 bytes, after compression: 2012 bytes\n> Time for select: head: 27 sec, patch: 26 sec, Performance Gain: 0%\n> table size: Head 612 MB, Patch: 872 MB\n\nI think the idea looks interesting and the results are also promising.\nI have a few initial comments, later I will do more detailed review.\n\n1.\n+ if (*value == orig_toast_value)\n+ toast_tuple_externalize(ttc, attribute, options);\n\nIsn't it looks cleaner to check whther the attribute is compressed or\nnot like this\nVARATT_IS_COMPRESSED(DatumGetPointer(*value)) ?\n\n2.\n+ /* Using the uncompressed data instead, deleting\ncompressed data. */\n+ pfree(DatumGetPointer(*value));\n\nchange as below\n/deleting compressed data/release memory for the compressed data\n\n3.\n+ memcpy(toast_attr_copy, toast_attr, sizeof(toast_attr));\n+ memcpy(toast_values_copy, toast_values, sizeof(toast_values));\n+\n\nAdd some comment here, what we are trying to copy and why?\n\n4.\n+ /* incompressible, ignore on subsequent compression passes. */\n+ orig_attr->tai_colflags |= TOASTCOL_INCOMPRESSIBLE;\n\nDo we need to set TOASTCOL_INCOMPRESSIBLE while trying to externalize\nit, the comment say \"ignore on subsequent compression passes\"\nbut after this will there be more compression passes? If we need to\nset this TOASTCOL_INCOMPRESSIBLE then comment should explain this.\n\n\n\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 8 Mar 2022 21:44:41 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"Thanks Dilip, I have fixed your comments, please find the updated patch.\n\nOn Tue, Mar 8, 2022 at 9:44 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:.\n\n> + /* incompressible, ignore on subsequent compression passes. */\n> + orig_attr->tai_colflags |= TOASTCOL_INCOMPRESSIBLE;\n>\n> Do we need to set TOASTCOL_INCOMPRESSIBLE while trying to externalize\n> it, the comment say \"ignore on subsequent compression passes\"\n> but after this will there be more compression passes? If we need to\n\nset this TOASTCOL_INCOMPRESSIBLE then comment should explain this.\n>\nThat was a mistake, this flag is not required at this point, as the\nattribute is externalized it will be marked as TOASTCOL_IGNORE, and\nsuch columns are not considered for compression, I removed it. Thanks for\npointing it out.\n\n-- \nRegards,\nDavinder\nEnterpriseDB: http://www.enterprisedb.com","msg_date":"Thu, 10 Mar 2022 14:04:06 +0530","msg_from":"davinder singh <davindersingh2692@gmail.com>","msg_from_op":true,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Thu, Mar 10, 2022 at 2:04 PM davinder singh\n<davindersingh2692@gmail.com> wrote:\n>\n> Thanks Dilip, I have fixed your comments, please find the updated patch.\n>\n\nSome more comments\n\n+ /*\n+ * For those cases where storing compressed data is not optimal,\nWe will use\n+ * this pointer copy for referring uncompressed data.\n+ */\n+ memcpy(toast_attr_copy, toast_attr, sizeof(toast_attr));\n+ memcpy(toast_values_copy, toast_values, sizeof(toast_values));\n\nI think we can change the comments like below\n\n\"Preserve references to the original uncompressed data before\nattempting the compression. So that during externalize if we decide\nnot to store the compressed data then we have the original data with\nus. For more details refer to comments atop\ntoast_tuple_opt_externalize\".\n\n+/*\n+ * Don't save compressed data to external storage unless it saves I/O while\n+ * accessing the same data by reducing the number of chunks.\n+ */\n+void\n+toast_tuple_opt_externalize(ToastTupleContext *ttc, int attribute, int options,\n+ Datum orig_toast_value, ToastAttrInfo *orig_attr)\n\nI think these comments are not explaining what is the problem with\nstoring the compressed data. So you need to expand this comment\nsaying if it is not reducing the number or chunks then there is no\npoint in storing the compressed data because then we will have\nadditional decompression cost whenever we are fetching that data.\n\n\n+\n+ /* Sanity check: if data is not compressed then we can proceed as usual. */\n+ if (!VARATT_IS_COMPRESSED(DatumGetPointer(*value)))\n+ toast_tuple_externalize(ttc, attribute, options);\n\nI think we don't need \"Sanity check:\" here, the remaining part of the\ncomment is fine.\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 10 Mar 2022 17:09:40 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"I think this is an interesting idea. My first thought is that it would be\nnice if we didn't have to first compress the data to see whether we wanted\nto store it compressed or not, but I don't think there is much we can do\nabout that. In any case, we aren't adding much work in the write path\ncompared to the potential savings in the read path, so that is probably\nokay.\n\nDo you think it is worth making this configurable? I don't think it is\noutside the realm of possibility for some users to care more about disk\nspace than read performance. And maybe the threshold for this optimization\ncould differ based on the workload.\n\n extern void toast_tuple_externalize(ToastTupleContext *ttc, int attribute,\n int options);\n+extern void toast_tuple_opt_externalize(ToastTupleContext *ttc, int attribute,\n+ int options, Datum old_toast_value,\n+ ToastAttrInfo *old_toast_attr);\n\nCould we bake this into toast_tuple_externalize() so that all existing\ncallers benefit from this optimization? Is there a reason to export both\nfunctions? Perhaps toast_tuple_externalize() should be renamed and made\nstatic, and then this new function could be called\ntoast_tuple_externalize() (which would be a wrapper around the internal\nfunction).\n\n+ /* Sanity check: if data is not compressed then we can proceed as usual. */\n+ if (!VARATT_IS_COMPRESSED(DatumGetPointer(*value)))\n+ toast_tuple_externalize(ttc, attribute, options);\n\nWith a --enable-cassert build, this line causes assertion failures in the\ncall to GetMemoryChunkContext() from pfree(). 'make check' is enough to\nreproduce it. Specifically, it fails the following assertion:\n\n\tAssertArg(MemoryContextIsValid(context));\n\nI didn't see an existing commitfest entry for this patch. I'd encourage\nyou to add one in the July commitfest:\n\n\thttps://commitfest.postgresql.org/38/\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Sat, 12 Mar 2022 14:13:47 -0800","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"Thanks, Nathan, for the review comments. Please find the updated patch.\nOn Sun, Mar 13, 2022 at 3:43 AM Nathan Bossart <nathandbossart@gmail.com>\nwrote:\n\n> Do you think it is worth making this configurable? I don't think it is\n> outside the realm of possibility for some users to care more about disk\n> space than read performance. And maybe the threshold for this optimization\n> could differ based on the workload.\n>\n\nI think here we can break it into two parts.\nThe first part if the user cares about the disk more than reading\nperformance, disable this?\nThat is a good idea, we can try this, lets see what others say.\n\nRegarding the 2nd part of configuring the threshold, Based on our\nexperiments, we have fixed it for the attributes with size > 2 * chunk_size.\nThe default chunk_size is 2KB and the page size is 8KB. While toasting each\nattribute is divided into chunks, and each page can hold a max of 4 such\nchunks.\nWe only need to think about the space used by the last chunk of the\nattribute.\nThis means with each value optimization, it might use extra space in the\nrange\n(0B,2KB]. I think this extra space is independent of attribute size. So we\ndon't\nneed to worry about configuring this threshold. Let me know if I missed\nsomething\nhere.\n\n\n> extern void toast_tuple_externalize(ToastTupleContext *ttc, int attribute,\n> int options);\n> +extern void toast_tuple_opt_externalize(ToastTupleContext *ttc, int\n> attribute,\n> + int options, Datum\n> old_toast_value,\n> + ToastAttrInfo *old_toast_attr);\n>\n> Could we bake this into toast_tuple_externalize() so that all existing\n> callers benefit from this optimization? Is there a reason to export both\n> functions? Perhaps toast_tuple_externalize() should be renamed and made\n> static, and then this new function could be called\n> toast_tuple_externalize() (which would be a wrapper around the internal\n> function).\n>\nThis function is used only in heap_toast_insert_or_update(), all existing\ncallers are\nusing new function only. As you suggested, I have renamed the new function\nas\nwrapper function and also only exporting the new function.\n\n\n>\n> + /* Sanity check: if data is not compressed then we can proceed as\n> usual. */\n> + if (!VARATT_IS_COMPRESSED(DatumGetPointer(*value)))\n> + toast_tuple_externalize(ttc, attribute, options);\n>\n> With a --enable-cassert build, this line causes assertion failures in the\n> call to GetMemoryChunkContext() from pfree(). 'make check' is enough to\n> reproduce it. Specifically, it fails the following assertion:\n>\n> AssertArg(MemoryContextIsValid(context));\n>\n\nThanks for pointing it out, this failure started because I was not handling\nthe\ncase when the data is already compressed even before toasting. Following\ncheck verifies if the data is compressed or not but that is not enough\nbecause\nwe can't optimize the toasting if we didn't get the data in the\nuncompressed form in the first place from the source.\n+ /* If data is not compressed then we can proceed as usual. */\n+ if (!VARATT_IS_COMPRESSED(DatumGetPointer(*value)))\n\nv1 patch didn't have this problem because it was verifying whether we have\ncompressed data in this toasting round or not. I have changed back to the\nearlier version.\n+ /* If data is not compressed then we can proceed as usual. */\n+ if (*value == orig_toast_value)\n\n\n\n-- \nRegards,\nDavinder\nEnterpriseDB: http://www.enterprisedb.com","msg_date":"Wed, 16 Mar 2022 01:03:38 +0530","msg_from":"davinder singh <davindersingh2692@gmail.com>","msg_from_op":true,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Wed, Mar 16, 2022 at 01:03:38AM +0530, davinder singh wrote:\n> Regarding the 2nd part of configuring the threshold, Based on our\n> experiments, we have fixed it for the attributes with size > 2 * chunk_size.\n> The default chunk_size is 2KB and the page size is 8KB. While toasting each\n> attribute is divided into chunks, and each page can hold a max of 4 such\n> chunks.\n> We only need to think about the space used by the last chunk of the\n> attribute.\n> This means with each value optimization, it might use extra space in the\n> range\n> (0B,2KB]. I think this extra space is independent of attribute size. So we\n> don't\n> need to worry about configuring this threshold. Let me know if I missed\n> something\n> here.\n\nLooking closer, ISTM there are two thresholds here. The first threshold is\nthe minimum size of the uncompressed data required to trigger this\noptimization. The purpose behind this threshold is to avoid increasing\ndisk space usage in return for very little gain in the read path (i.e.,\nskipping decompression). Specifically, this threshold helps avoid\nincreasing disk space usage when there are many small TOAST attributes.\nYou've chosen 2x the TOAST chunk size as the threshold.\n\nThe second threshold is the compression ratio required to trigger this\noptimization. The purpose behind this threshold is to avoid storing poorly\ncompressed data so that we can improve performance in the read path (i.e.,\nskipping decompression). You've chosen >=1 TOAST chunk as the threshold.\nThis means that when I have an attribute with size 3x a TOAST chunk, I'll\nonly store it compressed when the compressed attribute is 2x a TOAST chunk\nor smaller (a 33% improvement). If I have an attribute with size 100x a\nTOAST chunk, I'll store it compressed when the compressed attribute is 99x\na TOAST chunk or smaller (a 1% improvement).\n\nIIUC this means that it is much more likely that you will store a\ncompressed attribute when the uncompressed data is larger. The benefit of\nthis approach is that you know you'll never give up more than one TOAST\nchunk of disk space in the name of faster reads. The downside is that a\nhuge number of smaller attributes can require much more disk space than if\nthey were universally compressed.\n\nPerhaps this is missing one additional threshold. Perhaps there should be\na compression ratio for which the optimization is ignored. If my\nuncompressed data is 3x a TOAST chunk and compressing it would save half of\na TOAST chunk of disk space (a ~17% improvement), maybe we should still\nstore it compressed. I think this threshold would be most important for\nservers with many small TOAST attributes. I don't know whether it is\nnecessary to make all of this configurable or whether we should just use\nsome heuristics to find a good balance. In any case, I imagine the\nsettings should look different for the different TOAST compression methods.\n\nOr maybe we should just make the first threshold the configuration option\nfor this feature. If you set it to -1, the optimization is never used. If\nyou set it to 0, you always try to save on read time at the expense of disk\nspace. If you set it to 2x the TOAST chunk size, you might use up to 50%\nmore disk space for read path optimizations. If you set it to 100x the\nTOAST chunk size, you might use up to 1% more disk space for read\noptimizations. By default, this configuration option could be set\nconservatively (e.g., 10x a TOAST chunk, which could lead to a maximum of\n10% more disk space). I suspect the increase in disk space would be much\nless than these numbers in most cases, but it might be helpful to think of\nthese things in terms of the worst case scenario.\n\nI apologize for thinking out loud a bit here, but I hope this gives you\nsome insight into my perspective on this. In general, I am skeptical that\nwe can choose one threshold that will work for all PostgreSQL installations\nin the known universe.\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Tue, 15 Mar 2022 14:48:13 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Tue, Mar 15, 2022 at 5:48 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\n> I apologize for thinking out loud a bit here, but I hope this gives you\n> some insight into my perspective on this. In general, I am skeptical that\n> we can choose one threshold that will work for all PostgreSQL installations\n> in the known universe.\n\nI would like to take a slightly contrary position. I think that a\nsystem here that involves multiple knobs is going to be too\ncomplicated to be of any real-world use, because almost nobody will\nunderstand it or tune it properly for their installation. And who is\nto say that a single setting would even be appropriate for a whole\ninstallation, as opposed to something that needs to be tuned for each\nindividual table? A single tunable might be OK, but what I would\nreally like here is a heuristic that, while perhaps not optimal in\nevery environment, is good enough that a very high percentage of users\nwill never need to worry about it. In a case like this, a small gain\nthat happens automatically feels better than a large gain that\nrequires manual tweaking in every install.\n\nNow that doesn't answer the question of what that heuristic should be.\nI initially thought that if compression didn't end up reducing the\nnumber of TOAST chunks then there was no point, but that's not true,\nbecause having the last chunk be smaller than the rest can allow us to\nfit more than 4 into a page rather than exactly 4. However, this\neffect isn't likely to be important if most chunks are full-size\nchunks. If we insert 100 full-size chunks and 1 partial chunk at the\nend, we can't save much even if that chunk ends up being 1 byte\ninstead of a full-size chunk. 25 TOAST table pages out of 26 are going\nto be full of full-size chunks either way, so we can't save more than\n~4% and we might easily save nothing at all. As you say, the potential\nsavings are larger as the values get smaller, because a higher\nproportion of the TOAST table pages are not quite full. So I think the\nonly cases where it's even interesting to think about suppressing this\noptimization are those where the value is quite small -- and maybe the\nthing to do is just pick a conservatively large threshold and call it\ngood.\n\nFor example, say that instead of applying this technique when there\nare at least 2 TOAST chunks, we apply it when there are at least 500\n(i.e. normally 1MB). It's hard for me to imagine that we can ever\nlose, and in fact I think we could come down quite a bit from there\nand still end up winning pretty much all the time. Setting the\nthreshold to, say, 50 or 100 or 150 TOAST chunks instead of 2 may\nleave some money on the table, but if it means that we don't have\nmeaningful downsides and we don't have tunables for users to fiddle\nwith, it might be worth it.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 16 Mar 2022 10:08:45 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Wed, Mar 16, 2022 at 10:08:45AM -0400, Robert Haas wrote:\n> I would like to take a slightly contrary position. I think that a\n> system here that involves multiple knobs is going to be too\n> complicated to be of any real-world use, because almost nobody will\n> understand it or tune it properly for their installation. And who is\n> to say that a single setting would even be appropriate for a whole\n> installation, as opposed to something that needs to be tuned for each\n> individual table? A single tunable might be OK, but what I would\n> really like here is a heuristic that, while perhaps not optimal in\n> every environment, is good enough that a very high percentage of users\n> will never need to worry about it. In a case like this, a small gain\n> that happens automatically feels better than a large gain that\n> requires manual tweaking in every install.\n\nAgreed. If there is a tunable, it needs to be as simple as possible. And\nideally most users would never need to use it because the default would be\n\"good enough.\"\n\n> Now that doesn't answer the question of what that heuristic should be.\n> I initially thought that if compression didn't end up reducing the\n> number of TOAST chunks then there was no point, but that's not true,\n> because having the last chunk be smaller than the rest can allow us to\n> fit more than 4 into a page rather than exactly 4. However, this\n> effect isn't likely to be important if most chunks are full-size\n> chunks. If we insert 100 full-size chunks and 1 partial chunk at the\n> end, we can't save much even if that chunk ends up being 1 byte\n> instead of a full-size chunk. 25 TOAST table pages out of 26 are going\n> to be full of full-size chunks either way, so we can't save more than\n> ~4% and we might easily save nothing at all. As you say, the potential\n> savings are larger as the values get smaller, because a higher\n> proportion of the TOAST table pages are not quite full. So I think the\n> only cases where it's even interesting to think about suppressing this\n> optimization are those where the value is quite small -- and maybe the\n> thing to do is just pick a conservatively large threshold and call it\n> good.\n> \n> For example, say that instead of applying this technique when there\n> are at least 2 TOAST chunks, we apply it when there are at least 500\n> (i.e. normally 1MB). It's hard for me to imagine that we can ever\n> lose, and in fact I think we could come down quite a bit from there\n> and still end up winning pretty much all the time. Setting the\n> threshold to, say, 50 or 100 or 150 TOAST chunks instead of 2 may\n> leave some money on the table, but if it means that we don't have\n> meaningful downsides and we don't have tunables for users to fiddle\n> with, it might be worth it.\n\nAs long as we can demonstrate some decent improvements, I think using a\nconservative threshold is a good idea. I do wonder whether thiѕ could\nweaken the read performance gains quite a bit, though.\n\nThinking further, is simply reducing the number of TOAST chunks the right\nthing to look at? If I want to add a TOAST attribute that requires 100,000\nchunks, and you told me that I could save 10% in the read path for an extra\n250 chunks of disk space, I would probably choose read performance. If I\nwanted to add 100,000 attributes that were each 3 chunks, and you told me\nthat I could save 10% in the read path for an extra 75,000 chunks of disk\nspace, I might choose the extra disk space. These are admittedly extreme\n(and maybe even impossible) examples, but my point is that the amount of\ndisk space you are willing to give up may be related to the size of the\nattribute. And maybe one way to extract additional read performance with\nthis optimization is to use a variable threshold so that we are more likely\nto use it for large attributes.\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Wed, 16 Mar 2022 11:36:56 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Wed, Mar 16, 2022 at 11:36:56AM -0700, Nathan Bossart wrote:\n> Thinking further, is simply reducing the number of TOAST chunks the right\n> thing to look at? If I want to add a TOAST attribute that requires 100,000\n> chunks, and you told me that I could save 10% in the read path for an extra\n> 250 chunks of disk space, I would probably choose read performance. If I\n> wanted to add 100,000 attributes that were each 3 chunks, and you told me\n> that I could save 10% in the read path for an extra 75,000 chunks of disk\n> space, I might choose the extra disk space. These are admittedly extreme\n> (and maybe even impossible) examples, but my point is that the amount of\n> disk space you are willing to give up may be related to the size of the\n> attribute. And maybe one way to extract additional read performance with\n> this optimization is to use a variable threshold so that we are more likely\n> to use it for large attributes.\n\nI might be overthinking this. Maybe it is enough to skip compressing the\nattribute whenever compression saves no more than some percentage of the\nuncompressed attribute size. A conservative default setting might be\nsomething like 5% or 10%.\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Wed, 16 Mar 2022 11:46:37 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Wed, Mar 16, 2022 at 2:36 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\n> Thinking further, is simply reducing the number of TOAST chunks the right\n> thing to look at? If I want to add a TOAST attribute that requires 100,000\n> chunks, and you told me that I could save 10% in the read path for an extra\n> 250 chunks of disk space, I would probably choose read performance. If I\n> wanted to add 100,000 attributes that were each 3 chunks, and you told me\n> that I could save 10% in the read path for an extra 75,000 chunks of disk\n> space, I might choose the extra disk space. These are admittedly extreme\n> (and maybe even impossible) examples, but my point is that the amount of\n> disk space you are willing to give up may be related to the size of the\n> attribute. And maybe one way to extract additional read performance with\n> this optimization is to use a variable threshold so that we are more likely\n> to use it for large attributes.\n\nRight, so perhaps the ultimate thing here would be a more fine-grained\nknob than SET STORAGE EXTERNAL -- something that allows you to specify\nthat you want to compress only when it really helps. While some people\nmight find that useful, I think the current patch is less ambitious,\nand I think that's OK. It just wants to save something in the cases\nwhere it's basically free. Unfortunately we've learned that it's never\n*entirely* free because making the last TOAST chunk longer can always\ncost you something, even if it gets longer by only 1 byte. But for\nlarger values it's hard for that to be significant.\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Thu, 17 Mar 2022 13:04:17 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Thu, Mar 17, 2022 at 01:04:17PM -0400, Robert Haas wrote:\n> Right, so perhaps the ultimate thing here would be a more fine-grained\n> knob than SET STORAGE EXTERNAL -- something that allows you to specify\n> that you want to compress only when it really helps. While some people\n> might find that useful, I think the current patch is less ambitious,\n> and I think that's OK. It just wants to save something in the cases\n> where it's basically free. Unfortunately we've learned that it's never\n> *entirely* free because making the last TOAST chunk longer can always\n> cost you something, even if it gets longer by only 1 byte. But for\n> larger values it's hard for that to be significant.\n\nI guess I think we should be slightly more ambitious. One idea could be to\ncreate a default_toast_compression_ratio GUC with a default of 0.95. This\nmeans that, by default, a compressed attribute must be 0.95x or less of the\nsize of the uncompressed attribute to be stored compressed. Like\ndefault_toast_compression, this could also be overridden at the column\nlevel with something like\n\n\tALTER TABLE mytbl ALTER mycol SET COMPRESSION lz4 RATIO 0.9;\n\nIf the current \"basically free\" patch is intended for v15, then maybe all\nthis extra configurability stuff could wait for a bit, especially if we can\ndemonstrate a decent read performance boost with a more conservative\nsetting. However, I don't see anything terribly complicated about the\nproposed configurability changes (assuming my proposal makes some amount of\nsense to you and others).\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Thu, 17 Mar 2022 13:05:23 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Thu, Mar 17, 2022 at 4:05 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\n> On Thu, Mar 17, 2022 at 01:04:17PM -0400, Robert Haas wrote:\n> > Right, so perhaps the ultimate thing here would be a more fine-grained\n> > knob than SET STORAGE EXTERNAL -- something that allows you to specify\n> > that you want to compress only when it really helps. While some people\n> > might find that useful, I think the current patch is less ambitious,\n> > and I think that's OK. It just wants to save something in the cases\n> > where it's basically free. Unfortunately we've learned that it's never\n> > *entirely* free because making the last TOAST chunk longer can always\n> > cost you something, even if it gets longer by only 1 byte. But for\n> > larger values it's hard for that to be significant.\n>\n> I guess I think we should be slightly more ambitious. One idea could be to\n> create a default_toast_compression_ratio GUC with a default of 0.95. This\n> means that, by default, a compressed attribute must be 0.95x or less of the\n> size of the uncompressed attribute to be stored compressed. Like\n> default_toast_compression, this could also be overridden at the column\n> level with something like\n>\n> ALTER TABLE mytbl ALTER mycol SET COMPRESSION lz4 RATIO 0.9;\n>\n> If the current \"basically free\" patch is intended for v15, then maybe all\n> this extra configurability stuff could wait for a bit, especially if we can\n> demonstrate a decent read performance boost with a more conservative\n> setting. However, I don't see anything terribly complicated about the\n> proposed configurability changes (assuming my proposal makes some amount of\n> sense to you and others).\n\nWe seem to have a shortage of \"others\" showing up with opinions on\nthis topic, but I guess I'm not very confident about the general\nutility of such a setting. Just to be clear, I'm also not very\nconfident about the usefulness of the existing settings for\ncontrolling TOAST. Why is it useful default behavior to try to get\nrows down to 2kB by default, rather than 1.8kB or 3kB? Even more, why\ndon't we try to compress primarily based on the length of individual\nattributes and then compress further only if the resulting tuple\ndoesn't fit into a page at all? There doesn't seem to be anything\nmagic about fitting tuples into a quarter-page, yet the code acts as\nthough that's the primary goal - and then, when that didn't turn out\nto work well in all cases, we added a user-settable parameter\n(toast_tuple_target) to let you say you really want tuples in table X\nto fit into a third of a page or a fifth of a page instead of a\nquarter. And there's some justification for that: a proposal to\nfundamentally change the algorithm would likely have gotten bogged\ndown for years, and the parameter no doubt lets you solve some\nproblems. Yet if the whole algorithm is wrong, and I think maybe it\nis, then being able to change the constants is not really getting us\nwhere we need to be.\n\nThen, too, I'm not very confident about the usefulness of EXTENDED,\nEXTERNAL, and MAIN. I think it's useful to be able to categorically\ndisable compression (as EXTERNAL does), but you can't categorically\ndisable out-of-line storage because the value can be bigger than the\npage, so MAIN vs. EXTENDED is just changing the threshold for the use\nof out-of-line storage. However, it does so in a way that IMHO is not\nparticularly intuitive, which goes back to my earlier point about the\nalgorithm seeming wacky, and it's in any case unclear why we should\noffer exactly two possible thresholds and not any of the intermediate\nvalues.\n\nI think I'm prepared to concede that the setting you are proposing\nhere is easier to understand than either of those. Saying \"store this\nvalue compressed only if you an thereby reduce the size by at least\n10%\" is an understandable directive which I think will make sense even\nto people who know very little about PostgreSQL or databases\ngenerally, as long as they have some vague idea how compression works.\nHowever, I guess I am not confident that the setting would get much\nuse, or that it is the best way to measure what we care about. For\ninstance, if my value is 19kB in length, saving 10% isn't very\nimpactful, because it's still the same number of chunks, and the only\nbenefit I gain is that the last chunk is smaller, which is probably\nnot very impactful because I'm still going to have 8 full chunks. I\nwill gain more if the value is bigger or smaller. Going up, if I\nincrease the size to 19MB or 190MB, I'm really saving a very\nsignificant amount of cache space and, potentially, I/O. If I was\nhappy with a 10% savings at 19kB, I will likely be happy with a\nconsiderably smaller economy here. Going down, if I decrease the size\nto 1.9kB, I'm not storing any full-sized chunks any more, so reducing\nthe size of the one partial chunk that I'm storing seems like it could\nhave a real impact, provided of course that other toasted values in\nthe same table are similarly small. If the whole TOAST table is full\nof just partial chunks, then the smaller they are the more will fit\ninto a page, all things being equal, and thus the smaller the TOAST\ntable will be.\n\nMaybe the conclusion here is that more thought is needed before\nchanging anything in this area....\n\n-- \nRobert Haas\nEDB: http://www.enterprisedb.com\n\n\n","msg_date":"Tue, 22 Mar 2022 16:34:05 -0400","msg_from":"Robert Haas <robertmhaas@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Tue, Mar 22, 2022 at 04:34:05PM -0400, Robert Haas wrote:\n> We seem to have a shortage of \"others\" showing up with opinions on\n> this topic, but I guess I'm not very confident about the general\n> utility of such a setting. Just to be clear, I'm also not very\n> confident about the usefulness of the existing settings for\n> controlling TOAST. Why is it useful default behavior to try to get\n> rows down to 2kB by default, rather than 1.8kB or 3kB? Even more, why\n> don't we try to compress primarily based on the length of individual\n> attributes and then compress further only if the resulting tuple\n> doesn't fit into a page at all? There doesn't seem to be anything\n> magic about fitting tuples into a quarter-page, yet the code acts as\n> though that's the primary goal - and then, when that didn't turn out\n> to work well in all cases, we added a user-settable parameter\n> (toast_tuple_target) to let you say you really want tuples in table X\n> to fit into a third of a page or a fifth of a page instead of a\n> quarter. And there's some justification for that: a proposal to\n> fundamentally change the algorithm would likely have gotten bogged\n> down for years, and the parameter no doubt lets you solve some\n> problems. Yet if the whole algorithm is wrong, and I think maybe it\n> is, then being able to change the constants is not really getting us\n> where we need to be.\n> \n> Then, too, I'm not very confident about the usefulness of EXTENDED,\n> EXTERNAL, and MAIN. I think it's useful to be able to categorically\n> disable compression (as EXTERNAL does), but you can't categorically\n> disable out-of-line storage because the value can be bigger than the\n> page, so MAIN vs. EXTENDED is just changing the threshold for the use\n> of out-of-line storage. However, it does so in a way that IMHO is not\n> particularly intuitive, which goes back to my earlier point about the\n> algorithm seeming wacky, and it's in any case unclear why we should\n> offer exactly two possible thresholds and not any of the intermediate\n> values.\n\nI agree with all of this. Adding configurability for each constant might\nhelp folks in the short term, but using these knobs probably requires quite\na bit of expertise in Postgres internals as well as a good understanding of\nyour data. I think we ought to revist TOAST configurability from a user\nperspective. IOW what can be chosen automatically, and how do we enable\nusers to effectively configure the settings that cannot be chosen\nautomatically? IMO this is a worthwhile conversation to have as long as it\ndoesn't stray too far into the \"let's rewrite TOAST\" realm. I think there\nis always going to be some level of complexity with stuff like TOAST, but\nthere are probably all sorts of ways to simplify/improve it also.\n\n> Maybe the conclusion here is that more thought is needed before\n> changing anything in this area....\n\nYou've certainly got me thinking more about this. If the scope of this\nwork is going to expand, a few months before the first v16 commitfest is\nprobably the right time for it.\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Tue, 22 Mar 2022 14:42:53 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Tue, Mar 22, 2022 at 02:42:53PM -0700, Nathan Bossart wrote:\n> On Tue, Mar 22, 2022 at 04:34:05PM -0400, Robert Haas wrote:\n>> Then, too, I'm not very confident about the usefulness of EXTENDED,\n>> EXTERNAL, and MAIN. I think it's useful to be able to categorically\n>> disable compression (as EXTERNAL does), but you can't categorically\n>> disable out-of-line storage because the value can be bigger than the\n>> page, so MAIN vs. EXTENDED is just changing the threshold for the use\n>> of out-of-line storage. However, it does so in a way that IMHO is not\n>> particularly intuitive, which goes back to my earlier point about the\n>> algorithm seeming wacky, and it's in any case unclear why we should\n>> offer exactly two possible thresholds and not any of the intermediate\n>> values.\n> \n> I agree with all of this. Adding configurability for each constant might\n> help folks in the short term, but using these knobs probably requires quite\n> a bit of expertise in Postgres internals as well as a good understanding of\n> your data. I think we ought to revist TOAST configurability from a user\n> perspective. IOW what can be chosen automatically, and how do we enable\n> users to effectively configure the settings that cannot be chosen\n> automatically? IMO this is a worthwhile conversation to have as long as it\n> doesn't stray too far into the \"let's rewrite TOAST\" realm. I think there\n> is always going to be some level of complexity with stuff like TOAST, but\n> there are probably all sorts of ways to simplify/improve it also.\n\nI agree with this feeling. TOAST has already too many configuration\nparameters that have their own way of behaving slightling differently.\nIf we could reduce this number rather than increasing it, the better.\nI would be also curious to see how much those parameters become\nrelevant with more compression options possible with toast values.\n--\nMichael","msg_date":"Wed, 23 Mar 2022 13:28:46 +0900","msg_from":"Michael Paquier <michael@paquier.xyz>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Wed, Mar 23, 2022 at 9:58 AM Michael Paquier <michael@paquier.xyz> wrote:\n>\n> On Tue, Mar 22, 2022 at 02:42:53PM -0700, Nathan Bossart wrote:\n> > On Tue, Mar 22, 2022 at 04:34:05PM -0400, Robert Haas wrote:\n> >> Then, too, I'm not very confident about the usefulness of EXTENDED,\n> >> EXTERNAL, and MAIN. I think it's useful to be able to categorically\n> >> disable compression (as EXTERNAL does), but you can't categorically\n> >> disable out-of-line storage because the value can be bigger than the\n> >> page, so MAIN vs. EXTENDED is just changing the threshold for the use\n> >> of out-of-line storage. However, it does so in a way that IMHO is not\n> >> particularly intuitive, which goes back to my earlier point about the\n> >> algorithm seeming wacky, and it's in any case unclear why we should\n> >> offer exactly two possible thresholds and not any of the intermediate\n> >> values.\n> >\n> > I agree with all of this. Adding configurability for each constant might\n> > help folks in the short term, but using these knobs probably requires quite\n> > a bit of expertise in Postgres internals as well as a good understanding of\n> > your data. I think we ought to revist TOAST configurability from a user\n> > perspective. IOW what can be chosen automatically, and how do we enable\n> > users to effectively configure the settings that cannot be chosen\n> > automatically? IMO this is a worthwhile conversation to have as long as it\n> > doesn't stray too far into the \"let's rewrite TOAST\" realm. I think there\n> > is always going to be some level of complexity with stuff like TOAST, but\n> > there are probably all sorts of ways to simplify/improve it also.\n>\n> I agree with this feeling. TOAST has already too many configuration\n> parameters that have their own way of behaving slightling differently.\n> If we could reduce this number rather than increasing it, the better.\n> I would be also curious to see how much those parameters become\n> relevant with more compression options possible with toast values.\n\nAgree with you, it doesn't make much sense here to add an extra user\nvisible configuration parameter, where it is not really clear for\nusers how to really configure it to get the best out of it.\n\n-- \nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 23 Mar 2022 15:35:15 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"},{"msg_contents":"On Fri, Mar 18, 2022 at 1:35 AM Nathan Bossart <nathandbossart@gmail.com> wrote:\n>\n>\n> I guess I think we should be slightly more ambitious. One idea could be to\n> create a default_toast_compression_ratio GUC with a default of 0.95. This\n> means that, by default, a compressed attribute must be 0.95x or less of the\n> size of the uncompressed attribute to be stored compressed. Like\n> default_toast_compression, this could also be overridden at the column\n> level with something like\n\nI am not sure that we want a GUC to control that but we can certainly\nbe more ambitious. Basically, in the current patch if data is\nslightly large then we would always prefer to store the compressed\ndata, e.g. if the data size is 200kB then even if the compression\nratio is as low as 1% then we would choose to store then compressed\ndata.\n\nI think we can make it based on the compression ratio and then upper\nbound it with the number of chunk differences. For example if the\ncompression ratio < 10% then stored it uncompressed iff the chunk\ndifference < threshold. But with that we might see performance impact\non the smaller data which has a compressed ratio < 10% because their\nchunk difference will always be under the threshold. So maybe the\nchunk difference threshold can be a function based on the total\nnumbers of chunks required for the data, maybe a logarithmic function\nso that the threshold grows slowly along with the base data size.\n\n--\nRegards,\nDilip Kumar\nEnterpriseDB: http://www.enterprisedb.com\n\n\n","msg_date":"Wed, 23 Mar 2022 16:40:37 +0530","msg_from":"Dilip Kumar <dilipbalaut@gmail.com>","msg_from_op":false,"msg_subject":"Re: Optimize external TOAST storage"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nFor Toast storage [1] in PostgreSQL, first, the attribute value is\\ncompressed\\nand then divided into chunks. The problem with storing compressed value is,\\nif we\\nare not saving enough space such that it reduces the #chunks then we end up\\nadding extra decompression cost on every read.\\nBased on the discussion with Robert and Dilip, we tried to optimize this\\nprocess. The idea is to TOAST the compressed value only if we are saving at\\nleast\\n1 chunk(2KB default) of disk storage else use the uncompressed one. In this\\nway,\\nwe will save decompression costs without losing much on storage.\\n\\nIn our tests, we have observed improvement in the read performance up to\\n28% by\\ngiving up at most TOAST_MAX_CHUNK_SIZE (2KB) bytes for storage on disk. The\\ngain is more on large attributes (size > 4KB) because\\ncompression/decompression\\ncost increases with size.\\nHowever, We have found, this assumption is not true when the data\\ncompression\\nratio is more and the size is barely big enough to cause TOASTing. For\\nexample,\\nin the following test 4. b, we didn't get any performance advantage but the\\ntable\\nsize grew by 42% by storing uncompressed values.\\nTest Setup.\\n\\nCreate table t1_lz4 ( a text compression lz4, b text compression lz4);\\n-- Generate random data\\ncreate or replace function generate_att_data(len_info int)\\nreturns text\\nlanguage plpgsql\\nas\\n$$\\ndeclare\\n value text;\\nbegin\\n select array_agg(md5(g::text))\\n into value\\n from generate_series(1, round(len_info/33)::int) g;\\n return value;\\nend;\\n$$;\\n\\n--Test\\nSelect b from t1_lz4;\\n\\nTest 1:\\nData: rows 200000\\ninsert into t1_lz4(a, b) select generate_att_data(364), repeat\\n(generate_att_data(1980), 2);\\nSummary:\\nAttribute size: original: 7925 bytes, after compression: 7845 bytes\\nTime for select: head: 42 sec, patch: 37 sec, *Performance Gain: 11%*\\ntable size: Head 1662 MB, Patch: 1662 MB\\n\\nTest 2:\\nData: rows 100000\\ninsert into t1_lz4(a, b) select generate_att_data(364),\\ngenerate_att_data(16505);\\nSummary:\\nAttribute size: original: 16505 bytes, after compression: 16050 bytes\\nTime for select: head: 35.6 sec, patch: 30 sec, *Performance Gain: 14%*\\ntable size: Head 1636 MB, Patch: 1688 MB\\n\\nTest 3:\\nData: rows 50000\\ninsert into t1_lz4(a, b) select generate_att_data(364),\\ngenerate_att_data(31685);\\nSummary:\\nAttribute size: original: 31685 bytes, after compression: 30263 bytes\\nTime for select: head: 35.4 sec, patch: 25.5 sec, *Performance Gain: 28%*\\ntable size: Head 1601 MB, Patch: 1601 MB\\n\\nTest 4.a:\\nData: rows 200000\\ninsert into t1_lz4(a, b) select generate_att_data(11), repeat ('b', 250)\\n|| generate_att_data(3885);\\nSummary:\\nAttribute size: original: 3885 bytes, after compression: 3645 bytes\\nTime for select: head: 28 sec, patch: 26 sec, *Performance Gain: 7%*\\n*table size: Head 872 MB, Patch: 872 MB*\\n\\nTest 4.b (High compression):\\nData: rows 200000\\ninsert into t1_lz4(a, b) select generate_att_data(364), repeat\\n(generate_att_data(1980), 2);\\nSummary:\\nAttribute size: original: 3966 bytes, after compression: 2012 bytes\\nTime for select: head: 27 sec, patch: 26 sec, *Performance Gain: 0%*\\n*table size: Head 612 MB, Patch: 872 MB*\\n\\nThis is the worst case for this optimization because of the following 2\\nreasons.\\nFirst, the table size would grow by 50% when compressed size is half of the\\noriginal size and\\nyet barely large enough for TOASTing. Table size can't grow more than that\\nbecause If\\ncompression reduces the size even more then it will reduce the #chunks as\\nwell and it stores\\nthe compressed value in the table.\\n\\nSecond, not much gain in performance because of the small attribute size,\\nmore attributes fit\\n\\nin page (almost twice), on each page access it can access twice the number\\nof rows. And\\n\\nalso small value means low compression/decompression costs.\\n\\n\\nWe have avoided such cases by applying the optimization when attribute size\\n> 4KB.\\n\\nRegards,\\nDavinder\\nEnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Mon, 28 Feb 2022 15:22:11 +0530\",\n \"msg_from\": \"davinder singh <davindersingh2692@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 3:22 PM davinder singh\\n<davindersingh2692@gmail.com> wrote:\\n>\\n> Hi,\\n>\\n> For Toast storage [1] in PostgreSQL, first, the attribute value is compressed\\n> and then divided into chunks. The problem with storing compressed value is, if we\\n> are not saving enough space such that it reduces the #chunks then we end up\\n> adding extra decompression cost on every read.\\n> Based on the discussion with Robert and Dilip, we tried to optimize this\\n> process. The idea is to TOAST the compressed value only if we are saving at least\\n> 1 chunk(2KB default) of disk storage else use the uncompressed one. In this way,\\n> we will save decompression costs without losing much on storage.\\n>\\n> In our tests, we have observed improvement in the read performance up to 28% by\\n> giving up at most TOAST_MAX_CHUNK_SIZE (2KB) bytes for storage on disk. The\\n> gain is more on large attributes (size > 4KB) because compression/decompression\\n> cost increases with size.\\n> However, We have found, this assumption is not true when the data compression\\n> ratio is more and the size is barely big enough to cause TOASTing. For example,\\n> in the following test 4. b, we didn't get any performance advantage but the table\\n> size grew by 42% by storing uncompressed values.\\n> Test Setup.\\n>\\n> Create table t1_lz4 ( a text compression lz4, b text compression lz4);\\n> -- Generate random data\\n> create or replace function generate_att_data(len_info int)\\n> returns text\\n> language plpgsql\\n> as\\n> $$\\n> declare\\n> value text;\\n> begin\\n> select array_agg(md5(g::text))\\n> into value\\n> from generate_series(1, round(len_info/33)::int) g;\\n> return value;\\n> end;\\n> $$;\\n>\\n> --Test\\n> Select b from t1_lz4;\\n>\\n> Test 1:\\n> Data: rows 200000\\n> insert into t1_lz4(a, b) select generate_att_data(364), repeat (generate_att_data(1980), 2);\\n> Summary:\\n> Attribute size: original: 7925 bytes, after compression: 7845 bytes\\n> Time for select: head: 42 sec, patch: 37 sec, Performance Gain: 11%\\n> table size: Head 1662 MB, Patch: 1662 MB\\n>\\n> Test 2:\\n> Data: rows 100000\\n> insert into t1_lz4(a, b) select generate_att_data(364), generate_att_data(16505);\\n> Summary:\\n> Attribute size: original: 16505 bytes, after compression: 16050 bytes\\n> Time for select: head: 35.6 sec, patch: 30 sec, Performance Gain: 14%\\n> table size: Head 1636 MB, Patch: 1688 MB\\n>\\n> Test 3:\\n> Data: rows 50000\\n> insert into t1_lz4(a, b) select generate_att_data(364), generate_att_data(31685);\\n> Summary:\\n> Attribute size: original: 31685 bytes, after compression: 30263 bytes\\n> Time for select: head: 35.4 sec, patch: 25.5 sec, Performance Gain: 28%\\n> table size: Head 1601 MB, Patch: 1601 MB\\n>\\n> Test 4.a:\\n> Data: rows 200000\\n> insert into t1_lz4(a, b) select generate_att_data(11), repeat ('b', 250) || generate_att_data(3885);\\n> Summary:\\n> Attribute size: original: 3885 bytes, after compression: 3645 bytes\\n> Time for select: head: 28 sec, patch: 26 sec, Performance Gain: 7%\\n> table size: Head 872 MB, Patch: 872 MB\\n>\\n> Test 4.b (High compression):\\n> Data: rows 200000\\n> insert into t1_lz4(a, b) select generate_att_data(364), repeat (generate_att_data(1980), 2);\\n> Summary:\\n> Attribute size: original: 3966 bytes, after compression: 2012 bytes\\n> Time for select: head: 27 sec, patch: 26 sec, Performance Gain: 0%\\n> table size: Head 612 MB, Patch: 872 MB\\n\\nI think the idea looks interesting and the results are also promising.\\nI have a few initial comments, later I will do more detailed review.\\n\\n1.\\n+ if (*value == orig_toast_value)\\n+ toast_tuple_externalize(ttc, attribute, options);\\n\\nIsn't it looks cleaner to check whther the attribute is compressed or\\nnot like this\\nVARATT_IS_COMPRESSED(DatumGetPointer(*value)) ?\\n\\n2.\\n+ /* Using the uncompressed data instead, deleting\\ncompressed data. */\\n+ pfree(DatumGetPointer(*value));\\n\\nchange as below\\n/deleting compressed data/release memory for the compressed data\\n\\n3.\\n+ memcpy(toast_attr_copy, toast_attr, sizeof(toast_attr));\\n+ memcpy(toast_values_copy, toast_values, sizeof(toast_values));\\n+\\n\\nAdd some comment here, what we are trying to copy and why?\\n\\n4.\\n+ /* incompressible, ignore on subsequent compression passes. */\\n+ orig_attr->tai_colflags |= TOASTCOL_INCOMPRESSIBLE;\\n\\nDo we need to set TOASTCOL_INCOMPRESSIBLE while trying to externalize\\nit, the comment say \\\"ignore on subsequent compression passes\\\"\\nbut after this will there be more compression passes? If we need to\\nset this TOASTCOL_INCOMPRESSIBLE then comment should explain this.\\n\\n\\n\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Mar 2022 21:44:41 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"Thanks Dilip, I have fixed your comments, please find the updated patch.\\n\\nOn Tue, Mar 8, 2022 at 9:44 PM Dilip Kumar <dilipbalaut@gmail.com> wrote:.\\n\\n> + /* incompressible, ignore on subsequent compression passes. */\\n> + orig_attr->tai_colflags |= TOASTCOL_INCOMPRESSIBLE;\\n>\\n> Do we need to set TOASTCOL_INCOMPRESSIBLE while trying to externalize\\n> it, the comment say \\\"ignore on subsequent compression passes\\\"\\n> but after this will there be more compression passes? If we need to\\n\\nset this TOASTCOL_INCOMPRESSIBLE then comment should explain this.\\n>\\nThat was a mistake, this flag is not required at this point, as the\\nattribute is externalized it will be marked as TOASTCOL_IGNORE, and\\nsuch columns are not considered for compression, I removed it. Thanks for\\npointing it out.\\n\\n-- \\nRegards,\\nDavinder\\nEnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Thu, 10 Mar 2022 14:04:06 +0530\",\n \"msg_from\": \"davinder singh <davindersingh2692@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 10, 2022 at 2:04 PM davinder singh\\n<davindersingh2692@gmail.com> wrote:\\n>\\n> Thanks Dilip, I have fixed your comments, please find the updated patch.\\n>\\n\\nSome more comments\\n\\n+ /*\\n+ * For those cases where storing compressed data is not optimal,\\nWe will use\\n+ * this pointer copy for referring uncompressed data.\\n+ */\\n+ memcpy(toast_attr_copy, toast_attr, sizeof(toast_attr));\\n+ memcpy(toast_values_copy, toast_values, sizeof(toast_values));\\n\\nI think we can change the comments like below\\n\\n\\\"Preserve references to the original uncompressed data before\\nattempting the compression. So that during externalize if we decide\\nnot to store the compressed data then we have the original data with\\nus. For more details refer to comments atop\\ntoast_tuple_opt_externalize\\\".\\n\\n+/*\\n+ * Don't save compressed data to external storage unless it saves I/O while\\n+ * accessing the same data by reducing the number of chunks.\\n+ */\\n+void\\n+toast_tuple_opt_externalize(ToastTupleContext *ttc, int attribute, int options,\\n+ Datum orig_toast_value, ToastAttrInfo *orig_attr)\\n\\nI think these comments are not explaining what is the problem with\\nstoring the compressed data. So you need to expand this comment\\nsaying if it is not reducing the number or chunks then there is no\\npoint in storing the compressed data because then we will have\\nadditional decompression cost whenever we are fetching that data.\\n\\n\\n+\\n+ /* Sanity check: if data is not compressed then we can proceed as usual. */\\n+ if (!VARATT_IS_COMPRESSED(DatumGetPointer(*value)))\\n+ toast_tuple_externalize(ttc, attribute, options);\\n\\nI think we don't need \\\"Sanity check:\\\" here, the remaining part of the\\ncomment is fine.\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 10 Mar 2022 17:09:40 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"I think this is an interesting idea. My first thought is that it would be\\nnice if we didn't have to first compress the data to see whether we wanted\\nto store it compressed or not, but I don't think there is much we can do\\nabout that. In any case, we aren't adding much work in the write path\\ncompared to the potential savings in the read path, so that is probably\\nokay.\\n\\nDo you think it is worth making this configurable? I don't think it is\\noutside the realm of possibility for some users to care more about disk\\nspace than read performance. And maybe the threshold for this optimization\\ncould differ based on the workload.\\n\\n extern void toast_tuple_externalize(ToastTupleContext *ttc, int attribute,\\n int options);\\n+extern void toast_tuple_opt_externalize(ToastTupleContext *ttc, int attribute,\\n+ int options, Datum old_toast_value,\\n+ ToastAttrInfo *old_toast_attr);\\n\\nCould we bake this into toast_tuple_externalize() so that all existing\\ncallers benefit from this optimization? Is there a reason to export both\\nfunctions? Perhaps toast_tuple_externalize() should be renamed and made\\nstatic, and then this new function could be called\\ntoast_tuple_externalize() (which would be a wrapper around the internal\\nfunction).\\n\\n+ /* Sanity check: if data is not compressed then we can proceed as usual. */\\n+ if (!VARATT_IS_COMPRESSED(DatumGetPointer(*value)))\\n+ toast_tuple_externalize(ttc, attribute, options);\\n\\nWith a --enable-cassert build, this line causes assertion failures in the\\ncall to GetMemoryChunkContext() from pfree(). 'make check' is enough to\\nreproduce it. Specifically, it fails the following assertion:\\n\\n\\tAssertArg(MemoryContextIsValid(context));\\n\\nI didn't see an existing commitfest entry for this patch. I'd encourage\\nyou to add one in the July commitfest:\\n\\n\\thttps://commitfest.postgresql.org/38/\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Sat, 12 Mar 2022 14:13:47 -0800\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"Thanks, Nathan, for the review comments. Please find the updated patch.\\nOn Sun, Mar 13, 2022 at 3:43 AM Nathan Bossart <nathandbossart@gmail.com>\\nwrote:\\n\\n> Do you think it is worth making this configurable? I don't think it is\\n> outside the realm of possibility for some users to care more about disk\\n> space than read performance. And maybe the threshold for this optimization\\n> could differ based on the workload.\\n>\\n\\nI think here we can break it into two parts.\\nThe first part if the user cares about the disk more than reading\\nperformance, disable this?\\nThat is a good idea, we can try this, lets see what others say.\\n\\nRegarding the 2nd part of configuring the threshold, Based on our\\nexperiments, we have fixed it for the attributes with size > 2 * chunk_size.\\nThe default chunk_size is 2KB and the page size is 8KB. While toasting each\\nattribute is divided into chunks, and each page can hold a max of 4 such\\nchunks.\\nWe only need to think about the space used by the last chunk of the\\nattribute.\\nThis means with each value optimization, it might use extra space in the\\nrange\\n(0B,2KB]. I think this extra space is independent of attribute size. So we\\ndon't\\nneed to worry about configuring this threshold. Let me know if I missed\\nsomething\\nhere.\\n\\n\\n> extern void toast_tuple_externalize(ToastTupleContext *ttc, int attribute,\\n> int options);\\n> +extern void toast_tuple_opt_externalize(ToastTupleContext *ttc, int\\n> attribute,\\n> + int options, Datum\\n> old_toast_value,\\n> + ToastAttrInfo *old_toast_attr);\\n>\\n> Could we bake this into toast_tuple_externalize() so that all existing\\n> callers benefit from this optimization? Is there a reason to export both\\n> functions? Perhaps toast_tuple_externalize() should be renamed and made\\n> static, and then this new function could be called\\n> toast_tuple_externalize() (which would be a wrapper around the internal\\n> function).\\n>\\nThis function is used only in heap_toast_insert_or_update(), all existing\\ncallers are\\nusing new function only. As you suggested, I have renamed the new function\\nas\\nwrapper function and also only exporting the new function.\\n\\n\\n>\\n> + /* Sanity check: if data is not compressed then we can proceed as\\n> usual. */\\n> + if (!VARATT_IS_COMPRESSED(DatumGetPointer(*value)))\\n> + toast_tuple_externalize(ttc, attribute, options);\\n>\\n> With a --enable-cassert build, this line causes assertion failures in the\\n> call to GetMemoryChunkContext() from pfree(). 'make check' is enough to\\n> reproduce it. Specifically, it fails the following assertion:\\n>\\n> AssertArg(MemoryContextIsValid(context));\\n>\\n\\nThanks for pointing it out, this failure started because I was not handling\\nthe\\ncase when the data is already compressed even before toasting. Following\\ncheck verifies if the data is compressed or not but that is not enough\\nbecause\\nwe can't optimize the toasting if we didn't get the data in the\\nuncompressed form in the first place from the source.\\n+ /* If data is not compressed then we can proceed as usual. */\\n+ if (!VARATT_IS_COMPRESSED(DatumGetPointer(*value)))\\n\\nv1 patch didn't have this problem because it was verifying whether we have\\ncompressed data in this toasting round or not. I have changed back to the\\nearlier version.\\n+ /* If data is not compressed then we can proceed as usual. */\\n+ if (*value == orig_toast_value)\\n\\n\\n\\n-- \\nRegards,\\nDavinder\\nEnterpriseDB: http://www.enterprisedb.com\",\n \"msg_date\": \"Wed, 16 Mar 2022 01:03:38 +0530\",\n \"msg_from\": \"davinder singh <davindersingh2692@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 16, 2022 at 01:03:38AM +0530, davinder singh wrote:\\n> Regarding the 2nd part of configuring the threshold, Based on our\\n> experiments, we have fixed it for the attributes with size > 2 * chunk_size.\\n> The default chunk_size is 2KB and the page size is 8KB. While toasting each\\n> attribute is divided into chunks, and each page can hold a max of 4 such\\n> chunks.\\n> We only need to think about the space used by the last chunk of the\\n> attribute.\\n> This means with each value optimization, it might use extra space in the\\n> range\\n> (0B,2KB]. I think this extra space is independent of attribute size. So we\\n> don't\\n> need to worry about configuring this threshold. Let me know if I missed\\n> something\\n> here.\\n\\nLooking closer, ISTM there are two thresholds here. The first threshold is\\nthe minimum size of the uncompressed data required to trigger this\\noptimization. The purpose behind this threshold is to avoid increasing\\ndisk space usage in return for very little gain in the read path (i.e.,\\nskipping decompression). Specifically, this threshold helps avoid\\nincreasing disk space usage when there are many small TOAST attributes.\\nYou've chosen 2x the TOAST chunk size as the threshold.\\n\\nThe second threshold is the compression ratio required to trigger this\\noptimization. The purpose behind this threshold is to avoid storing poorly\\ncompressed data so that we can improve performance in the read path (i.e.,\\nskipping decompression). You've chosen >=1 TOAST chunk as the threshold.\\nThis means that when I have an attribute with size 3x a TOAST chunk, I'll\\nonly store it compressed when the compressed attribute is 2x a TOAST chunk\\nor smaller (a 33% improvement). If I have an attribute with size 100x a\\nTOAST chunk, I'll store it compressed when the compressed attribute is 99x\\na TOAST chunk or smaller (a 1% improvement).\\n\\nIIUC this means that it is much more likely that you will store a\\ncompressed attribute when the uncompressed data is larger. The benefit of\\nthis approach is that you know you'll never give up more than one TOAST\\nchunk of disk space in the name of faster reads. The downside is that a\\nhuge number of smaller attributes can require much more disk space than if\\nthey were universally compressed.\\n\\nPerhaps this is missing one additional threshold. Perhaps there should be\\na compression ratio for which the optimization is ignored. If my\\nuncompressed data is 3x a TOAST chunk and compressing it would save half of\\na TOAST chunk of disk space (a ~17% improvement), maybe we should still\\nstore it compressed. I think this threshold would be most important for\\nservers with many small TOAST attributes. I don't know whether it is\\nnecessary to make all of this configurable or whether we should just use\\nsome heuristics to find a good balance. In any case, I imagine the\\nsettings should look different for the different TOAST compression methods.\\n\\nOr maybe we should just make the first threshold the configuration option\\nfor this feature. If you set it to -1, the optimization is never used. If\\nyou set it to 0, you always try to save on read time at the expense of disk\\nspace. If you set it to 2x the TOAST chunk size, you might use up to 50%\\nmore disk space for read path optimizations. If you set it to 100x the\\nTOAST chunk size, you might use up to 1% more disk space for read\\noptimizations. By default, this configuration option could be set\\nconservatively (e.g., 10x a TOAST chunk, which could lead to a maximum of\\n10% more disk space). I suspect the increase in disk space would be much\\nless than these numbers in most cases, but it might be helpful to think of\\nthese things in terms of the worst case scenario.\\n\\nI apologize for thinking out loud a bit here, but I hope this gives you\\nsome insight into my perspective on this. In general, I am skeptical that\\nwe can choose one threshold that will work for all PostgreSQL installations\\nin the known universe.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 15 Mar 2022 14:48:13 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 15, 2022 at 5:48 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> I apologize for thinking out loud a bit here, but I hope this gives you\\n> some insight into my perspective on this. In general, I am skeptical that\\n> we can choose one threshold that will work for all PostgreSQL installations\\n> in the known universe.\\n\\nI would like to take a slightly contrary position. I think that a\\nsystem here that involves multiple knobs is going to be too\\ncomplicated to be of any real-world use, because almost nobody will\\nunderstand it or tune it properly for their installation. And who is\\nto say that a single setting would even be appropriate for a whole\\ninstallation, as opposed to something that needs to be tuned for each\\nindividual table? A single tunable might be OK, but what I would\\nreally like here is a heuristic that, while perhaps not optimal in\\nevery environment, is good enough that a very high percentage of users\\nwill never need to worry about it. In a case like this, a small gain\\nthat happens automatically feels better than a large gain that\\nrequires manual tweaking in every install.\\n\\nNow that doesn't answer the question of what that heuristic should be.\\nI initially thought that if compression didn't end up reducing the\\nnumber of TOAST chunks then there was no point, but that's not true,\\nbecause having the last chunk be smaller than the rest can allow us to\\nfit more than 4 into a page rather than exactly 4. However, this\\neffect isn't likely to be important if most chunks are full-size\\nchunks. If we insert 100 full-size chunks and 1 partial chunk at the\\nend, we can't save much even if that chunk ends up being 1 byte\\ninstead of a full-size chunk. 25 TOAST table pages out of 26 are going\\nto be full of full-size chunks either way, so we can't save more than\\n~4% and we might easily save nothing at all. As you say, the potential\\nsavings are larger as the values get smaller, because a higher\\nproportion of the TOAST table pages are not quite full. So I think the\\nonly cases where it's even interesting to think about suppressing this\\noptimization are those where the value is quite small -- and maybe the\\nthing to do is just pick a conservatively large threshold and call it\\ngood.\\n\\nFor example, say that instead of applying this technique when there\\nare at least 2 TOAST chunks, we apply it when there are at least 500\\n(i.e. normally 1MB). It's hard for me to imagine that we can ever\\nlose, and in fact I think we could come down quite a bit from there\\nand still end up winning pretty much all the time. Setting the\\nthreshold to, say, 50 or 100 or 150 TOAST chunks instead of 2 may\\nleave some money on the table, but if it means that we don't have\\nmeaningful downsides and we don't have tunables for users to fiddle\\nwith, it might be worth it.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 10:08:45 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 16, 2022 at 10:08:45AM -0400, Robert Haas wrote:\\n> I would like to take a slightly contrary position. I think that a\\n> system here that involves multiple knobs is going to be too\\n> complicated to be of any real-world use, because almost nobody will\\n> understand it or tune it properly for their installation. And who is\\n> to say that a single setting would even be appropriate for a whole\\n> installation, as opposed to something that needs to be tuned for each\\n> individual table? A single tunable might be OK, but what I would\\n> really like here is a heuristic that, while perhaps not optimal in\\n> every environment, is good enough that a very high percentage of users\\n> will never need to worry about it. In a case like this, a small gain\\n> that happens automatically feels better than a large gain that\\n> requires manual tweaking in every install.\\n\\nAgreed. If there is a tunable, it needs to be as simple as possible. And\\nideally most users would never need to use it because the default would be\\n\\\"good enough.\\\"\\n\\n> Now that doesn't answer the question of what that heuristic should be.\\n> I initially thought that if compression didn't end up reducing the\\n> number of TOAST chunks then there was no point, but that's not true,\\n> because having the last chunk be smaller than the rest can allow us to\\n> fit more than 4 into a page rather than exactly 4. However, this\\n> effect isn't likely to be important if most chunks are full-size\\n> chunks. If we insert 100 full-size chunks and 1 partial chunk at the\\n> end, we can't save much even if that chunk ends up being 1 byte\\n> instead of a full-size chunk. 25 TOAST table pages out of 26 are going\\n> to be full of full-size chunks either way, so we can't save more than\\n> ~4% and we might easily save nothing at all. As you say, the potential\\n> savings are larger as the values get smaller, because a higher\\n> proportion of the TOAST table pages are not quite full. So I think the\\n> only cases where it's even interesting to think about suppressing this\\n> optimization are those where the value is quite small -- and maybe the\\n> thing to do is just pick a conservatively large threshold and call it\\n> good.\\n> \\n> For example, say that instead of applying this technique when there\\n> are at least 2 TOAST chunks, we apply it when there are at least 500\\n> (i.e. normally 1MB). It's hard for me to imagine that we can ever\\n> lose, and in fact I think we could come down quite a bit from there\\n> and still end up winning pretty much all the time. Setting the\\n> threshold to, say, 50 or 100 or 150 TOAST chunks instead of 2 may\\n> leave some money on the table, but if it means that we don't have\\n> meaningful downsides and we don't have tunables for users to fiddle\\n> with, it might be worth it.\\n\\nAs long as we can demonstrate some decent improvements, I think using a\\nconservative threshold is a good idea. I do wonder whether thiѕ could\\nweaken the read performance gains quite a bit, though.\\n\\nThinking further, is simply reducing the number of TOAST chunks the right\\nthing to look at? If I want to add a TOAST attribute that requires 100,000\\nchunks, and you told me that I could save 10% in the read path for an extra\\n250 chunks of disk space, I would probably choose read performance. If I\\nwanted to add 100,000 attributes that were each 3 chunks, and you told me\\nthat I could save 10% in the read path for an extra 75,000 chunks of disk\\nspace, I might choose the extra disk space. These are admittedly extreme\\n(and maybe even impossible) examples, but my point is that the amount of\\ndisk space you are willing to give up may be related to the size of the\\nattribute. And maybe one way to extract additional read performance with\\nthis optimization is to use a variable threshold so that we are more likely\\nto use it for large attributes.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 11:36:56 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 16, 2022 at 11:36:56AM -0700, Nathan Bossart wrote:\\n> Thinking further, is simply reducing the number of TOAST chunks the right\\n> thing to look at? If I want to add a TOAST attribute that requires 100,000\\n> chunks, and you told me that I could save 10% in the read path for an extra\\n> 250 chunks of disk space, I would probably choose read performance. If I\\n> wanted to add 100,000 attributes that were each 3 chunks, and you told me\\n> that I could save 10% in the read path for an extra 75,000 chunks of disk\\n> space, I might choose the extra disk space. These are admittedly extreme\\n> (and maybe even impossible) examples, but my point is that the amount of\\n> disk space you are willing to give up may be related to the size of the\\n> attribute. And maybe one way to extract additional read performance with\\n> this optimization is to use a variable threshold so that we are more likely\\n> to use it for large attributes.\\n\\nI might be overthinking this. Maybe it is enough to skip compressing the\\nattribute whenever compression saves no more than some percentage of the\\nuncompressed attribute size. A conservative default setting might be\\nsomething like 5% or 10%.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 16 Mar 2022 11:46:37 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 16, 2022 at 2:36 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> Thinking further, is simply reducing the number of TOAST chunks the right\\n> thing to look at? If I want to add a TOAST attribute that requires 100,000\\n> chunks, and you told me that I could save 10% in the read path for an extra\\n> 250 chunks of disk space, I would probably choose read performance. If I\\n> wanted to add 100,000 attributes that were each 3 chunks, and you told me\\n> that I could save 10% in the read path for an extra 75,000 chunks of disk\\n> space, I might choose the extra disk space. These are admittedly extreme\\n> (and maybe even impossible) examples, but my point is that the amount of\\n> disk space you are willing to give up may be related to the size of the\\n> attribute. And maybe one way to extract additional read performance with\\n> this optimization is to use a variable threshold so that we are more likely\\n> to use it for large attributes.\\n\\nRight, so perhaps the ultimate thing here would be a more fine-grained\\nknob than SET STORAGE EXTERNAL -- something that allows you to specify\\nthat you want to compress only when it really helps. While some people\\nmight find that useful, I think the current patch is less ambitious,\\nand I think that's OK. It just wants to save something in the cases\\nwhere it's basically free. Unfortunately we've learned that it's never\\n*entirely* free because making the last TOAST chunk longer can always\\ncost you something, even if it gets longer by only 1 byte. But for\\nlarger values it's hard for that to be significant.\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 13:04:17 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 17, 2022 at 01:04:17PM -0400, Robert Haas wrote:\\n> Right, so perhaps the ultimate thing here would be a more fine-grained\\n> knob than SET STORAGE EXTERNAL -- something that allows you to specify\\n> that you want to compress only when it really helps. While some people\\n> might find that useful, I think the current patch is less ambitious,\\n> and I think that's OK. It just wants to save something in the cases\\n> where it's basically free. Unfortunately we've learned that it's never\\n> *entirely* free because making the last TOAST chunk longer can always\\n> cost you something, even if it gets longer by only 1 byte. But for\\n> larger values it's hard for that to be significant.\\n\\nI guess I think we should be slightly more ambitious. One idea could be to\\ncreate a default_toast_compression_ratio GUC with a default of 0.95. This\\nmeans that, by default, a compressed attribute must be 0.95x or less of the\\nsize of the uncompressed attribute to be stored compressed. Like\\ndefault_toast_compression, this could also be overridden at the column\\nlevel with something like\\n\\n\\tALTER TABLE mytbl ALTER mycol SET COMPRESSION lz4 RATIO 0.9;\\n\\nIf the current \\\"basically free\\\" patch is intended for v15, then maybe all\\nthis extra configurability stuff could wait for a bit, especially if we can\\ndemonstrate a decent read performance boost with a more conservative\\nsetting. However, I don't see anything terribly complicated about the\\nproposed configurability changes (assuming my proposal makes some amount of\\nsense to you and others).\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 17 Mar 2022 13:05:23 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 17, 2022 at 4:05 PM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n> On Thu, Mar 17, 2022 at 01:04:17PM -0400, Robert Haas wrote:\\n> > Right, so perhaps the ultimate thing here would be a more fine-grained\\n> > knob than SET STORAGE EXTERNAL -- something that allows you to specify\\n> > that you want to compress only when it really helps. While some people\\n> > might find that useful, I think the current patch is less ambitious,\\n> > and I think that's OK. It just wants to save something in the cases\\n> > where it's basically free. Unfortunately we've learned that it's never\\n> > *entirely* free because making the last TOAST chunk longer can always\\n> > cost you something, even if it gets longer by only 1 byte. But for\\n> > larger values it's hard for that to be significant.\\n>\\n> I guess I think we should be slightly more ambitious. One idea could be to\\n> create a default_toast_compression_ratio GUC with a default of 0.95. This\\n> means that, by default, a compressed attribute must be 0.95x or less of the\\n> size of the uncompressed attribute to be stored compressed. Like\\n> default_toast_compression, this could also be overridden at the column\\n> level with something like\\n>\\n> ALTER TABLE mytbl ALTER mycol SET COMPRESSION lz4 RATIO 0.9;\\n>\\n> If the current \\\"basically free\\\" patch is intended for v15, then maybe all\\n> this extra configurability stuff could wait for a bit, especially if we can\\n> demonstrate a decent read performance boost with a more conservative\\n> setting. However, I don't see anything terribly complicated about the\\n> proposed configurability changes (assuming my proposal makes some amount of\\n> sense to you and others).\\n\\nWe seem to have a shortage of \\\"others\\\" showing up with opinions on\\nthis topic, but I guess I'm not very confident about the general\\nutility of such a setting. Just to be clear, I'm also not very\\nconfident about the usefulness of the existing settings for\\ncontrolling TOAST. Why is it useful default behavior to try to get\\nrows down to 2kB by default, rather than 1.8kB or 3kB? Even more, why\\ndon't we try to compress primarily based on the length of individual\\nattributes and then compress further only if the resulting tuple\\ndoesn't fit into a page at all? There doesn't seem to be anything\\nmagic about fitting tuples into a quarter-page, yet the code acts as\\nthough that's the primary goal - and then, when that didn't turn out\\nto work well in all cases, we added a user-settable parameter\\n(toast_tuple_target) to let you say you really want tuples in table X\\nto fit into a third of a page or a fifth of a page instead of a\\nquarter. And there's some justification for that: a proposal to\\nfundamentally change the algorithm would likely have gotten bogged\\ndown for years, and the parameter no doubt lets you solve some\\nproblems. Yet if the whole algorithm is wrong, and I think maybe it\\nis, then being able to change the constants is not really getting us\\nwhere we need to be.\\n\\nThen, too, I'm not very confident about the usefulness of EXTENDED,\\nEXTERNAL, and MAIN. I think it's useful to be able to categorically\\ndisable compression (as EXTERNAL does), but you can't categorically\\ndisable out-of-line storage because the value can be bigger than the\\npage, so MAIN vs. EXTENDED is just changing the threshold for the use\\nof out-of-line storage. However, it does so in a way that IMHO is not\\nparticularly intuitive, which goes back to my earlier point about the\\nalgorithm seeming wacky, and it's in any case unclear why we should\\noffer exactly two possible thresholds and not any of the intermediate\\nvalues.\\n\\nI think I'm prepared to concede that the setting you are proposing\\nhere is easier to understand than either of those. Saying \\\"store this\\nvalue compressed only if you an thereby reduce the size by at least\\n10%\\\" is an understandable directive which I think will make sense even\\nto people who know very little about PostgreSQL or databases\\ngenerally, as long as they have some vague idea how compression works.\\nHowever, I guess I am not confident that the setting would get much\\nuse, or that it is the best way to measure what we care about. For\\ninstance, if my value is 19kB in length, saving 10% isn't very\\nimpactful, because it's still the same number of chunks, and the only\\nbenefit I gain is that the last chunk is smaller, which is probably\\nnot very impactful because I'm still going to have 8 full chunks. I\\nwill gain more if the value is bigger or smaller. Going up, if I\\nincrease the size to 19MB or 190MB, I'm really saving a very\\nsignificant amount of cache space and, potentially, I/O. If I was\\nhappy with a 10% savings at 19kB, I will likely be happy with a\\nconsiderably smaller economy here. Going down, if I decrease the size\\nto 1.9kB, I'm not storing any full-sized chunks any more, so reducing\\nthe size of the one partial chunk that I'm storing seems like it could\\nhave a real impact, provided of course that other toasted values in\\nthe same table are similarly small. If the whole TOAST table is full\\nof just partial chunks, then the smaller they are the more will fit\\ninto a page, all things being equal, and thus the smaller the TOAST\\ntable will be.\\n\\nMaybe the conclusion here is that more thought is needed before\\nchanging anything in this area....\\n\\n-- \\nRobert Haas\\nEDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 22 Mar 2022 16:34:05 -0400\",\n \"msg_from\": \"Robert Haas <robertmhaas@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 22, 2022 at 04:34:05PM -0400, Robert Haas wrote:\\n> We seem to have a shortage of \\\"others\\\" showing up with opinions on\\n> this topic, but I guess I'm not very confident about the general\\n> utility of such a setting. Just to be clear, I'm also not very\\n> confident about the usefulness of the existing settings for\\n> controlling TOAST. Why is it useful default behavior to try to get\\n> rows down to 2kB by default, rather than 1.8kB or 3kB? Even more, why\\n> don't we try to compress primarily based on the length of individual\\n> attributes and then compress further only if the resulting tuple\\n> doesn't fit into a page at all? There doesn't seem to be anything\\n> magic about fitting tuples into a quarter-page, yet the code acts as\\n> though that's the primary goal - and then, when that didn't turn out\\n> to work well in all cases, we added a user-settable parameter\\n> (toast_tuple_target) to let you say you really want tuples in table X\\n> to fit into a third of a page or a fifth of a page instead of a\\n> quarter. And there's some justification for that: a proposal to\\n> fundamentally change the algorithm would likely have gotten bogged\\n> down for years, and the parameter no doubt lets you solve some\\n> problems. Yet if the whole algorithm is wrong, and I think maybe it\\n> is, then being able to change the constants is not really getting us\\n> where we need to be.\\n> \\n> Then, too, I'm not very confident about the usefulness of EXTENDED,\\n> EXTERNAL, and MAIN. I think it's useful to be able to categorically\\n> disable compression (as EXTERNAL does), but you can't categorically\\n> disable out-of-line storage because the value can be bigger than the\\n> page, so MAIN vs. EXTENDED is just changing the threshold for the use\\n> of out-of-line storage. However, it does so in a way that IMHO is not\\n> particularly intuitive, which goes back to my earlier point about the\\n> algorithm seeming wacky, and it's in any case unclear why we should\\n> offer exactly two possible thresholds and not any of the intermediate\\n> values.\\n\\nI agree with all of this. Adding configurability for each constant might\\nhelp folks in the short term, but using these knobs probably requires quite\\na bit of expertise in Postgres internals as well as a good understanding of\\nyour data. I think we ought to revist TOAST configurability from a user\\nperspective. IOW what can be chosen automatically, and how do we enable\\nusers to effectively configure the settings that cannot be chosen\\nautomatically? IMO this is a worthwhile conversation to have as long as it\\ndoesn't stray too far into the \\\"let's rewrite TOAST\\\" realm. I think there\\nis always going to be some level of complexity with stuff like TOAST, but\\nthere are probably all sorts of ways to simplify/improve it also.\\n\\n> Maybe the conclusion here is that more thought is needed before\\n> changing anything in this area....\\n\\nYou've certainly got me thinking more about this. If the scope of this\\nwork is going to expand, a few months before the first v16 commitfest is\\nprobably the right time for it.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 22 Mar 2022 14:42:53 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 22, 2022 at 02:42:53PM -0700, Nathan Bossart wrote:\\n> On Tue, Mar 22, 2022 at 04:34:05PM -0400, Robert Haas wrote:\\n>> Then, too, I'm not very confident about the usefulness of EXTENDED,\\n>> EXTERNAL, and MAIN. I think it's useful to be able to categorically\\n>> disable compression (as EXTERNAL does), but you can't categorically\\n>> disable out-of-line storage because the value can be bigger than the\\n>> page, so MAIN vs. EXTENDED is just changing the threshold for the use\\n>> of out-of-line storage. However, it does so in a way that IMHO is not\\n>> particularly intuitive, which goes back to my earlier point about the\\n>> algorithm seeming wacky, and it's in any case unclear why we should\\n>> offer exactly two possible thresholds and not any of the intermediate\\n>> values.\\n> \\n> I agree with all of this. Adding configurability for each constant might\\n> help folks in the short term, but using these knobs probably requires quite\\n> a bit of expertise in Postgres internals as well as a good understanding of\\n> your data. I think we ought to revist TOAST configurability from a user\\n> perspective. IOW what can be chosen automatically, and how do we enable\\n> users to effectively configure the settings that cannot be chosen\\n> automatically? IMO this is a worthwhile conversation to have as long as it\\n> doesn't stray too far into the \\\"let's rewrite TOAST\\\" realm. I think there\\n> is always going to be some level of complexity with stuff like TOAST, but\\n> there are probably all sorts of ways to simplify/improve it also.\\n\\nI agree with this feeling. TOAST has already too many configuration\\nparameters that have their own way of behaving slightling differently.\\nIf we could reduce this number rather than increasing it, the better.\\nI would be also curious to see how much those parameters become\\nrelevant with more compression options possible with toast values.\\n--\\nMichael\",\n \"msg_date\": \"Wed, 23 Mar 2022 13:28:46 +0900\",\n \"msg_from\": \"Michael Paquier <michael@paquier.xyz>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 23, 2022 at 9:58 AM Michael Paquier <michael@paquier.xyz> wrote:\\n>\\n> On Tue, Mar 22, 2022 at 02:42:53PM -0700, Nathan Bossart wrote:\\n> > On Tue, Mar 22, 2022 at 04:34:05PM -0400, Robert Haas wrote:\\n> >> Then, too, I'm not very confident about the usefulness of EXTENDED,\\n> >> EXTERNAL, and MAIN. I think it's useful to be able to categorically\\n> >> disable compression (as EXTERNAL does), but you can't categorically\\n> >> disable out-of-line storage because the value can be bigger than the\\n> >> page, so MAIN vs. EXTENDED is just changing the threshold for the use\\n> >> of out-of-line storage. However, it does so in a way that IMHO is not\\n> >> particularly intuitive, which goes back to my earlier point about the\\n> >> algorithm seeming wacky, and it's in any case unclear why we should\\n> >> offer exactly two possible thresholds and not any of the intermediate\\n> >> values.\\n> >\\n> > I agree with all of this. Adding configurability for each constant might\\n> > help folks in the short term, but using these knobs probably requires quite\\n> > a bit of expertise in Postgres internals as well as a good understanding of\\n> > your data. I think we ought to revist TOAST configurability from a user\\n> > perspective. IOW what can be chosen automatically, and how do we enable\\n> > users to effectively configure the settings that cannot be chosen\\n> > automatically? IMO this is a worthwhile conversation to have as long as it\\n> > doesn't stray too far into the \\\"let's rewrite TOAST\\\" realm. I think there\\n> > is always going to be some level of complexity with stuff like TOAST, but\\n> > there are probably all sorts of ways to simplify/improve it also.\\n>\\n> I agree with this feeling. TOAST has already too many configuration\\n> parameters that have their own way of behaving slightling differently.\\n> If we could reduce this number rather than increasing it, the better.\\n> I would be also curious to see how much those parameters become\\n> relevant with more compression options possible with toast values.\\n\\nAgree with you, it doesn't make much sense here to add an extra user\\nvisible configuration parameter, where it is not really clear for\\nusers how to really configure it to get the best out of it.\\n\\n-- \\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Mar 2022 15:35:15 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 18, 2022 at 1:35 AM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n>\\n>\\n> I guess I think we should be slightly more ambitious. One idea could be to\\n> create a default_toast_compression_ratio GUC with a default of 0.95. This\\n> means that, by default, a compressed attribute must be 0.95x or less of the\\n> size of the uncompressed attribute to be stored compressed. Like\\n> default_toast_compression, this could also be overridden at the column\\n> level with something like\\n\\nI am not sure that we want a GUC to control that but we can certainly\\nbe more ambitious. Basically, in the current patch if data is\\nslightly large then we would always prefer to store the compressed\\ndata, e.g. if the data size is 200kB then even if the compression\\nratio is as low as 1% then we would choose to store then compressed\\ndata.\\n\\nI think we can make it based on the compression ratio and then upper\\nbound it with the number of chunk differences. For example if the\\ncompression ratio < 10% then stored it uncompressed iff the chunk\\ndifference < threshold. But with that we might see performance impact\\non the smaller data which has a compressed ratio < 10% because their\\nchunk difference will always be under the threshold. So maybe the\\nchunk difference threshold can be a function based on the total\\nnumbers of chunks required for the data, maybe a logarithmic function\\nso that the threshold grows slowly along with the base data size.\\n\\n--\\nRegards,\\nDilip Kumar\\nEnterpriseDB: http://www.enterprisedb.com\\n\\n\\n\",\n \"msg_date\": \"Wed, 23 Mar 2022 16:40:37 +0530\",\n \"msg_from\": \"Dilip Kumar <dilipbalaut@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Optimize external TOAST storage\"\n }\n]","isConversational":false}}},{"rowIdx":53595,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hia,\n\nWe've added some information such as the command and the timestamp to\nthe error context message by commit abc0910e2. This patch adds further\ninformation to it: replication origin name and commit-LSN.\n\nThis will be helpful for users to set the origin name and LSN to\npg_replication_origin_advance().\n\nThe errcontext message would become like follows:\n\n*Before\nERROR: duplicate key value violates unique constraint \"test_pkey\"\nDETAIL: Key (c)=(1) already exists.\nCONTEXT: processing remote data during \"INSERT\" for replication\ntarget relation \"public.test\" in transaction 726 at 2022-02-28\n20:59:56.005909+09\n\n* After\nERROR: duplicate key value violates unique constraint \"test_pkey\"\nDETAIL: Key (c)=(1) already exists.\nCONTEXT: processing remote data during \"INSERT\" for replication\ntarget relation \"public.test\" in transaction 726 committed at LSN\n0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\norigin \"pg_16395\"\n\nI'm a bit concerned that the message may be too long.\n\nI've attached two patches: the first one changes\napply_error_callback() so that it uses complete sentences with if-else\nblocks in order to have a translation work, the second patch adds the\norigin name and commit-LSN to the errcontext message.\n\nRegards,\n\n--\nMasahiko Sawada\nEDB: https://www.enterprisedb.com/","msg_date":"Mon, 28 Feb 2022 21:15:23 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Add the replication origin name and commit-LSN to logical replication\n worker errcontext"},{"msg_contents":"On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n> We've added some information such as the command and the timestamp to\n> the error context message by commit abc0910e2. This patch adds further\n> information to it: replication origin name and commit-LSN.\n>\n> This will be helpful for users to set the origin name and LSN to\n> pg_replication_origin_advance().\n>\n\n+1. This will make the use of pg_replication_origin_advance() relatively easy.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Mon, 28 Feb 2022 18:49:03 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n> Hia,\n>\n> We've added some information such as the command and the timestamp to\n> the error context message by commit abc0910e2. This patch adds further\n> information to it: replication origin name and commit-LSN.\n>\n> This will be helpful for users to set the origin name and LSN to\n> pg_replication_origin_advance().\n>\n> The errcontext message would become like follows:\n>\n> *Before\n> ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> DETAIL: Key (c)=(1) already exists.\n> CONTEXT: processing remote data during \"INSERT\" for replication\n> target relation \"public.test\" in transaction 726 at 2022-02-28\n> 20:59:56.005909+09\n>\n> * After\n> ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> DETAIL: Key (c)=(1) already exists.\n> CONTEXT: processing remote data during \"INSERT\" for replication\n> target relation \"public.test\" in transaction 726 committed at LSN\n> 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\n> origin \"pg_16395\"\n>\n> I'm a bit concerned that the message may be too long.\n\nIf you are willing to use abbreviations instead of full\nwords/sentences perhaps you can shorten the long CONTEXT part like\nbelow?\n\nBefore:\nCONTEXT: processing remote data during \"INSERT\" for replication\ntarget relation \"public.test\" in transaction 726 committed at LSN\n0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\nreplication origin \"pg_16395\"\n\nAfter:\nCONTEXT: processing remote data during \"INSERT\" for replication target\nrelation \"public.test\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\n20:58:27.964238+09, origin \"pg_16395\")\n\n------\nKind Regards,\nPeter Smith.\nFujitsu Australia.\n\n\n","msg_date":"Wed, 2 Mar 2022 13:54:31 +1100","msg_from":"Peter Smith <smithpb2250@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n> I've attached two patches: the first one changes\n> apply_error_callback() so that it uses complete sentences with if-else\n> blocks in order to have a translation work,\n>\n\nThis is an improvement over what we have now but I think this is still\nnot completely correct as per message translation rules:\n+ else\n+ errcontext(\"processing remote data during \\\"%s\\\" in transaction %u at %s\",\n+ logicalrep_message_type(errarg->command),\n+ errarg->remote_xid,\n+ (errarg->ts != 0) ? timestamptz_to_str(errarg->ts) : \"(not-set)\");\n\nAs per guidelines [1][2], we don't prefer to construct messages at\nrun-time aka we can do better for the following part: + (errarg->ts\n!= 0) ? timestamptz_to_str(errarg->ts) : \"(not-set)\". I think we need\nto use if-else here to split it further. If you agree, then the same\nneeds to be dealt with in other parts of the patch as well.\n\n[1] - https://www.postgresql.org/docs/devel/nls-programmer.html#NLS-GUIDELINES\n[2] - Do not construct sentences at run-time, like:\nprintf(\"Files were %s.\\n\", flag ? \"copied\" : \"removed\");\nThe word order within the sentence might be different in other languages.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 2 Mar 2022 08:25:13 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Wed, Mar 2, 2022 at 8:25 AM Peter Smith <smithpb2250@gmail.com> wrote:\n>\n> On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> >\n> > The errcontext message would become like follows:\n> >\n> > *Before\n> > ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> > DETAIL: Key (c)=(1) already exists.\n> > CONTEXT: processing remote data during \"INSERT\" for replication\n> > target relation \"public.test\" in transaction 726 at 2022-02-28\n> > 20:59:56.005909+09\n> >\n> > * After\n> > ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> > DETAIL: Key (c)=(1) already exists.\n> > CONTEXT: processing remote data during \"INSERT\" for replication\n> > target relation \"public.test\" in transaction 726 committed at LSN\n> > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\n> > origin \"pg_16395\"\n> >\n> > I'm a bit concerned that the message may be too long.\n>\n> If you are willing to use abbreviations instead of full\n> words/sentences perhaps you can shorten the long CONTEXT part like\n> below?\n>\n> Before:\n> CONTEXT: processing remote data during \"INSERT\" for replication\n> target relation \"public.test\" in transaction 726 committed at LSN\n> 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\n> replication origin \"pg_16395\"\n>\n> After:\n> CONTEXT: processing remote data during \"INSERT\" for replication target\n> relation \"public.test\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\n> 20:58:27.964238+09, origin \"pg_16395\")\n>\n\nI am wondering whether we can avoid having a timestamp in the message?\nIf one wants, it can be retrieved from the errors otherwise as well.\nFor example, I see the below error in my machine:\n2022-02-26 07:45:25.092 IST [17644] ERROR: duplicate key value\nviolates unique constraint \"t1_pkey\"\n2022-02-26 07:45:25.092 IST [17644] DETAIL: Key (c1)=(1) already exists.\n2022-02-26 07:45:25.092 IST [17644] CONTEXT: processing remote data\nduring \"INSERT\" for replication target relation \"public.t1\" in\ntransaction 724 at 2022-02-26 07:45:09.083848+05:30\n\nNow, here, won't the time at the starting of CONTEXT serves our\npurpose. If we can remove the timestamp, I think the message won't\nappear too long. What do you think?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 2 Mar 2022 08:51:17 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Wed, Mar 2, 2022 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Mar 2, 2022 at 8:25 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> >\n> > On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> > >\n> > > The errcontext message would become like follows:\n> > >\n> > > *Before\n> > > ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> > > DETAIL: Key (c)=(1) already exists.\n> > > CONTEXT: processing remote data during \"INSERT\" for replication\n> > > target relation \"public.test\" in transaction 726 at 2022-02-28\n> > > 20:59:56.005909+09\n> > >\n> > > * After\n> > > ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> > > DETAIL: Key (c)=(1) already exists.\n> > > CONTEXT: processing remote data during \"INSERT\" for replication\n> > > target relation \"public.test\" in transaction 726 committed at LSN\n> > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\n> > > origin \"pg_16395\"\n> > >\n> > > I'm a bit concerned that the message may be too long.\n> >\n> > If you are willing to use abbreviations instead of full\n> > words/sentences perhaps you can shorten the long CONTEXT part like\n> > below?\n> >\n> > Before:\n> > CONTEXT: processing remote data during \"INSERT\" for replication\n> > target relation \"public.test\" in transaction 726 committed at LSN\n> > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\n> > replication origin \"pg_16395\"\n> >\n> > After:\n> > CONTEXT: processing remote data during \"INSERT\" for replication target\n> > relation \"public.test\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\n> > 20:58:27.964238+09, origin \"pg_16395\")\n> >\n>\n> I am wondering whether we can avoid having a timestamp in the message?\n> If one wants, it can be retrieved from the errors otherwise as well.\n> For example, I see the below error in my machine:\n> 2022-02-26 07:45:25.092 IST [17644] ERROR: duplicate key value\n> violates unique constraint \"t1_pkey\"\n> 2022-02-26 07:45:25.092 IST [17644] DETAIL: Key (c1)=(1) already exists.\n> 2022-02-26 07:45:25.092 IST [17644] CONTEXT: processing remote data\n> during \"INSERT\" for replication target relation \"public.t1\" in\n> transaction 724 at 2022-02-26 07:45:09.083848+05:30\n>\n> Now, here, won't the time at the starting of CONTEXT serves our\n> purpose. If we can remove the timestamp, I think the message won't\n> appear too long. What do you think?\n\nThe time of the CONTEXT log message and the time in the message would\nlargely vary when the subscriber is much behind the publisher. But I\nbasically agree that the timestamp in the message might not be\nimportant, at least for now. If we will support conflict resolution\nthat resolves based on the commit timestamp in the future, we might\nwant it again.\n\nRegards,\n\n-- \nMasahiko Sawada\nEDB: https://www.enterprisedb.com/\n\n\n","msg_date":"Wed, 2 Mar 2022 13:02:45 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Wed, Mar 2, 2022 at 11:55 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> >\n> > I've attached two patches: the first one changes\n> > apply_error_callback() so that it uses complete sentences with if-else\n> > blocks in order to have a translation work,\n> >\n>\n> This is an improvement over what we have now but I think this is still\n> not completely correct as per message translation rules:\n> + else\n> + errcontext(\"processing remote data during \\\"%s\\\" in transaction %u at %s\",\n> + logicalrep_message_type(errarg->command),\n> + errarg->remote_xid,\n> + (errarg->ts != 0) ? timestamptz_to_str(errarg->ts) : \"(not-set)\");\n>\n> As per guidelines [1][2], we don't prefer to construct messages at\n> run-time aka we can do better for the following part: + (errarg->ts\n> != 0) ? timestamptz_to_str(errarg->ts) : \"(not-set)\". I think we need\n> to use if-else here to split it further. If you agree, then the same\n> needs to be dealt with in other parts of the patch as well.\n\nI intended to use \"(not-set)\" as a value rather than a word in the\nsentence so I think it doesn't violate the guidelines. We can split it\nfurther as you suggested but we will end up having more if-else\nbranches.\n\nRegards,\n\n--\nMasahiko Sawada\nEDB: https://www.enterprisedb.com/\n\n\n","msg_date":"Wed, 2 Mar 2022 14:39:54 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"At Wed, 2 Mar 2022 14:39:54 +0900, Masahiko Sawada <sawada.mshk@gmail.com> wrote in \n> On Wed, Mar 2, 2022 at 11:55 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> > >\n> > > I've attached two patches: the first one changes\n> > > apply_error_callback() so that it uses complete sentences with if-else\n> > > blocks in order to have a translation work,\n> > >\n> >\n> > This is an improvement over what we have now but I think this is still\n> > not completely correct as per message translation rules:\n> > + else\n> > + errcontext(\"processing remote data during \\\"%s\\\" in transaction %u at %s\",\n> > + logicalrep_message_type(errarg->command),\n> > + errarg->remote_xid,\n> > + (errarg->ts != 0) ? timestamptz_to_str(errarg->ts) : \"(not-set)\");\n> >\n> > As per guidelines [1][2], we don't prefer to construct messages at\n> > run-time aka we can do better for the following part: + (errarg->ts\n> > != 0) ? timestamptz_to_str(errarg->ts) : \"(not-set)\". I think we need\n> > to use if-else here to split it further. If you agree, then the same\n> > needs to be dealt with in other parts of the patch as well.\n> \n> I intended to use \"(not-set)\" as a value rather than a word in the\n> sentence so I think it doesn't violate the guidelines. We can split it\n> further as you suggested but we will end up having more if-else\n> branches.\n\nIt seems to me exactly our way. In the first place I doubt\n\"(not-set)\" fits the place for timestamp even in English.\n\nMoreover, if we (I?) translated the message into Japanese, it would\nlook like;\n\nCONTEXT: (not-set)にトランザクション 2352314 内で \"TRUNCATE\" でのリモートデータの処理中\n\nI don't think that looks fine. Translating \"(not-set)\" makes things\neven worse.\n\nCONTEXT: (非設定)にトランザクション 2352314 内で \"TRUNCATE\" でのリモートデータの処理中\n\nYes, I can alleviate that strangeness a bit by modulating it, but it\nstill looks odd.\n\nCONTEXT: 時刻(非設定)、トランザクション 2352314 内で \"TRUNCATE\" でのリモートデータの処理中\n\nRather, I'd prefer simply to list the attributes.\n\nCONTEXT: processing remote data during \"MESSAGE\". Transaction (unknown). Time (unknown), replication target relation (unknown), column (unknown)\nCONTEXT: \"MESSAGE\"でのリモートデータの処理中. トランザクション (不明). 時刻 (不明), レプリケーション対象リレーション (不明), column (不明)\n\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Wed, 02 Mar 2022 16:07:16 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Wed, Mar 2, 2022 at 9:33 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n> On Wed, Mar 2, 2022 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Wed, Mar 2, 2022 at 8:25 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > >\n> > > On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> > > >\n> > > > The errcontext message would become like follows:\n> > > >\n> > > > *Before\n> > > > ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> > > > DETAIL: Key (c)=(1) already exists.\n> > > > CONTEXT: processing remote data during \"INSERT\" for replication\n> > > > target relation \"public.test\" in transaction 726 at 2022-02-28\n> > > > 20:59:56.005909+09\n> > > >\n> > > > * After\n> > > > ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> > > > DETAIL: Key (c)=(1) already exists.\n> > > > CONTEXT: processing remote data during \"INSERT\" for replication\n> > > > target relation \"public.test\" in transaction 726 committed at LSN\n> > > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\n> > > > origin \"pg_16395\"\n> > > >\n> > > > I'm a bit concerned that the message may be too long.\n> > >\n> > > If you are willing to use abbreviations instead of full\n> > > words/sentences perhaps you can shorten the long CONTEXT part like\n> > > below?\n> > >\n> > > Before:\n> > > CONTEXT: processing remote data during \"INSERT\" for replication\n> > > target relation \"public.test\" in transaction 726 committed at LSN\n> > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\n> > > replication origin \"pg_16395\"\n> > >\n> > > After:\n> > > CONTEXT: processing remote data during \"INSERT\" for replication target\n> > > relation \"public.test\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\n> > > 20:58:27.964238+09, origin \"pg_16395\")\n> > >\n> >\n> > I am wondering whether we can avoid having a timestamp in the message?\n> > If one wants, it can be retrieved from the errors otherwise as well.\n> > For example, I see the below error in my machine:\n> > 2022-02-26 07:45:25.092 IST [17644] ERROR: duplicate key value\n> > violates unique constraint \"t1_pkey\"\n> > 2022-02-26 07:45:25.092 IST [17644] DETAIL: Key (c1)=(1) already exists.\n> > 2022-02-26 07:45:25.092 IST [17644] CONTEXT: processing remote data\n> > during \"INSERT\" for replication target relation \"public.t1\" in\n> > transaction 724 at 2022-02-26 07:45:09.083848+05:30\n> >\n> > Now, here, won't the time at the starting of CONTEXT serves our\n> > purpose. If we can remove the timestamp, I think the message won't\n> > appear too long. What do you think?\n>\n> The time of the CONTEXT log message and the time in the message would\n> largely vary when the subscriber is much behind the publisher. But I\n> basically agree that the timestamp in the message might not be\n> important, at least for now. If we will support conflict resolution\n> that resolves based on the commit timestamp in the future, we might\n> want it again.\n>\n\nPossible, but let's remove it for now as it will simplify the message\nand the need for additional branches. What do you think?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Wed, 2 Mar 2022 12:44:31 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Wed, Mar 2, 2022 at 4:14 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Mar 2, 2022 at 9:33 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> >\n> > On Wed, Mar 2, 2022 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > >\n> > > On Wed, Mar 2, 2022 at 8:25 AM Peter Smith <smithpb2250@gmail.com> wrote:\n> > > >\n> > > > On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> > > > >\n> > > > > The errcontext message would become like follows:\n> > > > >\n> > > > > *Before\n> > > > > ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> > > > > DETAIL: Key (c)=(1) already exists.\n> > > > > CONTEXT: processing remote data during \"INSERT\" for replication\n> > > > > target relation \"public.test\" in transaction 726 at 2022-02-28\n> > > > > 20:59:56.005909+09\n> > > > >\n> > > > > * After\n> > > > > ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> > > > > DETAIL: Key (c)=(1) already exists.\n> > > > > CONTEXT: processing remote data during \"INSERT\" for replication\n> > > > > target relation \"public.test\" in transaction 726 committed at LSN\n> > > > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\n> > > > > origin \"pg_16395\"\n> > > > >\n> > > > > I'm a bit concerned that the message may be too long.\n> > > >\n> > > > If you are willing to use abbreviations instead of full\n> > > > words/sentences perhaps you can shorten the long CONTEXT part like\n> > > > below?\n> > > >\n> > > > Before:\n> > > > CONTEXT: processing remote data during \"INSERT\" for replication\n> > > > target relation \"public.test\" in transaction 726 committed at LSN\n> > > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\n> > > > replication origin \"pg_16395\"\n> > > >\n> > > > After:\n> > > > CONTEXT: processing remote data during \"INSERT\" for replication target\n> > > > relation \"public.test\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\n> > > > 20:58:27.964238+09, origin \"pg_16395\")\n> > > >\n> > >\n> > > I am wondering whether we can avoid having a timestamp in the message?\n> > > If one wants, it can be retrieved from the errors otherwise as well.\n> > > For example, I see the below error in my machine:\n> > > 2022-02-26 07:45:25.092 IST [17644] ERROR: duplicate key value\n> > > violates unique constraint \"t1_pkey\"\n> > > 2022-02-26 07:45:25.092 IST [17644] DETAIL: Key (c1)=(1) already exists.\n> > > 2022-02-26 07:45:25.092 IST [17644] CONTEXT: processing remote data\n> > > during \"INSERT\" for replication target relation \"public.t1\" in\n> > > transaction 724 at 2022-02-26 07:45:09.083848+05:30\n> > >\n> > > Now, here, won't the time at the starting of CONTEXT serves our\n> > > purpose. If we can remove the timestamp, I think the message won't\n> > > appear too long. What do you think?\n> >\n> > The time of the CONTEXT log message and the time in the message would\n> > largely vary when the subscriber is much behind the publisher. But I\n> > basically agree that the timestamp in the message might not be\n> > important, at least for now. If we will support conflict resolution\n> > that resolves based on the commit timestamp in the future, we might\n> > want it again.\n> >\n>\n> Possible, but let's remove it for now as it will simplify the message\n> and the need for additional branches. What do you think?\n\nAgreed.\n\nI've attached updated patches.\n\nRegards,\n\n-- \nMasahiko Sawada\nEDB: https://www.enterprisedb.com/","msg_date":"Wed, 2 Mar 2022 16:34:59 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Wed, Mar 2, 2022 at 4:07 PM Kyotaro Horiguchi\n<horikyota.ntt@gmail.com> wrote:\n>\n> At Wed, 2 Mar 2022 14:39:54 +0900, Masahiko Sawada <sawada.mshk@gmail.com> wrote in\n> > On Wed, Mar 2, 2022 at 11:55 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> > >\n> > > On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> > > >\n> > > > I've attached two patches: the first one changes\n> > > > apply_error_callback() so that it uses complete sentences with if-else\n> > > > blocks in order to have a translation work,\n> > > >\n> > >\n> > > This is an improvement over what we have now but I think this is still\n> > > not completely correct as per message translation rules:\n> > > + else\n> > > + errcontext(\"processing remote data during \\\"%s\\\" in transaction %u at %s\",\n> > > + logicalrep_message_type(errarg->command),\n> > > + errarg->remote_xid,\n> > > + (errarg->ts != 0) ? timestamptz_to_str(errarg->ts) : \"(not-set)\");\n> > >\n> > > As per guidelines [1][2], we don't prefer to construct messages at\n> > > run-time aka we can do better for the following part: + (errarg->ts\n> > > != 0) ? timestamptz_to_str(errarg->ts) : \"(not-set)\". I think we need\n> > > to use if-else here to split it further. If you agree, then the same\n> > > needs to be dealt with in other parts of the patch as well.\n> >\n> > I intended to use \"(not-set)\" as a value rather than a word in the\n> > sentence so I think it doesn't violate the guidelines. We can split it\n> > further as you suggested but we will end up having more if-else\n> > branches.\n>\n> It seems to me exactly our way. In the first place I doubt\n> \"(not-set)\" fits the place for timestamp even in English.\n>\n> Moreover, if we (I?) translated the message into Japanese, it would\n> look like;\n>\n> CONTEXT: (not-set)にトランザクション 2352314 内で \"TRUNCATE\" でのリモートデータの処理中\n>\n> I don't think that looks fine. Translating \"(not-set)\" makes things\n> even worse.\n>\n> CONTEXT: (非設定)にトランザクション 2352314 内で \"TRUNCATE\" でのリモートデータの処理中\n>\n> Yes, I can alleviate that strangeness a bit by modulating it, but it\n> still looks odd.\n\nIndeed. But the timestamp is removed in the latest version patch.\n\n>\n> CONTEXT: 時刻(非設定)、トランザクション 2352314 内で \"TRUNCATE\" でのリモートデータの処理中\n>\n> Rather, I'd prefer simply to list the attributes.\n>\n> CONTEXT: processing remote data during \"MESSAGE\". Transaction (unknown). Time (unknown), replication target relation (unknown), column (unknown)\n> CONTEXT: \"MESSAGE\"でのリモートデータの処理中. トランザクション (不明). 時刻 (不明), レプリケーション対象リレーション (不明), column (不明)\n\nPeter Smith also seems to prefer this style. Looking at existing error\ncontext messages, we use this list style in logical.c:\n\nstatic void\noutput_plugin_error_callback(void *arg)\n{\n LogicalErrorCallbackState *state = (LogicalErrorCallbackState *) arg;\n\n /* not all callbacks have an associated LSN */\n if (state->report_location != InvalidXLogRecPtr)\n errcontext(\"slot \\\"%s\\\", output plugin \\\"%s\\\", in the %s\ncallback, associated LSN %X/%X\",\n NameStr(state->ctx->slot->data.name),\n NameStr(state->ctx->slot->data.plugin),\n state->callback_name,\n LSN_FORMAT_ARGS(state->report_location));\n else\n errcontext(\"slot \\\"%s\\\", output plugin \\\"%s\\\", in the %s callback\",\n NameStr(state->ctx->slot->data.name),\n NameStr(state->ctx->slot->data.plugin),\n state->callback_name);\n\n}\n\nRegards,\n\n--\nMasahiko Sawada\nEDB: https://www.enterprisedb.com/\n\n\n","msg_date":"Wed, 2 Mar 2022 17:11:44 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Wed, Mar 2, 2022 at 1:05 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n> On Wed, Mar 2, 2022 at 4:14 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> I've attached updated patches.\n>\n\nThe first patch LGTM. Some comments on the second patch:\n\n1.\n@@ -3499,6 +3503,17 @@ ApplyWorkerMain(Datum main_arg)\n myslotname = MemoryContextStrdup(ApplyContext, syncslotname);\n\n pfree(syncslotname);\n+\n+ /*\n+ * Allocate the origin name in long-lived context for error context\n+ * message\n+ */\n+ ReplicationOriginNameForTablesync(MySubscription->oid,\n+ MyLogicalRepWorker->relid,\n+ originname,\n+ sizeof(originname));\n+ apply_error_callback_arg.origin_name = MemoryContextStrdup(ApplyContext,\n+ originname);\n\nCan we assign this in LogicalRepSyncTableStart() where we already\nforming the origin name? That will avoid this extra call to\nReplicationOriginNameForTablesync.\n\n2.\n@@ -3657,30 +3679,37 @@ apply_error_callback(void *arg)\n errcontext(\"processing remote data during \\\"%s\\\"\",\n logicalrep_message_type(errarg->command));\n else\n- errcontext(\"processing remote data during \\\"%s\\\" in transaction %u\",\n+ errcontext(\"processing remote data during \\\"%s\\\" in transaction %u\ncommitted at LSN %X/%X from replication origin \\\"%s\\\"\",\n logicalrep_message_type(errarg->command),\n- errarg->remote_xid);\n+ errarg->remote_xid,\n+ LSN_FORMAT_ARGS(errarg->commit_lsn),\n+ errarg->origin_name);\n\nWon't we set the origin name before the command? If so, it can be used\nin the first message as well and we can change the condition in the\nbeginning such that if the origin or command is not set then we can\nreturn without adding additional context information.\n\nIsn't this error message used for rollback prepared failure as well?\nIf so, do we need to say \"... finished at LSN ...\" instead of \"...\ncommitted at LSN ...\"?\n\nThe other point about this message is that saying \".... from\nreplication origin ...\" sounds more like remote information similar to\npublication but the origin is of the local node. How about slightly\nchanging it to \"processing remote data for replication origin \\\"%s\\\"\nduring \\\"%s\\\" in transaction ...\"?\n\n3.\n+</screen>\n+ The LSN of the transaction that contains the change violating the\nconstraint and\n+ the replication origin name can be found from those outputs (LSN\n0/14C0378 and\n+ replication origin <literal>pg_16395</literal> in the above case).\nThe transaction\n+ can be skipped by calling the <link linkend=\"pg-replication-origin-advance\">\n <function>pg_replication_origin_advance()</function></link> function with\n- a <parameter>node_name</parameter> corresponding to the subscription name,\n- and a position. The current position of origins can be seen in the\n- <link linkend=\"view-pg-replication-origin-status\">\n+ the <parameter>node_name</parameter> and the next LSN of the commit LSN\n+ (i.e., 0/14C0379) from those outputs.\n\nAfter node_name, can we specify origin_name similar to what we do for\nLSN as that will make things more clear? Also, shall we mention that\nusers need to disable subscriptions to perform replication origin\nadvance?\n\nI think for prepared transactions, the user might need to use it twice\nbecause after skipping prepared xact, it will get an error again\nduring the processing of commit prepared (no such prepare exists). I\nthought of mentioning it but felt that might lead to specifying too\nmany details which can confuse users as well. What do you think?\n\n4. There are places in the patch like apply_handle_stream_start()\nwhich sets commit_lsn in callback arg as InvalidXLogRecPtr but the\ncallback function seems to be assuming that it is always a valid\nvalue. Shouldn't we need to avoid appending LSN for such cases?\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Thu, 3 Mar 2022 12:07:05 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Thu, Mar 3, 2022 at 3:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Wed, Mar 2, 2022 at 1:05 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> >\n> > On Wed, Mar 2, 2022 at 4:14 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > I've attached updated patches.\n> >\n>\n> The first patch LGTM. Some comments on the second patch:\n>\n> 1.\n> @@ -3499,6 +3503,17 @@ ApplyWorkerMain(Datum main_arg)\n> myslotname = MemoryContextStrdup(ApplyContext, syncslotname);\n>\n> pfree(syncslotname);\n> +\n> + /*\n> + * Allocate the origin name in long-lived context for error context\n> + * message\n> + */\n> + ReplicationOriginNameForTablesync(MySubscription->oid,\n> + MyLogicalRepWorker->relid,\n> + originname,\n> + sizeof(originname));\n> + apply_error_callback_arg.origin_name = MemoryContextStrdup(ApplyContext,\n> + originname);\n>\n> Can we assign this in LogicalRepSyncTableStart() where we already\n> forming the origin name? That will avoid this extra call to\n> ReplicationOriginNameForTablesync.\n\nYes, but it requires to expose either apply_error_callback_arg or the\ntablesync's origin name since the tablesync worker sets its origin\nname in tablesync.c. I think it's better to avoid exposing them.\n\n>\n> 2.\n> @@ -3657,30 +3679,37 @@ apply_error_callback(void *arg)\n> errcontext(\"processing remote data during \\\"%s\\\"\",\n> logicalrep_message_type(errarg->command));\n> else\n> - errcontext(\"processing remote data during \\\"%s\\\" in transaction %u\",\n> + errcontext(\"processing remote data during \\\"%s\\\" in transaction %u\n> committed at LSN %X/%X from replication origin \\\"%s\\\"\",\n> logicalrep_message_type(errarg->command),\n> - errarg->remote_xid);\n> + errarg->remote_xid,\n> + LSN_FORMAT_ARGS(errarg->commit_lsn),\n> + errarg->origin_name);\n>\n> Won't we set the origin name before the command? If so, it can be used\n> in the first message as well and we can change the condition in the\n> beginning such that if the origin or command is not set then we can\n> return without adding additional context information.\n\nGood point.\n\n>\n> Isn't this error message used for rollback prepared failure as well?\n> If so, do we need to say \"... finished at LSN ...\" instead of \"...\n> committed at LSN ...\"?\n\nRight. Or can we just remove \"committed\" since the current message is\n\"transaction %u at %s\"? That is , just replace the timestamp with LSN.\n\n>\n> The other point about this message is that saying \".... from\n> replication origin ...\" sounds more like remote information similar to\n> publication but the origin is of the local node. How about slightly\n> changing it to \"processing remote data for replication origin \\\"%s\\\"\n> during \\\"%s\\\" in transaction ...\"?\n\nOkay, so the modified message would be like:\n\n\"processing remote data for replication origin \\\"%s\\\" during \\\"%s\\\"\nfor replication target relation \\\"%s.%s\\\" column \\\"%s\\\" in transaction\n%u finished at LSN %X/%X\"\n\n>\n> 3.\n> +</screen>\n> + The LSN of the transaction that contains the change violating the\n> constraint and\n> + the replication origin name can be found from those outputs (LSN\n> 0/14C0378 and\n> + replication origin <literal>pg_16395</literal> in the above case).\n> The transaction\n> + can be skipped by calling the <link linkend=\"pg-replication-origin-advance\">\n> <function>pg_replication_origin_advance()</function></link> function with\n> - a <parameter>node_name</parameter> corresponding to the subscription name,\n> - and a position. The current position of origins can be seen in the\n> - <link linkend=\"view-pg-replication-origin-status\">\n> + the <parameter>node_name</parameter> and the next LSN of the commit LSN\n> + (i.e., 0/14C0379) from those outputs.\n>\n> After node_name, can we specify origin_name similar to what we do for\n> LSN as that will make things more clear? Also, shall we mention that\n> users need to disable subscriptions to perform replication origin\n> advance?\n\nAgreed.\n\n>\n> I think for prepared transactions, the user might need to use it twice\n> because after skipping prepared xact, it will get an error again\n> during the processing of commit prepared (no such prepare exists).\n\nGood point.\n\n> I\n> thought of mentioning it but felt that might lead to specifying too\n> many details which can confuse users as well. What do you think?\n\nGiven that this method of using pg_replication_origin_advance() is\nnormally not a preferable way, I think we might not need to mention\nit. Also, it needs twice for one transaction but the steps are the\nsame.\n\n>\n> 4. There are places in the patch like apply_handle_stream_start()\n> which sets commit_lsn in callback arg as InvalidXLogRecPtr but the\n> callback function seems to be assuming that it is always a valid\n> value. Shouldn't we need to avoid appending LSN for such cases?\n\nAgreed. Will fix it in the next version patch.\n\nI'm updating the patches and will submit them.\n\nRegards,\n\n--\nMasahiko Sawada\nEDB: https://www.enterprisedb.com/\n\n\n","msg_date":"Thu, 3 Mar 2022 22:02:49 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Thu, Mar 3, 2022 at 10:02 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n>\n> I'm updating the patches and will submit them.\n\nAttached updated version patches.\n\nRegards,\n\n-- \nMasahiko Sawada\nEDB: https://www.enterprisedb.com/","msg_date":"Fri, 4 Mar 2022 10:09:28 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Friday, March 4, 2022 10:09 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\r\n> On Thu, Mar 3, 2022 at 10:02 PM Masahiko Sawada\r\n> <sawada.mshk@gmail.com> wrote:\r\n> >\r\n> >\r\n> > I'm updating the patches and will submit them.\r\n> \r\n> Attached updated version patches.\r\nThank you for sharing the patch v3.\r\n\r\nFew minor comments.\r\n\r\n\r\n(1) v03-0001, apply_error_callback function\r\n\r\n- /* append transaction information */\r\n- if (TransactionIdIsNormal(errarg->remote_xid))\r\n+ if (errarg->rel == NULL)\r\n {\r\n- appendStringInfo(&buf, _(\" in transaction %u\"), errarg->remote_xid);\r\n\r\nShould write !errarg->rel ?\r\n\r\n(2) v03-0002, doc/src/sgml/logical-replication.sgml\r\n\r\n\r\n+ transaction that conflicts with the existing data. When a conflict produces\r\n+ an error, it is shown in the subscriber's server logs as follows:\r\n+<screen>\r\n+ERROR: duplicate key value violates unique constraint \"test_pkey\"\r\n+DETAIL: Key (c)=(1) already exists.\r\n+CONTEXT: processing remote data during \"INSERT\" for replication target relation \"public.test\" in transaction 725 committed at LSN 0/14BFA88\r\n+</screen>\r\n\r\n\r\nWe should update the CONTEXT message by using the v3-0001.\r\n\r\n(3) v03-0002, doc/src/sgml/logical-replication.sgml\r\n\r\n\r\n+ The LSN of the transaction that contains the change violating the constraint and\r\n+ the replication origin name can be found from those outputs (LSN 0/14C0378 and\r\n+ replication origin <literal>pg_16395</literal> in the above case). To skip the\r\n+ transaction, the subscription needs to be disabled temporarily by\r\n+ <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the transaction\r\n+ can be skipped by calling the <link linkend=\"pg-replication-origin-advance\">\r\n\r\nThe LSN(0/14C0378) is not same as the one in the above error context.\r\nIt's recommended to check LSNs directly written in the documentation.\r\n\r\n(4) one confirmation\r\n\r\nWe don't have a TAP test of pg_replication_origin_advance()\r\nfor v3, that utilizes this new log in a logical replication setup.\r\nThis is because existing tests for this function (in test_decoding) is only for permission check\r\nand argument validation, and we're just changing error message itself.\r\nIs this correct ?\r\n\r\n\r\nBest Regards,\r\n\tTakamichi Osumi\r\n\r\n","msg_date":"Fri, 4 Mar 2022 02:27:51 +0000","msg_from":"\"osumi.takamichi@fujitsu.com\" <osumi.takamichi@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Fri, Mar 4, 2022 at 11:27 AM osumi.takamichi@fujitsu.com\n<osumi.takamichi@fujitsu.com> wrote:\n>\n> On Friday, March 4, 2022 10:09 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> > On Thu, Mar 3, 2022 at 10:02 PM Masahiko Sawada\n> > <sawada.mshk@gmail.com> wrote:\n> > >\n> > >\n> > > I'm updating the patches and will submit them.\n> >\n> > Attached updated version patches.\n> Thank you for sharing the patch v3.\n>\n> Few minor comments.\n>\n>\n> (1) v03-0001, apply_error_callback function\n>\n> - /* append transaction information */\n> - if (TransactionIdIsNormal(errarg->remote_xid))\n> + if (errarg->rel == NULL)\n> {\n> - appendStringInfo(&buf, _(\" in transaction %u\"), errarg->remote_xid);\n>\n> Should write !errarg->rel ?\n\nI think either should be fine.\n\n>\n> (2) v03-0002, doc/src/sgml/logical-replication.sgml\n>\n>\n> + transaction that conflicts with the existing data. When a conflict produces\n> + an error, it is shown in the subscriber's server logs as follows:\n> +<screen>\n> +ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> +DETAIL: Key (c)=(1) already exists.\n> +CONTEXT: processing remote data during \"INSERT\" for replication target relation \"public.test\" in transaction 725 committed at LSN 0/14BFA88\n> +</screen>\n>\n>\n> We should update the CONTEXT message by using the v3-0001.\n>\n> (3) v03-0002, doc/src/sgml/logical-replication.sgml\n>\n>\n> + The LSN of the transaction that contains the change violating the constraint and\n> + the replication origin name can be found from those outputs (LSN 0/14C0378 and\n> + replication origin <literal>pg_16395</literal> in the above case). To skip the\n> + transaction, the subscription needs to be disabled temporarily by\n> + <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the transaction\n> + can be skipped by calling the <link linkend=\"pg-replication-origin-advance\">\n>\n> The LSN(0/14C0378) is not same as the one in the above error context.\n> It's recommended to check LSNs directly written in the documentation.\n\nRight, I missed checking and updating it.\n\n>\n> (4) one confirmation\n>\n> We don't have a TAP test of pg_replication_origin_advance()\n> for v3, that utilizes this new log in a logical replication setup.\n> This is because existing tests for this function (in test_decoding) is only for permission check\n> and argument validation, and we're just changing error message itself.\n> Is this correct ?\n\nYeah, I think it’s a good idea to add test in general but I don't\nthink we should include the tests for skipping a transaction by using\npg_replication_origin() in this patch because it's an existing way and\nupcoming ALTER SUBSCRIPTION SKIP patch includes the tests and it's\nmore appropriate way. But if others also think it should do that too\nalong with this update, I'm happy to add tests.\n\nI've attached updated patches.\n\nRegards,\n\n--\nMasahiko Sawada\nEDB: https://www.enterprisedb.com/","msg_date":"Fri, 4 Mar 2022 14:22:54 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Fri, Mar 4, 2022 at 6:40 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n> Attached updated version patches.\n>\n\nThe patch looks mostly good to me. Few minor comments:\n1. I think we can have an Assert for errarg->origin_name in\napply_error_callback after checking the command as this function\nassumes that it will always be set.\n2. I suggest minor changes in the documentation change:\nWhen a conflict produces an error, the replication won't proceed, and\nthe apply worker will emit the following kind of message to the\nsubscriber's server log:\n+<screen>\n+ERROR: duplicate key value violates unique constraint \"test_pkey\"\n+DETAIL: Key (c)=(1) already exists.\n+CONTEXT: processing remote data during \"INSERT\" for replication\ntarget relation \"public.test\" in transaction 725 committed at LSN\n0/14BFA88 from replication origin \"pg_16395\"\n+</screen>\n\nThe LSN of the transaction that contains the change violating the\nconstraint and the replication origin name can be found from the\nserver log (LSN 0/14C0378 and replication origin\n<literal>pg_16395</literal> in the above case). To skip the\ntransaction, the subscription needs to be disabled temporarily by\n<command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\ntransaction can be skipped by calling the <link\nlinkend=\"pg-replication-origin-advance\">\n<function>pg_replication_origin_advance()</function></link> function\nwith the <parameter>node_name</parameter> (i.e.,\n<literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\nLSN 0/14C0379).\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 4 Mar 2022 11:24:53 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Fri, Mar 4, 2022 at 10:53 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n> On Fri, Mar 4, 2022 at 11:27 AM osumi.takamichi@fujitsu.com\n> <osumi.takamichi@fujitsu.com> wrote:\n> >\n> > (4) one confirmation\n> >\n> > We don't have a TAP test of pg_replication_origin_advance()\n> > for v3, that utilizes this new log in a logical replication setup.\n> > This is because existing tests for this function (in test_decoding) is only for permission check\n> > and argument validation, and we're just changing error message itself.\n> > Is this correct ?\n>\n> Yeah, I think it’s a good idea to add test in general but I don't\n> think we should include the tests for skipping a transaction by using\n> pg_replication_origin() in this patch because it's an existing way and\n> upcoming ALTER SUBSCRIPTION SKIP patch includes the tests and it's\n> more appropriate way. But if others also think it should do that too\n> along with this update, I'm happy to add tests.\n>\n\nI also don't see a reason why this patch should add any tests related\nto origin.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 4 Mar 2022 11:26:20 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Friday, March 4, 2022 2:23 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\r\n> I've attached updated patches.\r\nHi, thank you for updating the patch.\r\n\r\nOne comment on v4.\r\n\r\nIn v4-0002, we introduce 'commit_lsn' in the ApplyErrorCallbackArg.\r\nThis member is set for prepare, rollback prepared and stream_abort as well.\r\nThe new log message format is useful when we have a prepare transaction\r\nthat keeps failing on the subscriber and want to know the target transaction\r\nfor the pg_replication_origin_advance(), right ?\r\nIf this is true, I wasn't sure if the name 'commit_lsn' is the\r\nmost accurate name for this variable. Should we adjust the name a bit ?\r\n\r\nEven when we decide to continue to use 'commit_lsn',\r\nit might be better to add some comments near the definition, I feel.\r\n\r\n\r\nBest Regards,\r\n\tTakamichi Osumi\r\n\r\n","msg_date":"Fri, 4 Mar 2022 06:15:49 +0000","msg_from":"\"osumi.takamichi@fujitsu.com\" <osumi.takamichi@fujitsu.com>","msg_from_op":false,"msg_subject":"RE: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Fri, Mar 4, 2022 at 11:45 AM osumi.takamichi@fujitsu.com\n<osumi.takamichi@fujitsu.com> wrote:\n>\n> On Friday, March 4, 2022 2:23 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> > I've attached updated patches.\n> Hi, thank you for updating the patch.\n>\n> One comment on v4.\n>\n> In v4-0002, we introduce 'commit_lsn' in the ApplyErrorCallbackArg.\n> This member is set for prepare, rollback prepared and stream_abort as well.\n> The new log message format is useful when we have a prepare transaction\n> that keeps failing on the subscriber and want to know the target transaction\n> for the pg_replication_origin_advance(), right ?\n> If this is true, I wasn't sure if the name 'commit_lsn' is the\n> most accurate name for this variable. Should we adjust the name a bit ?\n>\n\nIf we want to change this variable name, the other options could be\n'end_lsn', or 'finish_lsn'.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Fri, 4 Mar 2022 11:51:41 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Fri, Mar 4, 2022 at 2:55 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, Mar 4, 2022 at 6:40 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> >\n> > Attached updated version patches.\n> >\n>\n> The patch looks mostly good to me. Few minor comments:\n\nThank you for the comments!\n\n> 1. I think we can have an Assert for errarg->origin_name in\n> apply_error_callback after checking the command as this function\n> assumes that it will always be set.\n\nAdded.\n\n> 2. I suggest minor changes in the documentation change:\n> When a conflict produces an error, the replication won't proceed, and\n> the apply worker will emit the following kind of message to the\n> subscriber's server log:\n> +<screen>\n> +ERROR: duplicate key value violates unique constraint \"test_pkey\"\n> +DETAIL: Key (c)=(1) already exists.\n> +CONTEXT: processing remote data during \"INSERT\" for replication\n> target relation \"public.test\" in transaction 725 committed at LSN\n> 0/14BFA88 from replication origin \"pg_16395\"\n> +</screen>\n>\n> The LSN of the transaction that contains the change violating the\n> constraint and the replication origin name can be found from the\n> server log (LSN 0/14C0378 and replication origin\n> <literal>pg_16395</literal> in the above case). To skip the\n> transaction, the subscription needs to be disabled temporarily by\n> <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\n> transaction can be skipped by calling the <link\n> linkend=\"pg-replication-origin-advance\">\n> <function>pg_replication_origin_advance()</function></link> function\n> with the <parameter>node_name</parameter> (i.e.,\n> <literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\n> LSN 0/14C0379).\n\nFixed.\n\nOn Fri, Mar 4, 2022 at 3:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Fri, Mar 4, 2022 at 11:45 AM osumi.takamichi@fujitsu.com\n> <osumi.takamichi@fujitsu.com> wrote:\n> >\n> > On Friday, March 4, 2022 2:23 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> > > I've attached updated patches.\n> > Hi, thank you for updating the patch.\n> >\n> > One comment on v4.\n> >\n> > In v4-0002, we introduce 'commit_lsn' in the ApplyErrorCallbackArg.\n> > This member is set for prepare, rollback prepared and stream_abort as well.\n> > The new log message format is useful when we have a prepare transaction\n> > that keeps failing on the subscriber and want to know the target transaction\n> > for the pg_replication_origin_advance(), right ?\n> > If this is true, I wasn't sure if the name 'commit_lsn' is the\n> > most accurate name for this variable. Should we adjust the name a bit ?\n> >\n>\n> If we want to change this variable name, the other options could be\n> 'end_lsn', or 'finish_lsn'.\n>\n\nAgreed with 'finish_lsn'.\n\nI've attached updated patches. Please review them.\n\nRegards,\n\n-- \nMasahiko Sawada\nEDB: https://www.enterprisedb.com/","msg_date":"Fri, 4 Mar 2022 20:43:50 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Fri, Mar 4, 2022, at 2:54 AM, Amit Kapila wrote:\n> The LSN of the transaction that contains the change violating the\n> constraint and the replication origin name can be found from the\n> server log (LSN 0/14C0378 and replication origin\n> <literal>pg_16395</literal> in the above case). To skip the\n> transaction, the subscription needs to be disabled temporarily by\n> <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\n> transaction can be skipped by calling the <link\n> linkend=\"pg-replication-origin-advance\">\n> <function>pg_replication_origin_advance()</function></link> function\n> with the <parameter>node_name</parameter> (i.e.,\n> <literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\n> LSN 0/14C0379).\nYou could also add:\n\nAfter that the replication can be resumed by <command>ALTER SUBSCRIPTION ...\nENABLE</command>.\n\nLet's provide complete instructions.\n\n\n--\nEuler Taveira\nEDB https://www.enterprisedb.com/\n\nOn Fri, Mar 4, 2022, at 2:54 AM, Amit Kapila wrote:The LSN of the transaction that contains the change violating theconstraint and the replication origin name can be found from theserver log (LSN 0/14C0378 and replication origin<literal>pg_16395</literal> in the above case). To skip thetransaction, the subscription needs to be disabled temporarily by<command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, thetransaction can be skipped by calling the <linklinkend=\"pg-replication-origin-advance\"><function>pg_replication_origin_advance()</function></link> functionwith the <parameter>node_name</parameter> (i.e.,<literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,LSN 0/14C0379).You could also add:After that the replication can be resumed by <command>ALTER SUBSCRIPTION ...ENABLE</command>.Let's provide complete instructions.--Euler TaveiraEDB https://www.enterprisedb.com/","msg_date":"Fri, 04 Mar 2022 09:32:18 -0300","msg_from":"\"Euler Taveira\" <euler@eulerto.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication\n worker errcontext"},{"msg_contents":"On Fri, Mar 4, 2022 at 6:02 PM Euler Taveira <euler@eulerto.com> wrote:\n>\n> On Fri, Mar 4, 2022, at 2:54 AM, Amit Kapila wrote:\n>\n> The LSN of the transaction that contains the change violating the\n> constraint and the replication origin name can be found from the\n> server log (LSN 0/14C0378 and replication origin\n> <literal>pg_16395</literal> in the above case). To skip the\n> transaction, the subscription needs to be disabled temporarily by\n> <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\n> transaction can be skipped by calling the <link\n> linkend=\"pg-replication-origin-advance\">\n> <function>pg_replication_origin_advance()</function></link> function\n> with the <parameter>node_name</parameter> (i.e.,\n> <literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\n> LSN 0/14C0379).\n>\n> You could also add:\n>\n> After that the replication can be resumed by <command>ALTER SUBSCRIPTION ...\n> ENABLE</command>.\n>\n> Let's provide complete instructions.\n>\n\n+1. That sounds reasonable to me.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Sat, 5 Mar 2022 06:42:09 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Fri, Mar 4, 2022 at 9:32 PM Euler Taveira <euler@eulerto.com> wrote:\n>\n> On Fri, Mar 4, 2022, at 2:54 AM, Amit Kapila wrote:\n>\n> The LSN of the transaction that contains the change violating the\n> constraint and the replication origin name can be found from the\n> server log (LSN 0/14C0378 and replication origin\n> <literal>pg_16395</literal> in the above case). To skip the\n> transaction, the subscription needs to be disabled temporarily by\n> <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\n> transaction can be skipped by calling the <link\n> linkend=\"pg-replication-origin-advance\">\n> <function>pg_replication_origin_advance()</function></link> function\n> with the <parameter>node_name</parameter> (i.e.,\n> <literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\n> LSN 0/14C0379).\n>\n> You could also add:\n>\n> After that the replication can be resumed by <command>ALTER SUBSCRIPTION ...\n> ENABLE</command>.\n>\n> Let's provide complete instructions.\n\nThank you for the comment. +1.\n\nI've attached updated patches.\n\nRegards,\n\n-- \nMasahiko Sawada\nEDB: https://www.enterprisedb.com/","msg_date":"Mon, 7 Mar 2022 10:05:25 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Mon, Mar 7, 2022 at 6:36 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n>\n> Thank you for the comment. +1.\n>\n> I've attached updated patches.\n>\n\nPushed the first patch. Fixed one typo in the second patch and\nslightly changed the commit message, otherwise, it looks good to me.\nI'll push this tomorrow unless there are more comments.\n\n-- \nWith Regards,\nAmit Kapila.","msg_date":"Mon, 7 Mar 2022 10:06:37 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Mon, Mar 7, 2022 at 10:06 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 6:36 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> >\n> > Thank you for the comment. +1.\n> >\n> > I've attached updated patches.\n> >\n>\n> Pushed the first patch. Fixed one typo in the second patch and\n> slightly changed the commit message, otherwise, it looks good to me.\n> I'll push this tomorrow unless there are more comments.\n>\n\nPushed.\n\n-- \nWith Regards,\nAmit Kapila.\n\n\n","msg_date":"Tue, 8 Mar 2022 16:26:16 +0530","msg_from":"Amit Kapila <amit.kapila16@gmail.com>","msg_from_op":false,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"},{"msg_contents":"On Tue, Mar 8, 2022 at 7:56 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\n>\n> On Mon, Mar 7, 2022 at 10:06 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\n> >\n> > On Mon, Mar 7, 2022 at 6:36 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\n> > >\n> > > Thank you for the comment. +1.\n> > >\n> > > I've attached updated patches.\n> > >\n> >\n> > Pushed the first patch. Fixed one typo in the second patch and\n> > slightly changed the commit message, otherwise, it looks good to me.\n> > I'll push this tomorrow unless there are more comments.\n> >\n>\n> Pushed.\n\nThank you!\n\nRegards,\n\n-- \nMasahiko Sawada\nEDB: https://www.enterprisedb.com/\n\n\n","msg_date":"Wed, 9 Mar 2022 09:32:47 +0900","msg_from":"Masahiko Sawada <sawada.mshk@gmail.com>","msg_from_op":true,"msg_subject":"Re: Add the replication origin name and commit-LSN to logical\n replication worker errcontext"}],"string":"[\n {\n \"msg_contents\": \"Hia,\\n\\nWe've added some information such as the command and the timestamp to\\nthe error context message by commit abc0910e2. This patch adds further\\ninformation to it: replication origin name and commit-LSN.\\n\\nThis will be helpful for users to set the origin name and LSN to\\npg_replication_origin_advance().\\n\\nThe errcontext message would become like follows:\\n\\n*Before\\nERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\nDETAIL: Key (c)=(1) already exists.\\nCONTEXT: processing remote data during \\\"INSERT\\\" for replication\\ntarget relation \\\"public.test\\\" in transaction 726 at 2022-02-28\\n20:59:56.005909+09\\n\\n* After\\nERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\nDETAIL: Key (c)=(1) already exists.\\nCONTEXT: processing remote data during \\\"INSERT\\\" for replication\\ntarget relation \\\"public.test\\\" in transaction 726 committed at LSN\\n0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\\norigin \\\"pg_16395\\\"\\n\\nI'm a bit concerned that the message may be too long.\\n\\nI've attached two patches: the first one changes\\napply_error_callback() so that it uses complete sentences with if-else\\nblocks in order to have a translation work, the second patch adds the\\norigin name and commit-LSN to the errcontext message.\\n\\nRegards,\\n\\n--\\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\",\n \"msg_date\": \"Mon, 28 Feb 2022 21:15:23 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Add the replication origin name and commit-LSN to logical replication\\n worker errcontext\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n> We've added some information such as the command and the timestamp to\\n> the error context message by commit abc0910e2. This patch adds further\\n> information to it: replication origin name and commit-LSN.\\n>\\n> This will be helpful for users to set the origin name and LSN to\\n> pg_replication_origin_advance().\\n>\\n\\n+1. This will make the use of pg_replication_origin_advance() relatively easy.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 18:49:03 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n> Hia,\\n>\\n> We've added some information such as the command and the timestamp to\\n> the error context message by commit abc0910e2. This patch adds further\\n> information to it: replication origin name and commit-LSN.\\n>\\n> This will be helpful for users to set the origin name and LSN to\\n> pg_replication_origin_advance().\\n>\\n> The errcontext message would become like follows:\\n>\\n> *Before\\n> ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> DETAIL: Key (c)=(1) already exists.\\n> CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> target relation \\\"public.test\\\" in transaction 726 at 2022-02-28\\n> 20:59:56.005909+09\\n>\\n> * After\\n> ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> DETAIL: Key (c)=(1) already exists.\\n> CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> target relation \\\"public.test\\\" in transaction 726 committed at LSN\\n> 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\\n> origin \\\"pg_16395\\\"\\n>\\n> I'm a bit concerned that the message may be too long.\\n\\nIf you are willing to use abbreviations instead of full\\nwords/sentences perhaps you can shorten the long CONTEXT part like\\nbelow?\\n\\nBefore:\\nCONTEXT: processing remote data during \\\"INSERT\\\" for replication\\ntarget relation \\\"public.test\\\" in transaction 726 committed at LSN\\n0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\\nreplication origin \\\"pg_16395\\\"\\n\\nAfter:\\nCONTEXT: processing remote data during \\\"INSERT\\\" for replication target\\nrelation \\\"public.test\\\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\\n20:58:27.964238+09, origin \\\"pg_16395\\\")\\n\\n------\\nKind Regards,\\nPeter Smith.\\nFujitsu Australia.\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 13:54:31 +1100\",\n \"msg_from\": \"Peter Smith <smithpb2250@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n> I've attached two patches: the first one changes\\n> apply_error_callback() so that it uses complete sentences with if-else\\n> blocks in order to have a translation work,\\n>\\n\\nThis is an improvement over what we have now but I think this is still\\nnot completely correct as per message translation rules:\\n+ else\\n+ errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\" in transaction %u at %s\\\",\\n+ logicalrep_message_type(errarg->command),\\n+ errarg->remote_xid,\\n+ (errarg->ts != 0) ? timestamptz_to_str(errarg->ts) : \\\"(not-set)\\\");\\n\\nAs per guidelines [1][2], we don't prefer to construct messages at\\nrun-time aka we can do better for the following part: + (errarg->ts\\n!= 0) ? timestamptz_to_str(errarg->ts) : \\\"(not-set)\\\". I think we need\\nto use if-else here to split it further. If you agree, then the same\\nneeds to be dealt with in other parts of the patch as well.\\n\\n[1] - https://www.postgresql.org/docs/devel/nls-programmer.html#NLS-GUIDELINES\\n[2] - Do not construct sentences at run-time, like:\\nprintf(\\\"Files were %s.\\\\n\\\", flag ? \\\"copied\\\" : \\\"removed\\\");\\nThe word order within the sentence might be different in other languages.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 08:25:13 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 8:25 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n>\\n> On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> >\\n> > The errcontext message would become like follows:\\n> >\\n> > *Before\\n> > ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> > DETAIL: Key (c)=(1) already exists.\\n> > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > target relation \\\"public.test\\\" in transaction 726 at 2022-02-28\\n> > 20:59:56.005909+09\\n> >\\n> > * After\\n> > ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> > DETAIL: Key (c)=(1) already exists.\\n> > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > target relation \\\"public.test\\\" in transaction 726 committed at LSN\\n> > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\\n> > origin \\\"pg_16395\\\"\\n> >\\n> > I'm a bit concerned that the message may be too long.\\n>\\n> If you are willing to use abbreviations instead of full\\n> words/sentences perhaps you can shorten the long CONTEXT part like\\n> below?\\n>\\n> Before:\\n> CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> target relation \\\"public.test\\\" in transaction 726 committed at LSN\\n> 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\\n> replication origin \\\"pg_16395\\\"\\n>\\n> After:\\n> CONTEXT: processing remote data during \\\"INSERT\\\" for replication target\\n> relation \\\"public.test\\\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\\n> 20:58:27.964238+09, origin \\\"pg_16395\\\")\\n>\\n\\nI am wondering whether we can avoid having a timestamp in the message?\\nIf one wants, it can be retrieved from the errors otherwise as well.\\nFor example, I see the below error in my machine:\\n2022-02-26 07:45:25.092 IST [17644] ERROR: duplicate key value\\nviolates unique constraint \\\"t1_pkey\\\"\\n2022-02-26 07:45:25.092 IST [17644] DETAIL: Key (c1)=(1) already exists.\\n2022-02-26 07:45:25.092 IST [17644] CONTEXT: processing remote data\\nduring \\\"INSERT\\\" for replication target relation \\\"public.t1\\\" in\\ntransaction 724 at 2022-02-26 07:45:09.083848+05:30\\n\\nNow, here, won't the time at the starting of CONTEXT serves our\\npurpose. If we can remove the timestamp, I think the message won't\\nappear too long. What do you think?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 08:51:17 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Mar 2, 2022 at 8:25 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> >\\n> > On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> > >\\n> > > The errcontext message would become like follows:\\n> > >\\n> > > *Before\\n> > > ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> > > DETAIL: Key (c)=(1) already exists.\\n> > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > > target relation \\\"public.test\\\" in transaction 726 at 2022-02-28\\n> > > 20:59:56.005909+09\\n> > >\\n> > > * After\\n> > > ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> > > DETAIL: Key (c)=(1) already exists.\\n> > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > > target relation \\\"public.test\\\" in transaction 726 committed at LSN\\n> > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\\n> > > origin \\\"pg_16395\\\"\\n> > >\\n> > > I'm a bit concerned that the message may be too long.\\n> >\\n> > If you are willing to use abbreviations instead of full\\n> > words/sentences perhaps you can shorten the long CONTEXT part like\\n> > below?\\n> >\\n> > Before:\\n> > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > target relation \\\"public.test\\\" in transaction 726 committed at LSN\\n> > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\\n> > replication origin \\\"pg_16395\\\"\\n> >\\n> > After:\\n> > CONTEXT: processing remote data during \\\"INSERT\\\" for replication target\\n> > relation \\\"public.test\\\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\\n> > 20:58:27.964238+09, origin \\\"pg_16395\\\")\\n> >\\n>\\n> I am wondering whether we can avoid having a timestamp in the message?\\n> If one wants, it can be retrieved from the errors otherwise as well.\\n> For example, I see the below error in my machine:\\n> 2022-02-26 07:45:25.092 IST [17644] ERROR: duplicate key value\\n> violates unique constraint \\\"t1_pkey\\\"\\n> 2022-02-26 07:45:25.092 IST [17644] DETAIL: Key (c1)=(1) already exists.\\n> 2022-02-26 07:45:25.092 IST [17644] CONTEXT: processing remote data\\n> during \\\"INSERT\\\" for replication target relation \\\"public.t1\\\" in\\n> transaction 724 at 2022-02-26 07:45:09.083848+05:30\\n>\\n> Now, here, won't the time at the starting of CONTEXT serves our\\n> purpose. If we can remove the timestamp, I think the message won't\\n> appear too long. What do you think?\\n\\nThe time of the CONTEXT log message and the time in the message would\\nlargely vary when the subscriber is much behind the publisher. But I\\nbasically agree that the timestamp in the message might not be\\nimportant, at least for now. If we will support conflict resolution\\nthat resolves based on the commit timestamp in the future, we might\\nwant it again.\\n\\nRegards,\\n\\n-- \\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 13:02:45 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 11:55 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> >\\n> > I've attached two patches: the first one changes\\n> > apply_error_callback() so that it uses complete sentences with if-else\\n> > blocks in order to have a translation work,\\n> >\\n>\\n> This is an improvement over what we have now but I think this is still\\n> not completely correct as per message translation rules:\\n> + else\\n> + errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\" in transaction %u at %s\\\",\\n> + logicalrep_message_type(errarg->command),\\n> + errarg->remote_xid,\\n> + (errarg->ts != 0) ? timestamptz_to_str(errarg->ts) : \\\"(not-set)\\\");\\n>\\n> As per guidelines [1][2], we don't prefer to construct messages at\\n> run-time aka we can do better for the following part: + (errarg->ts\\n> != 0) ? timestamptz_to_str(errarg->ts) : \\\"(not-set)\\\". I think we need\\n> to use if-else here to split it further. If you agree, then the same\\n> needs to be dealt with in other parts of the patch as well.\\n\\nI intended to use \\\"(not-set)\\\" as a value rather than a word in the\\nsentence so I think it doesn't violate the guidelines. We can split it\\nfurther as you suggested but we will end up having more if-else\\nbranches.\\n\\nRegards,\\n\\n--\\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 14:39:54 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"At Wed, 2 Mar 2022 14:39:54 +0900, Masahiko Sawada <sawada.mshk@gmail.com> wrote in \\n> On Wed, Mar 2, 2022 at 11:55 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> > >\\n> > > I've attached two patches: the first one changes\\n> > > apply_error_callback() so that it uses complete sentences with if-else\\n> > > blocks in order to have a translation work,\\n> > >\\n> >\\n> > This is an improvement over what we have now but I think this is still\\n> > not completely correct as per message translation rules:\\n> > + else\\n> > + errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\" in transaction %u at %s\\\",\\n> > + logicalrep_message_type(errarg->command),\\n> > + errarg->remote_xid,\\n> > + (errarg->ts != 0) ? timestamptz_to_str(errarg->ts) : \\\"(not-set)\\\");\\n> >\\n> > As per guidelines [1][2], we don't prefer to construct messages at\\n> > run-time aka we can do better for the following part: + (errarg->ts\\n> > != 0) ? timestamptz_to_str(errarg->ts) : \\\"(not-set)\\\". I think we need\\n> > to use if-else here to split it further. If you agree, then the same\\n> > needs to be dealt with in other parts of the patch as well.\\n> \\n> I intended to use \\\"(not-set)\\\" as a value rather than a word in the\\n> sentence so I think it doesn't violate the guidelines. We can split it\\n> further as you suggested but we will end up having more if-else\\n> branches.\\n\\nIt seems to me exactly our way. In the first place I doubt\\n\\\"(not-set)\\\" fits the place for timestamp even in English.\\n\\nMoreover, if we (I?) translated the message into Japanese, it would\\nlook like;\\n\\nCONTEXT: (not-set)にトランザクション 2352314 内で \\\"TRUNCATE\\\" でのリモートデータの処理中\\n\\nI don't think that looks fine. Translating \\\"(not-set)\\\" makes things\\neven worse.\\n\\nCONTEXT: (非設定)にトランザクション 2352314 内で \\\"TRUNCATE\\\" でのリモートデータの処理中\\n\\nYes, I can alleviate that strangeness a bit by modulating it, but it\\nstill looks odd.\\n\\nCONTEXT: 時刻(非設定)、トランザクション 2352314 内で \\\"TRUNCATE\\\" でのリモートデータの処理中\\n\\nRather, I'd prefer simply to list the attributes.\\n\\nCONTEXT: processing remote data during \\\"MESSAGE\\\". Transaction (unknown). Time (unknown), replication target relation (unknown), column (unknown)\\nCONTEXT: \\\"MESSAGE\\\"でのリモートデータの処理中. トランザクション (不明). 時刻 (不明), レプリケーション対象リレーション (不明), column (不明)\\n\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Wed, 02 Mar 2022 16:07:16 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 9:33 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n> On Wed, Mar 2, 2022 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Wed, Mar 2, 2022 at 8:25 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > >\\n> > > On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> > > >\\n> > > > The errcontext message would become like follows:\\n> > > >\\n> > > > *Before\\n> > > > ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> > > > DETAIL: Key (c)=(1) already exists.\\n> > > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > > > target relation \\\"public.test\\\" in transaction 726 at 2022-02-28\\n> > > > 20:59:56.005909+09\\n> > > >\\n> > > > * After\\n> > > > ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> > > > DETAIL: Key (c)=(1) already exists.\\n> > > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > > > target relation \\\"public.test\\\" in transaction 726 committed at LSN\\n> > > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\\n> > > > origin \\\"pg_16395\\\"\\n> > > >\\n> > > > I'm a bit concerned that the message may be too long.\\n> > >\\n> > > If you are willing to use abbreviations instead of full\\n> > > words/sentences perhaps you can shorten the long CONTEXT part like\\n> > > below?\\n> > >\\n> > > Before:\\n> > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > > target relation \\\"public.test\\\" in transaction 726 committed at LSN\\n> > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\\n> > > replication origin \\\"pg_16395\\\"\\n> > >\\n> > > After:\\n> > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication target\\n> > > relation \\\"public.test\\\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\\n> > > 20:58:27.964238+09, origin \\\"pg_16395\\\")\\n> > >\\n> >\\n> > I am wondering whether we can avoid having a timestamp in the message?\\n> > If one wants, it can be retrieved from the errors otherwise as well.\\n> > For example, I see the below error in my machine:\\n> > 2022-02-26 07:45:25.092 IST [17644] ERROR: duplicate key value\\n> > violates unique constraint \\\"t1_pkey\\\"\\n> > 2022-02-26 07:45:25.092 IST [17644] DETAIL: Key (c1)=(1) already exists.\\n> > 2022-02-26 07:45:25.092 IST [17644] CONTEXT: processing remote data\\n> > during \\\"INSERT\\\" for replication target relation \\\"public.t1\\\" in\\n> > transaction 724 at 2022-02-26 07:45:09.083848+05:30\\n> >\\n> > Now, here, won't the time at the starting of CONTEXT serves our\\n> > purpose. If we can remove the timestamp, I think the message won't\\n> > appear too long. What do you think?\\n>\\n> The time of the CONTEXT log message and the time in the message would\\n> largely vary when the subscriber is much behind the publisher. But I\\n> basically agree that the timestamp in the message might not be\\n> important, at least for now. If we will support conflict resolution\\n> that resolves based on the commit timestamp in the future, we might\\n> want it again.\\n>\\n\\nPossible, but let's remove it for now as it will simplify the message\\nand the need for additional branches. What do you think?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 12:44:31 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 4:14 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Mar 2, 2022 at 9:33 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> >\\n> > On Wed, Mar 2, 2022 at 12:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > >\\n> > > On Wed, Mar 2, 2022 at 8:25 AM Peter Smith <smithpb2250@gmail.com> wrote:\\n> > > >\\n> > > > On Mon, Feb 28, 2022 at 11:16 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> > > > >\\n> > > > > The errcontext message would become like follows:\\n> > > > >\\n> > > > > *Before\\n> > > > > ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> > > > > DETAIL: Key (c)=(1) already exists.\\n> > > > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > > > > target relation \\\"public.test\\\" in transaction 726 at 2022-02-28\\n> > > > > 20:59:56.005909+09\\n> > > > >\\n> > > > > * After\\n> > > > > ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> > > > > DETAIL: Key (c)=(1) already exists.\\n> > > > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > > > > target relation \\\"public.test\\\" in transaction 726 committed at LSN\\n> > > > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from replication\\n> > > > > origin \\\"pg_16395\\\"\\n> > > > >\\n> > > > > I'm a bit concerned that the message may be too long.\\n> > > >\\n> > > > If you are willing to use abbreviations instead of full\\n> > > > words/sentences perhaps you can shorten the long CONTEXT part like\\n> > > > below?\\n> > > >\\n> > > > Before:\\n> > > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> > > > target relation \\\"public.test\\\" in transaction 726 committed at LSN\\n> > > > 0/14BFA88 and timestamp 2022-02-28 20:58:27.964238+09 from\\n> > > > replication origin \\\"pg_16395\\\"\\n> > > >\\n> > > > After:\\n> > > > CONTEXT: processing remote data during \\\"INSERT\\\" for replication target\\n> > > > relation \\\"public.test\\\" (txid 726, LSN 0/14BFA88, ts 2022-02-28\\n> > > > 20:58:27.964238+09, origin \\\"pg_16395\\\")\\n> > > >\\n> > >\\n> > > I am wondering whether we can avoid having a timestamp in the message?\\n> > > If one wants, it can be retrieved from the errors otherwise as well.\\n> > > For example, I see the below error in my machine:\\n> > > 2022-02-26 07:45:25.092 IST [17644] ERROR: duplicate key value\\n> > > violates unique constraint \\\"t1_pkey\\\"\\n> > > 2022-02-26 07:45:25.092 IST [17644] DETAIL: Key (c1)=(1) already exists.\\n> > > 2022-02-26 07:45:25.092 IST [17644] CONTEXT: processing remote data\\n> > > during \\\"INSERT\\\" for replication target relation \\\"public.t1\\\" in\\n> > > transaction 724 at 2022-02-26 07:45:09.083848+05:30\\n> > >\\n> > > Now, here, won't the time at the starting of CONTEXT serves our\\n> > > purpose. If we can remove the timestamp, I think the message won't\\n> > > appear too long. What do you think?\\n> >\\n> > The time of the CONTEXT log message and the time in the message would\\n> > largely vary when the subscriber is much behind the publisher. But I\\n> > basically agree that the timestamp in the message might not be\\n> > important, at least for now. If we will support conflict resolution\\n> > that resolves based on the commit timestamp in the future, we might\\n> > want it again.\\n> >\\n>\\n> Possible, but let's remove it for now as it will simplify the message\\n> and the need for additional branches. What do you think?\\n\\nAgreed.\\n\\nI've attached updated patches.\\n\\nRegards,\\n\\n-- \\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\",\n \"msg_date\": \"Wed, 2 Mar 2022 16:34:59 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 4:07 PM Kyotaro Horiguchi\\n<horikyota.ntt@gmail.com> wrote:\\n>\\n> At Wed, 2 Mar 2022 14:39:54 +0900, Masahiko Sawada <sawada.mshk@gmail.com> wrote in\\n> > On Wed, Mar 2, 2022 at 11:55 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> > >\\n> > > On Mon, Feb 28, 2022 at 5:46 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> > > >\\n> > > > I've attached two patches: the first one changes\\n> > > > apply_error_callback() so that it uses complete sentences with if-else\\n> > > > blocks in order to have a translation work,\\n> > > >\\n> > >\\n> > > This is an improvement over what we have now but I think this is still\\n> > > not completely correct as per message translation rules:\\n> > > + else\\n> > > + errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\" in transaction %u at %s\\\",\\n> > > + logicalrep_message_type(errarg->command),\\n> > > + errarg->remote_xid,\\n> > > + (errarg->ts != 0) ? timestamptz_to_str(errarg->ts) : \\\"(not-set)\\\");\\n> > >\\n> > > As per guidelines [1][2], we don't prefer to construct messages at\\n> > > run-time aka we can do better for the following part: + (errarg->ts\\n> > > != 0) ? timestamptz_to_str(errarg->ts) : \\\"(not-set)\\\". I think we need\\n> > > to use if-else here to split it further. If you agree, then the same\\n> > > needs to be dealt with in other parts of the patch as well.\\n> >\\n> > I intended to use \\\"(not-set)\\\" as a value rather than a word in the\\n> > sentence so I think it doesn't violate the guidelines. We can split it\\n> > further as you suggested but we will end up having more if-else\\n> > branches.\\n>\\n> It seems to me exactly our way. In the first place I doubt\\n> \\\"(not-set)\\\" fits the place for timestamp even in English.\\n>\\n> Moreover, if we (I?) translated the message into Japanese, it would\\n> look like;\\n>\\n> CONTEXT: (not-set)にトランザクション 2352314 内で \\\"TRUNCATE\\\" でのリモートデータの処理中\\n>\\n> I don't think that looks fine. Translating \\\"(not-set)\\\" makes things\\n> even worse.\\n>\\n> CONTEXT: (非設定)にトランザクション 2352314 内で \\\"TRUNCATE\\\" でのリモートデータの処理中\\n>\\n> Yes, I can alleviate that strangeness a bit by modulating it, but it\\n> still looks odd.\\n\\nIndeed. But the timestamp is removed in the latest version patch.\\n\\n>\\n> CONTEXT: 時刻(非設定)、トランザクション 2352314 内で \\\"TRUNCATE\\\" でのリモートデータの処理中\\n>\\n> Rather, I'd prefer simply to list the attributes.\\n>\\n> CONTEXT: processing remote data during \\\"MESSAGE\\\". Transaction (unknown). Time (unknown), replication target relation (unknown), column (unknown)\\n> CONTEXT: \\\"MESSAGE\\\"でのリモートデータの処理中. トランザクション (不明). 時刻 (不明), レプリケーション対象リレーション (不明), column (不明)\\n\\nPeter Smith also seems to prefer this style. Looking at existing error\\ncontext messages, we use this list style in logical.c:\\n\\nstatic void\\noutput_plugin_error_callback(void *arg)\\n{\\n LogicalErrorCallbackState *state = (LogicalErrorCallbackState *) arg;\\n\\n /* not all callbacks have an associated LSN */\\n if (state->report_location != InvalidXLogRecPtr)\\n errcontext(\\\"slot \\\\\\\"%s\\\\\\\", output plugin \\\\\\\"%s\\\\\\\", in the %s\\ncallback, associated LSN %X/%X\\\",\\n NameStr(state->ctx->slot->data.name),\\n NameStr(state->ctx->slot->data.plugin),\\n state->callback_name,\\n LSN_FORMAT_ARGS(state->report_location));\\n else\\n errcontext(\\\"slot \\\\\\\"%s\\\\\\\", output plugin \\\\\\\"%s\\\\\\\", in the %s callback\\\",\\n NameStr(state->ctx->slot->data.name),\\n NameStr(state->ctx->slot->data.plugin),\\n state->callback_name);\\n\\n}\\n\\nRegards,\\n\\n--\\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\\n\\n\\n\",\n \"msg_date\": \"Wed, 2 Mar 2022 17:11:44 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 1:05 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n> On Wed, Mar 2, 2022 at 4:14 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> I've attached updated patches.\\n>\\n\\nThe first patch LGTM. Some comments on the second patch:\\n\\n1.\\n@@ -3499,6 +3503,17 @@ ApplyWorkerMain(Datum main_arg)\\n myslotname = MemoryContextStrdup(ApplyContext, syncslotname);\\n\\n pfree(syncslotname);\\n+\\n+ /*\\n+ * Allocate the origin name in long-lived context for error context\\n+ * message\\n+ */\\n+ ReplicationOriginNameForTablesync(MySubscription->oid,\\n+ MyLogicalRepWorker->relid,\\n+ originname,\\n+ sizeof(originname));\\n+ apply_error_callback_arg.origin_name = MemoryContextStrdup(ApplyContext,\\n+ originname);\\n\\nCan we assign this in LogicalRepSyncTableStart() where we already\\nforming the origin name? That will avoid this extra call to\\nReplicationOriginNameForTablesync.\\n\\n2.\\n@@ -3657,30 +3679,37 @@ apply_error_callback(void *arg)\\n errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\"\\\",\\n logicalrep_message_type(errarg->command));\\n else\\n- errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\" in transaction %u\\\",\\n+ errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\" in transaction %u\\ncommitted at LSN %X/%X from replication origin \\\\\\\"%s\\\\\\\"\\\",\\n logicalrep_message_type(errarg->command),\\n- errarg->remote_xid);\\n+ errarg->remote_xid,\\n+ LSN_FORMAT_ARGS(errarg->commit_lsn),\\n+ errarg->origin_name);\\n\\nWon't we set the origin name before the command? If so, it can be used\\nin the first message as well and we can change the condition in the\\nbeginning such that if the origin or command is not set then we can\\nreturn without adding additional context information.\\n\\nIsn't this error message used for rollback prepared failure as well?\\nIf so, do we need to say \\\"... finished at LSN ...\\\" instead of \\\"...\\ncommitted at LSN ...\\\"?\\n\\nThe other point about this message is that saying \\\".... from\\nreplication origin ...\\\" sounds more like remote information similar to\\npublication but the origin is of the local node. How about slightly\\nchanging it to \\\"processing remote data for replication origin \\\\\\\"%s\\\\\\\"\\nduring \\\\\\\"%s\\\\\\\" in transaction ...\\\"?\\n\\n3.\\n+</screen>\\n+ The LSN of the transaction that contains the change violating the\\nconstraint and\\n+ the replication origin name can be found from those outputs (LSN\\n0/14C0378 and\\n+ replication origin <literal>pg_16395</literal> in the above case).\\nThe transaction\\n+ can be skipped by calling the <link linkend=\\\"pg-replication-origin-advance\\\">\\n <function>pg_replication_origin_advance()</function></link> function with\\n- a <parameter>node_name</parameter> corresponding to the subscription name,\\n- and a position. The current position of origins can be seen in the\\n- <link linkend=\\\"view-pg-replication-origin-status\\\">\\n+ the <parameter>node_name</parameter> and the next LSN of the commit LSN\\n+ (i.e., 0/14C0379) from those outputs.\\n\\nAfter node_name, can we specify origin_name similar to what we do for\\nLSN as that will make things more clear? Also, shall we mention that\\nusers need to disable subscriptions to perform replication origin\\nadvance?\\n\\nI think for prepared transactions, the user might need to use it twice\\nbecause after skipping prepared xact, it will get an error again\\nduring the processing of commit prepared (no such prepare exists). I\\nthought of mentioning it but felt that might lead to specifying too\\nmany details which can confuse users as well. What do you think?\\n\\n4. There are places in the patch like apply_handle_stream_start()\\nwhich sets commit_lsn in callback arg as InvalidXLogRecPtr but the\\ncallback function seems to be assuming that it is always a valid\\nvalue. Shouldn't we need to avoid appending LSN for such cases?\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Mar 2022 12:07:05 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 3, 2022 at 3:37 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Wed, Mar 2, 2022 at 1:05 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> >\\n> > On Wed, Mar 2, 2022 at 4:14 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > I've attached updated patches.\\n> >\\n>\\n> The first patch LGTM. Some comments on the second patch:\\n>\\n> 1.\\n> @@ -3499,6 +3503,17 @@ ApplyWorkerMain(Datum main_arg)\\n> myslotname = MemoryContextStrdup(ApplyContext, syncslotname);\\n>\\n> pfree(syncslotname);\\n> +\\n> + /*\\n> + * Allocate the origin name in long-lived context for error context\\n> + * message\\n> + */\\n> + ReplicationOriginNameForTablesync(MySubscription->oid,\\n> + MyLogicalRepWorker->relid,\\n> + originname,\\n> + sizeof(originname));\\n> + apply_error_callback_arg.origin_name = MemoryContextStrdup(ApplyContext,\\n> + originname);\\n>\\n> Can we assign this in LogicalRepSyncTableStart() where we already\\n> forming the origin name? That will avoid this extra call to\\n> ReplicationOriginNameForTablesync.\\n\\nYes, but it requires to expose either apply_error_callback_arg or the\\ntablesync's origin name since the tablesync worker sets its origin\\nname in tablesync.c. I think it's better to avoid exposing them.\\n\\n>\\n> 2.\\n> @@ -3657,30 +3679,37 @@ apply_error_callback(void *arg)\\n> errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\"\\\",\\n> logicalrep_message_type(errarg->command));\\n> else\\n> - errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\" in transaction %u\\\",\\n> + errcontext(\\\"processing remote data during \\\\\\\"%s\\\\\\\" in transaction %u\\n> committed at LSN %X/%X from replication origin \\\\\\\"%s\\\\\\\"\\\",\\n> logicalrep_message_type(errarg->command),\\n> - errarg->remote_xid);\\n> + errarg->remote_xid,\\n> + LSN_FORMAT_ARGS(errarg->commit_lsn),\\n> + errarg->origin_name);\\n>\\n> Won't we set the origin name before the command? If so, it can be used\\n> in the first message as well and we can change the condition in the\\n> beginning such that if the origin or command is not set then we can\\n> return without adding additional context information.\\n\\nGood point.\\n\\n>\\n> Isn't this error message used for rollback prepared failure as well?\\n> If so, do we need to say \\\"... finished at LSN ...\\\" instead of \\\"...\\n> committed at LSN ...\\\"?\\n\\nRight. Or can we just remove \\\"committed\\\" since the current message is\\n\\\"transaction %u at %s\\\"? That is , just replace the timestamp with LSN.\\n\\n>\\n> The other point about this message is that saying \\\".... from\\n> replication origin ...\\\" sounds more like remote information similar to\\n> publication but the origin is of the local node. How about slightly\\n> changing it to \\\"processing remote data for replication origin \\\\\\\"%s\\\\\\\"\\n> during \\\\\\\"%s\\\\\\\" in transaction ...\\\"?\\n\\nOkay, so the modified message would be like:\\n\\n\\\"processing remote data for replication origin \\\\\\\"%s\\\\\\\" during \\\\\\\"%s\\\\\\\"\\nfor replication target relation \\\\\\\"%s.%s\\\\\\\" column \\\\\\\"%s\\\\\\\" in transaction\\n%u finished at LSN %X/%X\\\"\\n\\n>\\n> 3.\\n> +</screen>\\n> + The LSN of the transaction that contains the change violating the\\n> constraint and\\n> + the replication origin name can be found from those outputs (LSN\\n> 0/14C0378 and\\n> + replication origin <literal>pg_16395</literal> in the above case).\\n> The transaction\\n> + can be skipped by calling the <link linkend=\\\"pg-replication-origin-advance\\\">\\n> <function>pg_replication_origin_advance()</function></link> function with\\n> - a <parameter>node_name</parameter> corresponding to the subscription name,\\n> - and a position. The current position of origins can be seen in the\\n> - <link linkend=\\\"view-pg-replication-origin-status\\\">\\n> + the <parameter>node_name</parameter> and the next LSN of the commit LSN\\n> + (i.e., 0/14C0379) from those outputs.\\n>\\n> After node_name, can we specify origin_name similar to what we do for\\n> LSN as that will make things more clear? Also, shall we mention that\\n> users need to disable subscriptions to perform replication origin\\n> advance?\\n\\nAgreed.\\n\\n>\\n> I think for prepared transactions, the user might need to use it twice\\n> because after skipping prepared xact, it will get an error again\\n> during the processing of commit prepared (no such prepare exists).\\n\\nGood point.\\n\\n> I\\n> thought of mentioning it but felt that might lead to specifying too\\n> many details which can confuse users as well. What do you think?\\n\\nGiven that this method of using pg_replication_origin_advance() is\\nnormally not a preferable way, I think we might not need to mention\\nit. Also, it needs twice for one transaction but the steps are the\\nsame.\\n\\n>\\n> 4. There are places in the patch like apply_handle_stream_start()\\n> which sets commit_lsn in callback arg as InvalidXLogRecPtr but the\\n> callback function seems to be assuming that it is always a valid\\n> value. Shouldn't we need to avoid appending LSN for such cases?\\n\\nAgreed. Will fix it in the next version patch.\\n\\nI'm updating the patches and will submit them.\\n\\nRegards,\\n\\n--\\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\\n\\n\\n\",\n \"msg_date\": \"Thu, 3 Mar 2022 22:02:49 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Thu, Mar 3, 2022 at 10:02 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n>\\n> I'm updating the patches and will submit them.\\n\\nAttached updated version patches.\\n\\nRegards,\\n\\n-- \\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\",\n \"msg_date\": \"Fri, 4 Mar 2022 10:09:28 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Friday, March 4, 2022 10:09 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\r\\n> On Thu, Mar 3, 2022 at 10:02 PM Masahiko Sawada\\r\\n> <sawada.mshk@gmail.com> wrote:\\r\\n> >\\r\\n> >\\r\\n> > I'm updating the patches and will submit them.\\r\\n> \\r\\n> Attached updated version patches.\\r\\nThank you for sharing the patch v3.\\r\\n\\r\\nFew minor comments.\\r\\n\\r\\n\\r\\n(1) v03-0001, apply_error_callback function\\r\\n\\r\\n- /* append transaction information */\\r\\n- if (TransactionIdIsNormal(errarg->remote_xid))\\r\\n+ if (errarg->rel == NULL)\\r\\n {\\r\\n- appendStringInfo(&buf, _(\\\" in transaction %u\\\"), errarg->remote_xid);\\r\\n\\r\\nShould write !errarg->rel ?\\r\\n\\r\\n(2) v03-0002, doc/src/sgml/logical-replication.sgml\\r\\n\\r\\n\\r\\n+ transaction that conflicts with the existing data. When a conflict produces\\r\\n+ an error, it is shown in the subscriber's server logs as follows:\\r\\n+<screen>\\r\\n+ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\r\\n+DETAIL: Key (c)=(1) already exists.\\r\\n+CONTEXT: processing remote data during \\\"INSERT\\\" for replication target relation \\\"public.test\\\" in transaction 725 committed at LSN 0/14BFA88\\r\\n+</screen>\\r\\n\\r\\n\\r\\nWe should update the CONTEXT message by using the v3-0001.\\r\\n\\r\\n(3) v03-0002, doc/src/sgml/logical-replication.sgml\\r\\n\\r\\n\\r\\n+ The LSN of the transaction that contains the change violating the constraint and\\r\\n+ the replication origin name can be found from those outputs (LSN 0/14C0378 and\\r\\n+ replication origin <literal>pg_16395</literal> in the above case). To skip the\\r\\n+ transaction, the subscription needs to be disabled temporarily by\\r\\n+ <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the transaction\\r\\n+ can be skipped by calling the <link linkend=\\\"pg-replication-origin-advance\\\">\\r\\n\\r\\nThe LSN(0/14C0378) is not same as the one in the above error context.\\r\\nIt's recommended to check LSNs directly written in the documentation.\\r\\n\\r\\n(4) one confirmation\\r\\n\\r\\nWe don't have a TAP test of pg_replication_origin_advance()\\r\\nfor v3, that utilizes this new log in a logical replication setup.\\r\\nThis is because existing tests for this function (in test_decoding) is only for permission check\\r\\nand argument validation, and we're just changing error message itself.\\r\\nIs this correct ?\\r\\n\\r\\n\\r\\nBest Regards,\\r\\n\\tTakamichi Osumi\\r\\n\\r\\n\",\n \"msg_date\": \"Fri, 4 Mar 2022 02:27:51 +0000\",\n \"msg_from\": \"\\\"osumi.takamichi@fujitsu.com\\\" <osumi.takamichi@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 4, 2022 at 11:27 AM osumi.takamichi@fujitsu.com\\n<osumi.takamichi@fujitsu.com> wrote:\\n>\\n> On Friday, March 4, 2022 10:09 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> > On Thu, Mar 3, 2022 at 10:02 PM Masahiko Sawada\\n> > <sawada.mshk@gmail.com> wrote:\\n> > >\\n> > >\\n> > > I'm updating the patches and will submit them.\\n> >\\n> > Attached updated version patches.\\n> Thank you for sharing the patch v3.\\n>\\n> Few minor comments.\\n>\\n>\\n> (1) v03-0001, apply_error_callback function\\n>\\n> - /* append transaction information */\\n> - if (TransactionIdIsNormal(errarg->remote_xid))\\n> + if (errarg->rel == NULL)\\n> {\\n> - appendStringInfo(&buf, _(\\\" in transaction %u\\\"), errarg->remote_xid);\\n>\\n> Should write !errarg->rel ?\\n\\nI think either should be fine.\\n\\n>\\n> (2) v03-0002, doc/src/sgml/logical-replication.sgml\\n>\\n>\\n> + transaction that conflicts with the existing data. When a conflict produces\\n> + an error, it is shown in the subscriber's server logs as follows:\\n> +<screen>\\n> +ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> +DETAIL: Key (c)=(1) already exists.\\n> +CONTEXT: processing remote data during \\\"INSERT\\\" for replication target relation \\\"public.test\\\" in transaction 725 committed at LSN 0/14BFA88\\n> +</screen>\\n>\\n>\\n> We should update the CONTEXT message by using the v3-0001.\\n>\\n> (3) v03-0002, doc/src/sgml/logical-replication.sgml\\n>\\n>\\n> + The LSN of the transaction that contains the change violating the constraint and\\n> + the replication origin name can be found from those outputs (LSN 0/14C0378 and\\n> + replication origin <literal>pg_16395</literal> in the above case). To skip the\\n> + transaction, the subscription needs to be disabled temporarily by\\n> + <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the transaction\\n> + can be skipped by calling the <link linkend=\\\"pg-replication-origin-advance\\\">\\n>\\n> The LSN(0/14C0378) is not same as the one in the above error context.\\n> It's recommended to check LSNs directly written in the documentation.\\n\\nRight, I missed checking and updating it.\\n\\n>\\n> (4) one confirmation\\n>\\n> We don't have a TAP test of pg_replication_origin_advance()\\n> for v3, that utilizes this new log in a logical replication setup.\\n> This is because existing tests for this function (in test_decoding) is only for permission check\\n> and argument validation, and we're just changing error message itself.\\n> Is this correct ?\\n\\nYeah, I think it’s a good idea to add test in general but I don't\\nthink we should include the tests for skipping a transaction by using\\npg_replication_origin() in this patch because it's an existing way and\\nupcoming ALTER SUBSCRIPTION SKIP patch includes the tests and it's\\nmore appropriate way. But if others also think it should do that too\\nalong with this update, I'm happy to add tests.\\n\\nI've attached updated patches.\\n\\nRegards,\\n\\n--\\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\",\n \"msg_date\": \"Fri, 4 Mar 2022 14:22:54 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 4, 2022 at 6:40 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n> Attached updated version patches.\\n>\\n\\nThe patch looks mostly good to me. Few minor comments:\\n1. I think we can have an Assert for errarg->origin_name in\\napply_error_callback after checking the command as this function\\nassumes that it will always be set.\\n2. I suggest minor changes in the documentation change:\\nWhen a conflict produces an error, the replication won't proceed, and\\nthe apply worker will emit the following kind of message to the\\nsubscriber's server log:\\n+<screen>\\n+ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n+DETAIL: Key (c)=(1) already exists.\\n+CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\ntarget relation \\\"public.test\\\" in transaction 725 committed at LSN\\n0/14BFA88 from replication origin \\\"pg_16395\\\"\\n+</screen>\\n\\nThe LSN of the transaction that contains the change violating the\\nconstraint and the replication origin name can be found from the\\nserver log (LSN 0/14C0378 and replication origin\\n<literal>pg_16395</literal> in the above case). To skip the\\ntransaction, the subscription needs to be disabled temporarily by\\n<command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\\ntransaction can be skipped by calling the <link\\nlinkend=\\\"pg-replication-origin-advance\\\">\\n<function>pg_replication_origin_advance()</function></link> function\\nwith the <parameter>node_name</parameter> (i.e.,\\n<literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\\nLSN 0/14C0379).\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Mar 2022 11:24:53 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 4, 2022 at 10:53 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n> On Fri, Mar 4, 2022 at 11:27 AM osumi.takamichi@fujitsu.com\\n> <osumi.takamichi@fujitsu.com> wrote:\\n> >\\n> > (4) one confirmation\\n> >\\n> > We don't have a TAP test of pg_replication_origin_advance()\\n> > for v3, that utilizes this new log in a logical replication setup.\\n> > This is because existing tests for this function (in test_decoding) is only for permission check\\n> > and argument validation, and we're just changing error message itself.\\n> > Is this correct ?\\n>\\n> Yeah, I think it’s a good idea to add test in general but I don't\\n> think we should include the tests for skipping a transaction by using\\n> pg_replication_origin() in this patch because it's an existing way and\\n> upcoming ALTER SUBSCRIPTION SKIP patch includes the tests and it's\\n> more appropriate way. But if others also think it should do that too\\n> along with this update, I'm happy to add tests.\\n>\\n\\nI also don't see a reason why this patch should add any tests related\\nto origin.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Mar 2022 11:26:20 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Friday, March 4, 2022 2:23 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\r\\n> I've attached updated patches.\\r\\nHi, thank you for updating the patch.\\r\\n\\r\\nOne comment on v4.\\r\\n\\r\\nIn v4-0002, we introduce 'commit_lsn' in the ApplyErrorCallbackArg.\\r\\nThis member is set for prepare, rollback prepared and stream_abort as well.\\r\\nThe new log message format is useful when we have a prepare transaction\\r\\nthat keeps failing on the subscriber and want to know the target transaction\\r\\nfor the pg_replication_origin_advance(), right ?\\r\\nIf this is true, I wasn't sure if the name 'commit_lsn' is the\\r\\nmost accurate name for this variable. Should we adjust the name a bit ?\\r\\n\\r\\nEven when we decide to continue to use 'commit_lsn',\\r\\nit might be better to add some comments near the definition, I feel.\\r\\n\\r\\n\\r\\nBest Regards,\\r\\n\\tTakamichi Osumi\\r\\n\\r\\n\",\n \"msg_date\": \"Fri, 4 Mar 2022 06:15:49 +0000\",\n \"msg_from\": \"\\\"osumi.takamichi@fujitsu.com\\\" <osumi.takamichi@fujitsu.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"RE: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 4, 2022 at 11:45 AM osumi.takamichi@fujitsu.com\\n<osumi.takamichi@fujitsu.com> wrote:\\n>\\n> On Friday, March 4, 2022 2:23 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> > I've attached updated patches.\\n> Hi, thank you for updating the patch.\\n>\\n> One comment on v4.\\n>\\n> In v4-0002, we introduce 'commit_lsn' in the ApplyErrorCallbackArg.\\n> This member is set for prepare, rollback prepared and stream_abort as well.\\n> The new log message format is useful when we have a prepare transaction\\n> that keeps failing on the subscriber and want to know the target transaction\\n> for the pg_replication_origin_advance(), right ?\\n> If this is true, I wasn't sure if the name 'commit_lsn' is the\\n> most accurate name for this variable. Should we adjust the name a bit ?\\n>\\n\\nIf we want to change this variable name, the other options could be\\n'end_lsn', or 'finish_lsn'.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Fri, 4 Mar 2022 11:51:41 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 4, 2022 at 2:55 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, Mar 4, 2022 at 6:40 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> >\\n> > Attached updated version patches.\\n> >\\n>\\n> The patch looks mostly good to me. Few minor comments:\\n\\nThank you for the comments!\\n\\n> 1. I think we can have an Assert for errarg->origin_name in\\n> apply_error_callback after checking the command as this function\\n> assumes that it will always be set.\\n\\nAdded.\\n\\n> 2. I suggest minor changes in the documentation change:\\n> When a conflict produces an error, the replication won't proceed, and\\n> the apply worker will emit the following kind of message to the\\n> subscriber's server log:\\n> +<screen>\\n> +ERROR: duplicate key value violates unique constraint \\\"test_pkey\\\"\\n> +DETAIL: Key (c)=(1) already exists.\\n> +CONTEXT: processing remote data during \\\"INSERT\\\" for replication\\n> target relation \\\"public.test\\\" in transaction 725 committed at LSN\\n> 0/14BFA88 from replication origin \\\"pg_16395\\\"\\n> +</screen>\\n>\\n> The LSN of the transaction that contains the change violating the\\n> constraint and the replication origin name can be found from the\\n> server log (LSN 0/14C0378 and replication origin\\n> <literal>pg_16395</literal> in the above case). To skip the\\n> transaction, the subscription needs to be disabled temporarily by\\n> <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\\n> transaction can be skipped by calling the <link\\n> linkend=\\\"pg-replication-origin-advance\\\">\\n> <function>pg_replication_origin_advance()</function></link> function\\n> with the <parameter>node_name</parameter> (i.e.,\\n> <literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\\n> LSN 0/14C0379).\\n\\nFixed.\\n\\nOn Fri, Mar 4, 2022 at 3:21 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Fri, Mar 4, 2022 at 11:45 AM osumi.takamichi@fujitsu.com\\n> <osumi.takamichi@fujitsu.com> wrote:\\n> >\\n> > On Friday, March 4, 2022 2:23 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> > > I've attached updated patches.\\n> > Hi, thank you for updating the patch.\\n> >\\n> > One comment on v4.\\n> >\\n> > In v4-0002, we introduce 'commit_lsn' in the ApplyErrorCallbackArg.\\n> > This member is set for prepare, rollback prepared and stream_abort as well.\\n> > The new log message format is useful when we have a prepare transaction\\n> > that keeps failing on the subscriber and want to know the target transaction\\n> > for the pg_replication_origin_advance(), right ?\\n> > If this is true, I wasn't sure if the name 'commit_lsn' is the\\n> > most accurate name for this variable. Should we adjust the name a bit ?\\n> >\\n>\\n> If we want to change this variable name, the other options could be\\n> 'end_lsn', or 'finish_lsn'.\\n>\\n\\nAgreed with 'finish_lsn'.\\n\\nI've attached updated patches. Please review them.\\n\\nRegards,\\n\\n-- \\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\",\n \"msg_date\": \"Fri, 4 Mar 2022 20:43:50 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 4, 2022, at 2:54 AM, Amit Kapila wrote:\\n> The LSN of the transaction that contains the change violating the\\n> constraint and the replication origin name can be found from the\\n> server log (LSN 0/14C0378 and replication origin\\n> <literal>pg_16395</literal> in the above case). To skip the\\n> transaction, the subscription needs to be disabled temporarily by\\n> <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\\n> transaction can be skipped by calling the <link\\n> linkend=\\\"pg-replication-origin-advance\\\">\\n> <function>pg_replication_origin_advance()</function></link> function\\n> with the <parameter>node_name</parameter> (i.e.,\\n> <literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\\n> LSN 0/14C0379).\\nYou could also add:\\n\\nAfter that the replication can be resumed by <command>ALTER SUBSCRIPTION ...\\nENABLE</command>.\\n\\nLet's provide complete instructions.\\n\\n\\n--\\nEuler Taveira\\nEDB https://www.enterprisedb.com/\\n\\nOn Fri, Mar 4, 2022, at 2:54 AM, Amit Kapila wrote:The LSN of the transaction that contains the change violating theconstraint and the replication origin name can be found from theserver log (LSN 0/14C0378 and replication origin<literal>pg_16395</literal> in the above case). To skip thetransaction, the subscription needs to be disabled temporarily by<command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, thetransaction can be skipped by calling the <linklinkend=\\\"pg-replication-origin-advance\\\"><function>pg_replication_origin_advance()</function></link> functionwith the <parameter>node_name</parameter> (i.e.,<literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,LSN 0/14C0379).You could also add:After that the replication can be resumed by <command>ALTER SUBSCRIPTION ...ENABLE</command>.Let's provide complete instructions.--Euler TaveiraEDB https://www.enterprisedb.com/\",\n \"msg_date\": \"Fri, 04 Mar 2022 09:32:18 -0300\",\n \"msg_from\": \"\\\"Euler Taveira\\\" <euler@eulerto.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication\\n worker errcontext\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 4, 2022 at 6:02 PM Euler Taveira <euler@eulerto.com> wrote:\\n>\\n> On Fri, Mar 4, 2022, at 2:54 AM, Amit Kapila wrote:\\n>\\n> The LSN of the transaction that contains the change violating the\\n> constraint and the replication origin name can be found from the\\n> server log (LSN 0/14C0378 and replication origin\\n> <literal>pg_16395</literal> in the above case). To skip the\\n> transaction, the subscription needs to be disabled temporarily by\\n> <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\\n> transaction can be skipped by calling the <link\\n> linkend=\\\"pg-replication-origin-advance\\\">\\n> <function>pg_replication_origin_advance()</function></link> function\\n> with the <parameter>node_name</parameter> (i.e.,\\n> <literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\\n> LSN 0/14C0379).\\n>\\n> You could also add:\\n>\\n> After that the replication can be resumed by <command>ALTER SUBSCRIPTION ...\\n> ENABLE</command>.\\n>\\n> Let's provide complete instructions.\\n>\\n\\n+1. That sounds reasonable to me.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Sat, 5 Mar 2022 06:42:09 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Fri, Mar 4, 2022 at 9:32 PM Euler Taveira <euler@eulerto.com> wrote:\\n>\\n> On Fri, Mar 4, 2022, at 2:54 AM, Amit Kapila wrote:\\n>\\n> The LSN of the transaction that contains the change violating the\\n> constraint and the replication origin name can be found from the\\n> server log (LSN 0/14C0378 and replication origin\\n> <literal>pg_16395</literal> in the above case). To skip the\\n> transaction, the subscription needs to be disabled temporarily by\\n> <command>ALTER SUBSCRIPTION ... DISABLE</command> first. Then, the\\n> transaction can be skipped by calling the <link\\n> linkend=\\\"pg-replication-origin-advance\\\">\\n> <function>pg_replication_origin_advance()</function></link> function\\n> with the <parameter>node_name</parameter> (i.e.,\\n> <literal>pg_16395</literal>) and the next LSN of the commit LSN (i.e.,\\n> LSN 0/14C0379).\\n>\\n> You could also add:\\n>\\n> After that the replication can be resumed by <command>ALTER SUBSCRIPTION ...\\n> ENABLE</command>.\\n>\\n> Let's provide complete instructions.\\n\\nThank you for the comment. +1.\\n\\nI've attached updated patches.\\n\\nRegards,\\n\\n-- \\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\",\n \"msg_date\": \"Mon, 7 Mar 2022 10:05:25 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 6:36 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n>\\n> Thank you for the comment. +1.\\n>\\n> I've attached updated patches.\\n>\\n\\nPushed the first patch. Fixed one typo in the second patch and\\nslightly changed the commit message, otherwise, it looks good to me.\\nI'll push this tomorrow unless there are more comments.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\",\n \"msg_date\": \"Mon, 7 Mar 2022 10:06:37 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Mon, Mar 7, 2022 at 10:06 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 6:36 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> >\\n> > Thank you for the comment. +1.\\n> >\\n> > I've attached updated patches.\\n> >\\n>\\n> Pushed the first patch. Fixed one typo in the second patch and\\n> slightly changed the commit message, otherwise, it looks good to me.\\n> I'll push this tomorrow unless there are more comments.\\n>\\n\\nPushed.\\n\\n-- \\nWith Regards,\\nAmit Kapila.\\n\\n\\n\",\n \"msg_date\": \"Tue, 8 Mar 2022 16:26:16 +0530\",\n \"msg_from\": \"Amit Kapila <amit.kapila16@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 8, 2022 at 7:56 PM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n>\\n> On Mon, Mar 7, 2022 at 10:06 AM Amit Kapila <amit.kapila16@gmail.com> wrote:\\n> >\\n> > On Mon, Mar 7, 2022 at 6:36 AM Masahiko Sawada <sawada.mshk@gmail.com> wrote:\\n> > >\\n> > > Thank you for the comment. +1.\\n> > >\\n> > > I've attached updated patches.\\n> > >\\n> >\\n> > Pushed the first patch. Fixed one typo in the second patch and\\n> > slightly changed the commit message, otherwise, it looks good to me.\\n> > I'll push this tomorrow unless there are more comments.\\n> >\\n>\\n> Pushed.\\n\\nThank you!\\n\\nRegards,\\n\\n-- \\nMasahiko Sawada\\nEDB: https://www.enterprisedb.com/\\n\\n\\n\",\n \"msg_date\": \"Wed, 9 Mar 2022 09:32:47 +0900\",\n \"msg_from\": \"Masahiko Sawada <sawada.mshk@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Add the replication origin name and commit-LSN to logical\\n replication worker errcontext\"\n }\n]","isConversational":false}}},{"rowIdx":53596,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hello\nI just spotted in src/include/access/xlog.h:\nextern XLogRecPtr CalculateMaxmumSafeLSN(void);\n\nThis function doesn't seem to be used anywhere or even defined? \"git grep CalculateMaxmumSafeLSN\" shows only this line.\nWas added in c6550776394e25c1620bc8258427c8f1d448080d \"Allow users to limit storage reserved by replication slots\"\n\nregards, Sergei\n\n\n","msg_date":"Mon, 28 Feb 2022 17:01:10 +0300","msg_from":"Sergei Kornilov <sk@zsrv.org>","msg_from_op":true,"msg_subject":"definition of CalculateMaxmumSafeLSN"},{"msg_contents":"At Mon, 28 Feb 2022 17:01:10 +0300, Sergei Kornilov <sk@zsrv.org> wrote in \n> Hello\n> I just spotted in src/include/access/xlog.h:\n> extern XLogRecPtr CalculateMaxmumSafeLSN(void);\n> \n> This function doesn't seem to be used anywhere or even defined? \"git grep CalculateMaxmumSafeLSN\" shows only this line.\n> Was added in c6550776394e25c1620bc8258427c8f1d448080d \"Allow users to limit storage reserved by replication slots\"\n\nHmm. I think I remember of that name. You're right, it is a function\nthat was once added during development of the commit but eventually\ngone.\n\nSo it should be removed.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Tue, 01 Mar 2022 10:11:03 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: definition of CalculateMaxmumSafeLSN"},{"msg_contents":"Hi,\n\nOn Tue, Mar 01, 2022 at 10:11:03AM +0900, Kyotaro Horiguchi wrote:\n> At Mon, 28 Feb 2022 17:01:10 +0300, Sergei Kornilov <sk@zsrv.org> wrote in \n> > Hello\n> > I just spotted in src/include/access/xlog.h:\n> > extern XLogRecPtr CalculateMaxmumSafeLSN(void);\n> > \n> > This function doesn't seem to be used anywhere or even defined? \"git grep CalculateMaxmumSafeLSN\" shows only this line.\n> > Was added in c6550776394e25c1620bc8258427c8f1d448080d \"Allow users to limit storage reserved by replication slots\"\n> \n> Hmm. I think I remember of that name. You're right, it is a function\n> that was once added during development of the commit but eventually\n> gone.\n> \n> So it should be removed.\n\n+1. I'm Ccing �lvaro as he's the committer.\n\n\n","msg_date":"Tue, 1 Mar 2022 12:44:40 +0800","msg_from":"Julien Rouhaud <rjuju123@gmail.com>","msg_from_op":false,"msg_subject":"Re: definition of CalculateMaxmumSafeLSN"},{"msg_contents":"On Mon, Feb 28, 2022 at 7:31 PM Sergei Kornilov <sk@zsrv.org> wrote:\n>\n> Hello\n> I just spotted in src/include/access/xlog.h:\n> extern XLogRecPtr CalculateMaxmumSafeLSN(void);\n>\n> This function doesn't seem to be used anywhere or even defined? \"git grep CalculateMaxmumSafeLSN\" shows only this line.\n> Was added in c6550776394e25c1620bc8258427c8f1d448080d \"Allow users to limit storage reserved by replication slots\"\n\n+1 to remove. Although the change is straightforward, I attached the\npatch just for the sake of convention. I've marked it as RfC\nhttps://commitfest.postgresql.org/38/3616/.\n\nJust for the record - there's another thread\nhttps://www.postgresql.org/message-id/CALj2ACVoQ7NEf43Xz0rfxsGOKYTN5r4VZp2DO2_5p%2BCMzsRPFw%40mail.gmail.com\n\nRegards,\nBharath Rupireddy.","msg_date":"Sat, 21 May 2022 16:42:20 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":false,"msg_subject":"Re: definition of CalculateMaxmumSafeLSN"},{"msg_contents":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> writes:\n> On Mon, Feb 28, 2022 at 7:31 PM Sergei Kornilov <sk@zsrv.org> wrote:\n>> I just spotted in src/include/access/xlog.h:\n>> extern XLogRecPtr CalculateMaxmumSafeLSN(void);\n>> This function doesn't seem to be used anywhere or even defined? \"git grep CalculateMaxmumSafeLSN\" shows only this line.\n\n> +1 to remove.\n\nRight, done.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Sat, 21 May 2022 13:27:12 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: definition of CalculateMaxmumSafeLSN"}],"string":"[\n {\n \"msg_contents\": \"Hello\\nI just spotted in src/include/access/xlog.h:\\nextern XLogRecPtr CalculateMaxmumSafeLSN(void);\\n\\nThis function doesn't seem to be used anywhere or even defined? \\\"git grep CalculateMaxmumSafeLSN\\\" shows only this line.\\nWas added in c6550776394e25c1620bc8258427c8f1d448080d \\\"Allow users to limit storage reserved by replication slots\\\"\\n\\nregards, Sergei\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 17:01:10 +0300\",\n \"msg_from\": \"Sergei Kornilov <sk@zsrv.org>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"definition of CalculateMaxmumSafeLSN\"\n },\n {\n \"msg_contents\": \"At Mon, 28 Feb 2022 17:01:10 +0300, Sergei Kornilov <sk@zsrv.org> wrote in \\n> Hello\\n> I just spotted in src/include/access/xlog.h:\\n> extern XLogRecPtr CalculateMaxmumSafeLSN(void);\\n> \\n> This function doesn't seem to be used anywhere or even defined? \\\"git grep CalculateMaxmumSafeLSN\\\" shows only this line.\\n> Was added in c6550776394e25c1620bc8258427c8f1d448080d \\\"Allow users to limit storage reserved by replication slots\\\"\\n\\nHmm. I think I remember of that name. You're right, it is a function\\nthat was once added during development of the commit but eventually\\ngone.\\n\\nSo it should be removed.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Tue, 01 Mar 2022 10:11:03 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: definition of CalculateMaxmumSafeLSN\"\n },\n {\n \"msg_contents\": \"Hi,\\n\\nOn Tue, Mar 01, 2022 at 10:11:03AM +0900, Kyotaro Horiguchi wrote:\\n> At Mon, 28 Feb 2022 17:01:10 +0300, Sergei Kornilov <sk@zsrv.org> wrote in \\n> > Hello\\n> > I just spotted in src/include/access/xlog.h:\\n> > extern XLogRecPtr CalculateMaxmumSafeLSN(void);\\n> > \\n> > This function doesn't seem to be used anywhere or even defined? \\\"git grep CalculateMaxmumSafeLSN\\\" shows only this line.\\n> > Was added in c6550776394e25c1620bc8258427c8f1d448080d \\\"Allow users to limit storage reserved by replication slots\\\"\\n> \\n> Hmm. I think I remember of that name. You're right, it is a function\\n> that was once added during development of the commit but eventually\\n> gone.\\n> \\n> So it should be removed.\\n\\n+1. I'm Ccing �lvaro as he's the committer.\\n\\n\\n\",\n \"msg_date\": \"Tue, 1 Mar 2022 12:44:40 +0800\",\n \"msg_from\": \"Julien Rouhaud <rjuju123@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: definition of CalculateMaxmumSafeLSN\"\n },\n {\n \"msg_contents\": \"On Mon, Feb 28, 2022 at 7:31 PM Sergei Kornilov <sk@zsrv.org> wrote:\\n>\\n> Hello\\n> I just spotted in src/include/access/xlog.h:\\n> extern XLogRecPtr CalculateMaxmumSafeLSN(void);\\n>\\n> This function doesn't seem to be used anywhere or even defined? \\\"git grep CalculateMaxmumSafeLSN\\\" shows only this line.\\n> Was added in c6550776394e25c1620bc8258427c8f1d448080d \\\"Allow users to limit storage reserved by replication slots\\\"\\n\\n+1 to remove. Although the change is straightforward, I attached the\\npatch just for the sake of convention. I've marked it as RfC\\nhttps://commitfest.postgresql.org/38/3616/.\\n\\nJust for the record - there's another thread\\nhttps://www.postgresql.org/message-id/CALj2ACVoQ7NEf43Xz0rfxsGOKYTN5r4VZp2DO2_5p%2BCMzsRPFw%40mail.gmail.com\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sat, 21 May 2022 16:42:20 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: definition of CalculateMaxmumSafeLSN\"\n },\n {\n \"msg_contents\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> writes:\\n> On Mon, Feb 28, 2022 at 7:31 PM Sergei Kornilov <sk@zsrv.org> wrote:\\n>> I just spotted in src/include/access/xlog.h:\\n>> extern XLogRecPtr CalculateMaxmumSafeLSN(void);\\n>> This function doesn't seem to be used anywhere or even defined? \\\"git grep CalculateMaxmumSafeLSN\\\" shows only this line.\\n\\n> +1 to remove.\\n\\nRight, done.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Sat, 21 May 2022 13:27:12 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: definition of CalculateMaxmumSafeLSN\"\n }\n]","isConversational":false}}},{"rowIdx":53597,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi,\n\nIt looks like we use \"log segment\" in various user-facing messages.\nThe term \"log\" can mean server logs as well. The \"WAL segment\" suits\nwell here and it is consistently used across the other user-facing\nmessages [1].\n\nHere's a small patch attempting to consistently use the \"WAL segment\".\n\nThoughts?\n\n[1]\npg_log_error(\"could not fetch WAL segment size: got %d rows and %d\nfields, expected %d rows and %d or more fields\",\npg_log_error(\"WAL segment size could not be parsed\");\npg_log_error(ngettext(\"WAL segment size must be a power of two between\n1 MB and 1 GB, but the remote server reported a value of %d byte\",\nprintf(_(\"WARNING: invalid WAL segment size\\n\"));\nprintf(_(\"Bytes per WAL segment: %u\\n\"),\nfatal_error(ngettext(\"WAL segment size must be a power of two between\n1 MB and 1 GB, but the WAL file \\\"%s\\\" header specifies %d byte\",\nerrmsg(\"requested WAL segment %s has already been removed\",\nelog(DEBUG2, \"removed temporary WAL segment \\\"%s\\\"\", path);\n\nRegards,\nBharath Rupireddy.","msg_date":"Mon, 28 Feb 2022 21:03:07 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"At Mon, 28 Feb 2022 21:03:07 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in \n> Hi,\n> \n> It looks like we use \"log segment\" in various user-facing messages.\n> The term \"log\" can mean server logs as well. The \"WAL segment\" suits\n> well here and it is consistently used across the other user-facing\n> messages [1].\n> \n> Here's a small patch attempting to consistently use the \"WAL segment\".\n> \n> Thoughts?\n\nI tend to agree to this. I also see \"log record(s)\" (without prefixed\nby \"write-ahead\") in many places especially in the documentation. I'm\nnot sure how we should treat \"WAL log\", though.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Tue, 01 Mar 2022 10:20:53 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"On Tue, Mar 1, 2022 at 6:50 AM Kyotaro Horiguchi\n<horikyota.ntt@gmail.com> wrote:\n>\n> At Mon, 28 Feb 2022 21:03:07 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in\n> > Hi,\n> >\n> > It looks like we use \"log segment\" in various user-facing messages.\n> > The term \"log\" can mean server logs as well. The \"WAL segment\" suits\n> > well here and it is consistently used across the other user-facing\n> > messages [1].\n> >\n> > Here's a small patch attempting to consistently use the \"WAL segment\".\n> >\n> > Thoughts?\n>\n> I tend to agree to this.\n\nThanks for taking a look at it. Here's the CF entry -\nhttps://commitfest.postgresql.org/38/3584/\n\n> I also see \"log record(s)\" (without prefixed\n> by \"write-ahead\") in many places especially in the documentation. I'm\n> not sure how we should treat \"WAL log\", though.\n\nYes, but the docs have a glossary term for 'Log record\" [1]. FWIW\nattaching docs change as v2-0002 patch. I found another place where\n\"log records\" is being used in pg_waldump.c, I changed that and\nattached v2-0001 patch.\n\nPlease review the v2 patch set.\n\n[1]\n <glossentry id=\"glossary-log-record\">\n <glossterm>Log record</glossterm>\n <glossdef>\n <para>\n Archaic term for a <glossterm linkend=\"glossary-wal-record\">WAL\nrecord</glossterm>.\n </para>\n </glossdef>\n </glossentry>\n\nRegards,\nBharath Rupireddy.","msg_date":"Wed, 2 Mar 2022 11:41:49 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"On Wed, Mar 2, 2022 at 11:41 AM Bharath Rupireddy\n<bharath.rupireddyforpostgres@gmail.com> wrote:\n>\n> Please review the v2 patch set.\n\nHad to rebase. Attaching v3 patch set.\n\nRegards,\nBharath Rupireddy.","msg_date":"Sat, 26 Mar 2022 11:26:16 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":" At all times, <productname>PostgreSQL</productname> maintains a\n <firstterm>write ahead log</firstterm> (WAL) in the <filename>pg_wal/</filename>\n- subdirectory of the cluster's data directory. The log records\n- every change made to the database's data files. This log exists\n+ subdirectory of the cluster's data directory. The WAL records\n+ capture every change made to the database's data files. This log exists\n\nI don't think this change really adds anything. The preceding sentence\nmakes it clear that we are discussing the write-ahead log, and IMO the\nchange in phrasing (\"the log records every change\" is changed to \"the\nrecords capture every change\") subtly changes the meaning of the sentence.\n\nThe rest looks good to me. \n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Thu, 31 Mar 2022 08:45:56 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"At Thu, 31 Mar 2022 08:45:56 -0700, Nathan Bossart <nathandbossart@gmail.com> wrote in \n> At all times, <productname>PostgreSQL</productname> maintains a\n> <firstterm>write ahead log</firstterm> (WAL) in the <filename>pg_wal/</filename>\n> - subdirectory of the cluster's data directory. The log records\n> - every change made to the database's data files. This log exists\n> + subdirectory of the cluster's data directory. The WAL records\n> + capture every change made to the database's data files. This log exists\n> \n> I don't think this change really adds anything. The preceding sentence\n> makes it clear that we are discussing the write-ahead log, and IMO the\n> change in phrasing (\"the log records every change\" is changed to \"the\n> records capture every change\") subtly changes the meaning of the sentence.\n> \n> The rest looks good to me. \n\n+1. It is not a composite noun \"log records\".\n\nThe original sentence is \"S(The log) V(records) O(every change that is\nmade to .. files)\". The proposed change looks like changing it to\n\"S(The log records) V(capture) O(every ..files)\". In that sense, the\noriginal one seem rather correct to me, since \"capture\" seems to have\nthe implication of \"write after log..\", to me.\n\n\nI looked though the document and found other use of \"log\nrecord|segment\". What do you think about the attached?\n\nThere're some uncertain point in the change.\n\n you should at least save the contents of the cluster's <filename>pg_wal</filename>\n- subdirectory, as it might contain logs which\n+ subdirectory, as it might contain WAL files which\n were not archived before the system went down.\n\nThe \"logs\" means acutally \"WAL segment (files)\" but the concept of\n\"segment\" is out of focus in the context. So just \"file\" is used\nthere. The same change is applied on dezon of places.\n\n\n- disk-space requirements for the <acronym>WAL</acronym> logs are met,\n+ disk-space requirements for the <acronym>WAL</acronym> are met,\n\nThis might be better be \"WAL files\" instead of just \"WAL\".\n\n\n- <acronym>WAL</acronym> logs are stored in the directory\n+ <acronym>WAL</acronym> is stored in the directory\n <filename>pg_wal</filename> under the data directory, as a set of\n\nI'm not sure which is better, use \"WAL\" as a collective noun, or \"WAL\nfiles\" as the cocrete objects.\n\n\n- The aim of <acronym>WAL</acronym> is to ensure that the log is\n+ The aim of <acronym>WAL</acronym> is to ensure that the WAL record is\n written before database records are altered, but this can be subverted by\n\nThis is not a mechanical change. But I think this is correct.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center","msg_date":"Fri, 01 Apr 2022 10:31:10 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"On Fri, Apr 01, 2022 at 10:31:10AM +0900, Kyotaro Horiguchi wrote:\n> you should at least save the contents of the cluster's <filename>pg_wal</filename>\n> - subdirectory, as it might contain logs which\n> + subdirectory, as it might contain WAL files which\n> were not archived before the system went down.\n> \n> The \"logs\" means acutally \"WAL segment (files)\" but the concept of\n> \"segment\" is out of focus in the context. So just \"file\" is used\n> there. The same change is applied on dezon of places.\n\nThis change seems reasonable to me.\n\n> - disk-space requirements for the <acronym>WAL</acronym> logs are met,\n> + disk-space requirements for the <acronym>WAL</acronym> are met,\n> \n> This might be better be \"WAL files\" instead of just \"WAL\".\n\n+1 for \"WAL files\"\n\n> - <acronym>WAL</acronym> logs are stored in the directory\n> + <acronym>WAL</acronym> is stored in the directory\n> <filename>pg_wal</filename> under the data directory, as a set of\n> \n> I'm not sure which is better, use \"WAL\" as a collective noun, or \"WAL\n> files\" as the cocrete objects.\n\nMy vote is for \"WAL files\" because it was previously \"WAL logs.\"\n\n> - The aim of <acronym>WAL</acronym> is to ensure that the log is\n> + The aim of <acronym>WAL</acronym> is to ensure that the WAL record is\n> written before database records are altered, but this can be subverted by\n> \n> This is not a mechanical change. But I think this is correct.\n\nIMO the original wording is fine. I think it is sufficiently clear that\n\"log\" refers to \"write-ahead log,\" and this sentence seems intended to\nconvey the basic rule of \"log before data.\" However, the rest of the\nsentence is a little weird. It's basically saying \"the aim of the log is\nto ensure that the log is written...\" Isn't the aim of the log to record\nthe database activity? Perhaps we should rewrite it to something like the\nfollowing:\n\n\tA basic rule of WAL is that the log must be written before the database\n\tfiles are altered, but this can be...\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Fri, 1 Apr 2022 07:24:48 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"It's been a few weeks, so I'm marking the commitfest entry as\nwaiting-on-author.\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Mon, 25 Apr 2022 12:54:43 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"On Tue, Apr 26, 2022 at 1:24 AM Nathan Bossart <nathandbossart@gmail.com> wrote:\n>\n> It's been a few weeks, so I'm marking the commitfest entry as\n> waiting-on-author.\n\nThanks. I'm attaching the updated v4 patches (also subsumed Kyotaro\nSan's patch at [1]). Please review it further.\n\n[1] https://www.postgresql.org/message-id/20220401.103110.1103213854487561781.horikyota.ntt%40gmail.com\n\nRegards,\nBharath Rupireddy.","msg_date":"Mon, 18 Jul 2022 16:24:12 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"Overall, these patches look reasonable.\n\nOn Mon, Jul 18, 2022 at 04:24:12PM +0530, Bharath Rupireddy wrote:\n> record. Because the entire content of data pages is saved in the\n> - log on the first page modification after a checkpoint (assuming\n> + WAL record on the first page modification after a checkpoint (assuming\n> <xref linkend=\"guc-full-page-writes\"/> is not disabled), all pages\n> changed since the checkpoint will be restored to a consistent\n> state.\n\nnitpick: I would remove the word \"record\" in this change.\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Mon, 18 Jul 2022 11:30:28 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"On Tue, Jul 19, 2022 at 12:00 AM Nathan Bossart\n<nathandbossart@gmail.com> wrote:\n>\n> Overall, these patches look reasonable.\n>\n> On Mon, Jul 18, 2022 at 04:24:12PM +0530, Bharath Rupireddy wrote:\n> > record. Because the entire content of data pages is saved in the\n> > - log on the first page modification after a checkpoint (assuming\n> > + WAL record on the first page modification after a checkpoint (assuming\n> > <xref linkend=\"guc-full-page-writes\"/> is not disabled), all pages\n> > changed since the checkpoint will be restored to a consistent\n> > state.\n>\n> nitpick: I would remove the word \"record\" in this change.\n\nDone. PSA v5 patch set.\n\nRegards,\nBharath Rupireddy.","msg_date":"Tue, 19 Jul 2022 14:43:59 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"On Tue, Jul 19, 2022 at 02:43:59PM +0530, Bharath Rupireddy wrote:\n> Done. PSA v5 patch set.\n\nLGTM. I've marked this as ready-for-committer.\n\n-- \nNathan Bossart\nAmazon Web Services: https://aws.amazon.com\n\n\n","msg_date":"Tue, 19 Jul 2022 09:58:28 -0700","msg_from":"Nathan Bossart <nathandbossart@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"At Tue, 19 Jul 2022 09:58:28 -0700, Nathan Bossart <nathandbossart@gmail.com> wrote in \n> On Tue, Jul 19, 2022 at 02:43:59PM +0530, Bharath Rupireddy wrote:\n> > Done. PSA v5 patch set.\n> \n> LGTM. I've marked this as ready-for-committer.\n\nI find the following sentense in config.sgml. \"Specifies the minimum\nsize of past log file segments kept in the pg_wal directory\"\n\npostgresql.conf.sample contains \"logfile segment\" in a few lines.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Wed, 20 Jul 2022 10:25:26 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"On Wed, Jul 20, 2022 at 6:55 AM Kyotaro Horiguchi\n<horikyota.ntt@gmail.com> wrote:\n>\n> At Tue, 19 Jul 2022 09:58:28 -0700, Nathan Bossart <nathandbossart@gmail.com> wrote in\n> > On Tue, Jul 19, 2022 at 02:43:59PM +0530, Bharath Rupireddy wrote:\n> > > Done. PSA v5 patch set.\n> >\n> > LGTM. I've marked this as ready-for-committer.\n>\n> I find the following sentense in config.sgml. \"Specifies the minimum\n> size of past log file segments kept in the pg_wal directory\"\n>\n> postgresql.conf.sample contains \"logfile segment\" in a few lines.\n\nDone. PSA v6 patch set.\n\nRegards,\nBharath Rupireddy.","msg_date":"Wed, 20 Jul 2022 10:02:22 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"At Wed, 20 Jul 2022 10:02:22 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in \n> Done. PSA v6 patch set.\n\nThanks!\n\n- Specifies the minimum size of past log file segments kept in the\n+ Specifies the minimum size of past WAL files kept in the\n\n- log file by reducing the value of this parameter. On a system with\n+ file by reducing the value of this parameter. On a system with\n\nLooks fine. And postgresq.conf.sample has the following lines:\n\n#archive_library = ''\t\t# library to use to archive a logfile segment\n\n#archive_command = ''\t\t# command to use to archive a logfile segment\n\n#archive_timeout = 0\t\t# force a logfile segment switch after this\n\n#restore_command = ''\t\t# command to use to restore an archived logfile segment\n\nAren't they need the same fix?\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center\n\n\n","msg_date":"Wed, 20 Jul 2022 16:25:06 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"On Wed, Jul 20, 2022 at 12:55 PM Kyotaro Horiguchi\n<horikyota.ntt@gmail.com> wrote:\n>\n> At Wed, 20 Jul 2022 10:02:22 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in\n> > Done. PSA v6 patch set.\n>\n> Thanks!\n>\n> - Specifies the minimum size of past log file segments kept in the\n> + Specifies the minimum size of past WAL files kept in the\n>\n> - log file by reducing the value of this parameter. On a system with\n> + file by reducing the value of this parameter. On a system with\n>\n> Looks fine. And postgresq.conf.sample has the following lines:\n>\n> #archive_library = '' # library to use to archive a logfile segment\n>\n> #archive_command = '' # command to use to archive a logfile segment\n>\n> #archive_timeout = 0 # force a logfile segment switch after this\n>\n> #restore_command = '' # command to use to restore an archived logfile segment\n>\n> Aren't they need the same fix?\n\nIndeed. Thanks. Now, they are in sync with their peers in .conf.sample\nfile as well as description in guc.c.\n\nPSA v7 patch set.\n\nRegards,\nBharath Rupireddy.","msg_date":"Wed, 20 Jul 2022 17:25:33 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"At Wed, 20 Jul 2022 17:25:33 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in \n> On Wed, Jul 20, 2022 at 12:55 PM Kyotaro Horiguchi\n> <horikyota.ntt@gmail.com> wrote:\n> PSA v7 patch set.\n\nThanks. Looks perfect, but (sorry..) in the final checking, I found\n\"log archive\" in the doc. If you agree to it please merge the\nattached (or refined one) and I'd call it a day.\n\nregards.\n\n-- \nKyotaro Horiguchi\nNTT Open Source Software Center","msg_date":"Thu, 21 Jul 2022 13:20:12 +0900 (JST)","msg_from":"Kyotaro Horiguchi <horikyota.ntt@gmail.com>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"On Thu, Jul 21, 2022 at 9:50 AM Kyotaro Horiguchi\n<horikyota.ntt@gmail.com> wrote:\n>\n> At Wed, 20 Jul 2022 17:25:33 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in\n> > On Wed, Jul 20, 2022 at 12:55 PM Kyotaro Horiguchi\n> > <horikyota.ntt@gmail.com> wrote:\n> > PSA v7 patch set.\n>\n> Thanks. Looks perfect, but (sorry..) in the final checking, I found\n> \"log archive\" in the doc. If you agree to it please merge the\n> attached (or refined one) and I'd call it a day.\n\nMerged. PSA v8 patch set.\n\nRegards,\nBharath Rupireddy.","msg_date":"Sat, 23 Jul 2022 15:28:19 +0530","msg_from":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>","msg_from_op":true,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"},{"msg_contents":"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> writes:\n> Merged. PSA v8 patch set.\n\nPushed, thanks.\n\n\t\t\tregards, tom lane\n\n\n","msg_date":"Wed, 14 Sep 2022 18:41:56 -0400","msg_from":"Tom Lane <tgl@sss.pgh.pa.us>","msg_from_op":false,"msg_subject":"Re: Use \"WAL segment\" instead of \"log segment\" consistently in\n user-facing messages"}],"string":"[\n {\n \"msg_contents\": \"Hi,\\n\\nIt looks like we use \\\"log segment\\\" in various user-facing messages.\\nThe term \\\"log\\\" can mean server logs as well. The \\\"WAL segment\\\" suits\\nwell here and it is consistently used across the other user-facing\\nmessages [1].\\n\\nHere's a small patch attempting to consistently use the \\\"WAL segment\\\".\\n\\nThoughts?\\n\\n[1]\\npg_log_error(\\\"could not fetch WAL segment size: got %d rows and %d\\nfields, expected %d rows and %d or more fields\\\",\\npg_log_error(\\\"WAL segment size could not be parsed\\\");\\npg_log_error(ngettext(\\\"WAL segment size must be a power of two between\\n1 MB and 1 GB, but the remote server reported a value of %d byte\\\",\\nprintf(_(\\\"WARNING: invalid WAL segment size\\\\n\\\"));\\nprintf(_(\\\"Bytes per WAL segment: %u\\\\n\\\"),\\nfatal_error(ngettext(\\\"WAL segment size must be a power of two between\\n1 MB and 1 GB, but the WAL file \\\\\\\"%s\\\\\\\" header specifies %d byte\\\",\\nerrmsg(\\\"requested WAL segment %s has already been removed\\\",\\nelog(DEBUG2, \\\"removed temporary WAL segment \\\\\\\"%s\\\\\\\"\\\", path);\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Mon, 28 Feb 2022 21:03:07 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"At Mon, 28 Feb 2022 21:03:07 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in \\n> Hi,\\n> \\n> It looks like we use \\\"log segment\\\" in various user-facing messages.\\n> The term \\\"log\\\" can mean server logs as well. The \\\"WAL segment\\\" suits\\n> well here and it is consistently used across the other user-facing\\n> messages [1].\\n> \\n> Here's a small patch attempting to consistently use the \\\"WAL segment\\\".\\n> \\n> Thoughts?\\n\\nI tend to agree to this. I also see \\\"log record(s)\\\" (without prefixed\\nby \\\"write-ahead\\\") in many places especially in the documentation. I'm\\nnot sure how we should treat \\\"WAL log\\\", though.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Tue, 01 Mar 2022 10:20:53 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"On Tue, Mar 1, 2022 at 6:50 AM Kyotaro Horiguchi\\n<horikyota.ntt@gmail.com> wrote:\\n>\\n> At Mon, 28 Feb 2022 21:03:07 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in\\n> > Hi,\\n> >\\n> > It looks like we use \\\"log segment\\\" in various user-facing messages.\\n> > The term \\\"log\\\" can mean server logs as well. The \\\"WAL segment\\\" suits\\n> > well here and it is consistently used across the other user-facing\\n> > messages [1].\\n> >\\n> > Here's a small patch attempting to consistently use the \\\"WAL segment\\\".\\n> >\\n> > Thoughts?\\n>\\n> I tend to agree to this.\\n\\nThanks for taking a look at it. Here's the CF entry -\\nhttps://commitfest.postgresql.org/38/3584/\\n\\n> I also see \\\"log record(s)\\\" (without prefixed\\n> by \\\"write-ahead\\\") in many places especially in the documentation. I'm\\n> not sure how we should treat \\\"WAL log\\\", though.\\n\\nYes, but the docs have a glossary term for 'Log record\\\" [1]. FWIW\\nattaching docs change as v2-0002 patch. I found another place where\\n\\\"log records\\\" is being used in pg_waldump.c, I changed that and\\nattached v2-0001 patch.\\n\\nPlease review the v2 patch set.\\n\\n[1]\\n <glossentry id=\\\"glossary-log-record\\\">\\n <glossterm>Log record</glossterm>\\n <glossdef>\\n <para>\\n Archaic term for a <glossterm linkend=\\\"glossary-wal-record\\\">WAL\\nrecord</glossterm>.\\n </para>\\n </glossdef>\\n </glossentry>\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Wed, 2 Mar 2022 11:41:49 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"On Wed, Mar 2, 2022 at 11:41 AM Bharath Rupireddy\\n<bharath.rupireddyforpostgres@gmail.com> wrote:\\n>\\n> Please review the v2 patch set.\\n\\nHad to rebase. Attaching v3 patch set.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sat, 26 Mar 2022 11:26:16 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \" At all times, <productname>PostgreSQL</productname> maintains a\\n <firstterm>write ahead log</firstterm> (WAL) in the <filename>pg_wal/</filename>\\n- subdirectory of the cluster's data directory. The log records\\n- every change made to the database's data files. This log exists\\n+ subdirectory of the cluster's data directory. The WAL records\\n+ capture every change made to the database's data files. This log exists\\n\\nI don't think this change really adds anything. The preceding sentence\\nmakes it clear that we are discussing the write-ahead log, and IMO the\\nchange in phrasing (\\\"the log records every change\\\" is changed to \\\"the\\nrecords capture every change\\\") subtly changes the meaning of the sentence.\\n\\nThe rest looks good to me. \\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Thu, 31 Mar 2022 08:45:56 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"At Thu, 31 Mar 2022 08:45:56 -0700, Nathan Bossart <nathandbossart@gmail.com> wrote in \\n> At all times, <productname>PostgreSQL</productname> maintains a\\n> <firstterm>write ahead log</firstterm> (WAL) in the <filename>pg_wal/</filename>\\n> - subdirectory of the cluster's data directory. The log records\\n> - every change made to the database's data files. This log exists\\n> + subdirectory of the cluster's data directory. The WAL records\\n> + capture every change made to the database's data files. This log exists\\n> \\n> I don't think this change really adds anything. The preceding sentence\\n> makes it clear that we are discussing the write-ahead log, and IMO the\\n> change in phrasing (\\\"the log records every change\\\" is changed to \\\"the\\n> records capture every change\\\") subtly changes the meaning of the sentence.\\n> \\n> The rest looks good to me. \\n\\n+1. It is not a composite noun \\\"log records\\\".\\n\\nThe original sentence is \\\"S(The log) V(records) O(every change that is\\nmade to .. files)\\\". The proposed change looks like changing it to\\n\\\"S(The log records) V(capture) O(every ..files)\\\". In that sense, the\\noriginal one seem rather correct to me, since \\\"capture\\\" seems to have\\nthe implication of \\\"write after log..\\\", to me.\\n\\n\\nI looked though the document and found other use of \\\"log\\nrecord|segment\\\". What do you think about the attached?\\n\\nThere're some uncertain point in the change.\\n\\n you should at least save the contents of the cluster's <filename>pg_wal</filename>\\n- subdirectory, as it might contain logs which\\n+ subdirectory, as it might contain WAL files which\\n were not archived before the system went down.\\n\\nThe \\\"logs\\\" means acutally \\\"WAL segment (files)\\\" but the concept of\\n\\\"segment\\\" is out of focus in the context. So just \\\"file\\\" is used\\nthere. The same change is applied on dezon of places.\\n\\n\\n- disk-space requirements for the <acronym>WAL</acronym> logs are met,\\n+ disk-space requirements for the <acronym>WAL</acronym> are met,\\n\\nThis might be better be \\\"WAL files\\\" instead of just \\\"WAL\\\".\\n\\n\\n- <acronym>WAL</acronym> logs are stored in the directory\\n+ <acronym>WAL</acronym> is stored in the directory\\n <filename>pg_wal</filename> under the data directory, as a set of\\n\\nI'm not sure which is better, use \\\"WAL\\\" as a collective noun, or \\\"WAL\\nfiles\\\" as the cocrete objects.\\n\\n\\n- The aim of <acronym>WAL</acronym> is to ensure that the log is\\n+ The aim of <acronym>WAL</acronym> is to ensure that the WAL record is\\n written before database records are altered, but this can be subverted by\\n\\nThis is not a mechanical change. But I think this is correct.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\",\n \"msg_date\": \"Fri, 01 Apr 2022 10:31:10 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"On Fri, Apr 01, 2022 at 10:31:10AM +0900, Kyotaro Horiguchi wrote:\\n> you should at least save the contents of the cluster's <filename>pg_wal</filename>\\n> - subdirectory, as it might contain logs which\\n> + subdirectory, as it might contain WAL files which\\n> were not archived before the system went down.\\n> \\n> The \\\"logs\\\" means acutally \\\"WAL segment (files)\\\" but the concept of\\n> \\\"segment\\\" is out of focus in the context. So just \\\"file\\\" is used\\n> there. The same change is applied on dezon of places.\\n\\nThis change seems reasonable to me.\\n\\n> - disk-space requirements for the <acronym>WAL</acronym> logs are met,\\n> + disk-space requirements for the <acronym>WAL</acronym> are met,\\n> \\n> This might be better be \\\"WAL files\\\" instead of just \\\"WAL\\\".\\n\\n+1 for \\\"WAL files\\\"\\n\\n> - <acronym>WAL</acronym> logs are stored in the directory\\n> + <acronym>WAL</acronym> is stored in the directory\\n> <filename>pg_wal</filename> under the data directory, as a set of\\n> \\n> I'm not sure which is better, use \\\"WAL\\\" as a collective noun, or \\\"WAL\\n> files\\\" as the cocrete objects.\\n\\nMy vote is for \\\"WAL files\\\" because it was previously \\\"WAL logs.\\\"\\n\\n> - The aim of <acronym>WAL</acronym> is to ensure that the log is\\n> + The aim of <acronym>WAL</acronym> is to ensure that the WAL record is\\n> written before database records are altered, but this can be subverted by\\n> \\n> This is not a mechanical change. But I think this is correct.\\n\\nIMO the original wording is fine. I think it is sufficiently clear that\\n\\\"log\\\" refers to \\\"write-ahead log,\\\" and this sentence seems intended to\\nconvey the basic rule of \\\"log before data.\\\" However, the rest of the\\nsentence is a little weird. It's basically saying \\\"the aim of the log is\\nto ensure that the log is written...\\\" Isn't the aim of the log to record\\nthe database activity? Perhaps we should rewrite it to something like the\\nfollowing:\\n\\n\\tA basic rule of WAL is that the log must be written before the database\\n\\tfiles are altered, but this can be...\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Fri, 1 Apr 2022 07:24:48 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"It's been a few weeks, so I'm marking the commitfest entry as\\nwaiting-on-author.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 25 Apr 2022 12:54:43 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"On Tue, Apr 26, 2022 at 1:24 AM Nathan Bossart <nathandbossart@gmail.com> wrote:\\n>\\n> It's been a few weeks, so I'm marking the commitfest entry as\\n> waiting-on-author.\\n\\nThanks. I'm attaching the updated v4 patches (also subsumed Kyotaro\\nSan's patch at [1]). Please review it further.\\n\\n[1] https://www.postgresql.org/message-id/20220401.103110.1103213854487561781.horikyota.ntt%40gmail.com\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Mon, 18 Jul 2022 16:24:12 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"Overall, these patches look reasonable.\\n\\nOn Mon, Jul 18, 2022 at 04:24:12PM +0530, Bharath Rupireddy wrote:\\n> record. Because the entire content of data pages is saved in the\\n> - log on the first page modification after a checkpoint (assuming\\n> + WAL record on the first page modification after a checkpoint (assuming\\n> <xref linkend=\\\"guc-full-page-writes\\\"/> is not disabled), all pages\\n> changed since the checkpoint will be restored to a consistent\\n> state.\\n\\nnitpick: I would remove the word \\\"record\\\" in this change.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Mon, 18 Jul 2022 11:30:28 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 19, 2022 at 12:00 AM Nathan Bossart\\n<nathandbossart@gmail.com> wrote:\\n>\\n> Overall, these patches look reasonable.\\n>\\n> On Mon, Jul 18, 2022 at 04:24:12PM +0530, Bharath Rupireddy wrote:\\n> > record. Because the entire content of data pages is saved in the\\n> > - log on the first page modification after a checkpoint (assuming\\n> > + WAL record on the first page modification after a checkpoint (assuming\\n> > <xref linkend=\\\"guc-full-page-writes\\\"/> is not disabled), all pages\\n> > changed since the checkpoint will be restored to a consistent\\n> > state.\\n>\\n> nitpick: I would remove the word \\\"record\\\" in this change.\\n\\nDone. PSA v5 patch set.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Tue, 19 Jul 2022 14:43:59 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"On Tue, Jul 19, 2022 at 02:43:59PM +0530, Bharath Rupireddy wrote:\\n> Done. PSA v5 patch set.\\n\\nLGTM. I've marked this as ready-for-committer.\\n\\n-- \\nNathan Bossart\\nAmazon Web Services: https://aws.amazon.com\\n\\n\\n\",\n \"msg_date\": \"Tue, 19 Jul 2022 09:58:28 -0700\",\n \"msg_from\": \"Nathan Bossart <nathandbossart@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"At Tue, 19 Jul 2022 09:58:28 -0700, Nathan Bossart <nathandbossart@gmail.com> wrote in \\n> On Tue, Jul 19, 2022 at 02:43:59PM +0530, Bharath Rupireddy wrote:\\n> > Done. PSA v5 patch set.\\n> \\n> LGTM. I've marked this as ready-for-committer.\\n\\nI find the following sentense in config.sgml. \\\"Specifies the minimum\\nsize of past log file segments kept in the pg_wal directory\\\"\\n\\npostgresql.conf.sample contains \\\"logfile segment\\\" in a few lines.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Wed, 20 Jul 2022 10:25:26 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"On Wed, Jul 20, 2022 at 6:55 AM Kyotaro Horiguchi\\n<horikyota.ntt@gmail.com> wrote:\\n>\\n> At Tue, 19 Jul 2022 09:58:28 -0700, Nathan Bossart <nathandbossart@gmail.com> wrote in\\n> > On Tue, Jul 19, 2022 at 02:43:59PM +0530, Bharath Rupireddy wrote:\\n> > > Done. PSA v5 patch set.\\n> >\\n> > LGTM. I've marked this as ready-for-committer.\\n>\\n> I find the following sentense in config.sgml. \\\"Specifies the minimum\\n> size of past log file segments kept in the pg_wal directory\\\"\\n>\\n> postgresql.conf.sample contains \\\"logfile segment\\\" in a few lines.\\n\\nDone. PSA v6 patch set.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Wed, 20 Jul 2022 10:02:22 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"At Wed, 20 Jul 2022 10:02:22 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in \\n> Done. PSA v6 patch set.\\n\\nThanks!\\n\\n- Specifies the minimum size of past log file segments kept in the\\n+ Specifies the minimum size of past WAL files kept in the\\n\\n- log file by reducing the value of this parameter. On a system with\\n+ file by reducing the value of this parameter. On a system with\\n\\nLooks fine. And postgresq.conf.sample has the following lines:\\n\\n#archive_library = ''\\t\\t# library to use to archive a logfile segment\\n\\n#archive_command = ''\\t\\t# command to use to archive a logfile segment\\n\\n#archive_timeout = 0\\t\\t# force a logfile segment switch after this\\n\\n#restore_command = ''\\t\\t# command to use to restore an archived logfile segment\\n\\nAren't they need the same fix?\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\\n\\n\\n\",\n \"msg_date\": \"Wed, 20 Jul 2022 16:25:06 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"On Wed, Jul 20, 2022 at 12:55 PM Kyotaro Horiguchi\\n<horikyota.ntt@gmail.com> wrote:\\n>\\n> At Wed, 20 Jul 2022 10:02:22 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in\\n> > Done. PSA v6 patch set.\\n>\\n> Thanks!\\n>\\n> - Specifies the minimum size of past log file segments kept in the\\n> + Specifies the minimum size of past WAL files kept in the\\n>\\n> - log file by reducing the value of this parameter. On a system with\\n> + file by reducing the value of this parameter. On a system with\\n>\\n> Looks fine. And postgresq.conf.sample has the following lines:\\n>\\n> #archive_library = '' # library to use to archive a logfile segment\\n>\\n> #archive_command = '' # command to use to archive a logfile segment\\n>\\n> #archive_timeout = 0 # force a logfile segment switch after this\\n>\\n> #restore_command = '' # command to use to restore an archived logfile segment\\n>\\n> Aren't they need the same fix?\\n\\nIndeed. Thanks. Now, they are in sync with their peers in .conf.sample\\nfile as well as description in guc.c.\\n\\nPSA v7 patch set.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Wed, 20 Jul 2022 17:25:33 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"At Wed, 20 Jul 2022 17:25:33 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in \\n> On Wed, Jul 20, 2022 at 12:55 PM Kyotaro Horiguchi\\n> <horikyota.ntt@gmail.com> wrote:\\n> PSA v7 patch set.\\n\\nThanks. Looks perfect, but (sorry..) in the final checking, I found\\n\\\"log archive\\\" in the doc. If you agree to it please merge the\\nattached (or refined one) and I'd call it a day.\\n\\nregards.\\n\\n-- \\nKyotaro Horiguchi\\nNTT Open Source Software Center\",\n \"msg_date\": \"Thu, 21 Jul 2022 13:20:12 +0900 (JST)\",\n \"msg_from\": \"Kyotaro Horiguchi <horikyota.ntt@gmail.com>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"On Thu, Jul 21, 2022 at 9:50 AM Kyotaro Horiguchi\\n<horikyota.ntt@gmail.com> wrote:\\n>\\n> At Wed, 20 Jul 2022 17:25:33 +0530, Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> wrote in\\n> > On Wed, Jul 20, 2022 at 12:55 PM Kyotaro Horiguchi\\n> > <horikyota.ntt@gmail.com> wrote:\\n> > PSA v7 patch set.\\n>\\n> Thanks. Looks perfect, but (sorry..) in the final checking, I found\\n> \\\"log archive\\\" in the doc. If you agree to it please merge the\\n> attached (or refined one) and I'd call it a day.\\n\\nMerged. PSA v8 patch set.\\n\\nRegards,\\nBharath Rupireddy.\",\n \"msg_date\": \"Sat, 23 Jul 2022 15:28:19 +0530\",\n \"msg_from\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n },\n {\n \"msg_contents\": \"Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com> writes:\\n> Merged. PSA v8 patch set.\\n\\nPushed, thanks.\\n\\n\\t\\t\\tregards, tom lane\\n\\n\\n\",\n \"msg_date\": \"Wed, 14 Sep 2022 18:41:56 -0400\",\n \"msg_from\": \"Tom Lane <tgl@sss.pgh.pa.us>\",\n \"msg_from_op\": false,\n \"msg_subject\": \"Re: Use \\\"WAL segment\\\" instead of \\\"log segment\\\" consistently in\\n user-facing messages\"\n }\n]","isConversational":false}}},{"rowIdx":53598,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"28 days has February ... So the March commitfest begins tomorrow. Meet\nyour friendly neighbourhood Commitfest Manager for March. Greetings!\n\nIf you have a patch you're hoping to get feedback on or you're\nexpecting to get committed this month make sure it's in the commitfest\nat https://commitfest.postgresql.org/37/\n\nCurrently there are 199 patches marked Needs Review and 25 Ready for Committer.\n\nThis is the last commitfest of the cycle so we'll be concentrating on\npatches that are likely to be committed this commitfest. I'll move any\npatches that are unlikely to be committed to target Postgres 16 and\nthe July commitfest shortly.\n\nIf you think your patch is really viable for Postgres 15 despite\nneeding feedback it can be helpful to post describing exactly what\nyou're blocked on and what remains to be done. It also sometimes helps\nget attention if you make clear that you have the time available this\nmonth to work on it in response to feedback.\n\nThe \"round robin reviewers\" system is no longer. Generally people can\njust grab any unassigned patch from the commitfest and review it but\nif for some reason you're reluctant feel free to email me -- please\nput \"commitfest\" in the subject line.\n\nIf you've submitted a patch and are feeling blocked waiting on\nfeedback please update the thread on the list. If you're still not\ngetting responses please let me know -- again, please put \"commitfest\"\nin the subject.\n\n-- \ngreg\n\n\n","msg_date":"Mon, 28 Feb 2022 11:45:17 -0500","msg_from":"Greg Stark <stark@mit.edu>","msg_from_op":true,"msg_subject":"CommitFest begins tomorrow... Get your patches in"}],"string":"[\n {\n \"msg_contents\": \"28 days has February ... So the March commitfest begins tomorrow. Meet\\nyour friendly neighbourhood Commitfest Manager for March. Greetings!\\n\\nIf you have a patch you're hoping to get feedback on or you're\\nexpecting to get committed this month make sure it's in the commitfest\\nat https://commitfest.postgresql.org/37/\\n\\nCurrently there are 199 patches marked Needs Review and 25 Ready for Committer.\\n\\nThis is the last commitfest of the cycle so we'll be concentrating on\\npatches that are likely to be committed this commitfest. I'll move any\\npatches that are unlikely to be committed to target Postgres 16 and\\nthe July commitfest shortly.\\n\\nIf you think your patch is really viable for Postgres 15 despite\\nneeding feedback it can be helpful to post describing exactly what\\nyou're blocked on and what remains to be done. It also sometimes helps\\nget attention if you make clear that you have the time available this\\nmonth to work on it in response to feedback.\\n\\nThe \\\"round robin reviewers\\\" system is no longer. Generally people can\\njust grab any unassigned patch from the commitfest and review it but\\nif for some reason you're reluctant feel free to email me -- please\\nput \\\"commitfest\\\" in the subject line.\\n\\nIf you've submitted a patch and are feeling blocked waiting on\\nfeedback please update the thread on the list. If you're still not\\ngetting responses please let me know -- again, please put \\\"commitfest\\\"\\nin the subject.\\n\\n-- \\ngreg\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 11:45:17 -0500\",\n \"msg_from\": \"Greg Stark <stark@mit.edu>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"CommitFest begins tomorrow... Get your patches in\"\n }\n]","isConversational":false}}},{"rowIdx":53599,"cells":{"threads":{"kind":"list like","value":[{"msg_contents":"Hi everyone,\n\nI have been doing some tests on a little server (at least compared to\nsome others around). It have 128 cores (64 physical), 128GB of RAM and\nagainst my will a 4-disk (10k RPM) RAID5.\n\nI have been using hammerdb testing from 9.6 to 14, currently on windows\n10. Obviously, I have been recording changes on performance.\n\nHammerdb results shows 2 numbers, NOPM and TPM the second one is\ncalculated using statistics.\n\nSadly, on 14 I saw a lowered number of TPM while NOPM kept on the\naverage (at least is the average since 11).\n\nThe reason that on 14 the TPM number dropped is because it's based \non the statistics[1] which of course are stalled[2].\n\nI consider this a regression because no other postgres version had \nthis problem on the same machine and the same OS. Anything I can do to\ntrack what caused this regression?\n\n[1] Query from hammerdb to get TPM number:\nselect sum(xact_commit + xact_rollback) \n from pg_stat_database\n\n[2] Message from the log, saying what is obvious\nLOG: 00000: using stale statistics instead of current ones because\nstats collector is not responding\n\n-- \nJaime Casanova\nDirector de Servicios Profesionales\nSystemGuards - Consultores de PostgreSQL\n\n\n","msg_date":"Mon, 28 Feb 2022 17:52:45 -0500","msg_from":"Jaime Casanova <jcasanov@systemguards.com.ec>","msg_from_op":true,"msg_subject":"stale statistics on postgres 14"}],"string":"[\n {\n \"msg_contents\": \"Hi everyone,\\n\\nI have been doing some tests on a little server (at least compared to\\nsome others around). It have 128 cores (64 physical), 128GB of RAM and\\nagainst my will a 4-disk (10k RPM) RAID5.\\n\\nI have been using hammerdb testing from 9.6 to 14, currently on windows\\n10. Obviously, I have been recording changes on performance.\\n\\nHammerdb results shows 2 numbers, NOPM and TPM the second one is\\ncalculated using statistics.\\n\\nSadly, on 14 I saw a lowered number of TPM while NOPM kept on the\\naverage (at least is the average since 11).\\n\\nThe reason that on 14 the TPM number dropped is because it's based \\non the statistics[1] which of course are stalled[2].\\n\\nI consider this a regression because no other postgres version had \\nthis problem on the same machine and the same OS. Anything I can do to\\ntrack what caused this regression?\\n\\n[1] Query from hammerdb to get TPM number:\\nselect sum(xact_commit + xact_rollback) \\n from pg_stat_database\\n\\n[2] Message from the log, saying what is obvious\\nLOG: 00000: using stale statistics instead of current ones because\\nstats collector is not responding\\n\\n-- \\nJaime Casanova\\nDirector de Servicios Profesionales\\nSystemGuards - Consultores de PostgreSQL\\n\\n\\n\",\n \"msg_date\": \"Mon, 28 Feb 2022 17:52:45 -0500\",\n \"msg_from\": \"Jaime Casanova <jcasanov@systemguards.com.ec>\",\n \"msg_from_op\": true,\n \"msg_subject\": \"stale statistics on postgres 14\"\n }\n]","isConversational":false}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":535,"numItemsPerPage":100,"numTotalItems":58371,"offset":53500,"length":100}},"jwt":"eyJhbGciOiJFZERTQSJ9.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc4MDYyMjQ4Mywic3ViIjoiL2RhdGFzZXRzL3dhbnNoZW5sL3Bnc3FsLWhhY2tlcnMiLCJleHAiOjE3ODA2MjYwODMsImlzcyI6Imh0dHBzOi8vaHVnZ2luZ2ZhY2UuY28ifQ.8qlEZAN5UpuX2AkZ9q_uxJe4DLea7YafEPWOglS53BVGVVG8m5SKg7lFdisx-9VXHDu0FODcK6r8qzUZ34zoCw","displayUrls":true,"splitSizeSummaries":[{"config":"default","split":"train","numRows":58371,"numBytesParquet":368254123}]},"discussionsStats":{"closed":0,"open":0,"total":0},"fullWidth":true,"hasGatedAccess":true,"hasFullAccess":true,"isEmbedded":false,"savedQueries":{"community":[],"user":[]}}">
threads listlengths 1 2.99k |
|---|
[
{
"msg_contents": "My recent commit 44fa8488 made vacuumlazy.c always scan the last page\nof every heap relation, even when it's all-visible, regardless of any\nother factor. The general idea is to always examine the last page in\nthe relation when that might prevent us from unsuccessfully attempting\nto trunca... |
[
{
"msg_contents": "Hi,\n\nA recent cfbot failure reminded me of a topic I wanted to raise: The use of\nBAIL in tap tests.\n\nWhen a tap test exits with BAIL it doesn't just consider that test failed, but\nit also stops running further tap tests in the same group:\nhttps://metacpan.org/pod/Test::More\n\"Indicate... |
[
{
"msg_contents": "Hi,\n\nWe see a successful compile and then a failure to read the file while\nlinking. We see that the animal got into that state recently and then\nfixed itself, and now it's back in that state. I don't know if it's\nsignificant, but it happened to fix itself when a configure change\ncame ... |
[
{
"msg_contents": "For a very long time, the AIX 7.1 buildfarm members (sungazer and tern)\nhave complained about\n\nip.c:236:17: warning: conversion from 'long unsigned int' to 'uchar_t' {aka 'unsigned char'} changes value from '1025' to '1' [-Woverflow]\n\nI'd ignored this as not being very interesting, but I... |
[
{
"msg_contents": "During development, I have been using the attached patch to simulate \nlibc collation versions on macOS. It just uses the internal major OS \nversion number. I don't know to what the extend the libc locales on \nmacOS are maintained or updated at all, so I don't know what practical \neffect... |
[
{
"msg_contents": "Add test case for an archive recovery corner case.\n\nWhile I was working on a patch to refactor things around xlog.c, I mixed\nup EndOfLogTLI and replayTLI at the end of recovery. As a result, if you\nrecovered to a point with a lower-numbered timeline in a WAL segment\nthat has a higher TLI... |
[
{
"msg_contents": "This is a rebase of the patch from [0]. It removes the internal padding \nimplementation in pgcrypto and lets OpenSSL do it. The internal \nimplementation was once applicable to the non-OpenSSL code paths, but \nthose have since been removed.\n\n\n[0]: \nhttps://www.postgresql.org/message-i... |
[
{
"msg_contents": "Hi.\n\nNow select ... for update ... [skip locked|nowait] options are not \npushed down to remote servers. I see the only reason is that we can \nspeak to pre-9.5 server, which doesn't understand skip locked option. \nAre there any other possible issues? Should we add foreign table option \nt... |
[
{
"msg_contents": "So I've been dealing a lot with building and maintaining dashboards\nfor (fleets of) Postgres servers. And it's a pain. I have a few\nstrongly held ideas about where the pain points are and what the right\nways to tackle them are. Some of which are going to be controversial I\nthink...\n\nThe... |
[
{
"msg_contents": "I tried to replicate the problem described in [1] about logical\nreplication from a v14 source DB to a v11 target. It fails as\ndescribed there; I've not yet tracked down why, but it looks like\nthe v11 apply worker fails and closes the connection after sending\nCREATE_REPLICATION_SLOT. The... |
[
{
"msg_contents": "Move scanint8() to numutils.c\n\nMove scanint8() to numutils.c and rename to pg_strtoint64(). We\nalready have a \"16\" and \"32\" version of that, and the code inside the\nfunctions was aligned, so this move makes all three versions\nconsistent. The API is also changed to no longer provide... |
[
{
"msg_contents": "I'm writing an index access method with its own unique file format. It\ninvolves storing large blobs that break across pages.\n\nThe file format itself doesn't need or use page headers. There's no need\nfor a checksum or to manage free space within the page.\n\nCan I treat pages as just a fla... |
[
{
"msg_contents": "In [1] there's a complaint that if you try to logically replicate\na partitioned table from v13-or-later to v12-or-earlier, you get\n\"table XXX not found on publisher\", which is pretty confusing\nbecause the publisher certainly does have such a table. That\nhappens because fetch_remote_tab... |
[
{
"msg_contents": "Hi,\n\nI've reported this issue, but without success in fixing it.\nNow I have installed 14.2 and used in development environment, and\nthe log still reports:\n2022-02-13 18:33:20.502 -03 [7976] LOG: could not rename temporary\nstatistics file \"pg_stat_tmp/global.tmp\" to \"pg_stat_tmp/glob... |
[
{
"msg_contents": "Hi,\n\nPer Coverity.\n\nLike the function pg_encoding_max_length_sql\n(src/backend/utils/mb/mbutils.c)\nOnly assertion is insufficient to avoid accessing array out-of-bounds.\n\nThis bug is live according Coverity at function: pg_verify_mbstr_len\n(src/backend/utils/mb/mbutils.c)\nCID 1469870... |
[
{
"msg_contents": "Hi,\n\nOver on the rather-long \"refactoring basebackup.c\" thread, there is a\nproposal, which I endorse, to add base-backup compression via zstd. To\ndo that, we'd need to patch configure.ac to create a new --with-zstd\nflag and appropriate supporting infrastructure. My colleague Jeevan\nLa... |
[
{
"msg_contents": "I've been fooling around with the duplicated-error-text issue\ndiscussed in [1], and I think I have a solution that is fairly\nbulletproof ... except that it cannot cope with this little gem\nat the bottom of PQgetResult:\n\n if (!res->events[i].proc(PGEVT_RESULTCREATE, &evt,\n ... |
[
{
"msg_contents": "While looking at recent failures in the new 028_pitr_timelines.pl \nrecovery test, I noticed that there have been a few failures in the \nbuildfarm in the recoveryCheck phase even before that, in the \n019_replslot_limit.pl test.\n\nFor example:\n\nhttps://buildfarm.postgresql.org/cgi-bin/sho... |
[
{
"msg_contents": "I've noticed that my WAL pre-allocation patches [0] routinely fail with\n\"tablespace is not empty\" errors on Linux [1]:\n\n\t DROP TABLESPACE regress_tblspace_renamed;\n\t+ERROR: tablespace \"regress_tblspace_renamed\" is not empty\n\nThis appears to have been discussed elsewhere (and fixe... |
[
{
"msg_contents": "Hi,\n\nWhile reading the code, I realized that the second sentence of the\nfollowing comment in worker.c is not correct:\n\n /*\n * Exit if the subscription was disabled. This normally should not happen\n * as the worker gets killed during ALTER SUBSCRIPTION ... DISABLE.\n */\n ... |
[
{
"msg_contents": "Hi,\n\n[1] reminded me of a topic that I wanted to bring up at some point:\n\nTo me the division of labor between initdb and bootstrap doesn't make much\nsense anymore:\n\n\ninitdb reads postgres.bki, replaces a few tokens, starts postgres in bootstrap\nmode, and then painstakenly feeds boots... |
[
{
"msg_contents": "Hi,\n\nI noticed the small typo in pg_logical_slot_get_changes_guts().\n\n======================================\n--- a/src/backend/replication/logical/logicalfuncs.c\n+++ b/src/backend/replication/logical/logicalfuncs.c\n@@ -295,7 +295,7 @@ pg_logical_slot_get_changes_guts(FunctionCallInfo\n... |
[
{
"msg_contents": "Hi,\n\nCurrently, numeric_pg_lsn is the only one that accepts the Numeric\nvalue and converts to uint64 and that is the reason all the type\nconversion code is embedded into it.\n\nI think it would be helpful if that numeric-to-uint64 conversion code\nis extracted as a separate function so th... |
[
{
"msg_contents": "While reviewing Thomas Munro's patchset to consider expanding the uses\nof specialized qsort [1], I wondered about some aspects of the current\nqsort implementation. For background, ours is based on Bentley &\nMcIlroy \"Engineering a Sort Function\" [2], which is a classic paper\nworth studyi... |
[
{
"msg_contents": "Hi,\n\nThe attached patch replaces the hard-coded type alignment value with\nthe defined macro for the same.\n\n--\nRegards,\nAmul Sul\nEDB: http://www.enterprisedb.com",
"msg_date": "Wed, 16 Feb 2022 17:47:11 +0530",
"msg_from": "Amul Sul <sulamul@gmail.com>",
"msg_from_op": true... |
[
{
"msg_contents": "$ make check-world >/dev/null\n2022-02-16 11:57:47.548 EST [3339702] LOG: database system was not properly shut down; automatic recovery in progress\n2022-02-16 11:57:47.612 EST [3339713] LOG: database system was not properly shut down; automatic recovery in progress\n2022-02-16 11:57:59.37... |
[
{
"msg_contents": "Now that lz4 also supports backup compression, decompression, and more\nfuture enhancements that can be done here, I think it is better to make\nthe --with-lz4 description more generic than adding specific details in\nthere.\n\nAttached is the patch to remove the specifics related to WAL and ... |
[
{
"msg_contents": "Largely following a recipe from Andres, I have migrated buildfarm\nanimals fairywren and jacana to a setup that shouldn't need (and in fact\nwon't be able to use) PostgreSQL::Test:Utils::perl2host(). AFAICT these\ntwo are the only buildfarm animals that run TAP tests under msys.\n\nSee discus... |
[
{
"msg_contents": "Hi,\n\nI noticed that effectively all indexes use the special region of a\npage to store some index-specific data on the page. In all cases I've\nnoticed, this is a constant-sized struct, located at what is\neffectively a fixed offset from the end of the page; indicated on the\npage by pd_spe... |
[
{
"msg_contents": "Seeing msys in TAP tests mentioned in a thread [1] tonight reminded me about\ntwo related (well, one of them) patches I had sitting around, so rather than\nforgetting again here are some small cleanups.\n\n0001 attempts to streamline how we detect Windows in the TAP tests (after that\nthere i... |
[
{
"msg_contents": "Hi, hackers!\n\nIf we trying to call pageinspect functions for pages which are filled \nwith nulls, we will get core dump. It happens with null pages for all \nindexes in pageinspect and for page_checksum. This problem was founded \nfirstly by Anastasia Lubennikova, and now I continue this ta... |
[
{
"msg_contents": "Hi,\n\nPer Coverity.\n\nMaybe this is a mistake, but,\nIs it necessary or not to free the memory allocated by nodeToString?\n\nIf yes, the patch attached fixes this.\n\nregards,\n\nRanier Vilela",
"msg_date": "Thu, 17 Feb 2022 09:35:06 -0300",
"msg_from": "Ranier Vilela <ranier.vf@gma... |
[
{
"msg_contents": "Hi,\n\nI've been looking at a report of a crash an in-place upgrade [1],\nrelated to a GiST index, and I tried to use the new GiST support added\nto pageinspect in 756ab29124 (so it's in 14). My knowledge of GiST is\nsomewhat limited, but after struggling with it for a while I wonder if\nit's... |
[
{
"msg_contents": "Hi all,\n\nI wanted to submit a patch to expose 2 new hooks (one for the\nauthentication check and another one for error reporting) in auth.c. These\nwill allow users to implement their own authentication methods for Postgres\nor add custom logic around authentication.\n\nA use case where thi... |
[
{
"msg_contents": "GCC 12, coming soon to a distro near you, complains like this:\n\npostgres.c: In function 'set_stack_base':\npostgres.c:3430:24: warning: storing the address of local variable 'stack_base' in 'stack_base_ptr' [-Wdangling-pointer=]\n 3430 | stack_base_ptr = &stack_base;\n | ... |
[
{
"msg_contents": "When formatting the output of an Interval, we call abs() on the hours\nfield of the Interval. Calling abs(INT_MIN) returns back INT_MIN\ncausing the output to contain two '-' characters. The attached patch\nfixes that issue by special casing INT_MIN hours.\n\nHere is an example of the issue:\... |
[
{
"msg_contents": "On Fri, Feb 05, 2021 at 03:55:20PM -0500, Tom Lane wrote:\n> We have, almost invariably, regretted it when we tried to use short\n> timeouts in test cases.\n\n> More generally, sometimes people want to do things like run a test\n> under valgrind. So it's not just \"underpowered machines\" th... |
[
{
"msg_contents": "Hi all,\n\n\nI find a potential memory leak in PostgresSQL 14.1, which is in the function describeOneTableDetails (./src/bin/psql/describe.c). The bug has been confirmed by an auditor of <pgsql-bugs>.\n\n\nSpecifically, at line 1603, a memory chunk is allocated with pg_strdup and assigned to ... |
[
{
"msg_contents": "I noticed that the JSON path lexer does not support the decimal literal \nsyntax forms\n\n.1\n1.\n\n(that is, there are no digits before or after the decimal point). This \nis allowed by the relevant ECMAScript standard \n(https://262.ecma-international.org/5.1/#sec-7.8.3) and of course SQL ... |
[
{
"msg_contents": "Hello,I added a patch to fix table_to_xmlschema, could you point me how to add a unit test to that?",
"msg_date": "Fri, 18 Feb 2022 17:47:14 -0000",
"msg_from": "Renan Soares Lopes <renans.l@icloud.com>",
"msg_from_op": true,
"msg_subject": "\n =?utf-8?B?W1BBVENIXSBBZGQgc3VwcG... |
[
{
"msg_contents": "Add support for building with ZSTD.\n\nThis commit doesn't actually add anything that uses ZSTD; that will be\ndone separately. It just puts the basic infrastructure into place.\n\nJeevan Ladhe, Robert Haas, and Michael Paquier. Reviewed by Justin\nPryzby and Andres Freund.\n\nDiscussion: htt... |
[
{
"msg_contents": "We've accumulated a few bit-twiddling hacks to get the compiler to\nemit a rotate instruction. Since we have a macro for that, let's use\nit, as in the attached. I thought the new call sites would look better\nwith a \"left\" version, so I added a new macro for that. That's not\nnecessary, ho... |
[
{
"msg_contents": "Hi,\nWhen i was looking at the code for ExecMergeJoin(), I noticed a few places\nwhere the join state name in the comment doesn't match the actual state.\n\nHere is a small patch with correction.\n\nCheers",
"msg_date": "Sat, 19 Feb 2022 05:20:14 -0800",
"msg_from": "Zhihong Yu <zyu@y... |
[
{
"msg_contents": "Some years ago we did a pass through the various worker processes to\nadd hibernation as a mechanism to reduce power consumption on an idle\nserver. Replication never got the memo, so power consumption on an\nidle server is not very effective on standby or logical subscribers.\nThe code and t... |
[
{
"msg_contents": "Forking: <E6CA9D45-96EA-4037-8BEB-344CCA75A38C@anarazel.de>\n\nOn Tue, Feb 15, 2022 at 10:42:09PM -0800, Andres Freund wrote:\n> >> I was thinking that we should make Utils.pm's INIT block responsible for\n> >> figuring out both the directory a test should run in and the log location,\n> >> i... |
[
{
"msg_contents": "Hi hackers,\nI've read in this blog (\nhttp://big-elephants.com/2015-10/writing-postgres-extensions-part-i/)\nand I wrote an extension about base36_encode with c code\nbut when I executed a query like this below I got a warning below.\n\n```\npostgres=# SELECT base36_encode(123);\nWARNING: p... |
[
{
"msg_contents": "Hi,\n\nLike how the commit\nhttps://github.com/postgres/postgres/commit/07daca53bfcad59618a9c6fad304e380cc9d2bc1\nThe are some paths that were missed:\n\n-At jsonfuncs.c\n The errors mgs do not report about the materialize mode.\n-At plperl.c\n The function plperl_return_next_internal does no... |
[
{
"msg_contents": "I need a way to separate the results of \\watch for each query execution.\n\nThere is only a blank line between the results of \\watch.\nHowever, there is also a blank line after the title, which complicates\nthe rules.\n\nMy suggestion is to insert a \"form feed(\\f)\" (preferably a newline)... |
[
{
"msg_contents": "Hi all,\n(Author and committer added in CC.)\n\nWhile reviewing the code of a bunch of SRF functions in the core code,\nI have noticed that the two functions mentioned in $subject are marked\nas proretset but both functions don't return a set of tuples, just one\nrecord for the object given i... |
[
{
"msg_contents": "Hi, \n\nThe comment of get_collation_name wrote: \n\n * NOTE: since collation name is not unique, be wary of code that uses this\n * for anything except preparing error messages.\n\nIn file explain.c, the function get_collation_name() is called, but the schema \nname won't be outputted.\n... |
[
{
"msg_contents": "Hi,\n\nLogical replication has been used to migration with minimal downtime. However,\nif you are dealing with a big database, the amount of required resources (disk\n-- due to WAL retention) increases as the backlog (WAL) increases. Unless you\nhave a generous amount of resources and can wai... |
[
{
"msg_contents": "check_query() in RewindTest.pm currently has this comment before handrolling\ntests for return code and stderr:\n\n # We don't use ok() for the exit code and stderr, because we want this\n # check to be just a single test.\n\nThe code came with the initial import of pg_rewind, and there is ... |
[
{
"msg_contents": "Hello,\n\nOne of the most frequently requested improvements from our customers\nis to reduce downtime associated with software updates (both major and\nminor versions). To do this, we have reviewed potential contributions to\nimproving logical replication.\n\nI’m working on a patch to support... |
[
{
"msg_contents": "Hi,\n\nWhen running check-world, a good chunk of the output is just pg_regress\nboilerplate. It doesn't matter when running tests individually or for tests\nwith a lot of individual tests like the main regression tests. But for lots of\nthe rest it is noisy. These days there are many more reg... |
[
{
"msg_contents": "It's turning out to be difficult to store the data for my custom index\naccess method in the main fork. Breaking up the data into pages with page\nheaders means a lot of extra work, a big performance hit, and disk space\nmanagement headaches. It's just not a good fit for my particular file\nf... |
[
{
"msg_contents": "Starting a new thread on the TAP patch from the \"[RFC] building postgres with\nmeson\" thread at 20220221165228.aqnfg45mceym7njm@alap3.anarazel.de to have\nsomewhere to discuss this patch.\n\n> On 21 Feb 2022, at 17:52, Andres Freund <andres@anarazel.de> wrote:\n> On 2021-10-13 13:54:10 +020... |
[
{
"msg_contents": "Hi hackers,\n\nWith the addition of archive modules (5ef1eef) and other archiving\nimprovements (e.g., beb4e9b), the amount of archiving overhead has been\nreduced. I spent some time measuring the remaining overhead, and the\nfollowing function stood out:\n\n\t/*\n\t * pgarch_archiveDone\n\t... |
[
{
"msg_contents": "Hello all, this is my first time posting here (I first posted on the IRC\nbut didn't get any response), so let me know if there's a different\nprocedure for asking questions.\n\nI'm interested in creating a Postgres extension that would enable\ndevelopers to write triggers in (modern) C++. Do... |
[
{
"msg_contents": "On 13.5 a wal flush PANIC is encountered after a standby is promoted.\r\n\r\nWith debugging, it was found that when a standby skips a missing continuation record on recovery, the missingContrecPtr is not invalidated after the record is skipped. Therefore, when the standby is promoted to a pri... |
[
{
"msg_contents": "I've been playing with an idea I had a while back. Basically that it\nwould be useful to have some \"noop\" operators that are used purely to\ninfluence the planner.\n\nFor context I've suggested in the past that there are two categories of hints:\n\n1 Hints that override the planner's decisi... |
[
{
"msg_contents": "Since 4035cd5d4, wal_compression=lz4 is supported. I think pg15 should also\nsupport wal_compression=zstd. There are free bits in the WAL header.\n\nThe last message on that thread includes a patch doing just that, which I've\nrebased.\nhttps://www.postgresql.org/message-id/YNqWd2GSMrnqWIfx... |
[
{
"msg_contents": "Hi,\n\nI recently had to work on some internal tool which, among other things, has to\nmerge pg_hba.conf and pg_ident.conf files between an existing (possibly\nalready updated) instance and some external repository.\n\nMy biggest concern is that any issue in either the external repository con... |
[
{
"msg_contents": "Hi,\n\nIn logical replication, currently Walsender sends the data that is\ngenerated locally and the data that are replicated from other\ninstances. This results in infinite recursion in circular logical\nreplication setup.\nHere the user is trying to have a 2-way replication setup with node ... |
[
{
"msg_contents": "Hello,\n\nI have created a patch that adds support for building Postgres for the\nwindows/arm64 platform using the MSVC toolchain. Following changes have\nbeen included\n\n1. Extend MSVC scripts to handle ARM64 platform.\n2. Add arm64 definition of spin_delay function.\n3. Exclude arm_acle.h ... |
[
{
"msg_contents": "Attached are 3 small patches that improve the performance of numeric\ndivision. Patch 0002 seems to have the biggest impact, but I think\nthey're all worth including, since they're quite simple changes, with\nnoticeable performance gains.\n\n\nPatch 0001 vectorises the inner loop of div_var_f... |
[
{
"msg_contents": "Hi!\n\nDuring a Patroni PR discussion \n(https://github.com/zalando/patroni/pull/2225), we realised that if one \nwants to use the \"-c\" option on a typical Debian/Ubuntu installation \n(where the config resides below /etc/postgresql/), pg_rewind needs a way \nto be told where the postgresql... |
[
{
"msg_contents": "Hello,\r\n\r\nAs part of the NSS work it came up [1] that clients don't have a good\r\nway to ask libpq what SSL library it was compiled with, unless they\r\nalready have a connection pointer so that they can call\r\nPQsslAttribute(conn, \"library\"). This poses a chicken-and-egg problem:\r\n... |
[
{
"msg_contents": "Hi,\n\nWhen verifying that the meson port actually runs all perl based tests I came\nacross src/interfaces/libpq/test/regress.pl. Instead of running tests yet\nanother way, it seems better to convert it to a tap test.\n\nI hope others agree?\n\nWhere would we want that test to live? Right no... |
[
{
"msg_contents": "I brought this up in [1] but wanted to start a dedicated thread.\n\nSince 16fa9b2b30a357 GiST indexes are not built in shared buffers.\nHowever, smgrimmedsync() is not done anywhere and skipFsync=true is\nalways passed to smgrwrite() and smgrextend(). So, if a checkpoint\nstarts and finishes ... |
[
{
"msg_contents": "In pg_regress we realloc() with the destination and source pointer being equal,\nwithout checking for OOM. While a fairly unlikely source of errors, is there a\nreason not to use pg_realloc() there for hygiene?\n\n--\nDaniel Gustafsson\t\thttps://vmware.com/",
"msg_date": "Wed, 23 Feb 20... |
[
{
"msg_contents": "Hi all,\r\n\r\nStephen pointed out [1] that the authenticated identity that's stored\r\nin MyProcPort can't be retrieved by extensions or triggers. Attached is\r\na patch that provides both a C API and a SQL function for retrieving\r\nit.\r\n\r\nGetAuthenticatedIdentityString() is a mouthful ... |
[
{
"msg_contents": "Hi all,\n\nI noticed something odd when going through some\nof the Interval code. The DAYS_PER_YEAR constant\nis defined in src/include/datatype/timestamp.h.\n> #define DAYS_PER_YEAR 365.25 /* assumes leap year every four years */\n\nWe execute the EXTRACT and date_part functions in\nsr... |
[
{
"msg_contents": "Hi,\n\nI found messages inserted a space before the \"%\" in pgbench.\nI think this is typo because there are no space before the \"%\" in other messages.\nWhat do you think? \n\ndiff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c\nindex f166a77e3a..4ebe5e6ea4 100644\n--- a/src... |
[
{
"msg_contents": "Thinking about poor performance in the case where the data fits in\nRAM, but the working set is too big for shared_buffers, I notice a\ncouple of things that seem bad in BufMgr, but don't understand why\nthey are like that.\n\n1. If we need to allocate a buffer to a new block we do this in on... |
[
{
"msg_contents": "Hi hackers,\n\nI noticed that create_index test (make installcheck) fails on my laptop\nbecause different query plans are chosen:\n\n- QUERY PLAN\n--------------------------------------------------------\n- Index Only Scan using tenk1_unique1 on tenk1\n- Index Cond: (un... |
[
{
"msg_contents": "Greetings,\n\nThis patch adds the ability to specify a RelFileNode and optional BlockNum to limit output of\npg_waldump records to only those which match the given criteria. This should be more performant\nthan `pg_waldump | grep` as well as more reliable given specific variations in output ... |
[
{
"msg_contents": "Hi,\n\nVictor Yegorov reported a crash related to a GiST index on ltree [1],\nfollowing a pg_upgrade from 12.x to 14.x, with a data set reproducing\nthis. I spent some time investigating this, and it seems this is a silly\nbug in commit\n\n commit 911e70207703799605f5a0e8aad9f06cff067c63 (HE... |
[
{
"msg_contents": "Hi everyone,\n\nWould adding additional examples to testlibpq3.c on handling more data\ntypes be helpful? I've attached a patch adding how to handle a REAL to\ncurrent example.\n\nRegards,\nMark",
"msg_date": "Thu, 24 Feb 2022 13:42:40 -0800",
"msg_from": "Mark Wong <markwkm@gmail.co... |
[
{
"msg_contents": "I'm trying to understand why hash indexes are built primarily in shared\nbuffers except when allocating a new splitpoint's worth of bucket pages\n-- which is done with smgrextend() directly in _hash_alloc_buckets().\n\nIs this just so that the value returned by smgrnblocks() includes the\nnew... |
[
{
"msg_contents": "Hi,\n\nThe final commitfest for pg15 will start in a few days, and I didn't see any\ndiscussion on it or anyone volunteering to be a CFM.\n\nI thought it would be a good idea to send this reminder now and avoid the same\nsituation as the last commitfest, to avoid unnecessary pain for the CFM(... |
[
{
"msg_contents": "Hi,\n\nplease find attached a patchset which adds lz4 compression in pg_dump.\n\nThe first commit does the heavy lifting required for additional compression methods.\nIt expands testing coverage for the already supported gzip compression. Commit\nbf9aa490db introduced cfp in compress_io.{c,h}... |
[
{
"msg_contents": "Here's a patch to add the sum of timings for JIT counters to\npg_stat_statements, as a way to follow-up on if JIT is doing a good or\na bad job in a configuration.\n\nI decided to only store the total time for the timings, since there\nare 4 different timings and storing max/min/etc for each ... |
[
{
"msg_contents": "Hi all,\n\nWhile doing some work using our functions [1] for calculate relations size\nI noticed an inconsistency between pg_total_relation_size and calculate\neverything separately, have a look in this example:\n\nfabrizio=# create table test_size (id bigserial primary key, toast_column\ntex... |
[
{
"msg_contents": "Hi,\n\nHere's a tiny patch to do $subject.\n\nRegards,\nBharath Rupireddy.",
"msg_date": "Fri, 25 Feb 2022 20:39:22 +0530",
"msg_from": "Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>",
"msg_from_op": true,
"msg_subject": "Fix typo in logicalfuncs.c - :%s/private d... |
[
{
"msg_contents": "This patch adds a configuration parameter jit_warn_above_fraction that\nwill cause a warning to be logged if the fraction of time spent on\ndoing JIT is bigger than the specified one. For example, this can be\nused to track down those cases where JIT ends up taking 90% of the\nquery runtime b... |
[
{
"msg_contents": "Hi All,\n\nI've been working on adding Windows+MinGW environment into cirrus-ci tasks\n(discussion about ci is here [1]).\nIt uses MSYS2 to set the environment. UCRT is chosen as C standard library,\ninstead of MSVCRT.\nThe task configures postgres with features that are available in MSYS2 (s... |
[
{
"msg_contents": "Hi all,\n\nI've seen various discussions around whether PG makes any guarantees on the\nordering of rows returned by the RETURNING clause (e.g. [1]). In a\nnutshell, when executing a statement such as the following:\n\nCREATE TABLE foo (id INT PRIMARY KEY GENERATED ALWAYS AS IDENTITY, data\nI... |
[
{
"msg_contents": "Hi,\n\nIt looks like there's a typo in pg_receivewal.c's\nget_destination_dir(), that is, use the function parameter dest_folder\ninstead of global variable basedir in the error message. Although\nthere's no harm in it as basedir is a global variable in\npg_receivewal.c, let's use the functio... |
[
{
"msg_contents": "About once a month over the last six months, my buildfarm animal\nflorican has gotten stuck while running the core regression tests.\nThe symptoms have looked very much the same each time: there is\na backend with two parallel worker processes that are just sitting\nand not consuming any CPU ... |
[
{
"msg_contents": "Hi,\n\nIn two recent investigations in occasional test failures\n(019_replslot_limit.pl failures, AIO rebase) the problems are somehow tied to\ncheckpointer.\n\nI don't yet know if actually causally related to precisely those failures, but\nwhen running e.g. 027_stream_regress.pl, I see phase... |
[
{
"msg_contents": "Hi,\n\nFirst time PostgreSQL contributor here :)\n\nI wanted to be able to allocate a bunch of numbers from a sequence at \nonce. Multiple people seem to be struggling with this \n(https://stackoverflow.com/questions/896274/select-multiple-ids-from-a-postgresql-sequence, \nhttps://www.depesz.... |
[
{
"msg_contents": "Hi,\n\nI found some redundant function calls in\npg_stat_statements.c/pg_stat_statements_internal(),\nThere is no need to call GetUserId() again because the value was\npreviously obtained.\nso I propose a patch to fix it.\n\n--- a/contrib/pg_stat_statements/pg_stat_statements.c\n+++ b/contrib... |
[
{
"msg_contents": "Hi, hackers\n\nWhen I try to use CREATE DATABASE IF NOT EXISTS in PostgreSQL, it complains\nthis syntax is not supported. We can use the following command to achieve\nthis, however, it's not straightforward.\n\n SELECT 'CREATE DATABASE mydb'\n WHERE NOT EXISTS (SELECT FROM pg_databas... |
[
{
"msg_contents": "Currently, when you connect with psql over SSL, you get a display like\nthis:\n\npsql (15devel)\nSSL connection (protocol: TLSv1.3, cipher: TLS_AES_256_GCM_SHA384, bits: 256, compression: off)\nType \"help\" for help.\n\nSince support for SSL compression has been removed from PostgreSQL, it\n... |
[
{
"msg_contents": "Hi,\n\nFor Toast storage [1] in PostgreSQL, first, the attribute value is\ncompressed\nand then divided into chunks. The problem with storing compressed value is,\nif we\nare not saving enough space such that it reduces the #chunks then we end up\nadding extra decompression cost on every read... |
[
{
"msg_contents": "Hia,\n\nWe've added some information such as the command and the timestamp to\nthe error context message by commit abc0910e2. This patch adds further\ninformation to it: replication origin name and commit-LSN.\n\nThis will be helpful for users to set the origin name and LSN to\npg_replication... |
[
{
"msg_contents": "Hello\nI just spotted in src/include/access/xlog.h:\nextern XLogRecPtr CalculateMaxmumSafeLSN(void);\n\nThis function doesn't seem to be used anywhere or even defined? \"git grep CalculateMaxmumSafeLSN\" shows only this line.\nWas added in c6550776394e25c1620bc8258427c8f1d448080d \"Allow user... |
[
{
"msg_contents": "Hi,\n\nIt looks like we use \"log segment\" in various user-facing messages.\nThe term \"log\" can mean server logs as well. The \"WAL segment\" suits\nwell here and it is consistently used across the other user-facing\nmessages [1].\n\nHere's a small patch attempting to consistently use the ... |
[
{
"msg_contents": "28 days has February ... So the March commitfest begins tomorrow. Meet\nyour friendly neighbourhood Commitfest Manager for March. Greetings!\n\nIf you have a patch you're hoping to get feedback on or you're\nexpecting to get committed this month make sure it's in the commitfest\nat https://c... |
[
{
"msg_contents": "Hi everyone,\n\nI have been doing some tests on a little server (at least compared to\nsome others around). It have 128 cores (64 physical), 128GB of RAM and\nagainst my will a 4-disk (10k RPM) RAID5.\n\nI have been using hammerdb testing from 9.6 to 14, currently on windows\n10. Obviously, I... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.